diff --git a/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024.json b/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024.json deleted file mode 100644 index 245782199eb9b9f63f916d9bc73db3c376afb434..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024", - "op_type": "conv2d", - "description": "2D convolution: 1x1 kernel, stride (1,1), dilation (1,1), 1024->1024 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 1024 - }, - "C_out": { - "type": "const", - "value": 1024 - }, - "Kh": { - "type": "const", - "value": 1 - }, - "Kw": { - "type": "const", - "value": 1 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512.json b/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512.json deleted file mode 100644 index dd5f2a1f3ef86e2378c7544760e8e62b06cf53c3..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512", - "op_type": "conv2d", - "description": "2D convolution: 1x1 kernel, stride (1,1), dilation (1,1), 128->512 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 128 - }, - "C_out": { - "type": "const", - "value": 512 - }, - "Kh": { - "type": "const", - "value": 1 - }, - "Kw": { - "type": "const", - "value": 1 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024.json b/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024.json deleted file mode 100644 index 564291af7e9817a81c019bbb768d64821f61abdf..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024", - "op_type": "conv2d", - "description": "2D convolution: 1x1 kernel, stride (1,1), dilation (1,1), 256->1024 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 256 - }, - "C_out": { - "type": "const", - "value": 1024 - }, - "Kh": { - "type": "const", - "value": 1 - }, - "Kw": { - "type": "const", - "value": 1 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048.json b/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048.json deleted file mode 100644 index 59dd74353e8f03af54dc2d768dbb59b32f32de19..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048", - "op_type": "conv2d", - "description": "2D convolution: 1x1 kernel, stride (1,1), dilation (1,1), 512->2048 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 512 - }, - "C_out": { - "type": "const", - "value": 2048 - }, - "Kh": { - "type": "const", - "value": 1 - }, - "Kw": { - "type": "const", - "value": 1 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256.json b/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256.json deleted file mode 100644 index e02bf757dae40ee419a70f5f5b9c1aed789e4faa..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256", - "op_type": "conv2d", - "description": "2D convolution: 1x1 kernel, stride (1,1), dilation (1,1), 64->256 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 64 - }, - "C_out": { - "type": "const", - "value": 256 - }, - "Kh": { - "type": "const", - "value": 1 - }, - "Kw": { - "type": "const", - "value": 1 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64.json b/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64.json deleted file mode 100644 index 000f3b80b1326ed7e343433d2941d2a83ee800c2..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64", - "op_type": "conv2d", - "description": "2D convolution: 1x1 kernel, stride (1,1), dilation (1,1), 64->64 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 64 - }, - "C_out": { - "type": "const", - "value": 64 - }, - "Kh": { - "type": "const", - "value": 1 - }, - "Kw": { - "type": "const", - "value": 1 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256.json b/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256.json deleted file mode 100644 index 6e30e71e06faf948f77e865d249d4d71c6903584..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256", - "op_type": "conv2d", - "description": "2D convolution: 3x3 kernel, stride (1,1), dilation (1,1), 128->256 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 128 - }, - "C_out": { - "type": "const", - "value": 256 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512.json b/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512.json deleted file mode 100644 index 0e64da85d9901170b3f86a8917dea242d7eb45fd..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512", - "op_type": "conv2d", - "description": "2D convolution: 3x3 kernel, stride (1,1), dilation (1,1), 256->512 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 256 - }, - "C_out": { - "type": "const", - "value": 512 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64.json b/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64.json deleted file mode 100644 index 2f950894c13b25a1bf56f8c5ca8c45f26ed9d225..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64", - "op_type": "conv2d", - "description": "2D convolution: 3x3 kernel, stride (1,1), dilation (1,1), 3->64 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 3 - }, - "C_out": { - "type": "const", - "value": 64 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024.json b/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024.json deleted file mode 100644 index f5652e098f9565acf74337e18e1dc7bd9167272d..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024", - "op_type": "conv2d", - "description": "2D convolution: 3x3 kernel, stride (1,1), dilation (1,1), 512->1024 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 512 - }, - "C_out": { - "type": "const", - "value": 1024 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128.json b/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128.json deleted file mode 100644 index cfa3f48139fda554dcaf0507be4a9cd88c634bfd..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128", - "op_type": "conv2d", - "description": "2D convolution: 3x3 kernel, stride (1,1), dilation (1,1), 64->128 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 64 - }, - "C_out": { - "type": "const", - "value": 128 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256.json b/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256.json deleted file mode 100644 index e87cc24fcc22583c1fb852d443bb98b76f55cb27..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256", - "op_type": "conv2d", - "description": "2D convolution: 3x3 kernel, stride (1,1), dilation (2,2), 128->256 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 128 - }, - "C_out": { - "type": "const", - "value": 256 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 2 - }, - "Dw": { - "type": "const", - "value": 2 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(2, 2))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256.json b/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256.json deleted file mode 100644 index 363fcf09bece0ecc7772a0ec5ed403f2f5165652..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256", - "op_type": "conv2d", - "description": "2D convolution: 3x3 kernel, stride (1,1), dilation (2,2), 256->256 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 256 - }, - "C_out": { - "type": "const", - "value": 256 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 2 - }, - "Dw": { - "type": "const", - "value": 2 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(2, 2))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512.json b/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512.json deleted file mode 100644 index a447d35ac4ae3c8ec8f3568ed9fd00f99908857e..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512", - "op_type": "conv2d", - "description": "2D convolution: 3x3 kernel, stride (1,1), dilation (2,2), 512->512 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 512 - }, - "C_out": { - "type": "const", - "value": 512 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 2 - }, - "Dw": { - "type": "const", - "value": 2 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(2, 2))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128.json b/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128.json deleted file mode 100644 index e712bf50576c6d1ce4d4045d3e96dd3691760943..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128", - "op_type": "conv2d", - "description": "2D convolution: 3x3 kernel, stride (1,1), dilation (2,2), 64->128 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 64 - }, - "C_out": { - "type": "const", - "value": 128 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 2 - }, - "Dw": { - "type": "const", - "value": 2 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(2, 2))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256.json b/definitions/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256.json deleted file mode 100644 index d455f1350065d14d4483d9cf19492c6aced0b585..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256", - "op_type": "conv2d", - "description": "2D convolution: 3x3 kernel, stride (2,2), dilation (1,1), 128->256 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 128 - }, - "C_out": { - "type": "const", - "value": 256 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(2, 2),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512.json b/definitions/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512.json deleted file mode 100644 index 7aae4d7ff8ac6d89301ea10588037b9930208883..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512", - "op_type": "conv2d", - "description": "2D convolution: 3x3 kernel, stride (2,2), dilation (1,1), 256->512 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 256 - }, - "C_out": { - "type": "const", - "value": 512 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(2, 2),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024.json b/definitions/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024.json deleted file mode 100644 index cf457d8b65b7ec77db99a78da9262c529f693b97..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024", - "op_type": "conv2d", - "description": "2D convolution: 3x3 kernel, stride (2,2), dilation (1,1), 512->1024 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 512 - }, - "C_out": { - "type": "const", - "value": 1024 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(2, 2),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128.json b/definitions/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128.json deleted file mode 100644 index 65bd859b33636522378c5803920914aa7d448af5..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128", - "op_type": "conv2d", - "description": "2D convolution: 3x3 kernel, stride (2,2), dilation (1,1), 64->128 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 64 - }, - "C_out": { - "type": "const", - "value": 128 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(2, 2),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128.json b/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128.json deleted file mode 100644 index 9429ca5236341e8c7252e30b73f16d3667e5995a..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128", - "op_type": "conv2d", - "description": "2D convolution: 5x5 kernel, stride (1,1), dilation (1,1), 128->128 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 128 - }, - "C_out": { - "type": "const", - "value": 128 - }, - "Kh": { - "type": "const", - "value": 5 - }, - "Kw": { - "type": "const", - "value": 5 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256.json b/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256.json deleted file mode 100644 index 0383937d9b188c6c2f1d2c11cfa52ad271e12476..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256", - "op_type": "conv2d", - "description": "2D convolution: 5x5 kernel, stride (1,1), dilation (1,1), 256->256 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 256 - }, - "C_out": { - "type": "const", - "value": 256 - }, - "Kh": { - "type": "const", - "value": 5 - }, - "Kw": { - "type": "const", - "value": 5 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64.json b/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64.json deleted file mode 100644 index 0e6e2f49ad2e45822403023627e0fadf338f6eea..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64", - "op_type": "conv2d", - "description": "2D convolution: 5x5 kernel, stride (1,1), dilation (1,1), 32->64 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 32 - }, - "C_out": { - "type": "const", - "value": 64 - }, - "Kh": { - "type": "const", - "value": 5 - }, - "Kw": { - "type": "const", - "value": 5 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32.json b/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32.json deleted file mode 100644 index 3669c17de7d8a3e7f1fa31a80e88d1f3542431ca..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32", - "op_type": "conv2d", - "description": "2D convolution: 5x5 kernel, stride (1,1), dilation (1,1), 3->32 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 3 - }, - "C_out": { - "type": "const", - "value": 32 - }, - "Kh": { - "type": "const", - "value": 5 - }, - "Kw": { - "type": "const", - "value": 5 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512.json b/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512.json deleted file mode 100644 index fd3e2f94d44cb810165ed408ec11cb3968af3be0..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512", - "op_type": "conv2d", - "description": "2D convolution: 5x5 kernel, stride (1,1), dilation (1,1), 512->512 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 512 - }, - "C_out": { - "type": "const", - "value": 512 - }, - "Kh": { - "type": "const", - "value": 5 - }, - "Kw": { - "type": "const", - "value": 5 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128.json b/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128.json deleted file mode 100644 index fdf90be63c3c3fd03558b25088f8b39e6816f0c5..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128", - "op_type": "conv2d", - "description": "2D convolution: 5x5 kernel, stride (1,1), dilation (1,1), 64->128 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 64 - }, - "C_out": { - "type": "const", - "value": 128 - }, - "Kh": { - "type": "const", - "value": 5 - }, - "Kw": { - "type": "const", - "value": 5 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(1, 1),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv/conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64.json b/definitions/conv/conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64.json deleted file mode 100644 index 333ac8ec87e4507a65fc67a69a6b8a37ec14d120..0000000000000000000000000000000000000000 --- a/definitions/conv/conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64.json +++ /dev/null @@ -1,116 +0,0 @@ -{ - "name": "conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64", - "op_type": "conv2d", - "description": "2D convolution: 7x7 kernel, stride (2,2), dilation (1,1), 3->64 channels. Extracted from tests/ncnn/candidate/convolution.cpp.", - "tags": [ - "status:phase2", - "isa:sve" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_h - (Kh-1)*Dh - 1) / Sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_w - (Kw-1)*Dw - 1) / Sw + 1" - }, - "C_in": { - "type": "const", - "value": 3 - }, - "C_out": { - "type": "const", - "value": 64 - }, - "Kh": { - "type": "const", - "value": 7 - }, - "Kw": { - "type": "const", - "value": 7 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "pad_top": { - "shape": null, - "dtype": "int32" - }, - "pad_left": { - "shape": null, - "dtype": "int32" - }, - "activation_type": { - "shape": null, - "dtype": "int32" - }, - "with_bias": { - "shape": null, - "dtype": "int32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*pad_top - (Kh-1)*Dh - 1) / Sh + 1", - "W_out == (W + 2*pad_left - (Kw-1)*Dw - 1) / Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\n\ndef run(input, weight, pad_top, pad_left, activation_type, with_bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight)\n y = F.conv2d(x, w, None,\n stride=(2, 2),\n padding=(int(pad_top), int(pad_left)),\n dilation=(1, 1))\n return y\n" -} diff --git a/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin1024_cout1024.json b/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin1024_cout1024.json deleted file mode 100644 index 2240cea785022479c5e94223ff7f54a8f520ea7f..0000000000000000000000000000000000000000 --- a/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin1024_cout1024.json +++ /dev/null @@ -1,78 +0,0 @@ -{ - "name": "conv1d_kw1_sw1_dw1_p0_cin1024_cout1024", - "op_type": "conv1d", - "description": "1D convolution: kw=1 stride=1 dilation=1 pad=0 C_in=1024 C_out=1024", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "W": { - "type": "var", - "parent": "N" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad - dw*(kw-1) - 1) / sw + 1" - }, - "C_in": { - "type": "const", - "value": 1024 - }, - "C_out": { - "type": "const", - "value": 1024 - }, - "Kw": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "C_in", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "C_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "W_out == (W + 2*0 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input).unsqueeze(0)\n w = torch.from_numpy(weight)\n b = torch.from_numpy(bias)\n return F.conv1d(x, w, b, stride=1, padding=0, dilation=1).squeeze(0).numpy()\n" -} diff --git a/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin128_cout512.json b/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin128_cout512.json deleted file mode 100644 index 39ce5a756a8117f2f4c4f1df8b694b4ed5c66365..0000000000000000000000000000000000000000 --- a/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin128_cout512.json +++ /dev/null @@ -1,78 +0,0 @@ -{ - "name": "conv1d_kw1_sw1_dw1_p0_cin128_cout512", - "op_type": "conv1d", - "description": "1D convolution: kw=1 stride=1 dilation=1 pad=0 C_in=128 C_out=512", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "W": { - "type": "var", - "parent": "N" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad - dw*(kw-1) - 1) / sw + 1" - }, - "C_in": { - "type": "const", - "value": 128 - }, - "C_out": { - "type": "const", - "value": 512 - }, - "Kw": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "C_in", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "C_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "W_out == (W + 2*0 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input).unsqueeze(0)\n w = torch.from_numpy(weight)\n b = torch.from_numpy(bias)\n return F.conv1d(x, w, b, stride=1, padding=0, dilation=1).squeeze(0).numpy()\n" -} diff --git a/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin256_cout1024.json b/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin256_cout1024.json deleted file mode 100644 index bf568ddd154bf2ff042c094371d281355d9e38e2..0000000000000000000000000000000000000000 --- a/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin256_cout1024.json +++ /dev/null @@ -1,78 +0,0 @@ -{ - "name": "conv1d_kw1_sw1_dw1_p0_cin256_cout1024", - "op_type": "conv1d", - "description": "1D convolution: kw=1 stride=1 dilation=1 pad=0 C_in=256 C_out=1024", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "W": { - "type": "var", - "parent": "N" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad - dw*(kw-1) - 1) / sw + 1" - }, - "C_in": { - "type": "const", - "value": 256 - }, - "C_out": { - "type": "const", - "value": 1024 - }, - "Kw": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "C_in", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "C_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "W_out == (W + 2*0 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input).unsqueeze(0)\n w = torch.from_numpy(weight)\n b = torch.from_numpy(bias)\n return F.conv1d(x, w, b, stride=1, padding=0, dilation=1).squeeze(0).numpy()\n" -} diff --git a/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin512_cout2048.json b/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin512_cout2048.json deleted file mode 100644 index d778b8377640a129233ca9d36066584caeaff0f4..0000000000000000000000000000000000000000 --- a/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin512_cout2048.json +++ /dev/null @@ -1,78 +0,0 @@ -{ - "name": "conv1d_kw1_sw1_dw1_p0_cin512_cout2048", - "op_type": "conv1d", - "description": "1D convolution: kw=1 stride=1 dilation=1 pad=0 C_in=512 C_out=2048", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "W": { - "type": "var", - "parent": "N" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad - dw*(kw-1) - 1) / sw + 1" - }, - "C_in": { - "type": "const", - "value": 512 - }, - "C_out": { - "type": "const", - "value": 2048 - }, - "Kw": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "C_in", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "C_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "W_out == (W + 2*0 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input).unsqueeze(0)\n w = torch.from_numpy(weight)\n b = torch.from_numpy(bias)\n return F.conv1d(x, w, b, stride=1, padding=0, dilation=1).squeeze(0).numpy()\n" -} diff --git a/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout256.json b/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout256.json deleted file mode 100644 index f164c3790b400b987b4bf3211255575b56fa610a..0000000000000000000000000000000000000000 --- a/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout256.json +++ /dev/null @@ -1,78 +0,0 @@ -{ - "name": "conv1d_kw1_sw1_dw1_p0_cin64_cout256", - "op_type": "conv1d", - "description": "1D convolution: kw=1 stride=1 dilation=1 pad=0 C_in=64 C_out=256", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "W": { - "type": "var", - "parent": "N" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad - dw*(kw-1) - 1) / sw + 1" - }, - "C_in": { - "type": "const", - "value": 64 - }, - "C_out": { - "type": "const", - "value": 256 - }, - "Kw": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "C_in", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "C_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "W_out == (W + 2*0 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input).unsqueeze(0)\n w = torch.from_numpy(weight)\n b = torch.from_numpy(bias)\n return F.conv1d(x, w, b, stride=1, padding=0, dilation=1).squeeze(0).numpy()\n" -} diff --git a/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout64.json b/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout64.json deleted file mode 100644 index b5a327202852d182b6c572985893ce26fe7a40f9..0000000000000000000000000000000000000000 --- a/definitions/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout64.json +++ /dev/null @@ -1,78 +0,0 @@ -{ - "name": "conv1d_kw1_sw1_dw1_p0_cin64_cout64", - "op_type": "conv1d", - "description": "1D convolution: kw=1 stride=1 dilation=1 pad=0 C_in=64 C_out=64", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "W": { - "type": "var", - "parent": "N" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad - dw*(kw-1) - 1) / sw + 1" - }, - "C_in": { - "type": "const", - "value": 64 - }, - "C_out": { - "type": "const", - "value": 64 - }, - "Kw": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "C_in", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "C_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "W_out == (W + 2*0 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input).unsqueeze(0)\n w = torch.from_numpy(weight)\n b = torch.from_numpy(bias)\n return F.conv1d(x, w, b, stride=1, padding=0, dilation=1).squeeze(0).numpy()\n" -} diff --git a/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin1024_cout1024.json b/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin1024_cout1024.json deleted file mode 100644 index 56385f1ec539f50e669d28b5626f5e39eae08342..0000000000000000000000000000000000000000 --- a/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin1024_cout1024.json +++ /dev/null @@ -1,78 +0,0 @@ -{ - "name": "conv1d_kw3_sw1_dw1_p1_cin1024_cout1024", - "op_type": "conv1d", - "description": "1D convolution: kw=3 stride=1 dilation=1 pad=1 C_in=1024 C_out=1024", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "W": { - "type": "var", - "parent": "N" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad - dw*(kw-1) - 1) / sw + 1" - }, - "C_in": { - "type": "const", - "value": 1024 - }, - "C_out": { - "type": "const", - "value": 1024 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "C_in", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "C_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "W_out == (W + 2*1 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input).unsqueeze(0)\n w = torch.from_numpy(weight)\n b = torch.from_numpy(bias)\n return F.conv1d(x, w, b, stride=1, padding=1, dilation=1).squeeze(0).numpy()\n" -} diff --git a/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin128_cout256.json b/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin128_cout256.json deleted file mode 100644 index 1ee92c35ca9b7c29d2d84a6de3cad829ecb85066..0000000000000000000000000000000000000000 --- a/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin128_cout256.json +++ /dev/null @@ -1,78 +0,0 @@ -{ - "name": "conv1d_kw3_sw1_dw1_p1_cin128_cout256", - "op_type": "conv1d", - "description": "1D convolution: kw=3 stride=1 dilation=1 pad=1 C_in=128 C_out=256", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "W": { - "type": "var", - "parent": "N" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad - dw*(kw-1) - 1) / sw + 1" - }, - "C_in": { - "type": "const", - "value": 128 - }, - "C_out": { - "type": "const", - "value": 256 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "C_in", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "C_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "W_out == (W + 2*1 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input).unsqueeze(0)\n w = torch.from_numpy(weight)\n b = torch.from_numpy(bias)\n return F.conv1d(x, w, b, stride=1, padding=1, dilation=1).squeeze(0).numpy()\n" -} diff --git a/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin256_cout512.json b/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin256_cout512.json deleted file mode 100644 index f289ff10b6eabe4d07341b58e3e8796e53c0158c..0000000000000000000000000000000000000000 --- a/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin256_cout512.json +++ /dev/null @@ -1,78 +0,0 @@ -{ - "name": "conv1d_kw3_sw1_dw1_p1_cin256_cout512", - "op_type": "conv1d", - "description": "1D convolution: kw=3 stride=1 dilation=1 pad=1 C_in=256 C_out=512", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "W": { - "type": "var", - "parent": "N" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad - dw*(kw-1) - 1) / sw + 1" - }, - "C_in": { - "type": "const", - "value": 256 - }, - "C_out": { - "type": "const", - "value": 512 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "C_in", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "C_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "W_out == (W + 2*1 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input).unsqueeze(0)\n w = torch.from_numpy(weight)\n b = torch.from_numpy(bias)\n return F.conv1d(x, w, b, stride=1, padding=1, dilation=1).squeeze(0).numpy()\n" -} diff --git a/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin3_cout64.json b/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin3_cout64.json deleted file mode 100644 index 661d967cc0e8d49abda67886dcab1000c7458592..0000000000000000000000000000000000000000 --- a/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin3_cout64.json +++ /dev/null @@ -1,78 +0,0 @@ -{ - "name": "conv1d_kw3_sw1_dw1_p1_cin3_cout64", - "op_type": "conv1d", - "description": "1D convolution: kw=3 stride=1 dilation=1 pad=1 C_in=3 C_out=64", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "W": { - "type": "var", - "parent": "N" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad - dw*(kw-1) - 1) / sw + 1" - }, - "C_in": { - "type": "const", - "value": 3 - }, - "C_out": { - "type": "const", - "value": 64 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "C_in", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "C_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "W_out == (W + 2*1 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input).unsqueeze(0)\n w = torch.from_numpy(weight)\n b = torch.from_numpy(bias)\n return F.conv1d(x, w, b, stride=1, padding=1, dilation=1).squeeze(0).numpy()\n" -} diff --git a/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin512_cout1024.json b/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin512_cout1024.json deleted file mode 100644 index 2b6818407a7974ba5f781670462671e8d8001a63..0000000000000000000000000000000000000000 --- a/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin512_cout1024.json +++ /dev/null @@ -1,78 +0,0 @@ -{ - "name": "conv1d_kw3_sw1_dw1_p1_cin512_cout1024", - "op_type": "conv1d", - "description": "1D convolution: kw=3 stride=1 dilation=1 pad=1 C_in=512 C_out=1024", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "W": { - "type": "var", - "parent": "N" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad - dw*(kw-1) - 1) / sw + 1" - }, - "C_in": { - "type": "const", - "value": 512 - }, - "C_out": { - "type": "const", - "value": 1024 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "C_in", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "C_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "W_out == (W + 2*1 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input).unsqueeze(0)\n w = torch.from_numpy(weight)\n b = torch.from_numpy(bias)\n return F.conv1d(x, w, b, stride=1, padding=1, dilation=1).squeeze(0).numpy()\n" -} diff --git a/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin64_cout128.json b/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin64_cout128.json deleted file mode 100644 index 0a3c031014f1ab7e6854ed47e8f47d9296ccaab4..0000000000000000000000000000000000000000 --- a/definitions/conv1d/conv1d_kw3_sw1_dw1_p1_cin64_cout128.json +++ /dev/null @@ -1,78 +0,0 @@ -{ - "name": "conv1d_kw3_sw1_dw1_p1_cin64_cout128", - "op_type": "conv1d", - "description": "1D convolution: kw=3 stride=1 dilation=1 pad=1 C_in=64 C_out=128", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "W": { - "type": "var", - "parent": "N" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad - dw*(kw-1) - 1) / sw + 1" - }, - "C_in": { - "type": "const", - "value": 64 - }, - "C_out": { - "type": "const", - "value": 128 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "C_in", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "C_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "W_out == (W + 2*1 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input).unsqueeze(0)\n w = torch.from_numpy(weight)\n b = torch.from_numpy(bias)\n return F.conv1d(x, w, b, stride=1, padding=1, dilation=1).squeeze(0).numpy()\n" -} diff --git a/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024.json b/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024.json deleted file mode 100644 index 755befd5378216791dd2ac0cf9febf3de5997d9d..0000000000000000000000000000000000000000 --- a/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024", - "op_type": "conv2d_depthwise", - "description": "Depthwise 2D conv: kh=3 kw=3 stride=(1,1) dilation=(1,1) pad=(1,1) C=1024", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_top - dh*(kh-1) - 1) / sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_left - dw*(kw-1) - 1) / sw + 1" - }, - "C": { - "type": "const", - "value": 1024 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*1 - 1*(Kh-1) - 1) // Sh + 1", - "W_out == (W + 2*1 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv2d(x, w, b, stride=(1, 1), padding=(1, 1), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128.json b/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128.json deleted file mode 100644 index 34217b5ca5d7a924287304484e625b8c0422ad98..0000000000000000000000000000000000000000 --- a/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128", - "op_type": "conv2d_depthwise", - "description": "Depthwise 2D conv: kh=3 kw=3 stride=(1,1) dilation=(1,1) pad=(1,1) C=128", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_top - dh*(kh-1) - 1) / sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_left - dw*(kw-1) - 1) / sw + 1" - }, - "C": { - "type": "const", - "value": 128 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*1 - 1*(Kh-1) - 1) // Sh + 1", - "W_out == (W + 2*1 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv2d(x, w, b, stride=(1, 1), padding=(1, 1), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048.json b/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048.json deleted file mode 100644 index 6ae0bb9d0c8659a2a937dfea16fb802da8132688..0000000000000000000000000000000000000000 --- a/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048", - "op_type": "conv2d_depthwise", - "description": "Depthwise 2D conv: kh=3 kw=3 stride=(1,1) dilation=(1,1) pad=(1,1) C=2048", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_top - dh*(kh-1) - 1) / sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_left - dw*(kw-1) - 1) / sw + 1" - }, - "C": { - "type": "const", - "value": 2048 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*1 - 1*(Kh-1) - 1) // Sh + 1", - "W_out == (W + 2*1 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv2d(x, w, b, stride=(1, 1), padding=(1, 1), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256.json b/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256.json deleted file mode 100644 index b1bcc193e6b1a17065edf26cc2c6aac75f057138..0000000000000000000000000000000000000000 --- a/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256", - "op_type": "conv2d_depthwise", - "description": "Depthwise 2D conv: kh=3 kw=3 stride=(1,1) dilation=(1,1) pad=(1,1) C=256", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_top - dh*(kh-1) - 1) / sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_left - dw*(kw-1) - 1) / sw + 1" - }, - "C": { - "type": "const", - "value": 256 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*1 - 1*(Kh-1) - 1) // Sh + 1", - "W_out == (W + 2*1 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv2d(x, w, b, stride=(1, 1), padding=(1, 1), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32.json b/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32.json deleted file mode 100644 index ae61591f08eb66bd5c74ccf71f7f68999859b04c..0000000000000000000000000000000000000000 --- a/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32", - "op_type": "conv2d_depthwise", - "description": "Depthwise 2D conv: kh=3 kw=3 stride=(1,1) dilation=(1,1) pad=(1,1) C=32", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_top - dh*(kh-1) - 1) / sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_left - dw*(kw-1) - 1) / sw + 1" - }, - "C": { - "type": "const", - "value": 32 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*1 - 1*(Kh-1) - 1) // Sh + 1", - "W_out == (W + 2*1 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv2d(x, w, b, stride=(1, 1), padding=(1, 1), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512.json b/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512.json deleted file mode 100644 index 268c8ef55bec402974d98e22cff21ebc2e6667e8..0000000000000000000000000000000000000000 --- a/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512", - "op_type": "conv2d_depthwise", - "description": "Depthwise 2D conv: kh=3 kw=3 stride=(1,1) dilation=(1,1) pad=(1,1) C=512", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_top - dh*(kh-1) - 1) / sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_left - dw*(kw-1) - 1) / sw + 1" - }, - "C": { - "type": "const", - "value": 512 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*1 - 1*(Kh-1) - 1) // Sh + 1", - "W_out == (W + 2*1 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv2d(x, w, b, stride=(1, 1), padding=(1, 1), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64.json b/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64.json deleted file mode 100644 index e0834e0f517caeb19b6b02cf7d49bec28f4c25eb..0000000000000000000000000000000000000000 --- a/definitions/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64", - "op_type": "conv2d_depthwise", - "description": "Depthwise 2D conv: kh=3 kw=3 stride=(1,1) dilation=(1,1) pad=(1,1) C=64", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_top - dh*(kh-1) - 1) / sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_left - dw*(kw-1) - 1) / sw + 1" - }, - "C": { - "type": "const", - "value": 64 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*1 - 1*(Kh-1) - 1) // Sh + 1", - "W_out == (W + 2*1 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv2d(x, w, b, stride=(1, 1), padding=(1, 1), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024.json b/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024.json deleted file mode 100644 index d38729c63df9e6898efab788ca9c6cd6cf932d1e..0000000000000000000000000000000000000000 --- a/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024", - "op_type": "conv2d_depthwise", - "description": "Depthwise 2D conv: kh=5 kw=5 stride=(1,1) dilation=(1,1) pad=(2,2) C=1024", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_top - dh*(kh-1) - 1) / sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_left - dw*(kw-1) - 1) / sw + 1" - }, - "C": { - "type": "const", - "value": 1024 - }, - "Kh": { - "type": "const", - "value": 5 - }, - "Kw": { - "type": "const", - "value": 5 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*2 - 1*(Kh-1) - 1) // Sh + 1", - "W_out == (W + 2*2 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv2d(x, w, b, stride=(1, 1), padding=(2, 2), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128.json b/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128.json deleted file mode 100644 index fbbb4bdd8ed8794ccc172cd4fef1117c44afdd86..0000000000000000000000000000000000000000 --- a/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128", - "op_type": "conv2d_depthwise", - "description": "Depthwise 2D conv: kh=5 kw=5 stride=(1,1) dilation=(1,1) pad=(2,2) C=128", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_top - dh*(kh-1) - 1) / sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_left - dw*(kw-1) - 1) / sw + 1" - }, - "C": { - "type": "const", - "value": 128 - }, - "Kh": { - "type": "const", - "value": 5 - }, - "Kw": { - "type": "const", - "value": 5 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*2 - 1*(Kh-1) - 1) // Sh + 1", - "W_out == (W + 2*2 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv2d(x, w, b, stride=(1, 1), padding=(2, 2), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256.json b/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256.json deleted file mode 100644 index a76bd73b99437617f6694f5e28234123ce1b93b1..0000000000000000000000000000000000000000 --- a/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256", - "op_type": "conv2d_depthwise", - "description": "Depthwise 2D conv: kh=5 kw=5 stride=(1,1) dilation=(1,1) pad=(2,2) C=256", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_top - dh*(kh-1) - 1) / sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_left - dw*(kw-1) - 1) / sw + 1" - }, - "C": { - "type": "const", - "value": 256 - }, - "Kh": { - "type": "const", - "value": 5 - }, - "Kw": { - "type": "const", - "value": 5 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*2 - 1*(Kh-1) - 1) // Sh + 1", - "W_out == (W + 2*2 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv2d(x, w, b, stride=(1, 1), padding=(2, 2), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32.json b/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32.json deleted file mode 100644 index 90d68bf7c61e0b04b444cf597c1001e1e869f5c7..0000000000000000000000000000000000000000 --- a/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32", - "op_type": "conv2d_depthwise", - "description": "Depthwise 2D conv: kh=5 kw=5 stride=(1,1) dilation=(1,1) pad=(2,2) C=32", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_top - dh*(kh-1) - 1) / sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_left - dw*(kw-1) - 1) / sw + 1" - }, - "C": { - "type": "const", - "value": 32 - }, - "Kh": { - "type": "const", - "value": 5 - }, - "Kw": { - "type": "const", - "value": 5 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*2 - 1*(Kh-1) - 1) // Sh + 1", - "W_out == (W + 2*2 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv2d(x, w, b, stride=(1, 1), padding=(2, 2), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512.json b/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512.json deleted file mode 100644 index 95e1f4cd1f1a9ea6f9d3695339ea5f397523897e..0000000000000000000000000000000000000000 --- a/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512", - "op_type": "conv2d_depthwise", - "description": "Depthwise 2D conv: kh=5 kw=5 stride=(1,1) dilation=(1,1) pad=(2,2) C=512", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_top - dh*(kh-1) - 1) / sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_left - dw*(kw-1) - 1) / sw + 1" - }, - "C": { - "type": "const", - "value": 512 - }, - "Kh": { - "type": "const", - "value": 5 - }, - "Kw": { - "type": "const", - "value": 5 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*2 - 1*(Kh-1) - 1) // Sh + 1", - "W_out == (W + 2*2 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv2d(x, w, b, stride=(1, 1), padding=(2, 2), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64.json b/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64.json deleted file mode 100644 index 8f25bd4cbfa3fa098c1b5c4d2430622f9338f01f..0000000000000000000000000000000000000000 --- a/definitions/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64", - "op_type": "conv2d_depthwise", - "description": "Depthwise 2D conv: kh=5 kw=5 stride=(1,1) dilation=(1,1) pad=(2,2) C=64", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H + 2*pad_top - dh*(kh-1) - 1) / sh + 1" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W + 2*pad_left - dw*(kw-1) - 1) / sw + 1" - }, - "C": { - "type": "const", - "value": 64 - }, - "Kh": { - "type": "const", - "value": 5 - }, - "Kw": { - "type": "const", - "value": 5 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H + 2*2 - 1*(Kh-1) - 1) // Sh + 1", - "W_out == (W + 2*2 - 1*(Kw-1) - 1) // Sw + 1" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv2d(x, w, b, stride=(1, 1), padding=(2, 2), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin128_cout256.json b/definitions/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin128_cout256.json deleted file mode 100644 index ab20fa28f81352a44ac0ba0c8b5f573e705043e2..0000000000000000000000000000000000000000 --- a/definitions/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin128_cout256.json +++ /dev/null @@ -1,105 +0,0 @@ -{ - "name": "deconv2d_kh3_kw3_sh1_sw1_cin128_cout256", - "op_type": "deconv2d", - "description": "Transposed 2D conv: kh=3 kw=3 stride=(1,1) dilation=(1,1) pad=0 C_in=128 C_out=256", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C_in": { - "type": "const", - "value": 128 - }, - "C_out": { - "type": "const", - "value": 256 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).permute(1, 0, 2, 3)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(1, 1), padding=(0, 0), dilation=(1, 1)).numpy()\n" -} diff --git a/definitions/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin256_cout512.json b/definitions/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin256_cout512.json deleted file mode 100644 index eb2f16533f756898bdf7c101d36c80e2173676c5..0000000000000000000000000000000000000000 --- a/definitions/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin256_cout512.json +++ /dev/null @@ -1,105 +0,0 @@ -{ - "name": "deconv2d_kh3_kw3_sh1_sw1_cin256_cout512", - "op_type": "deconv2d", - "description": "Transposed 2D conv: kh=3 kw=3 stride=(1,1) dilation=(1,1) pad=0 C_in=256 C_out=512", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C_in": { - "type": "const", - "value": 256 - }, - "C_out": { - "type": "const", - "value": 512 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).permute(1, 0, 2, 3)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(1, 1), padding=(0, 0), dilation=(1, 1)).numpy()\n" -} diff --git a/definitions/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024.json b/definitions/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024.json deleted file mode 100644 index f84a60a10068c1deee1378a9f9442e8a339e7465..0000000000000000000000000000000000000000 --- a/definitions/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024.json +++ /dev/null @@ -1,105 +0,0 @@ -{ - "name": "deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024", - "op_type": "deconv2d", - "description": "Transposed 2D conv: kh=3 kw=3 stride=(1,1) dilation=(1,1) pad=0 C_in=512 C_out=1024", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C_in": { - "type": "const", - "value": 512 - }, - "C_out": { - "type": "const", - "value": 1024 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).permute(1, 0, 2, 3)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(1, 1), padding=(0, 0), dilation=(1, 1)).numpy()\n" -} diff --git a/definitions/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin64_cout128.json b/definitions/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin64_cout128.json deleted file mode 100644 index 5583b6c646f58e10b0e7e5f29af3c9b380fb6d9a..0000000000000000000000000000000000000000 --- a/definitions/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin64_cout128.json +++ /dev/null @@ -1,105 +0,0 @@ -{ - "name": "deconv2d_kh3_kw3_sh1_sw1_cin64_cout128", - "op_type": "deconv2d", - "description": "Transposed 2D conv: kh=3 kw=3 stride=(1,1) dilation=(1,1) pad=0 C_in=64 C_out=128", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C_in": { - "type": "const", - "value": 64 - }, - "C_out": { - "type": "const", - "value": 128 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).permute(1, 0, 2, 3)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(1, 1), padding=(0, 0), dilation=(1, 1)).numpy()\n" -} diff --git a/definitions/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin128_cout256.json b/definitions/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin128_cout256.json deleted file mode 100644 index 2072d71ccb5589bc2ae5bf19fc3964b7207fd188..0000000000000000000000000000000000000000 --- a/definitions/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin128_cout256.json +++ /dev/null @@ -1,105 +0,0 @@ -{ - "name": "deconv2d_kh3_kw3_sh2_sw2_cin128_cout256", - "op_type": "deconv2d", - "description": "Transposed 2D conv: kh=3 kw=3 stride=(2,2) dilation=(1,1) pad=0 C_in=128 C_out=256", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C_in": { - "type": "const", - "value": 128 - }, - "C_out": { - "type": "const", - "value": 256 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).permute(1, 0, 2, 3)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(2, 2), padding=(0, 0), dilation=(1, 1)).numpy()\n" -} diff --git a/definitions/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin256_cout512.json b/definitions/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin256_cout512.json deleted file mode 100644 index 2e6454bb2cf61e345e34a4f5f5ad9d3914e3766c..0000000000000000000000000000000000000000 --- a/definitions/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin256_cout512.json +++ /dev/null @@ -1,105 +0,0 @@ -{ - "name": "deconv2d_kh3_kw3_sh2_sw2_cin256_cout512", - "op_type": "deconv2d", - "description": "Transposed 2D conv: kh=3 kw=3 stride=(2,2) dilation=(1,1) pad=0 C_in=256 C_out=512", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C_in": { - "type": "const", - "value": 256 - }, - "C_out": { - "type": "const", - "value": 512 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).permute(1, 0, 2, 3)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(2, 2), padding=(0, 0), dilation=(1, 1)).numpy()\n" -} diff --git a/definitions/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024.json b/definitions/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024.json deleted file mode 100644 index d56d0ceee2ccadc676d90b384a1a0d9f894de32e..0000000000000000000000000000000000000000 --- a/definitions/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024.json +++ /dev/null @@ -1,105 +0,0 @@ -{ - "name": "deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024", - "op_type": "deconv2d", - "description": "Transposed 2D conv: kh=3 kw=3 stride=(2,2) dilation=(1,1) pad=0 C_in=512 C_out=1024", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C_in": { - "type": "const", - "value": 512 - }, - "C_out": { - "type": "const", - "value": 1024 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).permute(1, 0, 2, 3)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(2, 2), padding=(0, 0), dilation=(1, 1)).numpy()\n" -} diff --git a/definitions/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin64_cout128.json b/definitions/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin64_cout128.json deleted file mode 100644 index 0442c66a30a4b5fc71e9dd49fd81f8c805edb1b4..0000000000000000000000000000000000000000 --- a/definitions/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin64_cout128.json +++ /dev/null @@ -1,105 +0,0 @@ -{ - "name": "deconv2d_kh3_kw3_sh2_sw2_cin64_cout128", - "op_type": "deconv2d", - "description": "Transposed 2D conv: kh=3 kw=3 stride=(2,2) dilation=(1,1) pad=0 C_in=64 C_out=128", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C_in": { - "type": "const", - "value": 64 - }, - "C_out": { - "type": "const", - "value": 128 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).permute(1, 0, 2, 3)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(2, 2), padding=(0, 0), dilation=(1, 1)).numpy()\n" -} diff --git a/definitions/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin128_cout256.json b/definitions/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin128_cout256.json deleted file mode 100644 index 13b234c65cb226e3cdcfba73886de69dd996c26a..0000000000000000000000000000000000000000 --- a/definitions/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin128_cout256.json +++ /dev/null @@ -1,105 +0,0 @@ -{ - "name": "deconv2d_kh4_kw4_sh1_sw1_cin128_cout256", - "op_type": "deconv2d", - "description": "Transposed 2D conv: kh=4 kw=4 stride=(1,1) dilation=(1,1) pad=0 C_in=128 C_out=256", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C_in": { - "type": "const", - "value": 128 - }, - "C_out": { - "type": "const", - "value": 256 - }, - "Kh": { - "type": "const", - "value": 4 - }, - "Kw": { - "type": "const", - "value": 4 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).permute(1, 0, 2, 3)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(1, 1), padding=(0, 0), dilation=(1, 1)).numpy()\n" -} diff --git a/definitions/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin256_cout512.json b/definitions/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin256_cout512.json deleted file mode 100644 index 708ccf0567bdcbf99d4b478c2b5b2065edbe87ae..0000000000000000000000000000000000000000 --- a/definitions/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin256_cout512.json +++ /dev/null @@ -1,105 +0,0 @@ -{ - "name": "deconv2d_kh4_kw4_sh1_sw1_cin256_cout512", - "op_type": "deconv2d", - "description": "Transposed 2D conv: kh=4 kw=4 stride=(1,1) dilation=(1,1) pad=0 C_in=256 C_out=512", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C_in": { - "type": "const", - "value": 256 - }, - "C_out": { - "type": "const", - "value": 512 - }, - "Kh": { - "type": "const", - "value": 4 - }, - "Kw": { - "type": "const", - "value": 4 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).permute(1, 0, 2, 3)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(1, 1), padding=(0, 0), dilation=(1, 1)).numpy()\n" -} diff --git a/definitions/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024.json b/definitions/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024.json deleted file mode 100644 index 3fbd328ae329abc83002f46365dcb04fe53b95cf..0000000000000000000000000000000000000000 --- a/definitions/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024.json +++ /dev/null @@ -1,105 +0,0 @@ -{ - "name": "deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024", - "op_type": "deconv2d", - "description": "Transposed 2D conv: kh=4 kw=4 stride=(1,1) dilation=(1,1) pad=0 C_in=512 C_out=1024", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C_in": { - "type": "const", - "value": 512 - }, - "C_out": { - "type": "const", - "value": 1024 - }, - "Kh": { - "type": "const", - "value": 4 - }, - "Kw": { - "type": "const", - "value": 4 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).permute(1, 0, 2, 3)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(1, 1), padding=(0, 0), dilation=(1, 1)).numpy()\n" -} diff --git a/definitions/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin64_cout128.json b/definitions/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin64_cout128.json deleted file mode 100644 index fd5abd46acf3572f6f1009b43896e841ce5f3d4a..0000000000000000000000000000000000000000 --- a/definitions/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin64_cout128.json +++ /dev/null @@ -1,105 +0,0 @@ -{ - "name": "deconv2d_kh4_kw4_sh1_sw1_cin64_cout128", - "op_type": "deconv2d", - "description": "Transposed 2D conv: kh=4 kw=4 stride=(1,1) dilation=(1,1) pad=0 C_in=64 C_out=128", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C_in": { - "type": "const", - "value": 64 - }, - "C_out": { - "type": "const", - "value": 128 - }, - "Kh": { - "type": "const", - "value": 4 - }, - "Kw": { - "type": "const", - "value": 4 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).permute(1, 0, 2, 3)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(1, 1), padding=(0, 0), dilation=(1, 1)).numpy()\n" -} diff --git a/definitions/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin128_cout256.json b/definitions/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin128_cout256.json deleted file mode 100644 index 0bf756790a9dd83385cd2a3d012cbe4d24d2e6c2..0000000000000000000000000000000000000000 --- a/definitions/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin128_cout256.json +++ /dev/null @@ -1,105 +0,0 @@ -{ - "name": "deconv2d_kh4_kw4_sh2_sw2_cin128_cout256", - "op_type": "deconv2d", - "description": "Transposed 2D conv: kh=4 kw=4 stride=(2,2) dilation=(1,1) pad=0 C_in=128 C_out=256", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C_in": { - "type": "const", - "value": 128 - }, - "C_out": { - "type": "const", - "value": 256 - }, - "Kh": { - "type": "const", - "value": 4 - }, - "Kw": { - "type": "const", - "value": 4 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).permute(1, 0, 2, 3)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(2, 2), padding=(0, 0), dilation=(1, 1)).numpy()\n" -} diff --git a/definitions/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin32_cout64.json b/definitions/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin32_cout64.json deleted file mode 100644 index c8046e4e1b9955f23422f39548d87d1226b47219..0000000000000000000000000000000000000000 --- a/definitions/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin32_cout64.json +++ /dev/null @@ -1,105 +0,0 @@ -{ - "name": "deconv2d_kh4_kw4_sh2_sw2_cin32_cout64", - "op_type": "deconv2d", - "description": "Transposed 2D conv: kh=4 kw=4 stride=(2,2) dilation=(1,1) pad=0 C_in=32 C_out=64", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C_in": { - "type": "const", - "value": 32 - }, - "C_out": { - "type": "const", - "value": 64 - }, - "Kh": { - "type": "const", - "value": 4 - }, - "Kw": { - "type": "const", - "value": 4 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).permute(1, 0, 2, 3)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(2, 2), padding=(0, 0), dilation=(1, 1)).numpy()\n" -} diff --git a/definitions/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin512_cout512.json b/definitions/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin512_cout512.json deleted file mode 100644 index 247633a6ff4f1e0b3eefc9ac60dcd295fe317d3a..0000000000000000000000000000000000000000 --- a/definitions/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin512_cout512.json +++ /dev/null @@ -1,105 +0,0 @@ -{ - "name": "deconv2d_kh4_kw4_sh2_sw2_cin512_cout512", - "op_type": "deconv2d", - "description": "Transposed 2D conv: kh=4 kw=4 stride=(2,2) dilation=(1,1) pad=0 C_in=512 C_out=512", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C_in": { - "type": "const", - "value": 512 - }, - "C_out": { - "type": "const", - "value": 512 - }, - "Kh": { - "type": "const", - "value": 4 - }, - "Kw": { - "type": "const", - "value": 4 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).permute(1, 0, 2, 3)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(2, 2), padding=(0, 0), dilation=(1, 1)).numpy()\n" -} diff --git a/definitions/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin64_cout128.json b/definitions/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin64_cout128.json deleted file mode 100644 index 6bd36afac5c269b698d9448f0eba5550b8385250..0000000000000000000000000000000000000000 --- a/definitions/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin64_cout128.json +++ /dev/null @@ -1,105 +0,0 @@ -{ - "name": "deconv2d_kh4_kw4_sh2_sw2_cin64_cout128", - "op_type": "deconv2d", - "description": "Transposed 2D conv: kh=4 kw=4 stride=(2,2) dilation=(1,1) pad=0 C_in=64 C_out=128", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C_in": { - "type": "const", - "value": 64 - }, - "C_out": { - "type": "const", - "value": 128 - }, - "Kh": { - "type": "const", - "value": 4 - }, - "Kw": { - "type": "const", - "value": 4 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C_in", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C_out", - "C_in", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C_out" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C_out", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).permute(1, 0, 2, 3)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(2, 2), padding=(0, 0), dilation=(1, 1)).numpy()\n" -} diff --git a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024.json b/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024.json deleted file mode 100644 index 0f8784dd8314cea42158d7ea55854e7f243022d5..0000000000000000000000000000000000000000 --- a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024", - "op_type": "deconv2d_depthwise", - "description": "Depthwise transposed 2D conv: kh=2 kw=2 stride=(2,2) dilation=(1,1) pad=0 C=1024", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C": { - "type": "const", - "value": 1024 - }, - "Kh": { - "type": "const", - "value": 2 - }, - "Kw": { - "type": "const", - "value": 2 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(2, 2), padding=(0, 0), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c128.json b/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c128.json deleted file mode 100644 index b1cad474dd040b255b48b8110d28add381cd904e..0000000000000000000000000000000000000000 --- a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c128.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c128", - "op_type": "deconv2d_depthwise", - "description": "Depthwise transposed 2D conv: kh=2 kw=2 stride=(2,2) dilation=(1,1) pad=0 C=128", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C": { - "type": "const", - "value": 128 - }, - "Kh": { - "type": "const", - "value": 2 - }, - "Kw": { - "type": "const", - "value": 2 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(2, 2), padding=(0, 0), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c256.json b/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c256.json deleted file mode 100644 index f19cb9f28099e3debba4c90debbd4a973656398b..0000000000000000000000000000000000000000 --- a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c256.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c256", - "op_type": "deconv2d_depthwise", - "description": "Depthwise transposed 2D conv: kh=2 kw=2 stride=(2,2) dilation=(1,1) pad=0 C=256", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C": { - "type": "const", - "value": 256 - }, - "Kh": { - "type": "const", - "value": 2 - }, - "Kw": { - "type": "const", - "value": 2 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(2, 2), padding=(0, 0), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c32.json b/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c32.json deleted file mode 100644 index 0b6c9e9975eab6de13105665e1850de67a1fb046..0000000000000000000000000000000000000000 --- a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c32.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c32", - "op_type": "deconv2d_depthwise", - "description": "Depthwise transposed 2D conv: kh=2 kw=2 stride=(2,2) dilation=(1,1) pad=0 C=32", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C": { - "type": "const", - "value": 32 - }, - "Kh": { - "type": "const", - "value": 2 - }, - "Kw": { - "type": "const", - "value": 2 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(2, 2), padding=(0, 0), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c512.json b/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c512.json deleted file mode 100644 index 98c3c7212fe907861fda9f4247cfc3a5acef06d3..0000000000000000000000000000000000000000 --- a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c512.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c512", - "op_type": "deconv2d_depthwise", - "description": "Depthwise transposed 2D conv: kh=2 kw=2 stride=(2,2) dilation=(1,1) pad=0 C=512", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C": { - "type": "const", - "value": 512 - }, - "Kh": { - "type": "const", - "value": 2 - }, - "Kw": { - "type": "const", - "value": 2 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(2, 2), padding=(0, 0), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c64.json b/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c64.json deleted file mode 100644 index b34039bc6dde29db318e8749151eccb1a110190f..0000000000000000000000000000000000000000 --- a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c64.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c64", - "op_type": "deconv2d_depthwise", - "description": "Depthwise transposed 2D conv: kh=2 kw=2 stride=(2,2) dilation=(1,1) pad=0 C=64", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C": { - "type": "const", - "value": 64 - }, - "Kh": { - "type": "const", - "value": 2 - }, - "Kw": { - "type": "const", - "value": 2 - }, - "Sh": { - "type": "const", - "value": 2 - }, - "Sw": { - "type": "const", - "value": 2 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(2, 2), padding=(0, 0), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c128.json b/definitions/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c128.json deleted file mode 100644 index 58ea264e1dd3759b1bc58ac661f73a858bd050b7..0000000000000000000000000000000000000000 --- a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c128.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "deconv2d_depthwise_kh3_kw3_sh1_sw1_c128", - "op_type": "deconv2d_depthwise", - "description": "Depthwise transposed 2D conv: kh=3 kw=3 stride=(1,1) dilation=(1,1) pad=0 C=128", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C": { - "type": "const", - "value": 128 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(1, 1), padding=(0, 0), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048.json b/definitions/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048.json deleted file mode 100644 index 2238abe0ba1d78e48c8883a715d954174a8874fb..0000000000000000000000000000000000000000 --- a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048", - "op_type": "deconv2d_depthwise", - "description": "Depthwise transposed 2D conv: kh=3 kw=3 stride=(1,1) dilation=(1,1) pad=0 C=2048", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C": { - "type": "const", - "value": 2048 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(1, 1), padding=(0, 0), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c32.json b/definitions/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c32.json deleted file mode 100644 index 99062a29d9cb109a9a85aff2a32fce7ff876d35e..0000000000000000000000000000000000000000 --- a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c32.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "deconv2d_depthwise_kh3_kw3_sh1_sw1_c32", - "op_type": "deconv2d_depthwise", - "description": "Depthwise transposed 2D conv: kh=3 kw=3 stride=(1,1) dilation=(1,1) pad=0 C=32", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C": { - "type": "const", - "value": 32 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(1, 1), padding=(0, 0), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c512.json b/definitions/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c512.json deleted file mode 100644 index 93c9f32ef7d3ae79a5413219ac6d171bafe2f844..0000000000000000000000000000000000000000 --- a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c512.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "deconv2d_depthwise_kh3_kw3_sh1_sw1_c512", - "op_type": "deconv2d_depthwise", - "description": "Depthwise transposed 2D conv: kh=3 kw=3 stride=(1,1) dilation=(1,1) pad=0 C=512", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C": { - "type": "const", - "value": 512 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(1, 1), padding=(0, 0), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c64.json b/definitions/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c64.json deleted file mode 100644 index a09831634cc732ab278eb6cd047505d31027bd2a..0000000000000000000000000000000000000000 --- a/definitions/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c64.json +++ /dev/null @@ -1,100 +0,0 @@ -{ - "name": "deconv2d_depthwise_kh3_kw3_sh1_sw1_c64", - "op_type": "deconv2d_depthwise", - "description": "Depthwise transposed 2D conv: kh=3 kw=3 stride=(1,1) dilation=(1,1) pad=0 C=64", - "tags": [ - "status:draft" - ], - "axes": { - "N": { - "type": "var" - }, - "H": { - "type": "var", - "parent": "N" - }, - "W": { - "type": "var", - "parent": "N" - }, - "H_out": { - "type": "var", - "parent": "N", - "description": "Derived: (H - 1) * sh + kh" - }, - "W_out": { - "type": "var", - "parent": "N", - "description": "Derived: (W - 1) * sw + kw" - }, - "C": { - "type": "const", - "value": 64 - }, - "Kh": { - "type": "const", - "value": 3 - }, - "Kw": { - "type": "const", - "value": 3 - }, - "Sh": { - "type": "const", - "value": 1 - }, - "Sw": { - "type": "const", - "value": 1 - }, - "Dh": { - "type": "const", - "value": 1 - }, - "Dw": { - "type": "const", - "value": 1 - } - }, - "inputs": { - "input": { - "shape": [ - "N", - "C", - "H", - "W" - ], - "dtype": "float32" - }, - "weight": { - "shape": [ - "C", - "Kh", - "Kw" - ], - "dtype": "float32" - }, - "bias": { - "shape": [ - "C" - ], - "dtype": "float32" - } - }, - "outputs": { - "output": { - "shape": [ - "N", - "C", - "H_out", - "W_out" - ], - "dtype": "float32" - } - }, - "constraints": [ - "H_out == (H - 1) * Sh + Kh", - "W_out == (W - 1) * Sw + Kw" - ], - "reference": "import torch\nimport torch.nn.functional as F\n\ndef run(input, weight, bias):\n n, c = input.shape[0], input.shape[1]\n x = torch.from_numpy(input)\n w = torch.from_numpy(weight).unsqueeze(1)\n b = torch.from_numpy(bias)\n return F.conv_transpose2d(x, w, b, stride=(1, 1), padding=(0, 0), dilation=(1, 1), groups=c).numpy()\n" -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin1024_cout1024.json b/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin1024_cout1024.json deleted file mode 100644 index 943bf60186e86a99ffb393015456cb75461e9cd4..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin1024_cout1024.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin1024_cout1024", - "definition": "conv1d_kw1_sw1_dw1_p0_cin1024_cout1024", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution1D_arm wrapper for conv1d_kw1_sw1_dw1_p0_cin1024_cout1024. Same kernel.cpp content across all conv1d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv1d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv1d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv1d Solution;\n// identical across all conv1d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in conv1d.h by delegating to\n// ncnn::Convolution1D_arm. The shim armbench_entry_conv1d (binding.cpp) pads\n// the 2D input (h=C_in, w=W) along w, allocates the 2D output\n// (h=C_out, w=out_w), then calls into this symbol. Const params come from the\n// Definition's const axes and are passed in as runtime args, so this kernel.cpp\n// is identical across all conv1d Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches the\n// conv2d baseline semantics.\n\n#include \"conv1d.h\"\n#include \"convolution1d_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution1D_arm conv;\n // conv1d output is a 2D Mat (w=out_w, h=out_c) \u2192 num_output is its height.\n conv.num_output = top_blob.h;\n conv.kernel_w = kernel_w;\n conv.stride_w = stride_w;\n conv.dilation_w = dilation_w;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution1D_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (2D, c==1).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv1d.h", - "content": "// bench/compile/builders/ncnn_harness/conv1d.h\n//\n// Calling-convention contract for every conv1d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv1d/ must define the\n// `ncnn::convolution1d_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv1d` in conv1d.cpp pads the 2D input via\n// copy_make_border, allocates the output 2D Mat, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolution1d.h:50-54 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV1D_H\n#define ARMBENCH_NCNN_HARNESS_CONV1D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded 2D input, layout (h=C_in, w=W), float32\n// top_blob: pre-allocated 2D output, layout (h=C_out, w=W_out), float32\n// weight_data: flat 1D Mat of size C_out * C_in * Kw\n// bias_data: flat 1D Mat of size C_out, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV1D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv1d \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv1d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolution1d_kernel (defined in kernel.cpp).\n//\n// The input is a 2D ncnn::Mat with w=sequence_length and h=channels.\n\n#include \"conv1d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 1024;\nconstexpr int kernel_w = 1;\nconstexpr int stride_w = 1;\nconstexpr int dilation_w = 1;\nconstexpr int pad_left = 0;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv1d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 2D input (h=C_in, w=W)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C_out*C_in*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (copy_make_border on 2D mat: top=0, bottom=0, left=pad, right=pad)\n ncnn::Mat bordered;\n if (pad_left > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n 0, 0, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims (matches Convolution1D::forward in convolution1d.h:107)\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bordered.w - ext_kw) / stride_w + 1;\n\n // 3. Allocate top as 2D Mat (w=out_w, h=out_c) \u2014 matches Convolution1D::forward:109\n top.create(out_w, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution1d_kernel(\n bordered, top, weight, bias,\n kernel_w, stride_w, dilation_w,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin128_cout512.json b/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin128_cout512.json deleted file mode 100644 index b3712b8b433bd87eff1b0b45afc483250e6629be..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin128_cout512.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin128_cout512", - "definition": "conv1d_kw1_sw1_dw1_p0_cin128_cout512", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution1D_arm wrapper for conv1d_kw1_sw1_dw1_p0_cin128_cout512. Same kernel.cpp content across all conv1d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv1d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv1d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv1d Solution;\n// identical across all conv1d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in conv1d.h by delegating to\n// ncnn::Convolution1D_arm. The shim armbench_entry_conv1d (binding.cpp) pads\n// the 2D input (h=C_in, w=W) along w, allocates the 2D output\n// (h=C_out, w=out_w), then calls into this symbol. Const params come from the\n// Definition's const axes and are passed in as runtime args, so this kernel.cpp\n// is identical across all conv1d Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches the\n// conv2d baseline semantics.\n\n#include \"conv1d.h\"\n#include \"convolution1d_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution1D_arm conv;\n // conv1d output is a 2D Mat (w=out_w, h=out_c) \u2192 num_output is its height.\n conv.num_output = top_blob.h;\n conv.kernel_w = kernel_w;\n conv.stride_w = stride_w;\n conv.dilation_w = dilation_w;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution1D_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (2D, c==1).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv1d.h", - "content": "// bench/compile/builders/ncnn_harness/conv1d.h\n//\n// Calling-convention contract for every conv1d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv1d/ must define the\n// `ncnn::convolution1d_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv1d` in conv1d.cpp pads the 2D input via\n// copy_make_border, allocates the output 2D Mat, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolution1d.h:50-54 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV1D_H\n#define ARMBENCH_NCNN_HARNESS_CONV1D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded 2D input, layout (h=C_in, w=W), float32\n// top_blob: pre-allocated 2D output, layout (h=C_out, w=W_out), float32\n// weight_data: flat 1D Mat of size C_out * C_in * Kw\n// bias_data: flat 1D Mat of size C_out, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV1D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv1d \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv1d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolution1d_kernel (defined in kernel.cpp).\n//\n// The input is a 2D ncnn::Mat with w=sequence_length and h=channels.\n\n#include \"conv1d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 512;\nconstexpr int kernel_w = 1;\nconstexpr int stride_w = 1;\nconstexpr int dilation_w = 1;\nconstexpr int pad_left = 0;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv1d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 2D input (h=C_in, w=W)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C_out*C_in*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (copy_make_border on 2D mat: top=0, bottom=0, left=pad, right=pad)\n ncnn::Mat bordered;\n if (pad_left > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n 0, 0, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims (matches Convolution1D::forward in convolution1d.h:107)\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bordered.w - ext_kw) / stride_w + 1;\n\n // 3. Allocate top as 2D Mat (w=out_w, h=out_c) \u2014 matches Convolution1D::forward:109\n top.create(out_w, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution1d_kernel(\n bordered, top, weight, bias,\n kernel_w, stride_w, dilation_w,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin256_cout1024.json b/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin256_cout1024.json deleted file mode 100644 index b44d7483151c29c1984963a3afe17a8ccb48a4e8..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin256_cout1024.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin256_cout1024", - "definition": "conv1d_kw1_sw1_dw1_p0_cin256_cout1024", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution1D_arm wrapper for conv1d_kw1_sw1_dw1_p0_cin256_cout1024. Same kernel.cpp content across all conv1d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv1d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv1d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv1d Solution;\n// identical across all conv1d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in conv1d.h by delegating to\n// ncnn::Convolution1D_arm. The shim armbench_entry_conv1d (binding.cpp) pads\n// the 2D input (h=C_in, w=W) along w, allocates the 2D output\n// (h=C_out, w=out_w), then calls into this symbol. Const params come from the\n// Definition's const axes and are passed in as runtime args, so this kernel.cpp\n// is identical across all conv1d Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches the\n// conv2d baseline semantics.\n\n#include \"conv1d.h\"\n#include \"convolution1d_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution1D_arm conv;\n // conv1d output is a 2D Mat (w=out_w, h=out_c) \u2192 num_output is its height.\n conv.num_output = top_blob.h;\n conv.kernel_w = kernel_w;\n conv.stride_w = stride_w;\n conv.dilation_w = dilation_w;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution1D_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (2D, c==1).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv1d.h", - "content": "// bench/compile/builders/ncnn_harness/conv1d.h\n//\n// Calling-convention contract for every conv1d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv1d/ must define the\n// `ncnn::convolution1d_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv1d` in conv1d.cpp pads the 2D input via\n// copy_make_border, allocates the output 2D Mat, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolution1d.h:50-54 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV1D_H\n#define ARMBENCH_NCNN_HARNESS_CONV1D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded 2D input, layout (h=C_in, w=W), float32\n// top_blob: pre-allocated 2D output, layout (h=C_out, w=W_out), float32\n// weight_data: flat 1D Mat of size C_out * C_in * Kw\n// bias_data: flat 1D Mat of size C_out, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV1D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv1d \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv1d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolution1d_kernel (defined in kernel.cpp).\n//\n// The input is a 2D ncnn::Mat with w=sequence_length and h=channels.\n\n#include \"conv1d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 1024;\nconstexpr int kernel_w = 1;\nconstexpr int stride_w = 1;\nconstexpr int dilation_w = 1;\nconstexpr int pad_left = 0;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv1d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 2D input (h=C_in, w=W)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C_out*C_in*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (copy_make_border on 2D mat: top=0, bottom=0, left=pad, right=pad)\n ncnn::Mat bordered;\n if (pad_left > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n 0, 0, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims (matches Convolution1D::forward in convolution1d.h:107)\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bordered.w - ext_kw) / stride_w + 1;\n\n // 3. Allocate top as 2D Mat (w=out_w, h=out_c) \u2014 matches Convolution1D::forward:109\n top.create(out_w, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution1d_kernel(\n bordered, top, weight, bias,\n kernel_w, stride_w, dilation_w,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin512_cout2048.json b/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin512_cout2048.json deleted file mode 100644 index 0ee0f9c3968a5b1c2690d375dd2530b4838aaad6..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin512_cout2048.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin512_cout2048", - "definition": "conv1d_kw1_sw1_dw1_p0_cin512_cout2048", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution1D_arm wrapper for conv1d_kw1_sw1_dw1_p0_cin512_cout2048. Same kernel.cpp content across all conv1d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv1d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv1d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv1d Solution;\n// identical across all conv1d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in conv1d.h by delegating to\n// ncnn::Convolution1D_arm. The shim armbench_entry_conv1d (binding.cpp) pads\n// the 2D input (h=C_in, w=W) along w, allocates the 2D output\n// (h=C_out, w=out_w), then calls into this symbol. Const params come from the\n// Definition's const axes and are passed in as runtime args, so this kernel.cpp\n// is identical across all conv1d Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches the\n// conv2d baseline semantics.\n\n#include \"conv1d.h\"\n#include \"convolution1d_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution1D_arm conv;\n // conv1d output is a 2D Mat (w=out_w, h=out_c) \u2192 num_output is its height.\n conv.num_output = top_blob.h;\n conv.kernel_w = kernel_w;\n conv.stride_w = stride_w;\n conv.dilation_w = dilation_w;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution1D_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (2D, c==1).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv1d.h", - "content": "// bench/compile/builders/ncnn_harness/conv1d.h\n//\n// Calling-convention contract for every conv1d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv1d/ must define the\n// `ncnn::convolution1d_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv1d` in conv1d.cpp pads the 2D input via\n// copy_make_border, allocates the output 2D Mat, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolution1d.h:50-54 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV1D_H\n#define ARMBENCH_NCNN_HARNESS_CONV1D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded 2D input, layout (h=C_in, w=W), float32\n// top_blob: pre-allocated 2D output, layout (h=C_out, w=W_out), float32\n// weight_data: flat 1D Mat of size C_out * C_in * Kw\n// bias_data: flat 1D Mat of size C_out, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV1D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv1d \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv1d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolution1d_kernel (defined in kernel.cpp).\n//\n// The input is a 2D ncnn::Mat with w=sequence_length and h=channels.\n\n#include \"conv1d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 2048;\nconstexpr int kernel_w = 1;\nconstexpr int stride_w = 1;\nconstexpr int dilation_w = 1;\nconstexpr int pad_left = 0;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv1d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 2D input (h=C_in, w=W)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C_out*C_in*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (copy_make_border on 2D mat: top=0, bottom=0, left=pad, right=pad)\n ncnn::Mat bordered;\n if (pad_left > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n 0, 0, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims (matches Convolution1D::forward in convolution1d.h:107)\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bordered.w - ext_kw) / stride_w + 1;\n\n // 3. Allocate top as 2D Mat (w=out_w, h=out_c) \u2014 matches Convolution1D::forward:109\n top.create(out_w, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution1d_kernel(\n bordered, top, weight, bias,\n kernel_w, stride_w, dilation_w,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout256.json b/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout256.json deleted file mode 100644 index 3de1a962b698394d6a09975123b5351ffd6d41cb..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout256.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin64_cout256", - "definition": "conv1d_kw1_sw1_dw1_p0_cin64_cout256", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution1D_arm wrapper for conv1d_kw1_sw1_dw1_p0_cin64_cout256. Same kernel.cpp content across all conv1d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv1d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv1d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv1d Solution;\n// identical across all conv1d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in conv1d.h by delegating to\n// ncnn::Convolution1D_arm. The shim armbench_entry_conv1d (binding.cpp) pads\n// the 2D input (h=C_in, w=W) along w, allocates the 2D output\n// (h=C_out, w=out_w), then calls into this symbol. Const params come from the\n// Definition's const axes and are passed in as runtime args, so this kernel.cpp\n// is identical across all conv1d Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches the\n// conv2d baseline semantics.\n\n#include \"conv1d.h\"\n#include \"convolution1d_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution1D_arm conv;\n // conv1d output is a 2D Mat (w=out_w, h=out_c) \u2192 num_output is its height.\n conv.num_output = top_blob.h;\n conv.kernel_w = kernel_w;\n conv.stride_w = stride_w;\n conv.dilation_w = dilation_w;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution1D_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (2D, c==1).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv1d.h", - "content": "// bench/compile/builders/ncnn_harness/conv1d.h\n//\n// Calling-convention contract for every conv1d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv1d/ must define the\n// `ncnn::convolution1d_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv1d` in conv1d.cpp pads the 2D input via\n// copy_make_border, allocates the output 2D Mat, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolution1d.h:50-54 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV1D_H\n#define ARMBENCH_NCNN_HARNESS_CONV1D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded 2D input, layout (h=C_in, w=W), float32\n// top_blob: pre-allocated 2D output, layout (h=C_out, w=W_out), float32\n// weight_data: flat 1D Mat of size C_out * C_in * Kw\n// bias_data: flat 1D Mat of size C_out, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV1D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv1d \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv1d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolution1d_kernel (defined in kernel.cpp).\n//\n// The input is a 2D ncnn::Mat with w=sequence_length and h=channels.\n\n#include \"conv1d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 256;\nconstexpr int kernel_w = 1;\nconstexpr int stride_w = 1;\nconstexpr int dilation_w = 1;\nconstexpr int pad_left = 0;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv1d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 2D input (h=C_in, w=W)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C_out*C_in*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (copy_make_border on 2D mat: top=0, bottom=0, left=pad, right=pad)\n ncnn::Mat bordered;\n if (pad_left > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n 0, 0, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims (matches Convolution1D::forward in convolution1d.h:107)\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bordered.w - ext_kw) / stride_w + 1;\n\n // 3. Allocate top as 2D Mat (w=out_w, h=out_c) \u2014 matches Convolution1D::forward:109\n top.create(out_w, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution1d_kernel(\n bordered, top, weight, bias,\n kernel_w, stride_w, dilation_w,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout64.json b/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout64.json deleted file mode 100644 index 1b665ffbc8441cd430d5432fd4462d922d4977a1..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout64.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin64_cout64", - "definition": "conv1d_kw1_sw1_dw1_p0_cin64_cout64", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution1D_arm wrapper for conv1d_kw1_sw1_dw1_p0_cin64_cout64. Same kernel.cpp content across all conv1d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv1d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv1d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv1d Solution;\n// identical across all conv1d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in conv1d.h by delegating to\n// ncnn::Convolution1D_arm. The shim armbench_entry_conv1d (binding.cpp) pads\n// the 2D input (h=C_in, w=W) along w, allocates the 2D output\n// (h=C_out, w=out_w), then calls into this symbol. Const params come from the\n// Definition's const axes and are passed in as runtime args, so this kernel.cpp\n// is identical across all conv1d Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches the\n// conv2d baseline semantics.\n\n#include \"conv1d.h\"\n#include \"convolution1d_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution1D_arm conv;\n // conv1d output is a 2D Mat (w=out_w, h=out_c) \u2192 num_output is its height.\n conv.num_output = top_blob.h;\n conv.kernel_w = kernel_w;\n conv.stride_w = stride_w;\n conv.dilation_w = dilation_w;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution1D_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (2D, c==1).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv1d.h", - "content": "// bench/compile/builders/ncnn_harness/conv1d.h\n//\n// Calling-convention contract for every conv1d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv1d/ must define the\n// `ncnn::convolution1d_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv1d` in conv1d.cpp pads the 2D input via\n// copy_make_border, allocates the output 2D Mat, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolution1d.h:50-54 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV1D_H\n#define ARMBENCH_NCNN_HARNESS_CONV1D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded 2D input, layout (h=C_in, w=W), float32\n// top_blob: pre-allocated 2D output, layout (h=C_out, w=W_out), float32\n// weight_data: flat 1D Mat of size C_out * C_in * Kw\n// bias_data: flat 1D Mat of size C_out, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV1D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv1d \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv1d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolution1d_kernel (defined in kernel.cpp).\n//\n// The input is a 2D ncnn::Mat with w=sequence_length and h=channels.\n\n#include \"conv1d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 64;\nconstexpr int kernel_w = 1;\nconstexpr int stride_w = 1;\nconstexpr int dilation_w = 1;\nconstexpr int pad_left = 0;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv1d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 2D input (h=C_in, w=W)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C_out*C_in*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (copy_make_border on 2D mat: top=0, bottom=0, left=pad, right=pad)\n ncnn::Mat bordered;\n if (pad_left > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n 0, 0, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims (matches Convolution1D::forward in convolution1d.h:107)\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bordered.w - ext_kw) / stride_w + 1;\n\n // 3. Allocate top as 2D Mat (w=out_w, h=out_c) \u2014 matches Convolution1D::forward:109\n top.create(out_w, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution1d_kernel(\n bordered, top, weight, bias,\n kernel_w, stride_w, dilation_w,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin1024_cout1024.json b/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin1024_cout1024.json deleted file mode 100644 index de8479cd97175cc8bd33ce0d5493e52840982118..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin1024_cout1024.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin1024_cout1024", - "definition": "conv1d_kw3_sw1_dw1_p1_cin1024_cout1024", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution1D_arm wrapper for conv1d_kw3_sw1_dw1_p1_cin1024_cout1024. Same kernel.cpp content across all conv1d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv1d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv1d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv1d Solution;\n// identical across all conv1d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in conv1d.h by delegating to\n// ncnn::Convolution1D_arm. The shim armbench_entry_conv1d (binding.cpp) pads\n// the 2D input (h=C_in, w=W) along w, allocates the 2D output\n// (h=C_out, w=out_w), then calls into this symbol. Const params come from the\n// Definition's const axes and are passed in as runtime args, so this kernel.cpp\n// is identical across all conv1d Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches the\n// conv2d baseline semantics.\n\n#include \"conv1d.h\"\n#include \"convolution1d_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution1D_arm conv;\n // conv1d output is a 2D Mat (w=out_w, h=out_c) \u2192 num_output is its height.\n conv.num_output = top_blob.h;\n conv.kernel_w = kernel_w;\n conv.stride_w = stride_w;\n conv.dilation_w = dilation_w;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution1D_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (2D, c==1).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv1d.h", - "content": "// bench/compile/builders/ncnn_harness/conv1d.h\n//\n// Calling-convention contract for every conv1d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv1d/ must define the\n// `ncnn::convolution1d_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv1d` in conv1d.cpp pads the 2D input via\n// copy_make_border, allocates the output 2D Mat, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolution1d.h:50-54 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV1D_H\n#define ARMBENCH_NCNN_HARNESS_CONV1D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded 2D input, layout (h=C_in, w=W), float32\n// top_blob: pre-allocated 2D output, layout (h=C_out, w=W_out), float32\n// weight_data: flat 1D Mat of size C_out * C_in * Kw\n// bias_data: flat 1D Mat of size C_out, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV1D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv1d \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv1d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolution1d_kernel (defined in kernel.cpp).\n//\n// The input is a 2D ncnn::Mat with w=sequence_length and h=channels.\n\n#include \"conv1d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 1024;\nconstexpr int kernel_w = 3;\nconstexpr int stride_w = 1;\nconstexpr int dilation_w = 1;\nconstexpr int pad_left = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv1d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 2D input (h=C_in, w=W)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C_out*C_in*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (copy_make_border on 2D mat: top=0, bottom=0, left=pad, right=pad)\n ncnn::Mat bordered;\n if (pad_left > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n 0, 0, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims (matches Convolution1D::forward in convolution1d.h:107)\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bordered.w - ext_kw) / stride_w + 1;\n\n // 3. Allocate top as 2D Mat (w=out_w, h=out_c) \u2014 matches Convolution1D::forward:109\n top.create(out_w, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution1d_kernel(\n bordered, top, weight, bias,\n kernel_w, stride_w, dilation_w,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin128_cout256.json b/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin128_cout256.json deleted file mode 100644 index 3640803bd721e4a294739c251bd109b8145c2fe1..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin128_cout256.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin128_cout256", - "definition": "conv1d_kw3_sw1_dw1_p1_cin128_cout256", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution1D_arm wrapper for conv1d_kw3_sw1_dw1_p1_cin128_cout256. Same kernel.cpp content across all conv1d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv1d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv1d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv1d Solution;\n// identical across all conv1d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in conv1d.h by delegating to\n// ncnn::Convolution1D_arm. The shim armbench_entry_conv1d (binding.cpp) pads\n// the 2D input (h=C_in, w=W) along w, allocates the 2D output\n// (h=C_out, w=out_w), then calls into this symbol. Const params come from the\n// Definition's const axes and are passed in as runtime args, so this kernel.cpp\n// is identical across all conv1d Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches the\n// conv2d baseline semantics.\n\n#include \"conv1d.h\"\n#include \"convolution1d_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution1D_arm conv;\n // conv1d output is a 2D Mat (w=out_w, h=out_c) \u2192 num_output is its height.\n conv.num_output = top_blob.h;\n conv.kernel_w = kernel_w;\n conv.stride_w = stride_w;\n conv.dilation_w = dilation_w;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution1D_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (2D, c==1).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv1d.h", - "content": "// bench/compile/builders/ncnn_harness/conv1d.h\n//\n// Calling-convention contract for every conv1d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv1d/ must define the\n// `ncnn::convolution1d_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv1d` in conv1d.cpp pads the 2D input via\n// copy_make_border, allocates the output 2D Mat, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolution1d.h:50-54 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV1D_H\n#define ARMBENCH_NCNN_HARNESS_CONV1D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded 2D input, layout (h=C_in, w=W), float32\n// top_blob: pre-allocated 2D output, layout (h=C_out, w=W_out), float32\n// weight_data: flat 1D Mat of size C_out * C_in * Kw\n// bias_data: flat 1D Mat of size C_out, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV1D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv1d \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv1d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolution1d_kernel (defined in kernel.cpp).\n//\n// The input is a 2D ncnn::Mat with w=sequence_length and h=channels.\n\n#include \"conv1d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 256;\nconstexpr int kernel_w = 3;\nconstexpr int stride_w = 1;\nconstexpr int dilation_w = 1;\nconstexpr int pad_left = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv1d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 2D input (h=C_in, w=W)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C_out*C_in*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (copy_make_border on 2D mat: top=0, bottom=0, left=pad, right=pad)\n ncnn::Mat bordered;\n if (pad_left > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n 0, 0, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims (matches Convolution1D::forward in convolution1d.h:107)\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bordered.w - ext_kw) / stride_w + 1;\n\n // 3. Allocate top as 2D Mat (w=out_w, h=out_c) \u2014 matches Convolution1D::forward:109\n top.create(out_w, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution1d_kernel(\n bordered, top, weight, bias,\n kernel_w, stride_w, dilation_w,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin256_cout512.json b/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin256_cout512.json deleted file mode 100644 index 686b9653f1041c3ad071a3e930cda8dd7ccce0cd..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin256_cout512.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin256_cout512", - "definition": "conv1d_kw3_sw1_dw1_p1_cin256_cout512", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution1D_arm wrapper for conv1d_kw3_sw1_dw1_p1_cin256_cout512. Same kernel.cpp content across all conv1d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv1d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv1d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv1d Solution;\n// identical across all conv1d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in conv1d.h by delegating to\n// ncnn::Convolution1D_arm. The shim armbench_entry_conv1d (binding.cpp) pads\n// the 2D input (h=C_in, w=W) along w, allocates the 2D output\n// (h=C_out, w=out_w), then calls into this symbol. Const params come from the\n// Definition's const axes and are passed in as runtime args, so this kernel.cpp\n// is identical across all conv1d Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches the\n// conv2d baseline semantics.\n\n#include \"conv1d.h\"\n#include \"convolution1d_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution1D_arm conv;\n // conv1d output is a 2D Mat (w=out_w, h=out_c) \u2192 num_output is its height.\n conv.num_output = top_blob.h;\n conv.kernel_w = kernel_w;\n conv.stride_w = stride_w;\n conv.dilation_w = dilation_w;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution1D_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (2D, c==1).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv1d.h", - "content": "// bench/compile/builders/ncnn_harness/conv1d.h\n//\n// Calling-convention contract for every conv1d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv1d/ must define the\n// `ncnn::convolution1d_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv1d` in conv1d.cpp pads the 2D input via\n// copy_make_border, allocates the output 2D Mat, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolution1d.h:50-54 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV1D_H\n#define ARMBENCH_NCNN_HARNESS_CONV1D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded 2D input, layout (h=C_in, w=W), float32\n// top_blob: pre-allocated 2D output, layout (h=C_out, w=W_out), float32\n// weight_data: flat 1D Mat of size C_out * C_in * Kw\n// bias_data: flat 1D Mat of size C_out, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV1D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv1d \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv1d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolution1d_kernel (defined in kernel.cpp).\n//\n// The input is a 2D ncnn::Mat with w=sequence_length and h=channels.\n\n#include \"conv1d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 512;\nconstexpr int kernel_w = 3;\nconstexpr int stride_w = 1;\nconstexpr int dilation_w = 1;\nconstexpr int pad_left = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv1d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 2D input (h=C_in, w=W)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C_out*C_in*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (copy_make_border on 2D mat: top=0, bottom=0, left=pad, right=pad)\n ncnn::Mat bordered;\n if (pad_left > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n 0, 0, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims (matches Convolution1D::forward in convolution1d.h:107)\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bordered.w - ext_kw) / stride_w + 1;\n\n // 3. Allocate top as 2D Mat (w=out_w, h=out_c) \u2014 matches Convolution1D::forward:109\n top.create(out_w, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution1d_kernel(\n bordered, top, weight, bias,\n kernel_w, stride_w, dilation_w,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin3_cout64.json b/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin3_cout64.json deleted file mode 100644 index f6c8740076b4a71bbd449c5050647996a5266474..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin3_cout64.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin3_cout64", - "definition": "conv1d_kw3_sw1_dw1_p1_cin3_cout64", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution1D_arm wrapper for conv1d_kw3_sw1_dw1_p1_cin3_cout64. Same kernel.cpp content across all conv1d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv1d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv1d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv1d Solution;\n// identical across all conv1d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in conv1d.h by delegating to\n// ncnn::Convolution1D_arm. The shim armbench_entry_conv1d (binding.cpp) pads\n// the 2D input (h=C_in, w=W) along w, allocates the 2D output\n// (h=C_out, w=out_w), then calls into this symbol. Const params come from the\n// Definition's const axes and are passed in as runtime args, so this kernel.cpp\n// is identical across all conv1d Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches the\n// conv2d baseline semantics.\n\n#include \"conv1d.h\"\n#include \"convolution1d_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution1D_arm conv;\n // conv1d output is a 2D Mat (w=out_w, h=out_c) \u2192 num_output is its height.\n conv.num_output = top_blob.h;\n conv.kernel_w = kernel_w;\n conv.stride_w = stride_w;\n conv.dilation_w = dilation_w;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution1D_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (2D, c==1).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv1d.h", - "content": "// bench/compile/builders/ncnn_harness/conv1d.h\n//\n// Calling-convention contract for every conv1d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv1d/ must define the\n// `ncnn::convolution1d_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv1d` in conv1d.cpp pads the 2D input via\n// copy_make_border, allocates the output 2D Mat, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolution1d.h:50-54 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV1D_H\n#define ARMBENCH_NCNN_HARNESS_CONV1D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded 2D input, layout (h=C_in, w=W), float32\n// top_blob: pre-allocated 2D output, layout (h=C_out, w=W_out), float32\n// weight_data: flat 1D Mat of size C_out * C_in * Kw\n// bias_data: flat 1D Mat of size C_out, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV1D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv1d \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv1d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolution1d_kernel (defined in kernel.cpp).\n//\n// The input is a 2D ncnn::Mat with w=sequence_length and h=channels.\n\n#include \"conv1d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 64;\nconstexpr int kernel_w = 3;\nconstexpr int stride_w = 1;\nconstexpr int dilation_w = 1;\nconstexpr int pad_left = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv1d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 2D input (h=C_in, w=W)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C_out*C_in*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (copy_make_border on 2D mat: top=0, bottom=0, left=pad, right=pad)\n ncnn::Mat bordered;\n if (pad_left > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n 0, 0, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims (matches Convolution1D::forward in convolution1d.h:107)\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bordered.w - ext_kw) / stride_w + 1;\n\n // 3. Allocate top as 2D Mat (w=out_w, h=out_c) \u2014 matches Convolution1D::forward:109\n top.create(out_w, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution1d_kernel(\n bordered, top, weight, bias,\n kernel_w, stride_w, dilation_w,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin512_cout1024.json b/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin512_cout1024.json deleted file mode 100644 index 92a6d31141d6e6f07c9d99f7d76f7e82d1505b97..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin512_cout1024.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin512_cout1024", - "definition": "conv1d_kw3_sw1_dw1_p1_cin512_cout1024", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution1D_arm wrapper for conv1d_kw3_sw1_dw1_p1_cin512_cout1024. Same kernel.cpp content across all conv1d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv1d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv1d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv1d Solution;\n// identical across all conv1d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in conv1d.h by delegating to\n// ncnn::Convolution1D_arm. The shim armbench_entry_conv1d (binding.cpp) pads\n// the 2D input (h=C_in, w=W) along w, allocates the 2D output\n// (h=C_out, w=out_w), then calls into this symbol. Const params come from the\n// Definition's const axes and are passed in as runtime args, so this kernel.cpp\n// is identical across all conv1d Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches the\n// conv2d baseline semantics.\n\n#include \"conv1d.h\"\n#include \"convolution1d_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution1D_arm conv;\n // conv1d output is a 2D Mat (w=out_w, h=out_c) \u2192 num_output is its height.\n conv.num_output = top_blob.h;\n conv.kernel_w = kernel_w;\n conv.stride_w = stride_w;\n conv.dilation_w = dilation_w;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution1D_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (2D, c==1).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv1d.h", - "content": "// bench/compile/builders/ncnn_harness/conv1d.h\n//\n// Calling-convention contract for every conv1d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv1d/ must define the\n// `ncnn::convolution1d_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv1d` in conv1d.cpp pads the 2D input via\n// copy_make_border, allocates the output 2D Mat, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolution1d.h:50-54 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV1D_H\n#define ARMBENCH_NCNN_HARNESS_CONV1D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded 2D input, layout (h=C_in, w=W), float32\n// top_blob: pre-allocated 2D output, layout (h=C_out, w=W_out), float32\n// weight_data: flat 1D Mat of size C_out * C_in * Kw\n// bias_data: flat 1D Mat of size C_out, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV1D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv1d \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv1d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolution1d_kernel (defined in kernel.cpp).\n//\n// The input is a 2D ncnn::Mat with w=sequence_length and h=channels.\n\n#include \"conv1d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 1024;\nconstexpr int kernel_w = 3;\nconstexpr int stride_w = 1;\nconstexpr int dilation_w = 1;\nconstexpr int pad_left = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv1d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 2D input (h=C_in, w=W)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C_out*C_in*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (copy_make_border on 2D mat: top=0, bottom=0, left=pad, right=pad)\n ncnn::Mat bordered;\n if (pad_left > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n 0, 0, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims (matches Convolution1D::forward in convolution1d.h:107)\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bordered.w - ext_kw) / stride_w + 1;\n\n // 3. Allocate top as 2D Mat (w=out_w, h=out_c) \u2014 matches Convolution1D::forward:109\n top.create(out_w, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution1d_kernel(\n bordered, top, weight, bias,\n kernel_w, stride_w, dilation_w,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin64_cout128.json b/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin64_cout128.json deleted file mode 100644 index 15b92c75c7d93b742db58ea83ce96f9ce9d4c3e9..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv1d/conv1d_kw3_sw1_dw1_p1_cin64_cout128.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin64_cout128", - "definition": "conv1d_kw3_sw1_dw1_p1_cin64_cout128", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution1D_arm wrapper for conv1d_kw3_sw1_dw1_p1_cin64_cout128. Same kernel.cpp content across all conv1d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv1d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv1d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv1d Solution;\n// identical across all conv1d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in conv1d.h by delegating to\n// ncnn::Convolution1D_arm. The shim armbench_entry_conv1d (binding.cpp) pads\n// the 2D input (h=C_in, w=W) along w, allocates the 2D output\n// (h=C_out, w=out_w), then calls into this symbol. Const params come from the\n// Definition's const axes and are passed in as runtime args, so this kernel.cpp\n// is identical across all conv1d Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches the\n// conv2d baseline semantics.\n\n#include \"conv1d.h\"\n#include \"convolution1d_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution1D_arm conv;\n // conv1d output is a 2D Mat (w=out_w, h=out_c) \u2192 num_output is its height.\n conv.num_output = top_blob.h;\n conv.kernel_w = kernel_w;\n conv.stride_w = stride_w;\n conv.dilation_w = dilation_w;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution1D_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (2D, c==1).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv1d.h", - "content": "// bench/compile/builders/ncnn_harness/conv1d.h\n//\n// Calling-convention contract for every conv1d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv1d/ must define the\n// `ncnn::convolution1d_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv1d` in conv1d.cpp pads the 2D input via\n// copy_make_border, allocates the output 2D Mat, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolution1d.h:50-54 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV1D_H\n#define ARMBENCH_NCNN_HARNESS_CONV1D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded 2D input, layout (h=C_in, w=W), float32\n// top_blob: pre-allocated 2D output, layout (h=C_out, w=W_out), float32\n// weight_data: flat 1D Mat of size C_out * C_in * Kw\n// bias_data: flat 1D Mat of size C_out, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution1d_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int stride_w, int dilation_w,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV1D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv1d \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv1d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolution1d_kernel (defined in kernel.cpp).\n//\n// The input is a 2D ncnn::Mat with w=sequence_length and h=channels.\n\n#include \"conv1d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 128;\nconstexpr int kernel_w = 3;\nconstexpr int stride_w = 1;\nconstexpr int dilation_w = 1;\nconstexpr int pad_left = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv1d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 2D input (h=C_in, w=W)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C_out*C_in*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (copy_make_border on 2D mat: top=0, bottom=0, left=pad, right=pad)\n ncnn::Mat bordered;\n if (pad_left > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n 0, 0, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims (matches Convolution1D::forward in convolution1d.h:107)\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bordered.w - ext_kw) / stride_w + 1;\n\n // 3. Allocate top as 2D Mat (w=out_w, h=out_c) \u2014 matches Convolution1D::forward:109\n top.create(out_w, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution1d_kernel(\n bordered, top, weight, bias,\n kernel_w, stride_w, dilation_w,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024.json deleted file mode 100644 index 0ee41152c9e235d1366f16e183a2d9567e66beca..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024", - "definition": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 1024;\nconstexpr int kernel_w = 1;\nconstexpr int kernel_h = 1;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 0;\nconstexpr int pad_top = 0;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512.json deleted file mode 100644 index c36957a689f491dbfd0998173e732be6532995b0..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512", - "definition": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 512;\nconstexpr int kernel_w = 1;\nconstexpr int kernel_h = 1;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 0;\nconstexpr int pad_top = 0;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024.json deleted file mode 100644 index ffb68c6f5e5eff26064fef2526abc0ad15cea994..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024", - "definition": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 1024;\nconstexpr int kernel_w = 1;\nconstexpr int kernel_h = 1;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 0;\nconstexpr int pad_top = 0;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048.json deleted file mode 100644 index 75bf6052ebb394c368e3265fce0abb8a17b539df..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048", - "definition": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 2048;\nconstexpr int kernel_w = 1;\nconstexpr int kernel_h = 1;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 0;\nconstexpr int pad_top = 0;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256.json deleted file mode 100644 index 001aea365dd6ddda75e5e70c7ae0181d5f1423fb..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256", - "definition": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 256;\nconstexpr int kernel_w = 1;\nconstexpr int kernel_h = 1;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 0;\nconstexpr int pad_top = 0;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64.json deleted file mode 100644 index c271ca59e58d2df58c8dd707420f052b3a41f6a6..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64", - "definition": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 64;\nconstexpr int kernel_w = 1;\nconstexpr int kernel_h = 1;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 0;\nconstexpr int pad_top = 0;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256.json deleted file mode 100644 index 3f8f009c1f0a22ea637e6407e820735b98090235..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 256;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 1;\nconstexpr int pad_top = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512.json deleted file mode 100644 index b8c37d18e7c12567aa64ff16b02b4e609cd21ef4..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 512;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 1;\nconstexpr int pad_top = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64.json deleted file mode 100644 index 97ef8ef4ac0fc4c0b67f1b552dd24ed83bfed6c0..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 64;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 1;\nconstexpr int pad_top = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024.json deleted file mode 100644 index 78989de8816f4b8368a90f77e320f81b561610ca..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 1024;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 1;\nconstexpr int pad_top = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128.json deleted file mode 100644 index de990c8ab766e471493223a8be94d941924ccadc..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 128;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 1;\nconstexpr int pad_top = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256.json deleted file mode 100644 index 359c059edf5f74186160a6df9e70870d78226504..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 256;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 2;\nconstexpr int dilation_h = 2;\nconstexpr int pad_left = 2;\nconstexpr int pad_top = 2;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256.json deleted file mode 100644 index 941a8089c53f6bd8184d4dbd7a12b38959a2fc85..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 256;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 2;\nconstexpr int dilation_h = 2;\nconstexpr int pad_left = 2;\nconstexpr int pad_top = 2;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512.json deleted file mode 100644 index a15d09bcf11b88ba276921811a2c5afc948d2351..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 512;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 2;\nconstexpr int dilation_h = 2;\nconstexpr int pad_left = 2;\nconstexpr int pad_top = 2;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128.json deleted file mode 100644 index 9cdcf6cb38bade8377e180b701646341dcfcf205..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 128;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 2;\nconstexpr int dilation_h = 2;\nconstexpr int pad_left = 2;\nconstexpr int pad_top = 2;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256.json deleted file mode 100644 index e03cbe8a1d2e08b8a4cc07528bc6b843894310d9..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256", - "definition": "conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 256;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 1;\nconstexpr int pad_top = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512.json deleted file mode 100644 index c7cc785de2903b9fc4e20bc55d95a5f2176bb059..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512", - "definition": "conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 512;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 1;\nconstexpr int pad_top = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024.json deleted file mode 100644 index 5b41c4f31387366af41dad8161d1a4c9cf11e815..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024", - "definition": "conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 1024;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 1;\nconstexpr int pad_top = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128.json deleted file mode 100644 index 12ce1af5e8aaa7a354fcd186b3ce42776f85bf30..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128", - "definition": "conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 128;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 1;\nconstexpr int pad_top = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128.json deleted file mode 100644 index dbc23b9264e1c405f8972eb2f705539d6be96b1a..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128", - "definition": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 128;\nconstexpr int kernel_w = 5;\nconstexpr int kernel_h = 5;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 2;\nconstexpr int pad_top = 2;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256.json deleted file mode 100644 index 4aa39ff564c79e9a8b95adef294ff37a3cec7ac9..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256", - "definition": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 256;\nconstexpr int kernel_w = 5;\nconstexpr int kernel_h = 5;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 2;\nconstexpr int pad_top = 2;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64.json deleted file mode 100644 index d879c2a8551205a45cb503b9355763e25ae4dd2b..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64", - "definition": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 64;\nconstexpr int kernel_w = 5;\nconstexpr int kernel_h = 5;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 2;\nconstexpr int pad_top = 2;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32.json deleted file mode 100644 index 3dd1c4d6865fb60e2a02c2db6d54943972aacc24..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32", - "definition": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 32;\nconstexpr int kernel_w = 5;\nconstexpr int kernel_h = 5;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 2;\nconstexpr int pad_top = 2;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512.json deleted file mode 100644 index d7d0ebb44bd172ccc9f94b1bea959b5b43c26dca..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512", - "definition": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 512;\nconstexpr int kernel_w = 5;\nconstexpr int kernel_h = 5;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 2;\nconstexpr int pad_top = 2;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128.json deleted file mode 100644 index 7170b97744d59d6345edae9fdbe1480f1715890b..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128", - "definition": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 128;\nconstexpr int kernel_w = 5;\nconstexpr int kernel_h = 5;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 2;\nconstexpr int pad_top = 2;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64.json b/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64.json deleted file mode 100644 index 943f67671eafd915dd15706f8e0e5ee8402d25d1..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d/conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64", - "definition": "conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Convolution_arm wrapper for conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64. Same kernel.cpp content across all Definitions; const params come from the Definition. Times create_pipeline + forward (matches today's baselines/ncnn.json semantics).", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d//kernel.cpp (TEMPLATE)\n//\n// Stamped verbatim into every baseline-ncnn-arm conv2d Solution by\n// scripts/extract_definitions.py (PHASE2.md deliverable #4).\n//\n// Implements the harness contract declared in\n// solutions/ncnn/_harness/conv2d.h by delegating to ncnn::Convolution_arm\n// (from libncnn_arm_heavy.a). Const params come from the Definition's\n// const axes (out_c, kw, kh, sw, sh, dw, dh) and are passed in by\n// armbench_entry_conv2d \u2192 convolution_kernel as runtime args, so this\n// kernel.cpp is identical across all Definitions.\n//\n// create_pipeline is included in the timed path on purpose \u2014 matches\n// today's baselines/ncnn.json semantics (run_conv2d_arm in\n// starter/ncnn/baseline/convolution_arm.h calls it the same way).\n// See PHASE2.md decision #4.\n\n#include \"conv2d.h\"\n#include \"convolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Convolution_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // Convolution_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob (the harness pre-allocated it\n // to the expected output shape).\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d.h", - "content": "\n//\n// Calling-convention contract for every conv2d Solution in the \"ncnn\" dataset.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c*in_c*kh*kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d \u2014 shipped in this Solution's own sources.\n// The input is a 3D ncnn::Mat with layout (c, h, w).\n\n#include \"conv2d.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int out_c = 64;\nconstexpr int kernel_w = 7;\nconstexpr int kernel_h = 7;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 3;\nconstexpr int pad_top = 3;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 original (un-padded) 3D input\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (top/bottom = pad_top, left/right = pad_left)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n\n // 3. Allocate top (3D: w=outw, h=outh, c=out_c)\n top.create(outw, outh, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to the Solution-supplied kernel\n return ncnn::convolution_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024.json b/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024.json deleted file mode 100644 index df70287b90c38714abd675f0acded358c1b70238..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024", - "definition": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::ConvolutionDepthWise_arm wrapper for conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024. Same kernel.cpp content across all conv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv2d_depthwise\n// Solution; identical across all conv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in conv2d_depthwise.h by delegating\n// to ncnn::ConvolutionDepthWise_arm. The shim armbench_entry_conv2d_depthwise\n// (binding.cpp) pads the 3D input, allocates the output (out_c == in_c), then\n// calls into this symbol with group == channels. Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all\n// conv2d_depthwise Definitions.\n\n#include \"conv2d_depthwise.h\"\n#include \"convolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n ConvolutionDepthWise_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // ConvolutionDepthWise_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.group = group;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob.\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/conv2d_depthwise.h\n//\n// Calling-convention contract for every conv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv2d_depthwise/ must define the\n// `ncnn::convolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv2d_depthwise` in conv2d_depthwise.cpp pads input via\n// copy_make_border, allocates output, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolutiondepthwise.h:58-65 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d_depthwise \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"conv2d_depthwise.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 1;\nconstexpr int pad_top = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C*Kh*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (matches ConvolutionDepthWise::make_padding)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n const int channels = bordered.c; // depthwise: out_c == in_c\n\n // 3. Allocate top\n top.create(outw, outh, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::convolutiondepthwise_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128.json b/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128.json deleted file mode 100644 index ee75b7e74fd7ef4405020dccd05ea6fc3ae1cf33..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128", - "definition": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::ConvolutionDepthWise_arm wrapper for conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128. Same kernel.cpp content across all conv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv2d_depthwise\n// Solution; identical across all conv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in conv2d_depthwise.h by delegating\n// to ncnn::ConvolutionDepthWise_arm. The shim armbench_entry_conv2d_depthwise\n// (binding.cpp) pads the 3D input, allocates the output (out_c == in_c), then\n// calls into this symbol with group == channels. Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all\n// conv2d_depthwise Definitions.\n\n#include \"conv2d_depthwise.h\"\n#include \"convolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n ConvolutionDepthWise_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // ConvolutionDepthWise_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.group = group;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob.\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/conv2d_depthwise.h\n//\n// Calling-convention contract for every conv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv2d_depthwise/ must define the\n// `ncnn::convolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv2d_depthwise` in conv2d_depthwise.cpp pads input via\n// copy_make_border, allocates output, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolutiondepthwise.h:58-65 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d_depthwise \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"conv2d_depthwise.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 1;\nconstexpr int pad_top = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C*Kh*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (matches ConvolutionDepthWise::make_padding)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n const int channels = bordered.c; // depthwise: out_c == in_c\n\n // 3. Allocate top\n top.create(outw, outh, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::convolutiondepthwise_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048.json b/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048.json deleted file mode 100644 index 98e7a806e2198fde15cf0c4183b6e894ec01473b..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048", - "definition": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::ConvolutionDepthWise_arm wrapper for conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048. Same kernel.cpp content across all conv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv2d_depthwise\n// Solution; identical across all conv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in conv2d_depthwise.h by delegating\n// to ncnn::ConvolutionDepthWise_arm. The shim armbench_entry_conv2d_depthwise\n// (binding.cpp) pads the 3D input, allocates the output (out_c == in_c), then\n// calls into this symbol with group == channels. Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all\n// conv2d_depthwise Definitions.\n\n#include \"conv2d_depthwise.h\"\n#include \"convolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n ConvolutionDepthWise_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // ConvolutionDepthWise_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.group = group;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob.\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/conv2d_depthwise.h\n//\n// Calling-convention contract for every conv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv2d_depthwise/ must define the\n// `ncnn::convolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv2d_depthwise` in conv2d_depthwise.cpp pads input via\n// copy_make_border, allocates output, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolutiondepthwise.h:58-65 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d_depthwise \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"conv2d_depthwise.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 1;\nconstexpr int pad_top = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C*Kh*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (matches ConvolutionDepthWise::make_padding)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n const int channels = bordered.c; // depthwise: out_c == in_c\n\n // 3. Allocate top\n top.create(outw, outh, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::convolutiondepthwise_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256.json b/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256.json deleted file mode 100644 index 6aab1906643bda995067ea218489d024ad8a4f9d..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256", - "definition": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::ConvolutionDepthWise_arm wrapper for conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256. Same kernel.cpp content across all conv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv2d_depthwise\n// Solution; identical across all conv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in conv2d_depthwise.h by delegating\n// to ncnn::ConvolutionDepthWise_arm. The shim armbench_entry_conv2d_depthwise\n// (binding.cpp) pads the 3D input, allocates the output (out_c == in_c), then\n// calls into this symbol with group == channels. Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all\n// conv2d_depthwise Definitions.\n\n#include \"conv2d_depthwise.h\"\n#include \"convolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n ConvolutionDepthWise_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // ConvolutionDepthWise_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.group = group;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob.\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/conv2d_depthwise.h\n//\n// Calling-convention contract for every conv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv2d_depthwise/ must define the\n// `ncnn::convolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv2d_depthwise` in conv2d_depthwise.cpp pads input via\n// copy_make_border, allocates output, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolutiondepthwise.h:58-65 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d_depthwise \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"conv2d_depthwise.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 1;\nconstexpr int pad_top = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C*Kh*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (matches ConvolutionDepthWise::make_padding)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n const int channels = bordered.c; // depthwise: out_c == in_c\n\n // 3. Allocate top\n top.create(outw, outh, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::convolutiondepthwise_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32.json b/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32.json deleted file mode 100644 index d0ea073d311029da1f9e6b6dbfbb81cc2562ec81..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32", - "definition": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::ConvolutionDepthWise_arm wrapper for conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32. Same kernel.cpp content across all conv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv2d_depthwise\n// Solution; identical across all conv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in conv2d_depthwise.h by delegating\n// to ncnn::ConvolutionDepthWise_arm. The shim armbench_entry_conv2d_depthwise\n// (binding.cpp) pads the 3D input, allocates the output (out_c == in_c), then\n// calls into this symbol with group == channels. Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all\n// conv2d_depthwise Definitions.\n\n#include \"conv2d_depthwise.h\"\n#include \"convolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n ConvolutionDepthWise_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // ConvolutionDepthWise_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.group = group;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob.\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/conv2d_depthwise.h\n//\n// Calling-convention contract for every conv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv2d_depthwise/ must define the\n// `ncnn::convolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv2d_depthwise` in conv2d_depthwise.cpp pads input via\n// copy_make_border, allocates output, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolutiondepthwise.h:58-65 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d_depthwise \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"conv2d_depthwise.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 1;\nconstexpr int pad_top = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C*Kh*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (matches ConvolutionDepthWise::make_padding)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n const int channels = bordered.c; // depthwise: out_c == in_c\n\n // 3. Allocate top\n top.create(outw, outh, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::convolutiondepthwise_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512.json b/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512.json deleted file mode 100644 index fab7d4e2c2249620bd1f18e773c1afedef67d93a..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512", - "definition": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::ConvolutionDepthWise_arm wrapper for conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512. Same kernel.cpp content across all conv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv2d_depthwise\n// Solution; identical across all conv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in conv2d_depthwise.h by delegating\n// to ncnn::ConvolutionDepthWise_arm. The shim armbench_entry_conv2d_depthwise\n// (binding.cpp) pads the 3D input, allocates the output (out_c == in_c), then\n// calls into this symbol with group == channels. Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all\n// conv2d_depthwise Definitions.\n\n#include \"conv2d_depthwise.h\"\n#include \"convolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n ConvolutionDepthWise_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // ConvolutionDepthWise_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.group = group;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob.\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/conv2d_depthwise.h\n//\n// Calling-convention contract for every conv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv2d_depthwise/ must define the\n// `ncnn::convolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv2d_depthwise` in conv2d_depthwise.cpp pads input via\n// copy_make_border, allocates output, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolutiondepthwise.h:58-65 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d_depthwise \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"conv2d_depthwise.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 1;\nconstexpr int pad_top = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C*Kh*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (matches ConvolutionDepthWise::make_padding)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n const int channels = bordered.c; // depthwise: out_c == in_c\n\n // 3. Allocate top\n top.create(outw, outh, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::convolutiondepthwise_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64.json b/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64.json deleted file mode 100644 index 394b5df45624fc89ffbc16539f53877ac7f0aac4..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64", - "definition": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::ConvolutionDepthWise_arm wrapper for conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64. Same kernel.cpp content across all conv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv2d_depthwise\n// Solution; identical across all conv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in conv2d_depthwise.h by delegating\n// to ncnn::ConvolutionDepthWise_arm. The shim armbench_entry_conv2d_depthwise\n// (binding.cpp) pads the 3D input, allocates the output (out_c == in_c), then\n// calls into this symbol with group == channels. Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all\n// conv2d_depthwise Definitions.\n\n#include \"conv2d_depthwise.h\"\n#include \"convolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n ConvolutionDepthWise_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // ConvolutionDepthWise_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.group = group;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob.\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/conv2d_depthwise.h\n//\n// Calling-convention contract for every conv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv2d_depthwise/ must define the\n// `ncnn::convolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv2d_depthwise` in conv2d_depthwise.cpp pads input via\n// copy_make_border, allocates output, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolutiondepthwise.h:58-65 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d_depthwise \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"conv2d_depthwise.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 1;\nconstexpr int pad_top = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C*Kh*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (matches ConvolutionDepthWise::make_padding)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n const int channels = bordered.c; // depthwise: out_c == in_c\n\n // 3. Allocate top\n top.create(outw, outh, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::convolutiondepthwise_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024.json b/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024.json deleted file mode 100644 index fa096d85ecc01e70aa898dc431bf5b8de16ec6cf..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024", - "definition": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::ConvolutionDepthWise_arm wrapper for conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024. Same kernel.cpp content across all conv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv2d_depthwise\n// Solution; identical across all conv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in conv2d_depthwise.h by delegating\n// to ncnn::ConvolutionDepthWise_arm. The shim armbench_entry_conv2d_depthwise\n// (binding.cpp) pads the 3D input, allocates the output (out_c == in_c), then\n// calls into this symbol with group == channels. Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all\n// conv2d_depthwise Definitions.\n\n#include \"conv2d_depthwise.h\"\n#include \"convolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n ConvolutionDepthWise_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // ConvolutionDepthWise_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.group = group;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob.\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/conv2d_depthwise.h\n//\n// Calling-convention contract for every conv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv2d_depthwise/ must define the\n// `ncnn::convolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv2d_depthwise` in conv2d_depthwise.cpp pads input via\n// copy_make_border, allocates output, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolutiondepthwise.h:58-65 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d_depthwise \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"conv2d_depthwise.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int kernel_w = 5;\nconstexpr int kernel_h = 5;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 2;\nconstexpr int pad_top = 2;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C*Kh*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (matches ConvolutionDepthWise::make_padding)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n const int channels = bordered.c; // depthwise: out_c == in_c\n\n // 3. Allocate top\n top.create(outw, outh, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::convolutiondepthwise_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128.json b/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128.json deleted file mode 100644 index fb0dc8c6209314a85b2a8a4958bba44ab09f1111..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128", - "definition": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::ConvolutionDepthWise_arm wrapper for conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128. Same kernel.cpp content across all conv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv2d_depthwise\n// Solution; identical across all conv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in conv2d_depthwise.h by delegating\n// to ncnn::ConvolutionDepthWise_arm. The shim armbench_entry_conv2d_depthwise\n// (binding.cpp) pads the 3D input, allocates the output (out_c == in_c), then\n// calls into this symbol with group == channels. Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all\n// conv2d_depthwise Definitions.\n\n#include \"conv2d_depthwise.h\"\n#include \"convolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n ConvolutionDepthWise_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // ConvolutionDepthWise_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.group = group;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob.\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/conv2d_depthwise.h\n//\n// Calling-convention contract for every conv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv2d_depthwise/ must define the\n// `ncnn::convolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv2d_depthwise` in conv2d_depthwise.cpp pads input via\n// copy_make_border, allocates output, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolutiondepthwise.h:58-65 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d_depthwise \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"conv2d_depthwise.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int kernel_w = 5;\nconstexpr int kernel_h = 5;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 2;\nconstexpr int pad_top = 2;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C*Kh*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (matches ConvolutionDepthWise::make_padding)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n const int channels = bordered.c; // depthwise: out_c == in_c\n\n // 3. Allocate top\n top.create(outw, outh, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::convolutiondepthwise_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256.json b/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256.json deleted file mode 100644 index c0e6914a4d9020761d437a7fa91dd2804ba4e6f8..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256", - "definition": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::ConvolutionDepthWise_arm wrapper for conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256. Same kernel.cpp content across all conv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv2d_depthwise\n// Solution; identical across all conv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in conv2d_depthwise.h by delegating\n// to ncnn::ConvolutionDepthWise_arm. The shim armbench_entry_conv2d_depthwise\n// (binding.cpp) pads the 3D input, allocates the output (out_c == in_c), then\n// calls into this symbol with group == channels. Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all\n// conv2d_depthwise Definitions.\n\n#include \"conv2d_depthwise.h\"\n#include \"convolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n ConvolutionDepthWise_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // ConvolutionDepthWise_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.group = group;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob.\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/conv2d_depthwise.h\n//\n// Calling-convention contract for every conv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv2d_depthwise/ must define the\n// `ncnn::convolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv2d_depthwise` in conv2d_depthwise.cpp pads input via\n// copy_make_border, allocates output, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolutiondepthwise.h:58-65 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d_depthwise \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"conv2d_depthwise.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int kernel_w = 5;\nconstexpr int kernel_h = 5;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 2;\nconstexpr int pad_top = 2;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C*Kh*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (matches ConvolutionDepthWise::make_padding)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n const int channels = bordered.c; // depthwise: out_c == in_c\n\n // 3. Allocate top\n top.create(outw, outh, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::convolutiondepthwise_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32.json b/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32.json deleted file mode 100644 index 6bf6c870220637b1babb97339a696403bfd1c78b..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32", - "definition": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::ConvolutionDepthWise_arm wrapper for conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32. Same kernel.cpp content across all conv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv2d_depthwise\n// Solution; identical across all conv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in conv2d_depthwise.h by delegating\n// to ncnn::ConvolutionDepthWise_arm. The shim armbench_entry_conv2d_depthwise\n// (binding.cpp) pads the 3D input, allocates the output (out_c == in_c), then\n// calls into this symbol with group == channels. Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all\n// conv2d_depthwise Definitions.\n\n#include \"conv2d_depthwise.h\"\n#include \"convolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n ConvolutionDepthWise_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // ConvolutionDepthWise_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.group = group;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob.\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/conv2d_depthwise.h\n//\n// Calling-convention contract for every conv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv2d_depthwise/ must define the\n// `ncnn::convolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv2d_depthwise` in conv2d_depthwise.cpp pads input via\n// copy_make_border, allocates output, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolutiondepthwise.h:58-65 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d_depthwise \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"conv2d_depthwise.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int kernel_w = 5;\nconstexpr int kernel_h = 5;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 2;\nconstexpr int pad_top = 2;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C*Kh*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (matches ConvolutionDepthWise::make_padding)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n const int channels = bordered.c; // depthwise: out_c == in_c\n\n // 3. Allocate top\n top.create(outw, outh, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::convolutiondepthwise_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512.json b/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512.json deleted file mode 100644 index 9bb456cbc2fe88e167c746029d47f1284a93cd83..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512", - "definition": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::ConvolutionDepthWise_arm wrapper for conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512. Same kernel.cpp content across all conv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv2d_depthwise\n// Solution; identical across all conv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in conv2d_depthwise.h by delegating\n// to ncnn::ConvolutionDepthWise_arm. The shim armbench_entry_conv2d_depthwise\n// (binding.cpp) pads the 3D input, allocates the output (out_c == in_c), then\n// calls into this symbol with group == channels. Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all\n// conv2d_depthwise Definitions.\n\n#include \"conv2d_depthwise.h\"\n#include \"convolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n ConvolutionDepthWise_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // ConvolutionDepthWise_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.group = group;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob.\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/conv2d_depthwise.h\n//\n// Calling-convention contract for every conv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv2d_depthwise/ must define the\n// `ncnn::convolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv2d_depthwise` in conv2d_depthwise.cpp pads input via\n// copy_make_border, allocates output, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolutiondepthwise.h:58-65 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d_depthwise \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"conv2d_depthwise.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int kernel_w = 5;\nconstexpr int kernel_h = 5;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 2;\nconstexpr int pad_top = 2;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C*Kh*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (matches ConvolutionDepthWise::make_padding)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n const int channels = bordered.c; // depthwise: out_c == in_c\n\n // 3. Allocate top\n top.create(outw, outh, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::convolutiondepthwise_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64.json b/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64.json deleted file mode 100644 index 66a7ad0ea10f96ecc8db36b89296fb0896be4c05..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64", - "definition": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::ConvolutionDepthWise_arm wrapper for conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64. Same kernel.cpp content across all conv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/conv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm conv2d_depthwise\n// Solution; identical across all conv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in conv2d_depthwise.h by delegating\n// to ncnn::ConvolutionDepthWise_arm. The shim armbench_entry_conv2d_depthwise\n// (binding.cpp) pads the 3D input, allocates the output (out_c == in_c), then\n// calls into this symbol with group == channels. Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all\n// conv2d_depthwise Definitions.\n\n#include \"conv2d_depthwise.h\"\n#include \"convolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n ConvolutionDepthWise_arm conv;\n conv.num_output = top_blob.c;\n conv.kernel_w = kernel_w; conv.kernel_h = kernel_h;\n conv.stride_w = stride_w; conv.stride_h = stride_h;\n conv.dilation_w = dilation_w; conv.dilation_h = dilation_h;\n // Harness already padded `bottom_blob` via copy_make_border, so disable\n // ConvolutionDepthWise_arm's internal padding.\n conv.pad_left = 0; conv.pad_right = 0;\n conv.pad_top = 0; conv.pad_bottom = 0;\n conv.pad_value = 0.f;\n conv.bias_term = bias_data.empty() ? 0 : 1;\n conv.weight_data_size = weight_data.w;\n conv.group = group;\n conv.int8_scale_term = 0;\n conv.activation_type = activation_type;\n conv.activation_params = const_cast(activation_params);\n conv.dynamic_weight = 0;\n conv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) conv.bias_data = const_cast(bias_data);\n\n if (conv.create_pipeline(opt) != 0) return -1;\n\n Mat local_top;\n int ret = conv.forward(bottom_blob, local_top, opt);\n if (ret != 0) return ret;\n if (local_top.empty()) return -1;\n\n // Copy into the caller-allocated top_blob.\n if (local_top.c != top_blob.c || local_top.h != top_blob.h || local_top.w != top_blob.w)\n return -1;\n for (int cc = 0; cc < local_top.c; ++cc) {\n std::memcpy(top_blob.channel(cc), local_top.channel(cc),\n local_top.h * local_top.w * sizeof(float));\n }\n return 0;\n}\n\n} // namespace ncnn\n" - }, - { - "path": "conv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/conv2d_depthwise.h\n//\n// Calling-convention contract for every conv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//conv2d_depthwise/ must define the\n// `ncnn::convolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_conv2d_depthwise` in conv2d_depthwise.cpp pads input via\n// copy_make_border, allocates output, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/convolutiondepthwise.h:58-65 signature\n// byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: already-padded input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Returns 0 on success, non-zero on error.\nint convolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_CONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for conv2d_depthwise \u2014 shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_conv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body pads the input, computes the output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// convolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"conv2d_depthwise.h\"\n\n#ifndef BORDER_CONSTANT\n#define BORDER_CONSTANT 0\n#endif\n\n// \u2500\u2500 Per-Definition constants (baked from the Definition's const axes) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\nnamespace {\nconstexpr int kernel_w = 5;\nconstexpr int kernel_h = 5;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int pad_left = 2;\nconstexpr int pad_top = 2;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_conv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* \u2014 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* \u2014 empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat, size = C*Kh*Kw\n void* bias_mat_v, // const ncnn::Mat* \u2014 flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* \u2014 empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Pad \u2192 bordered (matches ConvolutionDepthWise::make_padding)\n ncnn::Mat bordered;\n if (pad_left > 0 || pad_top > 0) {\n ncnn::Option pad_opt = opt;\n pad_opt.blob_allocator = opt.workspace_allocator;\n ncnn::copy_make_border(bottom, bordered,\n pad_top, pad_top, pad_left, pad_left,\n BORDER_CONSTANT, 0.f, pad_opt);\n if (bordered.empty()) return -100;\n } else {\n bordered = bottom;\n }\n\n // 2. Compute output dims\n const int ext_kw = dilation_w * (kernel_w - 1) + 1;\n const int ext_kh = dilation_h * (kernel_h - 1) + 1;\n const int outw = (bordered.w - ext_kw) / stride_w + 1;\n const int outh = (bordered.h - ext_kh) / stride_h + 1;\n const int channels = bordered.c; // depthwise: out_c == in_c\n\n // 3. Allocate top\n top.create(outw, outh, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 4. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::convolutiondepthwise_kernel(\n bordered, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin128_cout256.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin128_cout256.json deleted file mode 100644 index 34f9c225e0537c44f68b83fd6c396dc5015b7d7f..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin128_cout256.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_kh3_kw3_sh1_sw1_cin128_cout256", - "definition": "deconv2d_kh3_kw3_sh1_sw1_cin128_cout256", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Deconvolution_arm wrapper for deconv2d_kh3_kw3_sh1_sw1_cin128_cout256. Same kernel.cpp content across all deconv2d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d Solution;\n// identical across all deconv2d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d.h by delegating to\n// ncnn::Deconvolution_arm. The shim armbench_entry_deconv2d (binding.cpp)\n// allocates the full transposed-convolution output\n// (out_w = (in_w-1)*stride_w + dilation_w*(kernel_w-1)+1, likewise out_h) and\n// does NOT crop, so we disable Deconvolution_arm's internal cut_padding\n// (pad_* / output_pad_* / output_w/h all zero). Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all deconv2d\n// Definitions.\n\n#include \"deconv2d.h\"\n#include \"deconvolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Deconvolution_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of\n // Deconvolution_arm's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d.h\n//\n// Calling-convention contract for every deconv2d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d/ must define the\n// `ncnn::deconvolution_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d` in deconv2d.cpp allocates the output using\n// the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolution.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (in_c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c * in_c * Kh * Kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, dilation=1 case used in test definitions):\n// out_w = (in_w - 1) * stride_w + kernel_extent_w\n// out_h = (in_h - 1) * stride_h + kernel_extent_h\n//\n// Returns 0 on success, non-zero on error.\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body computes transposed-convolution output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// deconvolution_kernel (defined in kernel.cpp).\n//\n// Output formula (from ref_conv.h::ref_deconv2d):\n// ke_h = dilation_h * (kernel_h - 1) + 1\n// ke_w = dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + ke_h\n// out_w = (in_w - 1) * stride_w + ke_w\n\n#include \"deconv2d.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int out_c = 256;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (in_c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = out_c*in_c*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_deconv2d in ref_conv.h)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n\n // 2. Allocate top\n top.create(out_w, out_h, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to the Solution-supplied kernel\n return ncnn::deconvolution_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin256_cout512.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin256_cout512.json deleted file mode 100644 index fce6e2ea8f39a80d2752314b6275398d6cff1e66..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin256_cout512.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_kh3_kw3_sh1_sw1_cin256_cout512", - "definition": "deconv2d_kh3_kw3_sh1_sw1_cin256_cout512", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Deconvolution_arm wrapper for deconv2d_kh3_kw3_sh1_sw1_cin256_cout512. Same kernel.cpp content across all deconv2d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d Solution;\n// identical across all deconv2d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d.h by delegating to\n// ncnn::Deconvolution_arm. The shim armbench_entry_deconv2d (binding.cpp)\n// allocates the full transposed-convolution output\n// (out_w = (in_w-1)*stride_w + dilation_w*(kernel_w-1)+1, likewise out_h) and\n// does NOT crop, so we disable Deconvolution_arm's internal cut_padding\n// (pad_* / output_pad_* / output_w/h all zero). Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all deconv2d\n// Definitions.\n\n#include \"deconv2d.h\"\n#include \"deconvolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Deconvolution_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of\n // Deconvolution_arm's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d.h\n//\n// Calling-convention contract for every deconv2d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d/ must define the\n// `ncnn::deconvolution_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d` in deconv2d.cpp allocates the output using\n// the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolution.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (in_c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c * in_c * Kh * Kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, dilation=1 case used in test definitions):\n// out_w = (in_w - 1) * stride_w + kernel_extent_w\n// out_h = (in_h - 1) * stride_h + kernel_extent_h\n//\n// Returns 0 on success, non-zero on error.\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body computes transposed-convolution output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// deconvolution_kernel (defined in kernel.cpp).\n//\n// Output formula (from ref_conv.h::ref_deconv2d):\n// ke_h = dilation_h * (kernel_h - 1) + 1\n// ke_w = dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + ke_h\n// out_w = (in_w - 1) * stride_w + ke_w\n\n#include \"deconv2d.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int out_c = 512;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (in_c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = out_c*in_c*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_deconv2d in ref_conv.h)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n\n // 2. Allocate top\n top.create(out_w, out_h, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to the Solution-supplied kernel\n return ncnn::deconvolution_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024.json deleted file mode 100644 index cbb0d56256a954c1fec81c96657371a011ad15e0..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024", - "definition": "deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Deconvolution_arm wrapper for deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024. Same kernel.cpp content across all deconv2d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d Solution;\n// identical across all deconv2d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d.h by delegating to\n// ncnn::Deconvolution_arm. The shim armbench_entry_deconv2d (binding.cpp)\n// allocates the full transposed-convolution output\n// (out_w = (in_w-1)*stride_w + dilation_w*(kernel_w-1)+1, likewise out_h) and\n// does NOT crop, so we disable Deconvolution_arm's internal cut_padding\n// (pad_* / output_pad_* / output_w/h all zero). Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all deconv2d\n// Definitions.\n\n#include \"deconv2d.h\"\n#include \"deconvolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Deconvolution_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of\n // Deconvolution_arm's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d.h\n//\n// Calling-convention contract for every deconv2d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d/ must define the\n// `ncnn::deconvolution_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d` in deconv2d.cpp allocates the output using\n// the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolution.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (in_c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c * in_c * Kh * Kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, dilation=1 case used in test definitions):\n// out_w = (in_w - 1) * stride_w + kernel_extent_w\n// out_h = (in_h - 1) * stride_h + kernel_extent_h\n//\n// Returns 0 on success, non-zero on error.\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body computes transposed-convolution output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// deconvolution_kernel (defined in kernel.cpp).\n//\n// Output formula (from ref_conv.h::ref_deconv2d):\n// ke_h = dilation_h * (kernel_h - 1) + 1\n// ke_w = dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + ke_h\n// out_w = (in_w - 1) * stride_w + ke_w\n\n#include \"deconv2d.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int out_c = 1024;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (in_c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = out_c*in_c*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_deconv2d in ref_conv.h)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n\n // 2. Allocate top\n top.create(out_w, out_h, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to the Solution-supplied kernel\n return ncnn::deconvolution_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin64_cout128.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin64_cout128.json deleted file mode 100644 index e51342dfffde73e1c9da120574cd5f61a4fc6897..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin64_cout128.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_kh3_kw3_sh1_sw1_cin64_cout128", - "definition": "deconv2d_kh3_kw3_sh1_sw1_cin64_cout128", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Deconvolution_arm wrapper for deconv2d_kh3_kw3_sh1_sw1_cin64_cout128. Same kernel.cpp content across all deconv2d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d Solution;\n// identical across all deconv2d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d.h by delegating to\n// ncnn::Deconvolution_arm. The shim armbench_entry_deconv2d (binding.cpp)\n// allocates the full transposed-convolution output\n// (out_w = (in_w-1)*stride_w + dilation_w*(kernel_w-1)+1, likewise out_h) and\n// does NOT crop, so we disable Deconvolution_arm's internal cut_padding\n// (pad_* / output_pad_* / output_w/h all zero). Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all deconv2d\n// Definitions.\n\n#include \"deconv2d.h\"\n#include \"deconvolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Deconvolution_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of\n // Deconvolution_arm's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d.h\n//\n// Calling-convention contract for every deconv2d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d/ must define the\n// `ncnn::deconvolution_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d` in deconv2d.cpp allocates the output using\n// the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolution.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (in_c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c * in_c * Kh * Kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, dilation=1 case used in test definitions):\n// out_w = (in_w - 1) * stride_w + kernel_extent_w\n// out_h = (in_h - 1) * stride_h + kernel_extent_h\n//\n// Returns 0 on success, non-zero on error.\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body computes transposed-convolution output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// deconvolution_kernel (defined in kernel.cpp).\n//\n// Output formula (from ref_conv.h::ref_deconv2d):\n// ke_h = dilation_h * (kernel_h - 1) + 1\n// ke_w = dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + ke_h\n// out_w = (in_w - 1) * stride_w + ke_w\n\n#include \"deconv2d.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int out_c = 128;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (in_c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = out_c*in_c*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_deconv2d in ref_conv.h)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n\n // 2. Allocate top\n top.create(out_w, out_h, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to the Solution-supplied kernel\n return ncnn::deconvolution_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin128_cout256.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin128_cout256.json deleted file mode 100644 index 1d3070eb511645e92800760057f8a6f31184f8a1..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin128_cout256.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_kh3_kw3_sh2_sw2_cin128_cout256", - "definition": "deconv2d_kh3_kw3_sh2_sw2_cin128_cout256", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Deconvolution_arm wrapper for deconv2d_kh3_kw3_sh2_sw2_cin128_cout256. Same kernel.cpp content across all deconv2d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d Solution;\n// identical across all deconv2d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d.h by delegating to\n// ncnn::Deconvolution_arm. The shim armbench_entry_deconv2d (binding.cpp)\n// allocates the full transposed-convolution output\n// (out_w = (in_w-1)*stride_w + dilation_w*(kernel_w-1)+1, likewise out_h) and\n// does NOT crop, so we disable Deconvolution_arm's internal cut_padding\n// (pad_* / output_pad_* / output_w/h all zero). Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all deconv2d\n// Definitions.\n\n#include \"deconv2d.h\"\n#include \"deconvolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Deconvolution_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of\n // Deconvolution_arm's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d.h\n//\n// Calling-convention contract for every deconv2d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d/ must define the\n// `ncnn::deconvolution_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d` in deconv2d.cpp allocates the output using\n// the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolution.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (in_c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c * in_c * Kh * Kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, dilation=1 case used in test definitions):\n// out_w = (in_w - 1) * stride_w + kernel_extent_w\n// out_h = (in_h - 1) * stride_h + kernel_extent_h\n//\n// Returns 0 on success, non-zero on error.\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body computes transposed-convolution output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// deconvolution_kernel (defined in kernel.cpp).\n//\n// Output formula (from ref_conv.h::ref_deconv2d):\n// ke_h = dilation_h * (kernel_h - 1) + 1\n// ke_w = dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + ke_h\n// out_w = (in_w - 1) * stride_w + ke_w\n\n#include \"deconv2d.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int out_c = 256;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (in_c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = out_c*in_c*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_deconv2d in ref_conv.h)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n\n // 2. Allocate top\n top.create(out_w, out_h, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to the Solution-supplied kernel\n return ncnn::deconvolution_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin256_cout512.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin256_cout512.json deleted file mode 100644 index b2ca93ed785eed0b89b67a8ad7f85f2ab9c09092..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin256_cout512.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_kh3_kw3_sh2_sw2_cin256_cout512", - "definition": "deconv2d_kh3_kw3_sh2_sw2_cin256_cout512", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Deconvolution_arm wrapper for deconv2d_kh3_kw3_sh2_sw2_cin256_cout512. Same kernel.cpp content across all deconv2d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d Solution;\n// identical across all deconv2d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d.h by delegating to\n// ncnn::Deconvolution_arm. The shim armbench_entry_deconv2d (binding.cpp)\n// allocates the full transposed-convolution output\n// (out_w = (in_w-1)*stride_w + dilation_w*(kernel_w-1)+1, likewise out_h) and\n// does NOT crop, so we disable Deconvolution_arm's internal cut_padding\n// (pad_* / output_pad_* / output_w/h all zero). Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all deconv2d\n// Definitions.\n\n#include \"deconv2d.h\"\n#include \"deconvolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Deconvolution_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of\n // Deconvolution_arm's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d.h\n//\n// Calling-convention contract for every deconv2d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d/ must define the\n// `ncnn::deconvolution_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d` in deconv2d.cpp allocates the output using\n// the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolution.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (in_c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c * in_c * Kh * Kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, dilation=1 case used in test definitions):\n// out_w = (in_w - 1) * stride_w + kernel_extent_w\n// out_h = (in_h - 1) * stride_h + kernel_extent_h\n//\n// Returns 0 on success, non-zero on error.\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body computes transposed-convolution output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// deconvolution_kernel (defined in kernel.cpp).\n//\n// Output formula (from ref_conv.h::ref_deconv2d):\n// ke_h = dilation_h * (kernel_h - 1) + 1\n// ke_w = dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + ke_h\n// out_w = (in_w - 1) * stride_w + ke_w\n\n#include \"deconv2d.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int out_c = 512;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (in_c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = out_c*in_c*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_deconv2d in ref_conv.h)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n\n // 2. Allocate top\n top.create(out_w, out_h, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to the Solution-supplied kernel\n return ncnn::deconvolution_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024.json deleted file mode 100644 index 8a434d4d5738480de2df97f581e566440ffebcdb..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024", - "definition": "deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Deconvolution_arm wrapper for deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024. Same kernel.cpp content across all deconv2d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d Solution;\n// identical across all deconv2d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d.h by delegating to\n// ncnn::Deconvolution_arm. The shim armbench_entry_deconv2d (binding.cpp)\n// allocates the full transposed-convolution output\n// (out_w = (in_w-1)*stride_w + dilation_w*(kernel_w-1)+1, likewise out_h) and\n// does NOT crop, so we disable Deconvolution_arm's internal cut_padding\n// (pad_* / output_pad_* / output_w/h all zero). Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all deconv2d\n// Definitions.\n\n#include \"deconv2d.h\"\n#include \"deconvolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Deconvolution_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of\n // Deconvolution_arm's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d.h\n//\n// Calling-convention contract for every deconv2d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d/ must define the\n// `ncnn::deconvolution_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d` in deconv2d.cpp allocates the output using\n// the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolution.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (in_c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c * in_c * Kh * Kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, dilation=1 case used in test definitions):\n// out_w = (in_w - 1) * stride_w + kernel_extent_w\n// out_h = (in_h - 1) * stride_h + kernel_extent_h\n//\n// Returns 0 on success, non-zero on error.\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body computes transposed-convolution output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// deconvolution_kernel (defined in kernel.cpp).\n//\n// Output formula (from ref_conv.h::ref_deconv2d):\n// ke_h = dilation_h * (kernel_h - 1) + 1\n// ke_w = dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + ke_h\n// out_w = (in_w - 1) * stride_w + ke_w\n\n#include \"deconv2d.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int out_c = 1024;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (in_c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = out_c*in_c*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_deconv2d in ref_conv.h)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n\n // 2. Allocate top\n top.create(out_w, out_h, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to the Solution-supplied kernel\n return ncnn::deconvolution_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin64_cout128.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin64_cout128.json deleted file mode 100644 index 8f1ce646f6399713b07b2201ce67d96c32d703f4..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin64_cout128.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_kh3_kw3_sh2_sw2_cin64_cout128", - "definition": "deconv2d_kh3_kw3_sh2_sw2_cin64_cout128", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Deconvolution_arm wrapper for deconv2d_kh3_kw3_sh2_sw2_cin64_cout128. Same kernel.cpp content across all deconv2d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d Solution;\n// identical across all deconv2d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d.h by delegating to\n// ncnn::Deconvolution_arm. The shim armbench_entry_deconv2d (binding.cpp)\n// allocates the full transposed-convolution output\n// (out_w = (in_w-1)*stride_w + dilation_w*(kernel_w-1)+1, likewise out_h) and\n// does NOT crop, so we disable Deconvolution_arm's internal cut_padding\n// (pad_* / output_pad_* / output_w/h all zero). Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all deconv2d\n// Definitions.\n\n#include \"deconv2d.h\"\n#include \"deconvolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Deconvolution_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of\n // Deconvolution_arm's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d.h\n//\n// Calling-convention contract for every deconv2d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d/ must define the\n// `ncnn::deconvolution_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d` in deconv2d.cpp allocates the output using\n// the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolution.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (in_c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c * in_c * Kh * Kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, dilation=1 case used in test definitions):\n// out_w = (in_w - 1) * stride_w + kernel_extent_w\n// out_h = (in_h - 1) * stride_h + kernel_extent_h\n//\n// Returns 0 on success, non-zero on error.\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body computes transposed-convolution output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// deconvolution_kernel (defined in kernel.cpp).\n//\n// Output formula (from ref_conv.h::ref_deconv2d):\n// ke_h = dilation_h * (kernel_h - 1) + 1\n// ke_w = dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + ke_h\n// out_w = (in_w - 1) * stride_w + ke_w\n\n#include \"deconv2d.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int out_c = 128;\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (in_c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = out_c*in_c*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_deconv2d in ref_conv.h)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n\n // 2. Allocate top\n top.create(out_w, out_h, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to the Solution-supplied kernel\n return ncnn::deconvolution_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin128_cout256.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin128_cout256.json deleted file mode 100644 index 040e4a344886d0892453a181e4261770bd61b028..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin128_cout256.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_kh4_kw4_sh1_sw1_cin128_cout256", - "definition": "deconv2d_kh4_kw4_sh1_sw1_cin128_cout256", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Deconvolution_arm wrapper for deconv2d_kh4_kw4_sh1_sw1_cin128_cout256. Same kernel.cpp content across all deconv2d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d Solution;\n// identical across all deconv2d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d.h by delegating to\n// ncnn::Deconvolution_arm. The shim armbench_entry_deconv2d (binding.cpp)\n// allocates the full transposed-convolution output\n// (out_w = (in_w-1)*stride_w + dilation_w*(kernel_w-1)+1, likewise out_h) and\n// does NOT crop, so we disable Deconvolution_arm's internal cut_padding\n// (pad_* / output_pad_* / output_w/h all zero). Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all deconv2d\n// Definitions.\n\n#include \"deconv2d.h\"\n#include \"deconvolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Deconvolution_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of\n // Deconvolution_arm's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d.h\n//\n// Calling-convention contract for every deconv2d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d/ must define the\n// `ncnn::deconvolution_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d` in deconv2d.cpp allocates the output using\n// the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolution.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (in_c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c * in_c * Kh * Kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, dilation=1 case used in test definitions):\n// out_w = (in_w - 1) * stride_w + kernel_extent_w\n// out_h = (in_h - 1) * stride_h + kernel_extent_h\n//\n// Returns 0 on success, non-zero on error.\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body computes transposed-convolution output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// deconvolution_kernel (defined in kernel.cpp).\n//\n// Output formula (from ref_conv.h::ref_deconv2d):\n// ke_h = dilation_h * (kernel_h - 1) + 1\n// ke_w = dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + ke_h\n// out_w = (in_w - 1) * stride_w + ke_w\n\n#include \"deconv2d.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int out_c = 256;\nconstexpr int kernel_w = 4;\nconstexpr int kernel_h = 4;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (in_c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = out_c*in_c*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_deconv2d in ref_conv.h)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n\n // 2. Allocate top\n top.create(out_w, out_h, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to the Solution-supplied kernel\n return ncnn::deconvolution_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin256_cout512.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin256_cout512.json deleted file mode 100644 index 80b7c2047e9d8df4f025992708abaa84f1641748..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin256_cout512.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_kh4_kw4_sh1_sw1_cin256_cout512", - "definition": "deconv2d_kh4_kw4_sh1_sw1_cin256_cout512", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Deconvolution_arm wrapper for deconv2d_kh4_kw4_sh1_sw1_cin256_cout512. Same kernel.cpp content across all deconv2d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d Solution;\n// identical across all deconv2d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d.h by delegating to\n// ncnn::Deconvolution_arm. The shim armbench_entry_deconv2d (binding.cpp)\n// allocates the full transposed-convolution output\n// (out_w = (in_w-1)*stride_w + dilation_w*(kernel_w-1)+1, likewise out_h) and\n// does NOT crop, so we disable Deconvolution_arm's internal cut_padding\n// (pad_* / output_pad_* / output_w/h all zero). Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all deconv2d\n// Definitions.\n\n#include \"deconv2d.h\"\n#include \"deconvolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Deconvolution_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of\n // Deconvolution_arm's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d.h\n//\n// Calling-convention contract for every deconv2d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d/ must define the\n// `ncnn::deconvolution_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d` in deconv2d.cpp allocates the output using\n// the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolution.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (in_c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c * in_c * Kh * Kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, dilation=1 case used in test definitions):\n// out_w = (in_w - 1) * stride_w + kernel_extent_w\n// out_h = (in_h - 1) * stride_h + kernel_extent_h\n//\n// Returns 0 on success, non-zero on error.\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body computes transposed-convolution output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// deconvolution_kernel (defined in kernel.cpp).\n//\n// Output formula (from ref_conv.h::ref_deconv2d):\n// ke_h = dilation_h * (kernel_h - 1) + 1\n// ke_w = dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + ke_h\n// out_w = (in_w - 1) * stride_w + ke_w\n\n#include \"deconv2d.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int out_c = 512;\nconstexpr int kernel_w = 4;\nconstexpr int kernel_h = 4;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (in_c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = out_c*in_c*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_deconv2d in ref_conv.h)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n\n // 2. Allocate top\n top.create(out_w, out_h, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to the Solution-supplied kernel\n return ncnn::deconvolution_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024.json deleted file mode 100644 index f43c3c8c9af12d042c757d90baa5cc83f98d69e1..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024", - "definition": "deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Deconvolution_arm wrapper for deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024. Same kernel.cpp content across all deconv2d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d Solution;\n// identical across all deconv2d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d.h by delegating to\n// ncnn::Deconvolution_arm. The shim armbench_entry_deconv2d (binding.cpp)\n// allocates the full transposed-convolution output\n// (out_w = (in_w-1)*stride_w + dilation_w*(kernel_w-1)+1, likewise out_h) and\n// does NOT crop, so we disable Deconvolution_arm's internal cut_padding\n// (pad_* / output_pad_* / output_w/h all zero). Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all deconv2d\n// Definitions.\n\n#include \"deconv2d.h\"\n#include \"deconvolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Deconvolution_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of\n // Deconvolution_arm's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d.h\n//\n// Calling-convention contract for every deconv2d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d/ must define the\n// `ncnn::deconvolution_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d` in deconv2d.cpp allocates the output using\n// the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolution.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (in_c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c * in_c * Kh * Kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, dilation=1 case used in test definitions):\n// out_w = (in_w - 1) * stride_w + kernel_extent_w\n// out_h = (in_h - 1) * stride_h + kernel_extent_h\n//\n// Returns 0 on success, non-zero on error.\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body computes transposed-convolution output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// deconvolution_kernel (defined in kernel.cpp).\n//\n// Output formula (from ref_conv.h::ref_deconv2d):\n// ke_h = dilation_h * (kernel_h - 1) + 1\n// ke_w = dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + ke_h\n// out_w = (in_w - 1) * stride_w + ke_w\n\n#include \"deconv2d.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int out_c = 1024;\nconstexpr int kernel_w = 4;\nconstexpr int kernel_h = 4;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (in_c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = out_c*in_c*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_deconv2d in ref_conv.h)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n\n // 2. Allocate top\n top.create(out_w, out_h, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to the Solution-supplied kernel\n return ncnn::deconvolution_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin64_cout128.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin64_cout128.json deleted file mode 100644 index c0525e4af02e4cdaf36fab8c57c95eccb628dabe..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin64_cout128.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_kh4_kw4_sh1_sw1_cin64_cout128", - "definition": "deconv2d_kh4_kw4_sh1_sw1_cin64_cout128", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Deconvolution_arm wrapper for deconv2d_kh4_kw4_sh1_sw1_cin64_cout128. Same kernel.cpp content across all deconv2d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d Solution;\n// identical across all deconv2d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d.h by delegating to\n// ncnn::Deconvolution_arm. The shim armbench_entry_deconv2d (binding.cpp)\n// allocates the full transposed-convolution output\n// (out_w = (in_w-1)*stride_w + dilation_w*(kernel_w-1)+1, likewise out_h) and\n// does NOT crop, so we disable Deconvolution_arm's internal cut_padding\n// (pad_* / output_pad_* / output_w/h all zero). Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all deconv2d\n// Definitions.\n\n#include \"deconv2d.h\"\n#include \"deconvolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Deconvolution_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of\n // Deconvolution_arm's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d.h\n//\n// Calling-convention contract for every deconv2d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d/ must define the\n// `ncnn::deconvolution_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d` in deconv2d.cpp allocates the output using\n// the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolution.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (in_c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c * in_c * Kh * Kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, dilation=1 case used in test definitions):\n// out_w = (in_w - 1) * stride_w + kernel_extent_w\n// out_h = (in_h - 1) * stride_h + kernel_extent_h\n//\n// Returns 0 on success, non-zero on error.\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body computes transposed-convolution output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// deconvolution_kernel (defined in kernel.cpp).\n//\n// Output formula (from ref_conv.h::ref_deconv2d):\n// ke_h = dilation_h * (kernel_h - 1) + 1\n// ke_w = dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + ke_h\n// out_w = (in_w - 1) * stride_w + ke_w\n\n#include \"deconv2d.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int out_c = 128;\nconstexpr int kernel_w = 4;\nconstexpr int kernel_h = 4;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (in_c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = out_c*in_c*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_deconv2d in ref_conv.h)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n\n // 2. Allocate top\n top.create(out_w, out_h, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to the Solution-supplied kernel\n return ncnn::deconvolution_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin128_cout256.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin128_cout256.json deleted file mode 100644 index 6a8c6389cafb51031daa330de16f508c27b153a3..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin128_cout256.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_kh4_kw4_sh2_sw2_cin128_cout256", - "definition": "deconv2d_kh4_kw4_sh2_sw2_cin128_cout256", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Deconvolution_arm wrapper for deconv2d_kh4_kw4_sh2_sw2_cin128_cout256. Same kernel.cpp content across all deconv2d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d Solution;\n// identical across all deconv2d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d.h by delegating to\n// ncnn::Deconvolution_arm. The shim armbench_entry_deconv2d (binding.cpp)\n// allocates the full transposed-convolution output\n// (out_w = (in_w-1)*stride_w + dilation_w*(kernel_w-1)+1, likewise out_h) and\n// does NOT crop, so we disable Deconvolution_arm's internal cut_padding\n// (pad_* / output_pad_* / output_w/h all zero). Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all deconv2d\n// Definitions.\n\n#include \"deconv2d.h\"\n#include \"deconvolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Deconvolution_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of\n // Deconvolution_arm's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d.h\n//\n// Calling-convention contract for every deconv2d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d/ must define the\n// `ncnn::deconvolution_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d` in deconv2d.cpp allocates the output using\n// the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolution.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (in_c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c * in_c * Kh * Kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, dilation=1 case used in test definitions):\n// out_w = (in_w - 1) * stride_w + kernel_extent_w\n// out_h = (in_h - 1) * stride_h + kernel_extent_h\n//\n// Returns 0 on success, non-zero on error.\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body computes transposed-convolution output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// deconvolution_kernel (defined in kernel.cpp).\n//\n// Output formula (from ref_conv.h::ref_deconv2d):\n// ke_h = dilation_h * (kernel_h - 1) + 1\n// ke_w = dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + ke_h\n// out_w = (in_w - 1) * stride_w + ke_w\n\n#include \"deconv2d.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int out_c = 256;\nconstexpr int kernel_w = 4;\nconstexpr int kernel_h = 4;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (in_c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = out_c*in_c*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_deconv2d in ref_conv.h)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n\n // 2. Allocate top\n top.create(out_w, out_h, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to the Solution-supplied kernel\n return ncnn::deconvolution_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin32_cout64.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin32_cout64.json deleted file mode 100644 index d4713e2dafbd4207b4989147ee0f98cc2a83bf82..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin32_cout64.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_kh4_kw4_sh2_sw2_cin32_cout64", - "definition": "deconv2d_kh4_kw4_sh2_sw2_cin32_cout64", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Deconvolution_arm wrapper for deconv2d_kh4_kw4_sh2_sw2_cin32_cout64. Same kernel.cpp content across all deconv2d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d Solution;\n// identical across all deconv2d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d.h by delegating to\n// ncnn::Deconvolution_arm. The shim armbench_entry_deconv2d (binding.cpp)\n// allocates the full transposed-convolution output\n// (out_w = (in_w-1)*stride_w + dilation_w*(kernel_w-1)+1, likewise out_h) and\n// does NOT crop, so we disable Deconvolution_arm's internal cut_padding\n// (pad_* / output_pad_* / output_w/h all zero). Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all deconv2d\n// Definitions.\n\n#include \"deconv2d.h\"\n#include \"deconvolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Deconvolution_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of\n // Deconvolution_arm's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d.h\n//\n// Calling-convention contract for every deconv2d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d/ must define the\n// `ncnn::deconvolution_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d` in deconv2d.cpp allocates the output using\n// the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolution.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (in_c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c * in_c * Kh * Kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, dilation=1 case used in test definitions):\n// out_w = (in_w - 1) * stride_w + kernel_extent_w\n// out_h = (in_h - 1) * stride_h + kernel_extent_h\n//\n// Returns 0 on success, non-zero on error.\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body computes transposed-convolution output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// deconvolution_kernel (defined in kernel.cpp).\n//\n// Output formula (from ref_conv.h::ref_deconv2d):\n// ke_h = dilation_h * (kernel_h - 1) + 1\n// ke_w = dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + ke_h\n// out_w = (in_w - 1) * stride_w + ke_w\n\n#include \"deconv2d.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int out_c = 64;\nconstexpr int kernel_w = 4;\nconstexpr int kernel_h = 4;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (in_c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = out_c*in_c*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_deconv2d in ref_conv.h)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n\n // 2. Allocate top\n top.create(out_w, out_h, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to the Solution-supplied kernel\n return ncnn::deconvolution_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin512_cout512.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin512_cout512.json deleted file mode 100644 index add6fb40c5e81f533d5feea0c71e278c298dd2bb..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin512_cout512.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_kh4_kw4_sh2_sw2_cin512_cout512", - "definition": "deconv2d_kh4_kw4_sh2_sw2_cin512_cout512", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Deconvolution_arm wrapper for deconv2d_kh4_kw4_sh2_sw2_cin512_cout512. Same kernel.cpp content across all deconv2d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d Solution;\n// identical across all deconv2d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d.h by delegating to\n// ncnn::Deconvolution_arm. The shim armbench_entry_deconv2d (binding.cpp)\n// allocates the full transposed-convolution output\n// (out_w = (in_w-1)*stride_w + dilation_w*(kernel_w-1)+1, likewise out_h) and\n// does NOT crop, so we disable Deconvolution_arm's internal cut_padding\n// (pad_* / output_pad_* / output_w/h all zero). Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all deconv2d\n// Definitions.\n\n#include \"deconv2d.h\"\n#include \"deconvolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Deconvolution_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of\n // Deconvolution_arm's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d.h\n//\n// Calling-convention contract for every deconv2d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d/ must define the\n// `ncnn::deconvolution_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d` in deconv2d.cpp allocates the output using\n// the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolution.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (in_c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c * in_c * Kh * Kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, dilation=1 case used in test definitions):\n// out_w = (in_w - 1) * stride_w + kernel_extent_w\n// out_h = (in_h - 1) * stride_h + kernel_extent_h\n//\n// Returns 0 on success, non-zero on error.\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body computes transposed-convolution output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// deconvolution_kernel (defined in kernel.cpp).\n//\n// Output formula (from ref_conv.h::ref_deconv2d):\n// ke_h = dilation_h * (kernel_h - 1) + 1\n// ke_w = dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + ke_h\n// out_w = (in_w - 1) * stride_w + ke_w\n\n#include \"deconv2d.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int out_c = 512;\nconstexpr int kernel_w = 4;\nconstexpr int kernel_h = 4;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (in_c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = out_c*in_c*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_deconv2d in ref_conv.h)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n\n // 2. Allocate top\n top.create(out_w, out_h, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to the Solution-supplied kernel\n return ncnn::deconvolution_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin64_cout128.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin64_cout128.json deleted file mode 100644 index 8f0bdf622b3cfb7d99ad38cf5bb84ea7cb93d427..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin64_cout128.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_kh4_kw4_sh2_sw2_cin64_cout128", - "definition": "deconv2d_kh4_kw4_sh2_sw2_cin64_cout128", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::Deconvolution_arm wrapper for deconv2d_kh4_kw4_sh2_sw2_cin64_cout128. Same kernel.cpp content across all deconv2d Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d Solution;\n// identical across all deconv2d Definitions (const params arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d.h by delegating to\n// ncnn::Deconvolution_arm. The shim armbench_entry_deconv2d (binding.cpp)\n// allocates the full transposed-convolution output\n// (out_w = (in_w-1)*stride_w + dilation_w*(kernel_w-1)+1, likewise out_h) and\n// does NOT crop, so we disable Deconvolution_arm's internal cut_padding\n// (pad_* / output_pad_* / output_w/h all zero). Const params come from the\n// Definition's const axes, so this kernel.cpp is identical across all deconv2d\n// Definitions.\n\n#include \"deconv2d.h\"\n#include \"deconvolution_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n Deconvolution_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of\n // Deconvolution_arm's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d.h\n//\n// Calling-convention contract for every deconv2d Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d/ must define the\n// `ncnn::deconvolution_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d` in deconv2d.cpp allocates the output using\n// the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolution.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (in_c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (out_c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size out_c * in_c * Kh * Kw\n// bias_data: flat 1D Mat of size out_c, or empty Mat if no bias\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, dilation=1 case used in test definitions):\n// out_w = (in_w - 1) * stride_w + kernel_extent_w\n// out_h = (in_h - 1) * stride_h + kernel_extent_h\n//\n// Returns 0 on success, non-zero on error.\nint deconvolution_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d via ctypes with ONLY the Mat/Option pointers\n// (no scalar args). The body computes transposed-convolution output dims,\n// allocates the output Mat, then dispatches to the Solution's\n// deconvolution_kernel (defined in kernel.cpp).\n//\n// Output formula (from ref_conv.h::ref_deconv2d):\n// ke_h = dilation_h * (kernel_h - 1) + 1\n// ke_w = dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + ke_h\n// out_w = (in_w - 1) * stride_w + ke_w\n\n#include \"deconv2d.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int out_c = 128;\nconstexpr int kernel_w = 4;\nconstexpr int kernel_h = 4;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (in_c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = out_c*in_c*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_deconv2d in ref_conv.h)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n\n // 2. Allocate top\n top.create(out_w, out_h, out_c, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to the Solution-supplied kernel\n return ncnn::deconvolution_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024.json deleted file mode 100644 index 65e4495e59eb5168d1f44476889401fcc404b248..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024", - "definition": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::DeconvolutionDepthWise_arm wrapper for deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024. Same kernel.cpp content across all deconv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d_depthwise\n// Solution; identical across all deconv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d_depthwise.h by\n// delegating to ncnn::DeconvolutionDepthWise_arm. The shim\n// armbench_entry_deconv2d_depthwise (binding.cpp) allocates the full\n// transposed-convolution output (out_c == in_c) and does NOT crop, so we\n// disable the layer's internal cut_padding. group == channels (depthwise).\n// Const params come from the Definition's const axes, so this kernel.cpp is\n// identical across all deconv2d_depthwise Definitions.\n\n#include \"deconv2d_depthwise.h\"\n#include \"deconvolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n DeconvolutionDepthWise_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of the\n // layer's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.group = group;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d_depthwise.h\n//\n// Calling-convention contract for every deconv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d_depthwise/ must define the\n// `ncnn::deconvolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d_depthwise` in deconv2d_depthwise.cpp allocates\n// the output using the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolutiondepthwise.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, from ref_conv.h::ref_depthwise_deconv2d):\n// out_w = (in_w - 1) * stride_w + dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + dilation_h * (kernel_h - 1) + 1\n//\n// Returns 0 on success, non-zero on error.\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d_depthwise — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body computes transposed-convolution output\n// dims, allocates the output Mat, then dispatches to the Solution's\n// deconvolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"deconv2d_depthwise.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int kernel_w = 2;\nconstexpr int kernel_h = 2;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = C*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_depthwise_deconv2d)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n const int channels = bottom.c * bottom.elempack; // depthwise: out_c == in_c; elempack>1 when pack4\n\n // 2. Allocate top\n top.create(out_w, out_h, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::deconvolutiondepthwise_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c128.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c128.json deleted file mode 100644 index 1dfda16388db9d98ac604bf123cc0bede245489d..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c128.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c128", - "definition": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c128", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::DeconvolutionDepthWise_arm wrapper for deconv2d_depthwise_kh2_kw2_sh2_sw2_c128. Same kernel.cpp content across all deconv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d_depthwise\n// Solution; identical across all deconv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d_depthwise.h by\n// delegating to ncnn::DeconvolutionDepthWise_arm. The shim\n// armbench_entry_deconv2d_depthwise (binding.cpp) allocates the full\n// transposed-convolution output (out_c == in_c) and does NOT crop, so we\n// disable the layer's internal cut_padding. group == channels (depthwise).\n// Const params come from the Definition's const axes, so this kernel.cpp is\n// identical across all deconv2d_depthwise Definitions.\n\n#include \"deconv2d_depthwise.h\"\n#include \"deconvolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n DeconvolutionDepthWise_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of the\n // layer's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.group = group;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d_depthwise.h\n//\n// Calling-convention contract for every deconv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d_depthwise/ must define the\n// `ncnn::deconvolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d_depthwise` in deconv2d_depthwise.cpp allocates\n// the output using the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolutiondepthwise.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, from ref_conv.h::ref_depthwise_deconv2d):\n// out_w = (in_w - 1) * stride_w + dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + dilation_h * (kernel_h - 1) + 1\n//\n// Returns 0 on success, non-zero on error.\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d_depthwise — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body computes transposed-convolution output\n// dims, allocates the output Mat, then dispatches to the Solution's\n// deconvolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"deconv2d_depthwise.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int kernel_w = 2;\nconstexpr int kernel_h = 2;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = C*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_depthwise_deconv2d)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n const int channels = bottom.c * bottom.elempack; // depthwise: out_c == in_c; elempack>1 when pack4\n\n // 2. Allocate top\n top.create(out_w, out_h, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::deconvolutiondepthwise_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c256.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c256.json deleted file mode 100644 index 1bcb70736d0b58d1b21ccfe38dc3b6d41abc425a..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c256.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c256", - "definition": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c256", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::DeconvolutionDepthWise_arm wrapper for deconv2d_depthwise_kh2_kw2_sh2_sw2_c256. Same kernel.cpp content across all deconv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d_depthwise\n// Solution; identical across all deconv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d_depthwise.h by\n// delegating to ncnn::DeconvolutionDepthWise_arm. The shim\n// armbench_entry_deconv2d_depthwise (binding.cpp) allocates the full\n// transposed-convolution output (out_c == in_c) and does NOT crop, so we\n// disable the layer's internal cut_padding. group == channels (depthwise).\n// Const params come from the Definition's const axes, so this kernel.cpp is\n// identical across all deconv2d_depthwise Definitions.\n\n#include \"deconv2d_depthwise.h\"\n#include \"deconvolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n DeconvolutionDepthWise_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of the\n // layer's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.group = group;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d_depthwise.h\n//\n// Calling-convention contract for every deconv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d_depthwise/ must define the\n// `ncnn::deconvolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d_depthwise` in deconv2d_depthwise.cpp allocates\n// the output using the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolutiondepthwise.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, from ref_conv.h::ref_depthwise_deconv2d):\n// out_w = (in_w - 1) * stride_w + dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + dilation_h * (kernel_h - 1) + 1\n//\n// Returns 0 on success, non-zero on error.\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d_depthwise — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body computes transposed-convolution output\n// dims, allocates the output Mat, then dispatches to the Solution's\n// deconvolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"deconv2d_depthwise.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int kernel_w = 2;\nconstexpr int kernel_h = 2;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = C*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_depthwise_deconv2d)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n const int channels = bottom.c * bottom.elempack; // depthwise: out_c == in_c; elempack>1 when pack4\n\n // 2. Allocate top\n top.create(out_w, out_h, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::deconvolutiondepthwise_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c32.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c32.json deleted file mode 100644 index 317efd761e869c4914a275c849b391a938c26a5c..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c32.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c32", - "definition": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c32", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::DeconvolutionDepthWise_arm wrapper for deconv2d_depthwise_kh2_kw2_sh2_sw2_c32. Same kernel.cpp content across all deconv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d_depthwise\n// Solution; identical across all deconv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d_depthwise.h by\n// delegating to ncnn::DeconvolutionDepthWise_arm. The shim\n// armbench_entry_deconv2d_depthwise (binding.cpp) allocates the full\n// transposed-convolution output (out_c == in_c) and does NOT crop, so we\n// disable the layer's internal cut_padding. group == channels (depthwise).\n// Const params come from the Definition's const axes, so this kernel.cpp is\n// identical across all deconv2d_depthwise Definitions.\n\n#include \"deconv2d_depthwise.h\"\n#include \"deconvolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n DeconvolutionDepthWise_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of the\n // layer's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.group = group;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d_depthwise.h\n//\n// Calling-convention contract for every deconv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d_depthwise/ must define the\n// `ncnn::deconvolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d_depthwise` in deconv2d_depthwise.cpp allocates\n// the output using the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolutiondepthwise.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, from ref_conv.h::ref_depthwise_deconv2d):\n// out_w = (in_w - 1) * stride_w + dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + dilation_h * (kernel_h - 1) + 1\n//\n// Returns 0 on success, non-zero on error.\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d_depthwise — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body computes transposed-convolution output\n// dims, allocates the output Mat, then dispatches to the Solution's\n// deconvolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"deconv2d_depthwise.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int kernel_w = 2;\nconstexpr int kernel_h = 2;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = C*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_depthwise_deconv2d)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n const int channels = bottom.c * bottom.elempack; // depthwise: out_c == in_c; elempack>1 when pack4\n\n // 2. Allocate top\n top.create(out_w, out_h, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::deconvolutiondepthwise_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c512.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c512.json deleted file mode 100644 index 5b5f7e7ed487257292e5effe8801781abcd99344..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c512.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c512", - "definition": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c512", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::DeconvolutionDepthWise_arm wrapper for deconv2d_depthwise_kh2_kw2_sh2_sw2_c512. Same kernel.cpp content across all deconv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d_depthwise\n// Solution; identical across all deconv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d_depthwise.h by\n// delegating to ncnn::DeconvolutionDepthWise_arm. The shim\n// armbench_entry_deconv2d_depthwise (binding.cpp) allocates the full\n// transposed-convolution output (out_c == in_c) and does NOT crop, so we\n// disable the layer's internal cut_padding. group == channels (depthwise).\n// Const params come from the Definition's const axes, so this kernel.cpp is\n// identical across all deconv2d_depthwise Definitions.\n\n#include \"deconv2d_depthwise.h\"\n#include \"deconvolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n DeconvolutionDepthWise_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of the\n // layer's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.group = group;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d_depthwise.h\n//\n// Calling-convention contract for every deconv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d_depthwise/ must define the\n// `ncnn::deconvolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d_depthwise` in deconv2d_depthwise.cpp allocates\n// the output using the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolutiondepthwise.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, from ref_conv.h::ref_depthwise_deconv2d):\n// out_w = (in_w - 1) * stride_w + dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + dilation_h * (kernel_h - 1) + 1\n//\n// Returns 0 on success, non-zero on error.\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d_depthwise — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body computes transposed-convolution output\n// dims, allocates the output Mat, then dispatches to the Solution's\n// deconvolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"deconv2d_depthwise.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int kernel_w = 2;\nconstexpr int kernel_h = 2;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = C*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_depthwise_deconv2d)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n const int channels = bottom.c * bottom.elempack; // depthwise: out_c == in_c; elempack>1 when pack4\n\n // 2. Allocate top\n top.create(out_w, out_h, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::deconvolutiondepthwise_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c64.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c64.json deleted file mode 100644 index 73c977ea549a6c0be53da0a30d7f2cd664261bad..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c64.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c64", - "definition": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c64", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::DeconvolutionDepthWise_arm wrapper for deconv2d_depthwise_kh2_kw2_sh2_sw2_c64. Same kernel.cpp content across all deconv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d_depthwise\n// Solution; identical across all deconv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d_depthwise.h by\n// delegating to ncnn::DeconvolutionDepthWise_arm. The shim\n// armbench_entry_deconv2d_depthwise (binding.cpp) allocates the full\n// transposed-convolution output (out_c == in_c) and does NOT crop, so we\n// disable the layer's internal cut_padding. group == channels (depthwise).\n// Const params come from the Definition's const axes, so this kernel.cpp is\n// identical across all deconv2d_depthwise Definitions.\n\n#include \"deconv2d_depthwise.h\"\n#include \"deconvolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n DeconvolutionDepthWise_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of the\n // layer's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.group = group;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d_depthwise.h\n//\n// Calling-convention contract for every deconv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d_depthwise/ must define the\n// `ncnn::deconvolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d_depthwise` in deconv2d_depthwise.cpp allocates\n// the output using the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolutiondepthwise.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, from ref_conv.h::ref_depthwise_deconv2d):\n// out_w = (in_w - 1) * stride_w + dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + dilation_h * (kernel_h - 1) + 1\n//\n// Returns 0 on success, non-zero on error.\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d_depthwise — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body computes transposed-convolution output\n// dims, allocates the output Mat, then dispatches to the Solution's\n// deconvolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"deconv2d_depthwise.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int kernel_w = 2;\nconstexpr int kernel_h = 2;\nconstexpr int stride_w = 2;\nconstexpr int stride_h = 2;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = C*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_depthwise_deconv2d)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n const int channels = bottom.c * bottom.elempack; // depthwise: out_c == in_c; elempack>1 when pack4\n\n // 2. Allocate top\n top.create(out_w, out_h, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::deconvolutiondepthwise_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c128.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c128.json deleted file mode 100644 index f62731ccfd8c332d27fca36b5304cad7d8e4ca2e..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c128.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c128", - "definition": "deconv2d_depthwise_kh3_kw3_sh1_sw1_c128", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::DeconvolutionDepthWise_arm wrapper for deconv2d_depthwise_kh3_kw3_sh1_sw1_c128. Same kernel.cpp content across all deconv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d_depthwise\n// Solution; identical across all deconv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d_depthwise.h by\n// delegating to ncnn::DeconvolutionDepthWise_arm. The shim\n// armbench_entry_deconv2d_depthwise (binding.cpp) allocates the full\n// transposed-convolution output (out_c == in_c) and does NOT crop, so we\n// disable the layer's internal cut_padding. group == channels (depthwise).\n// Const params come from the Definition's const axes, so this kernel.cpp is\n// identical across all deconv2d_depthwise Definitions.\n\n#include \"deconv2d_depthwise.h\"\n#include \"deconvolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n DeconvolutionDepthWise_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of the\n // layer's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.group = group;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d_depthwise.h\n//\n// Calling-convention contract for every deconv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d_depthwise/ must define the\n// `ncnn::deconvolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d_depthwise` in deconv2d_depthwise.cpp allocates\n// the output using the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolutiondepthwise.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, from ref_conv.h::ref_depthwise_deconv2d):\n// out_w = (in_w - 1) * stride_w + dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + dilation_h * (kernel_h - 1) + 1\n//\n// Returns 0 on success, non-zero on error.\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d_depthwise — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body computes transposed-convolution output\n// dims, allocates the output Mat, then dispatches to the Solution's\n// deconvolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"deconv2d_depthwise.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = C*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_depthwise_deconv2d)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n const int channels = bottom.c * bottom.elempack; // depthwise: out_c == in_c; elempack>1 when pack4\n\n // 2. Allocate top\n top.create(out_w, out_h, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::deconvolutiondepthwise_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048.json deleted file mode 100644 index 7a9c6e3312bd7c1e707cc83e4d00e4dbba342c4a..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048", - "definition": "deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::DeconvolutionDepthWise_arm wrapper for deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048. Same kernel.cpp content across all deconv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d_depthwise\n// Solution; identical across all deconv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d_depthwise.h by\n// delegating to ncnn::DeconvolutionDepthWise_arm. The shim\n// armbench_entry_deconv2d_depthwise (binding.cpp) allocates the full\n// transposed-convolution output (out_c == in_c) and does NOT crop, so we\n// disable the layer's internal cut_padding. group == channels (depthwise).\n// Const params come from the Definition's const axes, so this kernel.cpp is\n// identical across all deconv2d_depthwise Definitions.\n\n#include \"deconv2d_depthwise.h\"\n#include \"deconvolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n DeconvolutionDepthWise_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of the\n // layer's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.group = group;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d_depthwise.h\n//\n// Calling-convention contract for every deconv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d_depthwise/ must define the\n// `ncnn::deconvolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d_depthwise` in deconv2d_depthwise.cpp allocates\n// the output using the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolutiondepthwise.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, from ref_conv.h::ref_depthwise_deconv2d):\n// out_w = (in_w - 1) * stride_w + dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + dilation_h * (kernel_h - 1) + 1\n//\n// Returns 0 on success, non-zero on error.\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d_depthwise — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body computes transposed-convolution output\n// dims, allocates the output Mat, then dispatches to the Solution's\n// deconvolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"deconv2d_depthwise.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = C*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_depthwise_deconv2d)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n const int channels = bottom.c * bottom.elempack; // depthwise: out_c == in_c; elempack>1 when pack4\n\n // 2. Allocate top\n top.create(out_w, out_h, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::deconvolutiondepthwise_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c32.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c32.json deleted file mode 100644 index a56527446ecbdb8373e551544044ff8b923959cb..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c32.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c32", - "definition": "deconv2d_depthwise_kh3_kw3_sh1_sw1_c32", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::DeconvolutionDepthWise_arm wrapper for deconv2d_depthwise_kh3_kw3_sh1_sw1_c32. Same kernel.cpp content across all deconv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d_depthwise\n// Solution; identical across all deconv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d_depthwise.h by\n// delegating to ncnn::DeconvolutionDepthWise_arm. The shim\n// armbench_entry_deconv2d_depthwise (binding.cpp) allocates the full\n// transposed-convolution output (out_c == in_c) and does NOT crop, so we\n// disable the layer's internal cut_padding. group == channels (depthwise).\n// Const params come from the Definition's const axes, so this kernel.cpp is\n// identical across all deconv2d_depthwise Definitions.\n\n#include \"deconv2d_depthwise.h\"\n#include \"deconvolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n DeconvolutionDepthWise_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of the\n // layer's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.group = group;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d_depthwise.h\n//\n// Calling-convention contract for every deconv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d_depthwise/ must define the\n// `ncnn::deconvolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d_depthwise` in deconv2d_depthwise.cpp allocates\n// the output using the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolutiondepthwise.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, from ref_conv.h::ref_depthwise_deconv2d):\n// out_w = (in_w - 1) * stride_w + dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + dilation_h * (kernel_h - 1) + 1\n//\n// Returns 0 on success, non-zero on error.\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d_depthwise — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body computes transposed-convolution output\n// dims, allocates the output Mat, then dispatches to the Solution's\n// deconvolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"deconv2d_depthwise.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = C*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_depthwise_deconv2d)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n const int channels = bottom.c * bottom.elempack; // depthwise: out_c == in_c; elempack>1 when pack4\n\n // 2. Allocate top\n top.create(out_w, out_h, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::deconvolutiondepthwise_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c512.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c512.json deleted file mode 100644 index 7d7f470806ee60b9137ec0a5d4501f10a6704af7..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c512.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c512", - "definition": "deconv2d_depthwise_kh3_kw3_sh1_sw1_c512", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::DeconvolutionDepthWise_arm wrapper for deconv2d_depthwise_kh3_kw3_sh1_sw1_c512. Same kernel.cpp content across all deconv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d_depthwise\n// Solution; identical across all deconv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d_depthwise.h by\n// delegating to ncnn::DeconvolutionDepthWise_arm. The shim\n// armbench_entry_deconv2d_depthwise (binding.cpp) allocates the full\n// transposed-convolution output (out_c == in_c) and does NOT crop, so we\n// disable the layer's internal cut_padding. group == channels (depthwise).\n// Const params come from the Definition's const axes, so this kernel.cpp is\n// identical across all deconv2d_depthwise Definitions.\n\n#include \"deconv2d_depthwise.h\"\n#include \"deconvolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n DeconvolutionDepthWise_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of the\n // layer's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.group = group;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d_depthwise.h\n//\n// Calling-convention contract for every deconv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d_depthwise/ must define the\n// `ncnn::deconvolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d_depthwise` in deconv2d_depthwise.cpp allocates\n// the output using the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolutiondepthwise.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, from ref_conv.h::ref_depthwise_deconv2d):\n// out_w = (in_w - 1) * stride_w + dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + dilation_h * (kernel_h - 1) + 1\n//\n// Returns 0 on success, non-zero on error.\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d_depthwise — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body computes transposed-convolution output\n// dims, allocates the output Mat, then dispatches to the Solution's\n// deconvolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"deconv2d_depthwise.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = C*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_depthwise_deconv2d)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n const int channels = bottom.c * bottom.elempack; // depthwise: out_c == in_c; elempack>1 when pack4\n\n // 2. Allocate top\n top.create(out_w, out_h, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::deconvolutiondepthwise_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c64.json b/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c64.json deleted file mode 100644 index 5ebd3e463904d2c2ee409060056fd97fc63af819..0000000000000000000000000000000000000000 --- a/solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c64.json +++ /dev/null @@ -1,43 +0,0 @@ -{ - "name": "baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c64", - "definition": "deconv2d_depthwise_kh3_kw3_sh1_sw1_c64", - "dataset": "ncnn", - "author": "baseline-ncnn-arm", - "description": "ncnn::DeconvolutionDepthWise_arm wrapper for deconv2d_depthwise_kh3_kw3_sh1_sw1_c64. Same kernel.cpp content across all deconv2d_depthwise Definitions; const params come from the Definition. Times create_pipeline + forward.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "binding.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [ - "ncnn", - "ncnn_arm_heavy", - "openmp" - ], - "isa_features": [ - "sve" - ], - "compile_flags": [ - "-O3", - "-march=armv8.2-a+fp16+dotprod+sve", - "-fopenmp" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "kernel.cpp", - "content": "// solutions/ncnn/baseline-ncnn-arm/deconv2d_depthwise//kernel.cpp\n//\n// Embedded verbatim as kernel.cpp in every baseline-ncnn-arm deconv2d_depthwise\n// Solution; identical across all deconv2d_depthwise Definitions (const params\n// arrive as runtime args).\n//\n// Implements the harness contract declared in deconv2d_depthwise.h by\n// delegating to ncnn::DeconvolutionDepthWise_arm. The shim\n// armbench_entry_deconv2d_depthwise (binding.cpp) allocates the full\n// transposed-convolution output (out_c == in_c) and does NOT crop, so we\n// disable the layer's internal cut_padding. group == channels (depthwise).\n// Const params come from the Definition's const axes, so this kernel.cpp is\n// identical across all deconv2d_depthwise Definitions.\n\n#include \"deconv2d_depthwise.h\"\n#include \"deconvolutiondepthwise_arm.h\"\n\n#include \n\nnamespace ncnn {\n\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt)\n{\n DeconvolutionDepthWise_arm deconv;\n deconv.num_output = top_blob.c;\n deconv.kernel_w = kernel_w; deconv.kernel_h = kernel_h;\n deconv.stride_w = stride_w; deconv.stride_h = stride_h;\n deconv.dilation_w = dilation_w; deconv.dilation_h = dilation_h;\n // Harness allocated the full (un-cropped) output, so disable all of the\n // layer's output padding / cropping.\n deconv.pad_left = 0; deconv.pad_right = 0;\n deconv.pad_top = 0; deconv.pad_bottom = 0;\n deconv.output_pad_right = 0; deconv.output_pad_bottom = 0;\n deconv.output_w = 0; deconv.output_h = 0;\n deconv.bias_term = bias_data.empty() ? 0 : 1;\n deconv.weight_data_size = weight_data.w;\n deconv.group = group;\n deconv.activation_type = activation_type;\n deconv.activation_params = const_cast(activation_params);\n deconv.dynamic_weight = 0;\n deconv.weight_data = const_cast(weight_data);\n if (!bias_data.empty()) deconv.bias_data = const_cast(bias_data);\n\n if (deconv.create_pipeline(opt) != 0) return -1;\n\n // forward() may produce pack4 output when use_packing_layout=true;\n // the Python harness unwraps pack4->pack1 in unwrap_output.\n return deconv.forward(bottom_blob, top_blob, opt);\n}\n\n} // namespace ncnn\n" - }, - { - "path": "deconv2d_depthwise.h", - "content": "// bench/compile/builders/ncnn_harness/deconv2d_depthwise.h\n//\n// Calling-convention contract for every deconv2d_depthwise Solution in the \"ncnn\" dataset.\n//\n// Every Solution under solutions/ncnn//deconv2d_depthwise/ must define the\n// `ncnn::deconvolutiondepthwise_kernel` symbol with EXACTLY this signature. The shared\n// shim `armbench_entry_deconv2d_depthwise` in deconv2d_depthwise.cpp allocates\n// the output using the transposed-convolution output formula, then calls into this symbol.\n// The Solution author writes the kernel body; everything else is the harness's job.\n//\n// This matches starter/ncnn/candidate/deconvolutiondepthwise.h signature byte-for-byte.\n\n#ifndef ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n#define ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n\n#include \"mat.h\"\n#include \"option.h\"\n\nnamespace ncnn {\n\n// The solution-supplied kernel.\n//\n// bottom_blob: input, dims=3, layout (c, h, w), float32\n// top_blob: pre-allocated output, dims=3, layout (c, out_h, out_w), float32\n// weight_data: flat 1D Mat of size C * Kh * Kw (depthwise: C_in == C_out)\n// bias_data: flat 1D Mat of size C, or empty Mat if no bias\n// group: number of groups == number of channels (depthwise)\n// activation_params: per-activation params Mat (empty for type=0/1=relu)\n// opt: ncnn options (num_threads=1 enforced by bench/runner.py)\n//\n// Output dims (no padding, from ref_conv.h::ref_depthwise_deconv2d):\n// out_w = (in_w - 1) * stride_w + dilation_w * (kernel_w - 1) + 1\n// out_h = (in_h - 1) * stride_h + dilation_h * (kernel_h - 1) + 1\n//\n// Returns 0 on success, non-zero on error.\nint deconvolutiondepthwise_kernel(const Mat& bottom_blob, Mat& top_blob,\n const Mat& weight_data, const Mat& bias_data,\n int kernel_w, int kernel_h,\n int stride_w, int stride_h,\n int dilation_w, int dilation_h,\n int group,\n int activation_type, const Mat& activation_params,\n const Option& opt);\n\n} // namespace ncnn\n\n#endif // ARMBENCH_NCNN_HARNESS_DECONV2D_DEPTHWISE_H\n" - }, - { - "path": "binding.cpp", - "content": "// ncnn::Mat shim for deconv2d_depthwise — shipped in this Solution's own sources.\n//\n// All per-Definition scalar params are baked as constexpr below, so the runner\n// calls armbench_entry_deconv2d_depthwise via ctypes with ONLY the Mat/Option\n// pointers (no scalar args). The body computes transposed-convolution output\n// dims, allocates the output Mat, then dispatches to the Solution's\n// deconvolutiondepthwise_kernel (defined in kernel.cpp).\n//\n// Depthwise: in_c == out_c == channels; group == channels (derived from input).\n\n#include \"deconv2d_depthwise.h\"\n\n// ── Per-Definition constants (baked from the Definition's const axes) ────────\nnamespace {\nconstexpr int kernel_w = 3;\nconstexpr int kernel_h = 3;\nconstexpr int stride_w = 1;\nconstexpr int stride_h = 1;\nconstexpr int dilation_w = 1;\nconstexpr int dilation_h = 1;\nconstexpr int activation_type = 0;\n} // namespace\n\nextern \"C\" {\n\nint armbench_entry_deconv2d_depthwise(\n // tensors (opaque ncnn::Mat*)\n void* bottom_mat_v, // const ncnn::Mat* — 3D input (c, h, w)\n void* top_mat_v, // ncnn::Mat* — empty Mat; we create() output here\n void* weight_mat_v, // const ncnn::Mat* — flat 1D Mat, size = C*kh*kw\n void* bias_mat_v, // const ncnn::Mat* — flat 1D Mat or empty\n void* activation_params_v, // const ncnn::Mat* — empty for type=0/1\n void* opt_v) // const ncnn::Option*\n{\n const auto& bottom = *reinterpret_cast(bottom_mat_v);\n auto& top = *reinterpret_cast(top_mat_v);\n const auto& weight = *reinterpret_cast(weight_mat_v);\n const auto& bias = *reinterpret_cast(bias_mat_v);\n const auto& act_par = *reinterpret_cast(activation_params_v);\n const auto& opt = *reinterpret_cast(opt_v);\n\n if (bottom.empty()) return -100;\n\n // 1. Compute transposed-convolution output dims (matches ref_depthwise_deconv2d)\n const int ke_h = dilation_h * (kernel_h - 1) + 1;\n const int ke_w = dilation_w * (kernel_w - 1) + 1;\n const int out_w = (bottom.w - 1) * stride_w + ke_w;\n const int out_h = (bottom.h - 1) * stride_h + ke_h;\n const int channels = bottom.c * bottom.elempack; // depthwise: out_c == in_c; elempack>1 when pack4\n\n // 2. Allocate top\n top.create(out_w, out_h, channels, (size_t)4u, opt.blob_allocator);\n if (top.empty()) return -100;\n\n // 3. Dispatch to Solution kernel with group == channels (depthwise)\n return ncnn::deconvolutiondepthwise_kernel(\n bottom, top, weight, bias,\n kernel_w, kernel_h,\n stride_w, stride_h,\n dilation_w, dilation_h,\n channels,\n activation_type, act_par,\n opt);\n}\n\n} // extern \"C\"\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin1024_cout1024.json b/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin1024_cout1024.json deleted file mode 100644 index 7f3f915511baa43f8f2edcca73bed8bc17d3d119..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin1024_cout1024.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv1d_kw1_sw1_dw1_p0_cin1024_cout1024", - "definition": "conv1d_kw1_sw1_dw1_p0_cin1024_cout1024", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv1d for conv1d_kw1_sw1_dw1_p0_cin1024_cout1024. Constexpr-baked dims; armbench_entry_conv1d calls inner_conv1d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv1d.cpp::armbench_entry_conv1d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv1d.h", - "content": "#pragma once\n\n// Per-definition constants for this conv1d specialisation.\nnamespace conv1d_def {\nconstexpr int Cin = 1024;\nconstexpr int Cout = 1024;\nconstexpr int Kw = 1;\nconstexpr int Sw = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 0;\n} // namespace conv1d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// W and W_out are var dims; pre-computed by the binding harness.\n// Input layout: (Cin, W), output layout: (Cout, W_out).\nvoid inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv1d.cpp", - "content": "// Binding harness: computes W_out from constexpr params, calls inner_conv1d.\n// ABI: armbench_entry_conv1d(input*, output*, weight*, bias*, W)\n// Input layout: (Cin, W); output layout: (Cout, W_out).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" int armbench_entry_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W)\n{\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv1d(input, output, weight, bias, W, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv1d.\n// LLM target: replace this file with an optimised inner_conv1d.\n// All per-definition constants live in conv1d_def:: (conv1d.h).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" void inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out)\n{\n for (int oc = 0; oc < Cout; ++oc) {\n float* out_c = output + (long)oc * W_out;\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[oc];\n for (int ic = 0; ic < Cin; ++ic) {\n const float* in_c = input + (long)ic * W;\n for (int kw = 0; kw < Kw; ++kw) {\n int iw = ow * Sw - pad + kw * Dw;\n if (iw >= 0 && iw < W)\n sum += in_c[iw] * weight[((long)oc * Cin + ic) * Kw + kw];\n }\n }\n out_c[ow] = sum;\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin128_cout512.json b/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin128_cout512.json deleted file mode 100644 index 26e0a232c8297af1fd4759ff7d7d296c215cfaf1..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin128_cout512.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv1d_kw1_sw1_dw1_p0_cin128_cout512", - "definition": "conv1d_kw1_sw1_dw1_p0_cin128_cout512", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv1d for conv1d_kw1_sw1_dw1_p0_cin128_cout512. Constexpr-baked dims; armbench_entry_conv1d calls inner_conv1d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv1d.cpp::armbench_entry_conv1d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv1d.h", - "content": "#pragma once\n\n// Per-definition constants for this conv1d specialisation.\nnamespace conv1d_def {\nconstexpr int Cin = 128;\nconstexpr int Cout = 512;\nconstexpr int Kw = 1;\nconstexpr int Sw = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 0;\n} // namespace conv1d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// W and W_out are var dims; pre-computed by the binding harness.\n// Input layout: (Cin, W), output layout: (Cout, W_out).\nvoid inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv1d.cpp", - "content": "// Binding harness: computes W_out from constexpr params, calls inner_conv1d.\n// ABI: armbench_entry_conv1d(input*, output*, weight*, bias*, W)\n// Input layout: (Cin, W); output layout: (Cout, W_out).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" int armbench_entry_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W)\n{\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv1d(input, output, weight, bias, W, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv1d.\n// LLM target: replace this file with an optimised inner_conv1d.\n// All per-definition constants live in conv1d_def:: (conv1d.h).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" void inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out)\n{\n for (int oc = 0; oc < Cout; ++oc) {\n float* out_c = output + (long)oc * W_out;\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[oc];\n for (int ic = 0; ic < Cin; ++ic) {\n const float* in_c = input + (long)ic * W;\n for (int kw = 0; kw < Kw; ++kw) {\n int iw = ow * Sw - pad + kw * Dw;\n if (iw >= 0 && iw < W)\n sum += in_c[iw] * weight[((long)oc * Cin + ic) * Kw + kw];\n }\n }\n out_c[ow] = sum;\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin256_cout1024.json b/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin256_cout1024.json deleted file mode 100644 index 0ca1200e70b34834482da1b3323ee192dd2de9ca..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin256_cout1024.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv1d_kw1_sw1_dw1_p0_cin256_cout1024", - "definition": "conv1d_kw1_sw1_dw1_p0_cin256_cout1024", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv1d for conv1d_kw1_sw1_dw1_p0_cin256_cout1024. Constexpr-baked dims; armbench_entry_conv1d calls inner_conv1d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv1d.cpp::armbench_entry_conv1d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv1d.h", - "content": "#pragma once\n\n// Per-definition constants for this conv1d specialisation.\nnamespace conv1d_def {\nconstexpr int Cin = 256;\nconstexpr int Cout = 1024;\nconstexpr int Kw = 1;\nconstexpr int Sw = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 0;\n} // namespace conv1d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// W and W_out are var dims; pre-computed by the binding harness.\n// Input layout: (Cin, W), output layout: (Cout, W_out).\nvoid inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv1d.cpp", - "content": "// Binding harness: computes W_out from constexpr params, calls inner_conv1d.\n// ABI: armbench_entry_conv1d(input*, output*, weight*, bias*, W)\n// Input layout: (Cin, W); output layout: (Cout, W_out).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" int armbench_entry_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W)\n{\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv1d(input, output, weight, bias, W, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv1d.\n// LLM target: replace this file with an optimised inner_conv1d.\n// All per-definition constants live in conv1d_def:: (conv1d.h).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" void inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out)\n{\n for (int oc = 0; oc < Cout; ++oc) {\n float* out_c = output + (long)oc * W_out;\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[oc];\n for (int ic = 0; ic < Cin; ++ic) {\n const float* in_c = input + (long)ic * W;\n for (int kw = 0; kw < Kw; ++kw) {\n int iw = ow * Sw - pad + kw * Dw;\n if (iw >= 0 && iw < W)\n sum += in_c[iw] * weight[((long)oc * Cin + ic) * Kw + kw];\n }\n }\n out_c[ow] = sum;\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin512_cout2048.json b/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin512_cout2048.json deleted file mode 100644 index 7760413cd03e62281ea4a392c4f453212b56bc03..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin512_cout2048.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv1d_kw1_sw1_dw1_p0_cin512_cout2048", - "definition": "conv1d_kw1_sw1_dw1_p0_cin512_cout2048", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv1d for conv1d_kw1_sw1_dw1_p0_cin512_cout2048. Constexpr-baked dims; armbench_entry_conv1d calls inner_conv1d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv1d.cpp::armbench_entry_conv1d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv1d.h", - "content": "#pragma once\n\n// Per-definition constants for this conv1d specialisation.\nnamespace conv1d_def {\nconstexpr int Cin = 512;\nconstexpr int Cout = 2048;\nconstexpr int Kw = 1;\nconstexpr int Sw = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 0;\n} // namespace conv1d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// W and W_out are var dims; pre-computed by the binding harness.\n// Input layout: (Cin, W), output layout: (Cout, W_out).\nvoid inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv1d.cpp", - "content": "// Binding harness: computes W_out from constexpr params, calls inner_conv1d.\n// ABI: armbench_entry_conv1d(input*, output*, weight*, bias*, W)\n// Input layout: (Cin, W); output layout: (Cout, W_out).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" int armbench_entry_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W)\n{\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv1d(input, output, weight, bias, W, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv1d.\n// LLM target: replace this file with an optimised inner_conv1d.\n// All per-definition constants live in conv1d_def:: (conv1d.h).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" void inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out)\n{\n for (int oc = 0; oc < Cout; ++oc) {\n float* out_c = output + (long)oc * W_out;\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[oc];\n for (int ic = 0; ic < Cin; ++ic) {\n const float* in_c = input + (long)ic * W;\n for (int kw = 0; kw < Kw; ++kw) {\n int iw = ow * Sw - pad + kw * Dw;\n if (iw >= 0 && iw < W)\n sum += in_c[iw] * weight[((long)oc * Cin + ic) * Kw + kw];\n }\n }\n out_c[ow] = sum;\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout256.json b/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout256.json deleted file mode 100644 index c68f47582e51aceb5bdb998b035bcc10a11bdc6a..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout256.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv1d_kw1_sw1_dw1_p0_cin64_cout256", - "definition": "conv1d_kw1_sw1_dw1_p0_cin64_cout256", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv1d for conv1d_kw1_sw1_dw1_p0_cin64_cout256. Constexpr-baked dims; armbench_entry_conv1d calls inner_conv1d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv1d.cpp::armbench_entry_conv1d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv1d.h", - "content": "#pragma once\n\n// Per-definition constants for this conv1d specialisation.\nnamespace conv1d_def {\nconstexpr int Cin = 64;\nconstexpr int Cout = 256;\nconstexpr int Kw = 1;\nconstexpr int Sw = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 0;\n} // namespace conv1d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// W and W_out are var dims; pre-computed by the binding harness.\n// Input layout: (Cin, W), output layout: (Cout, W_out).\nvoid inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv1d.cpp", - "content": "// Binding harness: computes W_out from constexpr params, calls inner_conv1d.\n// ABI: armbench_entry_conv1d(input*, output*, weight*, bias*, W)\n// Input layout: (Cin, W); output layout: (Cout, W_out).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" int armbench_entry_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W)\n{\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv1d(input, output, weight, bias, W, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv1d.\n// LLM target: replace this file with an optimised inner_conv1d.\n// All per-definition constants live in conv1d_def:: (conv1d.h).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" void inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out)\n{\n for (int oc = 0; oc < Cout; ++oc) {\n float* out_c = output + (long)oc * W_out;\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[oc];\n for (int ic = 0; ic < Cin; ++ic) {\n const float* in_c = input + (long)ic * W;\n for (int kw = 0; kw < Kw; ++kw) {\n int iw = ow * Sw - pad + kw * Dw;\n if (iw >= 0 && iw < W)\n sum += in_c[iw] * weight[((long)oc * Cin + ic) * Kw + kw];\n }\n }\n out_c[ow] = sum;\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout64.json b/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout64.json deleted file mode 100644 index 03d51e2f3df54aad23e86e4e1f5fc4d0cdc1ea77..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout64.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv1d_kw1_sw1_dw1_p0_cin64_cout64", - "definition": "conv1d_kw1_sw1_dw1_p0_cin64_cout64", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv1d for conv1d_kw1_sw1_dw1_p0_cin64_cout64. Constexpr-baked dims; armbench_entry_conv1d calls inner_conv1d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv1d.cpp::armbench_entry_conv1d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv1d.h", - "content": "#pragma once\n\n// Per-definition constants for this conv1d specialisation.\nnamespace conv1d_def {\nconstexpr int Cin = 64;\nconstexpr int Cout = 64;\nconstexpr int Kw = 1;\nconstexpr int Sw = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 0;\n} // namespace conv1d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// W and W_out are var dims; pre-computed by the binding harness.\n// Input layout: (Cin, W), output layout: (Cout, W_out).\nvoid inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv1d.cpp", - "content": "// Binding harness: computes W_out from constexpr params, calls inner_conv1d.\n// ABI: armbench_entry_conv1d(input*, output*, weight*, bias*, W)\n// Input layout: (Cin, W); output layout: (Cout, W_out).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" int armbench_entry_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W)\n{\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv1d(input, output, weight, bias, W, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv1d.\n// LLM target: replace this file with an optimised inner_conv1d.\n// All per-definition constants live in conv1d_def:: (conv1d.h).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" void inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out)\n{\n for (int oc = 0; oc < Cout; ++oc) {\n float* out_c = output + (long)oc * W_out;\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[oc];\n for (int ic = 0; ic < Cin; ++ic) {\n const float* in_c = input + (long)ic * W;\n for (int kw = 0; kw < Kw; ++kw) {\n int iw = ow * Sw - pad + kw * Dw;\n if (iw >= 0 && iw < W)\n sum += in_c[iw] * weight[((long)oc * Cin + ic) * Kw + kw];\n }\n }\n out_c[ow] = sum;\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin1024_cout1024.json b/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin1024_cout1024.json deleted file mode 100644 index 6813654d349e511a3b93bf469859248a550db41d..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin1024_cout1024.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv1d_kw3_sw1_dw1_p1_cin1024_cout1024", - "definition": "conv1d_kw3_sw1_dw1_p1_cin1024_cout1024", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv1d for conv1d_kw3_sw1_dw1_p1_cin1024_cout1024. Constexpr-baked dims; armbench_entry_conv1d calls inner_conv1d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv1d.cpp::armbench_entry_conv1d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv1d.h", - "content": "#pragma once\n\n// Per-definition constants for this conv1d specialisation.\nnamespace conv1d_def {\nconstexpr int Cin = 1024;\nconstexpr int Cout = 1024;\nconstexpr int Kw = 3;\nconstexpr int Sw = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 1;\n} // namespace conv1d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// W and W_out are var dims; pre-computed by the binding harness.\n// Input layout: (Cin, W), output layout: (Cout, W_out).\nvoid inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv1d.cpp", - "content": "// Binding harness: computes W_out from constexpr params, calls inner_conv1d.\n// ABI: armbench_entry_conv1d(input*, output*, weight*, bias*, W)\n// Input layout: (Cin, W); output layout: (Cout, W_out).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" int armbench_entry_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W)\n{\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv1d(input, output, weight, bias, W, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv1d.\n// LLM target: replace this file with an optimised inner_conv1d.\n// All per-definition constants live in conv1d_def:: (conv1d.h).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" void inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out)\n{\n for (int oc = 0; oc < Cout; ++oc) {\n float* out_c = output + (long)oc * W_out;\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[oc];\n for (int ic = 0; ic < Cin; ++ic) {\n const float* in_c = input + (long)ic * W;\n for (int kw = 0; kw < Kw; ++kw) {\n int iw = ow * Sw - pad + kw * Dw;\n if (iw >= 0 && iw < W)\n sum += in_c[iw] * weight[((long)oc * Cin + ic) * Kw + kw];\n }\n }\n out_c[ow] = sum;\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin128_cout256.json b/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin128_cout256.json deleted file mode 100644 index 7cf74d386f0a54060b46284fced8ec55a03ed0dd..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin128_cout256.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv1d_kw3_sw1_dw1_p1_cin128_cout256", - "definition": "conv1d_kw3_sw1_dw1_p1_cin128_cout256", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv1d for conv1d_kw3_sw1_dw1_p1_cin128_cout256. Constexpr-baked dims; armbench_entry_conv1d calls inner_conv1d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv1d.cpp::armbench_entry_conv1d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv1d.h", - "content": "#pragma once\n\n// Per-definition constants for this conv1d specialisation.\nnamespace conv1d_def {\nconstexpr int Cin = 128;\nconstexpr int Cout = 256;\nconstexpr int Kw = 3;\nconstexpr int Sw = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 1;\n} // namespace conv1d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// W and W_out are var dims; pre-computed by the binding harness.\n// Input layout: (Cin, W), output layout: (Cout, W_out).\nvoid inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv1d.cpp", - "content": "// Binding harness: computes W_out from constexpr params, calls inner_conv1d.\n// ABI: armbench_entry_conv1d(input*, output*, weight*, bias*, W)\n// Input layout: (Cin, W); output layout: (Cout, W_out).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" int armbench_entry_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W)\n{\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv1d(input, output, weight, bias, W, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv1d.\n// LLM target: replace this file with an optimised inner_conv1d.\n// All per-definition constants live in conv1d_def:: (conv1d.h).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" void inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out)\n{\n for (int oc = 0; oc < Cout; ++oc) {\n float* out_c = output + (long)oc * W_out;\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[oc];\n for (int ic = 0; ic < Cin; ++ic) {\n const float* in_c = input + (long)ic * W;\n for (int kw = 0; kw < Kw; ++kw) {\n int iw = ow * Sw - pad + kw * Dw;\n if (iw >= 0 && iw < W)\n sum += in_c[iw] * weight[((long)oc * Cin + ic) * Kw + kw];\n }\n }\n out_c[ow] = sum;\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin256_cout512.json b/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin256_cout512.json deleted file mode 100644 index b214b413b12145f41ec70ac9a45a1bfc257e05c3..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin256_cout512.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv1d_kw3_sw1_dw1_p1_cin256_cout512", - "definition": "conv1d_kw3_sw1_dw1_p1_cin256_cout512", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv1d for conv1d_kw3_sw1_dw1_p1_cin256_cout512. Constexpr-baked dims; armbench_entry_conv1d calls inner_conv1d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv1d.cpp::armbench_entry_conv1d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv1d.h", - "content": "#pragma once\n\n// Per-definition constants for this conv1d specialisation.\nnamespace conv1d_def {\nconstexpr int Cin = 256;\nconstexpr int Cout = 512;\nconstexpr int Kw = 3;\nconstexpr int Sw = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 1;\n} // namespace conv1d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// W and W_out are var dims; pre-computed by the binding harness.\n// Input layout: (Cin, W), output layout: (Cout, W_out).\nvoid inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv1d.cpp", - "content": "// Binding harness: computes W_out from constexpr params, calls inner_conv1d.\n// ABI: armbench_entry_conv1d(input*, output*, weight*, bias*, W)\n// Input layout: (Cin, W); output layout: (Cout, W_out).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" int armbench_entry_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W)\n{\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv1d(input, output, weight, bias, W, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv1d.\n// LLM target: replace this file with an optimised inner_conv1d.\n// All per-definition constants live in conv1d_def:: (conv1d.h).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" void inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out)\n{\n for (int oc = 0; oc < Cout; ++oc) {\n float* out_c = output + (long)oc * W_out;\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[oc];\n for (int ic = 0; ic < Cin; ++ic) {\n const float* in_c = input + (long)ic * W;\n for (int kw = 0; kw < Kw; ++kw) {\n int iw = ow * Sw - pad + kw * Dw;\n if (iw >= 0 && iw < W)\n sum += in_c[iw] * weight[((long)oc * Cin + ic) * Kw + kw];\n }\n }\n out_c[ow] = sum;\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin3_cout64.json b/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin3_cout64.json deleted file mode 100644 index 94ed7d7ef7f68876005a2bd2acad895e9db4bb57..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin3_cout64.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv1d_kw3_sw1_dw1_p1_cin3_cout64", - "definition": "conv1d_kw3_sw1_dw1_p1_cin3_cout64", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv1d for conv1d_kw3_sw1_dw1_p1_cin3_cout64. Constexpr-baked dims; armbench_entry_conv1d calls inner_conv1d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv1d.cpp::armbench_entry_conv1d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv1d.h", - "content": "#pragma once\n\n// Per-definition constants for this conv1d specialisation.\nnamespace conv1d_def {\nconstexpr int Cin = 3;\nconstexpr int Cout = 64;\nconstexpr int Kw = 3;\nconstexpr int Sw = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 1;\n} // namespace conv1d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// W and W_out are var dims; pre-computed by the binding harness.\n// Input layout: (Cin, W), output layout: (Cout, W_out).\nvoid inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv1d.cpp", - "content": "// Binding harness: computes W_out from constexpr params, calls inner_conv1d.\n// ABI: armbench_entry_conv1d(input*, output*, weight*, bias*, W)\n// Input layout: (Cin, W); output layout: (Cout, W_out).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" int armbench_entry_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W)\n{\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv1d(input, output, weight, bias, W, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv1d.\n// LLM target: replace this file with an optimised inner_conv1d.\n// All per-definition constants live in conv1d_def:: (conv1d.h).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" void inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out)\n{\n for (int oc = 0; oc < Cout; ++oc) {\n float* out_c = output + (long)oc * W_out;\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[oc];\n for (int ic = 0; ic < Cin; ++ic) {\n const float* in_c = input + (long)ic * W;\n for (int kw = 0; kw < Kw; ++kw) {\n int iw = ow * Sw - pad + kw * Dw;\n if (iw >= 0 && iw < W)\n sum += in_c[iw] * weight[((long)oc * Cin + ic) * Kw + kw];\n }\n }\n out_c[ow] = sum;\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin512_cout1024.json b/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin512_cout1024.json deleted file mode 100644 index c6cc172618cbe829e7138d38f0e84f10080947bd..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin512_cout1024.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv1d_kw3_sw1_dw1_p1_cin512_cout1024", - "definition": "conv1d_kw3_sw1_dw1_p1_cin512_cout1024", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv1d for conv1d_kw3_sw1_dw1_p1_cin512_cout1024. Constexpr-baked dims; armbench_entry_conv1d calls inner_conv1d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv1d.cpp::armbench_entry_conv1d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv1d.h", - "content": "#pragma once\n\n// Per-definition constants for this conv1d specialisation.\nnamespace conv1d_def {\nconstexpr int Cin = 512;\nconstexpr int Cout = 1024;\nconstexpr int Kw = 3;\nconstexpr int Sw = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 1;\n} // namespace conv1d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// W and W_out are var dims; pre-computed by the binding harness.\n// Input layout: (Cin, W), output layout: (Cout, W_out).\nvoid inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv1d.cpp", - "content": "// Binding harness: computes W_out from constexpr params, calls inner_conv1d.\n// ABI: armbench_entry_conv1d(input*, output*, weight*, bias*, W)\n// Input layout: (Cin, W); output layout: (Cout, W_out).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" int armbench_entry_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W)\n{\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv1d(input, output, weight, bias, W, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv1d.\n// LLM target: replace this file with an optimised inner_conv1d.\n// All per-definition constants live in conv1d_def:: (conv1d.h).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" void inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out)\n{\n for (int oc = 0; oc < Cout; ++oc) {\n float* out_c = output + (long)oc * W_out;\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[oc];\n for (int ic = 0; ic < Cin; ++ic) {\n const float* in_c = input + (long)ic * W;\n for (int kw = 0; kw < Kw; ++kw) {\n int iw = ow * Sw - pad + kw * Dw;\n if (iw >= 0 && iw < W)\n sum += in_c[iw] * weight[((long)oc * Cin + ic) * Kw + kw];\n }\n }\n out_c[ow] = sum;\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin64_cout128.json b/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin64_cout128.json deleted file mode 100644 index 3449f5f9cbeae0fe615051afa81c6689259edcc0..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv1d/conv1d_kw3_sw1_dw1_p1_cin64_cout128.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv1d_kw3_sw1_dw1_p1_cin64_cout128", - "definition": "conv1d_kw3_sw1_dw1_p1_cin64_cout128", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv1d for conv1d_kw3_sw1_dw1_p1_cin64_cout128. Constexpr-baked dims; armbench_entry_conv1d calls inner_conv1d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv1d.cpp::armbench_entry_conv1d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv1d.h", - "content": "#pragma once\n\n// Per-definition constants for this conv1d specialisation.\nnamespace conv1d_def {\nconstexpr int Cin = 64;\nconstexpr int Cout = 128;\nconstexpr int Kw = 3;\nconstexpr int Sw = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 1;\n} // namespace conv1d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// W and W_out are var dims; pre-computed by the binding harness.\n// Input layout: (Cin, W), output layout: (Cout, W_out).\nvoid inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv1d.cpp", - "content": "// Binding harness: computes W_out from constexpr params, calls inner_conv1d.\n// ABI: armbench_entry_conv1d(input*, output*, weight*, bias*, W)\n// Input layout: (Cin, W); output layout: (Cout, W_out).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" int armbench_entry_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W)\n{\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv1d(input, output, weight, bias, W, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv1d.\n// LLM target: replace this file with an optimised inner_conv1d.\n// All per-definition constants live in conv1d_def:: (conv1d.h).\n#include \"conv1d.h\"\nusing namespace conv1d_def;\n\nextern \"C\" void inner_conv1d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int W, int W_out)\n{\n for (int oc = 0; oc < Cout; ++oc) {\n float* out_c = output + (long)oc * W_out;\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[oc];\n for (int ic = 0; ic < Cin; ++ic) {\n const float* in_c = input + (long)ic * W;\n for (int kw = 0; kw < Kw; ++kw) {\n int iw = ow * Sw - pad + kw * Dw;\n if (iw >= 0 && iw < W)\n sum += in_c[iw] * weight[((long)oc * Cin + ic) * Kw + kw];\n }\n }\n out_c[ow] = sum;\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024.json deleted file mode 100644 index 40e4448745d18dbf1658e65e02f118dd1f978484..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024", - "definition": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 1024;\nconstexpr int Cout = 1024;\nconstexpr int Kh = 1;\nconstexpr int Kw = 1;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 0;\nconstexpr int pad_left = 0;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512.json deleted file mode 100644 index fb62a316d33482ca09838a392fca30946f83ef5f..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512", - "definition": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 128;\nconstexpr int Cout = 512;\nconstexpr int Kh = 1;\nconstexpr int Kw = 1;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 0;\nconstexpr int pad_left = 0;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024.json deleted file mode 100644 index ff2f3fff396cb030efc72520dc09cf8b4ea79914..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024", - "definition": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 256;\nconstexpr int Cout = 1024;\nconstexpr int Kh = 1;\nconstexpr int Kw = 1;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 0;\nconstexpr int pad_left = 0;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048.json deleted file mode 100644 index 4a344fb81bc8aee01f5031aaddb5d9c0182a66e6..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048", - "definition": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 512;\nconstexpr int Cout = 2048;\nconstexpr int Kh = 1;\nconstexpr int Kw = 1;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 0;\nconstexpr int pad_left = 0;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256.json deleted file mode 100644 index 20991a39fa84c944bd513b7727c7d8ad19714ecd..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256", - "definition": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 64;\nconstexpr int Cout = 256;\nconstexpr int Kh = 1;\nconstexpr int Kw = 1;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 0;\nconstexpr int pad_left = 0;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64.json deleted file mode 100644 index 3af72d28497fdc83941a4d8cc8fddebdb11c031d..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64", - "definition": "conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 64;\nconstexpr int Cout = 64;\nconstexpr int Kh = 1;\nconstexpr int Kw = 1;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 0;\nconstexpr int pad_left = 0;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256.json deleted file mode 100644 index f6dbfe8b3b4fd92f6def09f5eba714d72df92d14..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 128;\nconstexpr int Cout = 256;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 1;\nconstexpr int pad_left = 1;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512.json deleted file mode 100644 index 9b9db8eadd128c68d140772d90daec9622e1b71c..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 256;\nconstexpr int Cout = 512;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 1;\nconstexpr int pad_left = 1;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64.json deleted file mode 100644 index e92e265cb6b170ddf4f2cb88ae94a86475d56b15..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 3;\nconstexpr int Cout = 64;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 1;\nconstexpr int pad_left = 1;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024.json deleted file mode 100644 index 5dca5908c93bcfeca4c6531468e7e7d5dde3a81f..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 512;\nconstexpr int Cout = 1024;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 1;\nconstexpr int pad_left = 1;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128.json deleted file mode 100644 index bd92b416407815881311b400ba2db03e6c3ac8cc..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 64;\nconstexpr int Cout = 128;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 1;\nconstexpr int pad_left = 1;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256.json deleted file mode 100644 index c603c681019955f07881304d0f6be453cbe87d7a..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 128;\nconstexpr int Cout = 256;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 2;\nconstexpr int Dw = 2;\nconstexpr int pad_top = 2;\nconstexpr int pad_left = 2;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256.json deleted file mode 100644 index 7f2e64ac22590af466aca19c4b8098c2217943fc..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 256;\nconstexpr int Cout = 256;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 2;\nconstexpr int Dw = 2;\nconstexpr int pad_top = 2;\nconstexpr int pad_left = 2;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512.json deleted file mode 100644 index 56bab478dbe3e331f60721c688daa198b9a60f90..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 512;\nconstexpr int Cout = 512;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 2;\nconstexpr int Dw = 2;\nconstexpr int pad_top = 2;\nconstexpr int pad_left = 2;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128.json deleted file mode 100644 index 849d1830908eb0b7783163da02ed87c48aaaa7ed..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128", - "definition": "conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 64;\nconstexpr int Cout = 128;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 2;\nconstexpr int Dw = 2;\nconstexpr int pad_top = 2;\nconstexpr int pad_left = 2;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256.json deleted file mode 100644 index ad5c239573b94123ae16119fbe880df1d180b6af..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256", - "definition": "conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 128;\nconstexpr int Cout = 256;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 1;\nconstexpr int pad_left = 1;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512.json deleted file mode 100644 index 95bfff900cd902ab21b8323cb178cfba2e5189a3..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512", - "definition": "conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 256;\nconstexpr int Cout = 512;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 1;\nconstexpr int pad_left = 1;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024.json deleted file mode 100644 index d258833776988aefafa9b7e358a22890b6f063e7..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024", - "definition": "conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 512;\nconstexpr int Cout = 1024;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 1;\nconstexpr int pad_left = 1;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128.json deleted file mode 100644 index 1462addb6376cb321270f5cc68ee23ef5dca033b..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128", - "definition": "conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 64;\nconstexpr int Cout = 128;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 1;\nconstexpr int pad_left = 1;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128.json deleted file mode 100644 index a23f9069754dae60a843866d8ad273ed17ee0c8e..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128", - "definition": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 128;\nconstexpr int Cout = 128;\nconstexpr int Kh = 5;\nconstexpr int Kw = 5;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 2;\nconstexpr int pad_left = 2;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256.json deleted file mode 100644 index ff2f8644cead8357dc6015dd570f73ea4dc5d030..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256", - "definition": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 256;\nconstexpr int Cout = 256;\nconstexpr int Kh = 5;\nconstexpr int Kw = 5;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 2;\nconstexpr int pad_left = 2;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64.json deleted file mode 100644 index 659bea27b23420347dc7d047f9faf811308b22cd..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64", - "definition": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 32;\nconstexpr int Cout = 64;\nconstexpr int Kh = 5;\nconstexpr int Kw = 5;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 2;\nconstexpr int pad_left = 2;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32.json deleted file mode 100644 index 7fb1c4582364e99116bf48b3fa77d171f1e751bd..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32", - "definition": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 3;\nconstexpr int Cout = 32;\nconstexpr int Kh = 5;\nconstexpr int Kw = 5;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 2;\nconstexpr int pad_left = 2;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512.json deleted file mode 100644 index 26af50c1f6cda041925611efb8a6676cfe16532c..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512", - "definition": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 512;\nconstexpr int Cout = 512;\nconstexpr int Kh = 5;\nconstexpr int Kw = 5;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 2;\nconstexpr int pad_left = 2;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128.json deleted file mode 100644 index bb274e321bbe4911a4fdeb062c392816b2ea6a71..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128", - "definition": "conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 64;\nconstexpr int Cout = 128;\nconstexpr int Kh = 5;\nconstexpr int Kw = 5;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 2;\nconstexpr int pad_left = 2;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64.json b/solutions/ncnn/reference-scalar/conv2d/conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64.json deleted file mode 100644 index 1a7337c572e6238257f9d57a9f89424520bf12ca..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d/conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64", - "definition": "conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d for conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64. Port of ref_conv.h::ref_conv2d to the CandidateBuilder C-ABI; no ncnn dependency. Ground-truth correctness, slow perf. The load-bearing proof for the raw candidate path.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d.cpp::armbench_entry_conv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d.h", - "content": "// Auto-generated by scripts/gen_candidate_bindings.py \u2014 do not hand-edit.\n#pragma once\n\n// Per-definition constants for this conv2d specialisation.\nnamespace conv2d_def {\nconstexpr int Cin = 3;\nconstexpr int Cout = 64;\nconstexpr int Kh = 7;\nconstexpr int Kw = 7;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad_top = 3;\nconstexpr int pad_left = 3;\nconstexpr int activation_type = 0;\n} // namespace conv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\nvoid inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d.cpp", - "content": "// Binding harness: computes output dims from constexpr params, calls inner_conv2d.\n// The ctypes mirror of armbench_entry_conv2d is bench/datasets/raw.py\n// SIGNATURES[\"conv2d\"] \u2014 edit one, edit the other.\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" int armbench_entry_conv2d(\n const float* input, float* output,\n const float* weight,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad_top - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad_left - ext_kw) / Sw + 1;\n inner_conv2d(input, output, weight, N, H, W, H_out, W_out);\n return 0;\n}" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar conv2d \u2014 port of ref_conv.h::ref_conv2d.\n// LLM target: replace this file with an optimised inner_conv2d.\n// All per-definition constants live in conv2d_def:: (conv2d.h).\n#include \"conv2d.h\"\nusing namespace conv2d_def;\n\nextern \"C\" void inner_conv2d(\n const float* input, float* output, const float* weight,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int oc = 0; oc < Cout; ++oc) {\n float* outc = out_n + (long)oc * H_out * W_out;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = 0.0f;\n for (int ic = 0; ic < Cin; ++ic) {\n const float* inc = in_n + (long)ic * H * W;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad_top + kh * Dh;\n int iw = ow * Sw - pad_left + kw * Dw;\n float px = (ih >= 0 && ih < H && iw >= 0 && iw < W)\n ? inc[ih * W + iw] : 0.0f;\n int widx = ((oc * Cin + ic) * Kh + kh) * Kw + kw;\n sum += px * weight[widx];\n }\n }\n }\n if (activation_type == 1 && sum < 0.0f) sum = 0.0f;\n outc[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024.json b/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024.json deleted file mode 100644 index 0f22f5d374a050983e69a13ebd016927229c7b5c..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024", - "definition": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d_depthwise for conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024. Constexpr-baked dims; armbench_entry_conv2d_depthwise calls inner_conv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d_depthwise.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this conv2d_depthwise specialisation.\nnamespace conv2d_depthwise_def {\nconstexpr int C = 1024;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 1;\n} // namespace conv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\nvoid inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_conv2d_depthwise.\n// ABI: armbench_entry_conv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise conv2d.\n// LLM target: replace this file with an optimised inner_conv2d_depthwise.\n// All per-definition constants live in conv2d_depthwise_def:: (conv2d_depthwise.h).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" void inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[c];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad + kh * Dh;\n int iw = ow * Sw - pad + kw * Dw;\n if (ih >= 0 && ih < H && iw >= 0 && iw < W)\n sum += in_c[ih * W + iw] * w_c[kh * Kw + kw];\n }\n }\n out_c[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128.json b/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128.json deleted file mode 100644 index 3b28de5b02c3b4f6a08c8a244797239bd15f3776..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128", - "definition": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d_depthwise for conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128. Constexpr-baked dims; armbench_entry_conv2d_depthwise calls inner_conv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d_depthwise.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this conv2d_depthwise specialisation.\nnamespace conv2d_depthwise_def {\nconstexpr int C = 128;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 1;\n} // namespace conv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\nvoid inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_conv2d_depthwise.\n// ABI: armbench_entry_conv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise conv2d.\n// LLM target: replace this file with an optimised inner_conv2d_depthwise.\n// All per-definition constants live in conv2d_depthwise_def:: (conv2d_depthwise.h).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" void inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[c];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad + kh * Dh;\n int iw = ow * Sw - pad + kw * Dw;\n if (ih >= 0 && ih < H && iw >= 0 && iw < W)\n sum += in_c[ih * W + iw] * w_c[kh * Kw + kw];\n }\n }\n out_c[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048.json b/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048.json deleted file mode 100644 index e3de5b0ce0dada0cea8c295b800cec695e5d4e9a..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048", - "definition": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d_depthwise for conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048. Constexpr-baked dims; armbench_entry_conv2d_depthwise calls inner_conv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d_depthwise.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this conv2d_depthwise specialisation.\nnamespace conv2d_depthwise_def {\nconstexpr int C = 2048;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 1;\n} // namespace conv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\nvoid inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_conv2d_depthwise.\n// ABI: armbench_entry_conv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise conv2d.\n// LLM target: replace this file with an optimised inner_conv2d_depthwise.\n// All per-definition constants live in conv2d_depthwise_def:: (conv2d_depthwise.h).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" void inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[c];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad + kh * Dh;\n int iw = ow * Sw - pad + kw * Dw;\n if (ih >= 0 && ih < H && iw >= 0 && iw < W)\n sum += in_c[ih * W + iw] * w_c[kh * Kw + kw];\n }\n }\n out_c[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256.json b/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256.json deleted file mode 100644 index 8d647e1d762901898fa534e334b87758a8b865d3..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256", - "definition": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d_depthwise for conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256. Constexpr-baked dims; armbench_entry_conv2d_depthwise calls inner_conv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d_depthwise.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this conv2d_depthwise specialisation.\nnamespace conv2d_depthwise_def {\nconstexpr int C = 256;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 1;\n} // namespace conv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\nvoid inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_conv2d_depthwise.\n// ABI: armbench_entry_conv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise conv2d.\n// LLM target: replace this file with an optimised inner_conv2d_depthwise.\n// All per-definition constants live in conv2d_depthwise_def:: (conv2d_depthwise.h).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" void inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[c];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad + kh * Dh;\n int iw = ow * Sw - pad + kw * Dw;\n if (ih >= 0 && ih < H && iw >= 0 && iw < W)\n sum += in_c[ih * W + iw] * w_c[kh * Kw + kw];\n }\n }\n out_c[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32.json b/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32.json deleted file mode 100644 index d076df669a88a3d185934ad69a2c3aa2700e6242..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32", - "definition": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d_depthwise for conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32. Constexpr-baked dims; armbench_entry_conv2d_depthwise calls inner_conv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d_depthwise.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this conv2d_depthwise specialisation.\nnamespace conv2d_depthwise_def {\nconstexpr int C = 32;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 1;\n} // namespace conv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\nvoid inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_conv2d_depthwise.\n// ABI: armbench_entry_conv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise conv2d.\n// LLM target: replace this file with an optimised inner_conv2d_depthwise.\n// All per-definition constants live in conv2d_depthwise_def:: (conv2d_depthwise.h).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" void inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[c];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad + kh * Dh;\n int iw = ow * Sw - pad + kw * Dw;\n if (ih >= 0 && ih < H && iw >= 0 && iw < W)\n sum += in_c[ih * W + iw] * w_c[kh * Kw + kw];\n }\n }\n out_c[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512.json b/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512.json deleted file mode 100644 index b0f8ba25bdc32bf388529ebe9863dce93d53414d..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512", - "definition": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d_depthwise for conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512. Constexpr-baked dims; armbench_entry_conv2d_depthwise calls inner_conv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d_depthwise.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this conv2d_depthwise specialisation.\nnamespace conv2d_depthwise_def {\nconstexpr int C = 512;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 1;\n} // namespace conv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\nvoid inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_conv2d_depthwise.\n// ABI: armbench_entry_conv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise conv2d.\n// LLM target: replace this file with an optimised inner_conv2d_depthwise.\n// All per-definition constants live in conv2d_depthwise_def:: (conv2d_depthwise.h).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" void inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[c];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad + kh * Dh;\n int iw = ow * Sw - pad + kw * Dw;\n if (ih >= 0 && ih < H && iw >= 0 && iw < W)\n sum += in_c[ih * W + iw] * w_c[kh * Kw + kw];\n }\n }\n out_c[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64.json b/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64.json deleted file mode 100644 index 48f591e9efd8662e22ad34197c612c66972e6f84..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64", - "definition": "conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d_depthwise for conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64. Constexpr-baked dims; armbench_entry_conv2d_depthwise calls inner_conv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d_depthwise.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this conv2d_depthwise specialisation.\nnamespace conv2d_depthwise_def {\nconstexpr int C = 64;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 1;\n} // namespace conv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\nvoid inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_conv2d_depthwise.\n// ABI: armbench_entry_conv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise conv2d.\n// LLM target: replace this file with an optimised inner_conv2d_depthwise.\n// All per-definition constants live in conv2d_depthwise_def:: (conv2d_depthwise.h).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" void inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[c];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad + kh * Dh;\n int iw = ow * Sw - pad + kw * Dw;\n if (ih >= 0 && ih < H && iw >= 0 && iw < W)\n sum += in_c[ih * W + iw] * w_c[kh * Kw + kw];\n }\n }\n out_c[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024.json b/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024.json deleted file mode 100644 index 703d894729110172ed735d3d282170cfcdeae91a..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024", - "definition": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d_depthwise for conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024. Constexpr-baked dims; armbench_entry_conv2d_depthwise calls inner_conv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d_depthwise.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this conv2d_depthwise specialisation.\nnamespace conv2d_depthwise_def {\nconstexpr int C = 1024;\nconstexpr int Kh = 5;\nconstexpr int Kw = 5;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 2;\n} // namespace conv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\nvoid inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_conv2d_depthwise.\n// ABI: armbench_entry_conv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise conv2d.\n// LLM target: replace this file with an optimised inner_conv2d_depthwise.\n// All per-definition constants live in conv2d_depthwise_def:: (conv2d_depthwise.h).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" void inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[c];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad + kh * Dh;\n int iw = ow * Sw - pad + kw * Dw;\n if (ih >= 0 && ih < H && iw >= 0 && iw < W)\n sum += in_c[ih * W + iw] * w_c[kh * Kw + kw];\n }\n }\n out_c[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128.json b/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128.json deleted file mode 100644 index 598bbe83f063077c4f12c6086dbdafea9fe00c0f..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128", - "definition": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d_depthwise for conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128. Constexpr-baked dims; armbench_entry_conv2d_depthwise calls inner_conv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d_depthwise.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this conv2d_depthwise specialisation.\nnamespace conv2d_depthwise_def {\nconstexpr int C = 128;\nconstexpr int Kh = 5;\nconstexpr int Kw = 5;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 2;\n} // namespace conv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\nvoid inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_conv2d_depthwise.\n// ABI: armbench_entry_conv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise conv2d.\n// LLM target: replace this file with an optimised inner_conv2d_depthwise.\n// All per-definition constants live in conv2d_depthwise_def:: (conv2d_depthwise.h).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" void inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[c];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad + kh * Dh;\n int iw = ow * Sw - pad + kw * Dw;\n if (ih >= 0 && ih < H && iw >= 0 && iw < W)\n sum += in_c[ih * W + iw] * w_c[kh * Kw + kw];\n }\n }\n out_c[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256.json b/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256.json deleted file mode 100644 index 5ceb041d21b881ce7f0435ad145765def4f135f9..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256", - "definition": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d_depthwise for conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256. Constexpr-baked dims; armbench_entry_conv2d_depthwise calls inner_conv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d_depthwise.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this conv2d_depthwise specialisation.\nnamespace conv2d_depthwise_def {\nconstexpr int C = 256;\nconstexpr int Kh = 5;\nconstexpr int Kw = 5;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 2;\n} // namespace conv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\nvoid inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_conv2d_depthwise.\n// ABI: armbench_entry_conv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise conv2d.\n// LLM target: replace this file with an optimised inner_conv2d_depthwise.\n// All per-definition constants live in conv2d_depthwise_def:: (conv2d_depthwise.h).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" void inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[c];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad + kh * Dh;\n int iw = ow * Sw - pad + kw * Dw;\n if (ih >= 0 && ih < H && iw >= 0 && iw < W)\n sum += in_c[ih * W + iw] * w_c[kh * Kw + kw];\n }\n }\n out_c[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32.json b/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32.json deleted file mode 100644 index d9cc0da88535d2466fee3ee2599b6e0ab1253655..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32", - "definition": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d_depthwise for conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32. Constexpr-baked dims; armbench_entry_conv2d_depthwise calls inner_conv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d_depthwise.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this conv2d_depthwise specialisation.\nnamespace conv2d_depthwise_def {\nconstexpr int C = 32;\nconstexpr int Kh = 5;\nconstexpr int Kw = 5;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 2;\n} // namespace conv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\nvoid inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_conv2d_depthwise.\n// ABI: armbench_entry_conv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise conv2d.\n// LLM target: replace this file with an optimised inner_conv2d_depthwise.\n// All per-definition constants live in conv2d_depthwise_def:: (conv2d_depthwise.h).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" void inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[c];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad + kh * Dh;\n int iw = ow * Sw - pad + kw * Dw;\n if (ih >= 0 && ih < H && iw >= 0 && iw < W)\n sum += in_c[ih * W + iw] * w_c[kh * Kw + kw];\n }\n }\n out_c[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512.json b/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512.json deleted file mode 100644 index 682b0c75bcff82eed0d25735b308f020dcd0858c..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512", - "definition": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d_depthwise for conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512. Constexpr-baked dims; armbench_entry_conv2d_depthwise calls inner_conv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d_depthwise.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this conv2d_depthwise specialisation.\nnamespace conv2d_depthwise_def {\nconstexpr int C = 512;\nconstexpr int Kh = 5;\nconstexpr int Kw = 5;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 2;\n} // namespace conv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\nvoid inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_conv2d_depthwise.\n// ABI: armbench_entry_conv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise conv2d.\n// LLM target: replace this file with an optimised inner_conv2d_depthwise.\n// All per-definition constants live in conv2d_depthwise_def:: (conv2d_depthwise.h).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" void inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[c];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad + kh * Dh;\n int iw = ow * Sw - pad + kw * Dw;\n if (ih >= 0 && ih < H && iw >= 0 && iw < W)\n sum += in_c[ih * W + iw] * w_c[kh * Kw + kw];\n }\n }\n out_c[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64.json b/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64.json deleted file mode 100644 index 4581f58024fc3c5b203cd08f0aaef930cb4abfbf..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64", - "definition": "conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* conv2d_depthwise for conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64. Constexpr-baked dims; armbench_entry_conv2d_depthwise calls inner_conv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "conv2d_depthwise.cpp::armbench_entry_conv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "conv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this conv2d_depthwise specialisation.\nnamespace conv2d_depthwise_def {\nconstexpr int C = 64;\nconstexpr int Kh = 5;\nconstexpr int Kw = 5;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\nconstexpr int pad = 2;\n} // namespace conv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\nvoid inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "conv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_conv2d_depthwise.\n// ABI: armbench_entry_conv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n constexpr int ext_kh = Dh * (Kh - 1) + 1;\n constexpr int ext_kw = Dw * (Kw - 1) + 1;\n const int H_out = (H + 2 * pad - ext_kh) / Sh + 1;\n const int W_out = (W + 2 * pad - ext_kw) / Sw + 1;\n inner_conv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise conv2d.\n// LLM target: replace this file with an optimised inner_conv2d_depthwise.\n// All per-definition constants live in conv2d_depthwise_def:: (conv2d_depthwise.h).\n#include \"conv2d_depthwise.h\"\nusing namespace conv2d_depthwise_def;\n\nextern \"C\" void inner_conv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int oh = 0; oh < H_out; ++oh) {\n for (int ow = 0; ow < W_out; ++ow) {\n float sum = bias[c];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int ih = oh * Sh - pad + kh * Dh;\n int iw = ow * Sw - pad + kw * Dw;\n if (ih >= 0 && ih < H && iw >= 0 && iw < W)\n sum += in_c[ih * W + iw] * w_c[kh * Kw + kw];\n }\n }\n out_c[oh * W_out + ow] = sum;\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin128_cout256.json b/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin128_cout256.json deleted file mode 100644 index 43f1622c345b24e78162a5b5b4bf445cf75e0ad7..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin128_cout256.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_kh3_kw3_sh1_sw1_cin128_cout256", - "definition": "deconv2d_kh3_kw3_sh1_sw1_cin128_cout256", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d for deconv2d_kh3_kw3_sh1_sw1_cin128_cout256. Constexpr-baked dims; armbench_entry_deconv2d calls inner_deconv2d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d.cpp::armbench_entry_deconv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d specialisation.\nnamespace deconv2d_def {\nconstexpr int Cin = 128;\nconstexpr int Cout = 256;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, Cin, H, W), output layout: (N, Cout, H_out, W_out).\n// Weight layout: (Cout, Cin, Kh, Kw).\nvoid inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d.\n// ABI: armbench_entry_deconv2d(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, Cin, H, W); output layout: (N, Cout, H_out, W_out).\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" int armbench_entry_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar transposed conv2d (deconv2d).\n// LLM target: replace this file with an optimised inner_deconv2d.\n// All per-definition constants live in deconv2d_def:: (deconv2d.h).\n// Weight layout: (Cout, Cin, Kh, Kw); no padding applied.\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" void inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_co[i] = bias[co];\n }\n }\n\n // Scatter-accumulate from each input element\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int ci = 0; ci < Cin; ++ci) {\n const float* in_ci = in_n + (long)ci * H * W;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_ci[ih * W + iw];\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n const float* w_co_ci = weight +\n ((long)co * Cin + ci) * Kh * Kw;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_co[oh * W_out + ow] +=\n in_val * w_co_ci[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin256_cout512.json b/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin256_cout512.json deleted file mode 100644 index 4e0561e245d5446b3199ac9d4331a9e34dcaa23b..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin256_cout512.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_kh3_kw3_sh1_sw1_cin256_cout512", - "definition": "deconv2d_kh3_kw3_sh1_sw1_cin256_cout512", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d for deconv2d_kh3_kw3_sh1_sw1_cin256_cout512. Constexpr-baked dims; armbench_entry_deconv2d calls inner_deconv2d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d.cpp::armbench_entry_deconv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d specialisation.\nnamespace deconv2d_def {\nconstexpr int Cin = 256;\nconstexpr int Cout = 512;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, Cin, H, W), output layout: (N, Cout, H_out, W_out).\n// Weight layout: (Cout, Cin, Kh, Kw).\nvoid inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d.\n// ABI: armbench_entry_deconv2d(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, Cin, H, W); output layout: (N, Cout, H_out, W_out).\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" int armbench_entry_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar transposed conv2d (deconv2d).\n// LLM target: replace this file with an optimised inner_deconv2d.\n// All per-definition constants live in deconv2d_def:: (deconv2d.h).\n// Weight layout: (Cout, Cin, Kh, Kw); no padding applied.\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" void inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_co[i] = bias[co];\n }\n }\n\n // Scatter-accumulate from each input element\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int ci = 0; ci < Cin; ++ci) {\n const float* in_ci = in_n + (long)ci * H * W;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_ci[ih * W + iw];\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n const float* w_co_ci = weight +\n ((long)co * Cin + ci) * Kh * Kw;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_co[oh * W_out + ow] +=\n in_val * w_co_ci[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024.json b/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024.json deleted file mode 100644 index 79d06f618a23aab50a64212f3a5ff8f1e3ff98ae..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024", - "definition": "deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d for deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024. Constexpr-baked dims; armbench_entry_deconv2d calls inner_deconv2d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d.cpp::armbench_entry_deconv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d specialisation.\nnamespace deconv2d_def {\nconstexpr int Cin = 512;\nconstexpr int Cout = 1024;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, Cin, H, W), output layout: (N, Cout, H_out, W_out).\n// Weight layout: (Cout, Cin, Kh, Kw).\nvoid inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d.\n// ABI: armbench_entry_deconv2d(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, Cin, H, W); output layout: (N, Cout, H_out, W_out).\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" int armbench_entry_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar transposed conv2d (deconv2d).\n// LLM target: replace this file with an optimised inner_deconv2d.\n// All per-definition constants live in deconv2d_def:: (deconv2d.h).\n// Weight layout: (Cout, Cin, Kh, Kw); no padding applied.\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" void inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_co[i] = bias[co];\n }\n }\n\n // Scatter-accumulate from each input element\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int ci = 0; ci < Cin; ++ci) {\n const float* in_ci = in_n + (long)ci * H * W;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_ci[ih * W + iw];\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n const float* w_co_ci = weight +\n ((long)co * Cin + ci) * Kh * Kw;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_co[oh * W_out + ow] +=\n in_val * w_co_ci[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin64_cout128.json b/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin64_cout128.json deleted file mode 100644 index 21042dbad6f4560ba97d3f5213b5bbbe5aa36985..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin64_cout128.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_kh3_kw3_sh1_sw1_cin64_cout128", - "definition": "deconv2d_kh3_kw3_sh1_sw1_cin64_cout128", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d for deconv2d_kh3_kw3_sh1_sw1_cin64_cout128. Constexpr-baked dims; armbench_entry_deconv2d calls inner_deconv2d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d.cpp::armbench_entry_deconv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d specialisation.\nnamespace deconv2d_def {\nconstexpr int Cin = 64;\nconstexpr int Cout = 128;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, Cin, H, W), output layout: (N, Cout, H_out, W_out).\n// Weight layout: (Cout, Cin, Kh, Kw).\nvoid inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d.\n// ABI: armbench_entry_deconv2d(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, Cin, H, W); output layout: (N, Cout, H_out, W_out).\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" int armbench_entry_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar transposed conv2d (deconv2d).\n// LLM target: replace this file with an optimised inner_deconv2d.\n// All per-definition constants live in deconv2d_def:: (deconv2d.h).\n// Weight layout: (Cout, Cin, Kh, Kw); no padding applied.\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" void inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_co[i] = bias[co];\n }\n }\n\n // Scatter-accumulate from each input element\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int ci = 0; ci < Cin; ++ci) {\n const float* in_ci = in_n + (long)ci * H * W;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_ci[ih * W + iw];\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n const float* w_co_ci = weight +\n ((long)co * Cin + ci) * Kh * Kw;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_co[oh * W_out + ow] +=\n in_val * w_co_ci[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin128_cout256.json b/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin128_cout256.json deleted file mode 100644 index 8e42d413dfff4dbd31219758589da0225475659c..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin128_cout256.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_kh3_kw3_sh2_sw2_cin128_cout256", - "definition": "deconv2d_kh3_kw3_sh2_sw2_cin128_cout256", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d for deconv2d_kh3_kw3_sh2_sw2_cin128_cout256. Constexpr-baked dims; armbench_entry_deconv2d calls inner_deconv2d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d.cpp::armbench_entry_deconv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d specialisation.\nnamespace deconv2d_def {\nconstexpr int Cin = 128;\nconstexpr int Cout = 256;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, Cin, H, W), output layout: (N, Cout, H_out, W_out).\n// Weight layout: (Cout, Cin, Kh, Kw).\nvoid inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d.\n// ABI: armbench_entry_deconv2d(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, Cin, H, W); output layout: (N, Cout, H_out, W_out).\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" int armbench_entry_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar transposed conv2d (deconv2d).\n// LLM target: replace this file with an optimised inner_deconv2d.\n// All per-definition constants live in deconv2d_def:: (deconv2d.h).\n// Weight layout: (Cout, Cin, Kh, Kw); no padding applied.\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" void inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_co[i] = bias[co];\n }\n }\n\n // Scatter-accumulate from each input element\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int ci = 0; ci < Cin; ++ci) {\n const float* in_ci = in_n + (long)ci * H * W;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_ci[ih * W + iw];\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n const float* w_co_ci = weight +\n ((long)co * Cin + ci) * Kh * Kw;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_co[oh * W_out + ow] +=\n in_val * w_co_ci[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin256_cout512.json b/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin256_cout512.json deleted file mode 100644 index 6f643a308536bcae8e367627c3ca7f573a9ae103..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin256_cout512.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_kh3_kw3_sh2_sw2_cin256_cout512", - "definition": "deconv2d_kh3_kw3_sh2_sw2_cin256_cout512", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d for deconv2d_kh3_kw3_sh2_sw2_cin256_cout512. Constexpr-baked dims; armbench_entry_deconv2d calls inner_deconv2d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d.cpp::armbench_entry_deconv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d specialisation.\nnamespace deconv2d_def {\nconstexpr int Cin = 256;\nconstexpr int Cout = 512;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, Cin, H, W), output layout: (N, Cout, H_out, W_out).\n// Weight layout: (Cout, Cin, Kh, Kw).\nvoid inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d.\n// ABI: armbench_entry_deconv2d(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, Cin, H, W); output layout: (N, Cout, H_out, W_out).\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" int armbench_entry_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar transposed conv2d (deconv2d).\n// LLM target: replace this file with an optimised inner_deconv2d.\n// All per-definition constants live in deconv2d_def:: (deconv2d.h).\n// Weight layout: (Cout, Cin, Kh, Kw); no padding applied.\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" void inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_co[i] = bias[co];\n }\n }\n\n // Scatter-accumulate from each input element\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int ci = 0; ci < Cin; ++ci) {\n const float* in_ci = in_n + (long)ci * H * W;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_ci[ih * W + iw];\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n const float* w_co_ci = weight +\n ((long)co * Cin + ci) * Kh * Kw;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_co[oh * W_out + ow] +=\n in_val * w_co_ci[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024.json b/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024.json deleted file mode 100644 index cae063eadc0e0d8c12b3c1576bc5ebe3c7ca66cb..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024", - "definition": "deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d for deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024. Constexpr-baked dims; armbench_entry_deconv2d calls inner_deconv2d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d.cpp::armbench_entry_deconv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d specialisation.\nnamespace deconv2d_def {\nconstexpr int Cin = 512;\nconstexpr int Cout = 1024;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, Cin, H, W), output layout: (N, Cout, H_out, W_out).\n// Weight layout: (Cout, Cin, Kh, Kw).\nvoid inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d.\n// ABI: armbench_entry_deconv2d(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, Cin, H, W); output layout: (N, Cout, H_out, W_out).\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" int armbench_entry_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar transposed conv2d (deconv2d).\n// LLM target: replace this file with an optimised inner_deconv2d.\n// All per-definition constants live in deconv2d_def:: (deconv2d.h).\n// Weight layout: (Cout, Cin, Kh, Kw); no padding applied.\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" void inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_co[i] = bias[co];\n }\n }\n\n // Scatter-accumulate from each input element\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int ci = 0; ci < Cin; ++ci) {\n const float* in_ci = in_n + (long)ci * H * W;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_ci[ih * W + iw];\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n const float* w_co_ci = weight +\n ((long)co * Cin + ci) * Kh * Kw;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_co[oh * W_out + ow] +=\n in_val * w_co_ci[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin64_cout128.json b/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin64_cout128.json deleted file mode 100644 index 7a002fadb3e0e70156766ce489beb1e1b6d8dc4f..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin64_cout128.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_kh3_kw3_sh2_sw2_cin64_cout128", - "definition": "deconv2d_kh3_kw3_sh2_sw2_cin64_cout128", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d for deconv2d_kh3_kw3_sh2_sw2_cin64_cout128. Constexpr-baked dims; armbench_entry_deconv2d calls inner_deconv2d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d.cpp::armbench_entry_deconv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d specialisation.\nnamespace deconv2d_def {\nconstexpr int Cin = 64;\nconstexpr int Cout = 128;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, Cin, H, W), output layout: (N, Cout, H_out, W_out).\n// Weight layout: (Cout, Cin, Kh, Kw).\nvoid inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d.\n// ABI: armbench_entry_deconv2d(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, Cin, H, W); output layout: (N, Cout, H_out, W_out).\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" int armbench_entry_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar transposed conv2d (deconv2d).\n// LLM target: replace this file with an optimised inner_deconv2d.\n// All per-definition constants live in deconv2d_def:: (deconv2d.h).\n// Weight layout: (Cout, Cin, Kh, Kw); no padding applied.\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" void inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_co[i] = bias[co];\n }\n }\n\n // Scatter-accumulate from each input element\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int ci = 0; ci < Cin; ++ci) {\n const float* in_ci = in_n + (long)ci * H * W;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_ci[ih * W + iw];\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n const float* w_co_ci = weight +\n ((long)co * Cin + ci) * Kh * Kw;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_co[oh * W_out + ow] +=\n in_val * w_co_ci[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin128_cout256.json b/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin128_cout256.json deleted file mode 100644 index 0eaec68c01c8ac554841fc852179daa2b4a78f5a..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin128_cout256.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_kh4_kw4_sh1_sw1_cin128_cout256", - "definition": "deconv2d_kh4_kw4_sh1_sw1_cin128_cout256", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d for deconv2d_kh4_kw4_sh1_sw1_cin128_cout256. Constexpr-baked dims; armbench_entry_deconv2d calls inner_deconv2d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d.cpp::armbench_entry_deconv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d specialisation.\nnamespace deconv2d_def {\nconstexpr int Cin = 128;\nconstexpr int Cout = 256;\nconstexpr int Kh = 4;\nconstexpr int Kw = 4;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, Cin, H, W), output layout: (N, Cout, H_out, W_out).\n// Weight layout: (Cout, Cin, Kh, Kw).\nvoid inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d.\n// ABI: armbench_entry_deconv2d(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, Cin, H, W); output layout: (N, Cout, H_out, W_out).\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" int armbench_entry_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar transposed conv2d (deconv2d).\n// LLM target: replace this file with an optimised inner_deconv2d.\n// All per-definition constants live in deconv2d_def:: (deconv2d.h).\n// Weight layout: (Cout, Cin, Kh, Kw); no padding applied.\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" void inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_co[i] = bias[co];\n }\n }\n\n // Scatter-accumulate from each input element\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int ci = 0; ci < Cin; ++ci) {\n const float* in_ci = in_n + (long)ci * H * W;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_ci[ih * W + iw];\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n const float* w_co_ci = weight +\n ((long)co * Cin + ci) * Kh * Kw;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_co[oh * W_out + ow] +=\n in_val * w_co_ci[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin256_cout512.json b/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin256_cout512.json deleted file mode 100644 index ca8469dab3eced45e97371d78d271b06e77692d2..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin256_cout512.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_kh4_kw4_sh1_sw1_cin256_cout512", - "definition": "deconv2d_kh4_kw4_sh1_sw1_cin256_cout512", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d for deconv2d_kh4_kw4_sh1_sw1_cin256_cout512. Constexpr-baked dims; armbench_entry_deconv2d calls inner_deconv2d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d.cpp::armbench_entry_deconv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d specialisation.\nnamespace deconv2d_def {\nconstexpr int Cin = 256;\nconstexpr int Cout = 512;\nconstexpr int Kh = 4;\nconstexpr int Kw = 4;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, Cin, H, W), output layout: (N, Cout, H_out, W_out).\n// Weight layout: (Cout, Cin, Kh, Kw).\nvoid inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d.\n// ABI: armbench_entry_deconv2d(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, Cin, H, W); output layout: (N, Cout, H_out, W_out).\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" int armbench_entry_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar transposed conv2d (deconv2d).\n// LLM target: replace this file with an optimised inner_deconv2d.\n// All per-definition constants live in deconv2d_def:: (deconv2d.h).\n// Weight layout: (Cout, Cin, Kh, Kw); no padding applied.\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" void inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_co[i] = bias[co];\n }\n }\n\n // Scatter-accumulate from each input element\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int ci = 0; ci < Cin; ++ci) {\n const float* in_ci = in_n + (long)ci * H * W;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_ci[ih * W + iw];\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n const float* w_co_ci = weight +\n ((long)co * Cin + ci) * Kh * Kw;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_co[oh * W_out + ow] +=\n in_val * w_co_ci[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024.json b/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024.json deleted file mode 100644 index e9f17b7847df72b45b0a3b24c3a93d659cf738ff..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024", - "definition": "deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d for deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024. Constexpr-baked dims; armbench_entry_deconv2d calls inner_deconv2d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d.cpp::armbench_entry_deconv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d specialisation.\nnamespace deconv2d_def {\nconstexpr int Cin = 512;\nconstexpr int Cout = 1024;\nconstexpr int Kh = 4;\nconstexpr int Kw = 4;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, Cin, H, W), output layout: (N, Cout, H_out, W_out).\n// Weight layout: (Cout, Cin, Kh, Kw).\nvoid inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d.\n// ABI: armbench_entry_deconv2d(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, Cin, H, W); output layout: (N, Cout, H_out, W_out).\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" int armbench_entry_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar transposed conv2d (deconv2d).\n// LLM target: replace this file with an optimised inner_deconv2d.\n// All per-definition constants live in deconv2d_def:: (deconv2d.h).\n// Weight layout: (Cout, Cin, Kh, Kw); no padding applied.\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" void inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_co[i] = bias[co];\n }\n }\n\n // Scatter-accumulate from each input element\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int ci = 0; ci < Cin; ++ci) {\n const float* in_ci = in_n + (long)ci * H * W;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_ci[ih * W + iw];\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n const float* w_co_ci = weight +\n ((long)co * Cin + ci) * Kh * Kw;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_co[oh * W_out + ow] +=\n in_val * w_co_ci[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin64_cout128.json b/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin64_cout128.json deleted file mode 100644 index 9e9ccba35045a615f29d76357d50ce1a3abf94b3..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin64_cout128.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_kh4_kw4_sh1_sw1_cin64_cout128", - "definition": "deconv2d_kh4_kw4_sh1_sw1_cin64_cout128", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d for deconv2d_kh4_kw4_sh1_sw1_cin64_cout128. Constexpr-baked dims; armbench_entry_deconv2d calls inner_deconv2d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d.cpp::armbench_entry_deconv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d specialisation.\nnamespace deconv2d_def {\nconstexpr int Cin = 64;\nconstexpr int Cout = 128;\nconstexpr int Kh = 4;\nconstexpr int Kw = 4;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, Cin, H, W), output layout: (N, Cout, H_out, W_out).\n// Weight layout: (Cout, Cin, Kh, Kw).\nvoid inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d.\n// ABI: armbench_entry_deconv2d(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, Cin, H, W); output layout: (N, Cout, H_out, W_out).\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" int armbench_entry_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar transposed conv2d (deconv2d).\n// LLM target: replace this file with an optimised inner_deconv2d.\n// All per-definition constants live in deconv2d_def:: (deconv2d.h).\n// Weight layout: (Cout, Cin, Kh, Kw); no padding applied.\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" void inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_co[i] = bias[co];\n }\n }\n\n // Scatter-accumulate from each input element\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int ci = 0; ci < Cin; ++ci) {\n const float* in_ci = in_n + (long)ci * H * W;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_ci[ih * W + iw];\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n const float* w_co_ci = weight +\n ((long)co * Cin + ci) * Kh * Kw;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_co[oh * W_out + ow] +=\n in_val * w_co_ci[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin128_cout256.json b/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin128_cout256.json deleted file mode 100644 index 606b39d38c8d17b7d417f92a607cf5b43c23fdc8..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin128_cout256.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_kh4_kw4_sh2_sw2_cin128_cout256", - "definition": "deconv2d_kh4_kw4_sh2_sw2_cin128_cout256", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d for deconv2d_kh4_kw4_sh2_sw2_cin128_cout256. Constexpr-baked dims; armbench_entry_deconv2d calls inner_deconv2d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d.cpp::armbench_entry_deconv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d specialisation.\nnamespace deconv2d_def {\nconstexpr int Cin = 128;\nconstexpr int Cout = 256;\nconstexpr int Kh = 4;\nconstexpr int Kw = 4;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, Cin, H, W), output layout: (N, Cout, H_out, W_out).\n// Weight layout: (Cout, Cin, Kh, Kw).\nvoid inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d.\n// ABI: armbench_entry_deconv2d(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, Cin, H, W); output layout: (N, Cout, H_out, W_out).\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" int armbench_entry_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar transposed conv2d (deconv2d).\n// LLM target: replace this file with an optimised inner_deconv2d.\n// All per-definition constants live in deconv2d_def:: (deconv2d.h).\n// Weight layout: (Cout, Cin, Kh, Kw); no padding applied.\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" void inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_co[i] = bias[co];\n }\n }\n\n // Scatter-accumulate from each input element\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int ci = 0; ci < Cin; ++ci) {\n const float* in_ci = in_n + (long)ci * H * W;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_ci[ih * W + iw];\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n const float* w_co_ci = weight +\n ((long)co * Cin + ci) * Kh * Kw;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_co[oh * W_out + ow] +=\n in_val * w_co_ci[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin32_cout64.json b/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin32_cout64.json deleted file mode 100644 index 72f061ab5955fea518add627da2084914441a30e..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin32_cout64.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_kh4_kw4_sh2_sw2_cin32_cout64", - "definition": "deconv2d_kh4_kw4_sh2_sw2_cin32_cout64", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d for deconv2d_kh4_kw4_sh2_sw2_cin32_cout64. Constexpr-baked dims; armbench_entry_deconv2d calls inner_deconv2d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d.cpp::armbench_entry_deconv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d specialisation.\nnamespace deconv2d_def {\nconstexpr int Cin = 32;\nconstexpr int Cout = 64;\nconstexpr int Kh = 4;\nconstexpr int Kw = 4;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, Cin, H, W), output layout: (N, Cout, H_out, W_out).\n// Weight layout: (Cout, Cin, Kh, Kw).\nvoid inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d.\n// ABI: armbench_entry_deconv2d(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, Cin, H, W); output layout: (N, Cout, H_out, W_out).\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" int armbench_entry_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar transposed conv2d (deconv2d).\n// LLM target: replace this file with an optimised inner_deconv2d.\n// All per-definition constants live in deconv2d_def:: (deconv2d.h).\n// Weight layout: (Cout, Cin, Kh, Kw); no padding applied.\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" void inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_co[i] = bias[co];\n }\n }\n\n // Scatter-accumulate from each input element\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int ci = 0; ci < Cin; ++ci) {\n const float* in_ci = in_n + (long)ci * H * W;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_ci[ih * W + iw];\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n const float* w_co_ci = weight +\n ((long)co * Cin + ci) * Kh * Kw;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_co[oh * W_out + ow] +=\n in_val * w_co_ci[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin512_cout512.json b/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin512_cout512.json deleted file mode 100644 index fecd33a72b103faabdc24750c04973c83dbaa228..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin512_cout512.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_kh4_kw4_sh2_sw2_cin512_cout512", - "definition": "deconv2d_kh4_kw4_sh2_sw2_cin512_cout512", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d for deconv2d_kh4_kw4_sh2_sw2_cin512_cout512. Constexpr-baked dims; armbench_entry_deconv2d calls inner_deconv2d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d.cpp::armbench_entry_deconv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d specialisation.\nnamespace deconv2d_def {\nconstexpr int Cin = 512;\nconstexpr int Cout = 512;\nconstexpr int Kh = 4;\nconstexpr int Kw = 4;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, Cin, H, W), output layout: (N, Cout, H_out, W_out).\n// Weight layout: (Cout, Cin, Kh, Kw).\nvoid inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d.\n// ABI: armbench_entry_deconv2d(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, Cin, H, W); output layout: (N, Cout, H_out, W_out).\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" int armbench_entry_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar transposed conv2d (deconv2d).\n// LLM target: replace this file with an optimised inner_deconv2d.\n// All per-definition constants live in deconv2d_def:: (deconv2d.h).\n// Weight layout: (Cout, Cin, Kh, Kw); no padding applied.\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" void inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_co[i] = bias[co];\n }\n }\n\n // Scatter-accumulate from each input element\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int ci = 0; ci < Cin; ++ci) {\n const float* in_ci = in_n + (long)ci * H * W;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_ci[ih * W + iw];\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n const float* w_co_ci = weight +\n ((long)co * Cin + ci) * Kh * Kw;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_co[oh * W_out + ow] +=\n in_val * w_co_ci[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin64_cout128.json b/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin64_cout128.json deleted file mode 100644 index aa20e3597bbd656431109260b36103dabc847792..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin64_cout128.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_kh4_kw4_sh2_sw2_cin64_cout128", - "definition": "deconv2d_kh4_kw4_sh2_sw2_cin64_cout128", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d for deconv2d_kh4_kw4_sh2_sw2_cin64_cout128. Constexpr-baked dims; armbench_entry_deconv2d calls inner_deconv2d. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d.cpp::armbench_entry_deconv2d", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d specialisation.\nnamespace deconv2d_def {\nconstexpr int Cin = 64;\nconstexpr int Cout = 128;\nconstexpr int Kh = 4;\nconstexpr int Kw = 4;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, Cin, H, W), output layout: (N, Cout, H_out, W_out).\n// Weight layout: (Cout, Cin, Kh, Kw).\nvoid inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d.\n// ABI: armbench_entry_deconv2d(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, Cin, H, W); output layout: (N, Cout, H_out, W_out).\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" int armbench_entry_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar transposed conv2d (deconv2d).\n// LLM target: replace this file with an optimised inner_deconv2d.\n// All per-definition constants live in deconv2d_def:: (deconv2d.h).\n// Weight layout: (Cout, Cin, Kh, Kw); no padding applied.\n#include \"deconv2d.h\"\nusing namespace deconv2d_def;\n\nextern \"C\" void inner_deconv2d(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_co[i] = bias[co];\n }\n }\n\n // Scatter-accumulate from each input element\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * Cin * H * W;\n float* out_n = output + (long)n * Cout * H_out * W_out;\n for (int ci = 0; ci < Cin; ++ci) {\n const float* in_ci = in_n + (long)ci * H * W;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_ci[ih * W + iw];\n for (int co = 0; co < Cout; ++co) {\n float* out_co = out_n + (long)co * H_out * W_out;\n const float* w_co_ci = weight +\n ((long)co * Cin + ci) * Kh * Kw;\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_co[oh * W_out + ow] +=\n in_val * w_co_ci[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024.json b/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024.json deleted file mode 100644 index 3db838e79349a42731953c2521185c31988adbd7..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024", - "definition": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d_depthwise for deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024. Constexpr-baked dims; armbench_entry_deconv2d_depthwise calls inner_deconv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d_depthwise.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d_depthwise specialisation.\nnamespace deconv2d_depthwise_def {\nconstexpr int C = 1024;\nconstexpr int Kh = 2;\nconstexpr int Kw = 2;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\n// Weight layout: (C, Kh, Kw) \u2014 one filter per channel.\nvoid inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d_depthwise.\n// ABI: armbench_entry_deconv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise transposed conv2d (deconv2d_depthwise).\n// LLM target: replace this file with an optimised inner_deconv2d_depthwise.\n// All per-definition constants live in deconv2d_depthwise_def:: (deconv2d_depthwise.h).\n// Weight layout: (C, Kh, Kw); no padding applied.\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" void inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n float* out_c = out_n + (long)c * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_c[i] = bias[c];\n }\n }\n\n // Scatter-accumulate from each input element (depthwise: channels independent)\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_c[ih * W + iw];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_c[oh * W_out + ow] += in_val * w_c[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c128.json b/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c128.json deleted file mode 100644 index 506916c5603b1da500653936bcef681f563144af..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c128.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c128", - "definition": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c128", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d_depthwise for deconv2d_depthwise_kh2_kw2_sh2_sw2_c128. Constexpr-baked dims; armbench_entry_deconv2d_depthwise calls inner_deconv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d_depthwise.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d_depthwise specialisation.\nnamespace deconv2d_depthwise_def {\nconstexpr int C = 128;\nconstexpr int Kh = 2;\nconstexpr int Kw = 2;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\n// Weight layout: (C, Kh, Kw) \u2014 one filter per channel.\nvoid inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d_depthwise.\n// ABI: armbench_entry_deconv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise transposed conv2d (deconv2d_depthwise).\n// LLM target: replace this file with an optimised inner_deconv2d_depthwise.\n// All per-definition constants live in deconv2d_depthwise_def:: (deconv2d_depthwise.h).\n// Weight layout: (C, Kh, Kw); no padding applied.\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" void inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n float* out_c = out_n + (long)c * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_c[i] = bias[c];\n }\n }\n\n // Scatter-accumulate from each input element (depthwise: channels independent)\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_c[ih * W + iw];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_c[oh * W_out + ow] += in_val * w_c[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c256.json b/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c256.json deleted file mode 100644 index 0e3ed8f6b2300e9b54f41e7f030c2e6b994baf8b..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c256.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c256", - "definition": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c256", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d_depthwise for deconv2d_depthwise_kh2_kw2_sh2_sw2_c256. Constexpr-baked dims; armbench_entry_deconv2d_depthwise calls inner_deconv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d_depthwise.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d_depthwise specialisation.\nnamespace deconv2d_depthwise_def {\nconstexpr int C = 256;\nconstexpr int Kh = 2;\nconstexpr int Kw = 2;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\n// Weight layout: (C, Kh, Kw) \u2014 one filter per channel.\nvoid inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d_depthwise.\n// ABI: armbench_entry_deconv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise transposed conv2d (deconv2d_depthwise).\n// LLM target: replace this file with an optimised inner_deconv2d_depthwise.\n// All per-definition constants live in deconv2d_depthwise_def:: (deconv2d_depthwise.h).\n// Weight layout: (C, Kh, Kw); no padding applied.\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" void inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n float* out_c = out_n + (long)c * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_c[i] = bias[c];\n }\n }\n\n // Scatter-accumulate from each input element (depthwise: channels independent)\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_c[ih * W + iw];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_c[oh * W_out + ow] += in_val * w_c[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c32.json b/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c32.json deleted file mode 100644 index 2fc33b93038519a1c38b27c990d90471ce0cf13d..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c32.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c32", - "definition": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c32", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d_depthwise for deconv2d_depthwise_kh2_kw2_sh2_sw2_c32. Constexpr-baked dims; armbench_entry_deconv2d_depthwise calls inner_deconv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d_depthwise.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d_depthwise specialisation.\nnamespace deconv2d_depthwise_def {\nconstexpr int C = 32;\nconstexpr int Kh = 2;\nconstexpr int Kw = 2;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\n// Weight layout: (C, Kh, Kw) \u2014 one filter per channel.\nvoid inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d_depthwise.\n// ABI: armbench_entry_deconv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise transposed conv2d (deconv2d_depthwise).\n// LLM target: replace this file with an optimised inner_deconv2d_depthwise.\n// All per-definition constants live in deconv2d_depthwise_def:: (deconv2d_depthwise.h).\n// Weight layout: (C, Kh, Kw); no padding applied.\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" void inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n float* out_c = out_n + (long)c * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_c[i] = bias[c];\n }\n }\n\n // Scatter-accumulate from each input element (depthwise: channels independent)\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_c[ih * W + iw];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_c[oh * W_out + ow] += in_val * w_c[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c512.json b/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c512.json deleted file mode 100644 index 85814f37922c44852fe4156ffd4c67b4a6aad053..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c512.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c512", - "definition": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c512", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d_depthwise for deconv2d_depthwise_kh2_kw2_sh2_sw2_c512. Constexpr-baked dims; armbench_entry_deconv2d_depthwise calls inner_deconv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d_depthwise.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d_depthwise specialisation.\nnamespace deconv2d_depthwise_def {\nconstexpr int C = 512;\nconstexpr int Kh = 2;\nconstexpr int Kw = 2;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\n// Weight layout: (C, Kh, Kw) \u2014 one filter per channel.\nvoid inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d_depthwise.\n// ABI: armbench_entry_deconv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise transposed conv2d (deconv2d_depthwise).\n// LLM target: replace this file with an optimised inner_deconv2d_depthwise.\n// All per-definition constants live in deconv2d_depthwise_def:: (deconv2d_depthwise.h).\n// Weight layout: (C, Kh, Kw); no padding applied.\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" void inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n float* out_c = out_n + (long)c * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_c[i] = bias[c];\n }\n }\n\n // Scatter-accumulate from each input element (depthwise: channels independent)\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_c[ih * W + iw];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_c[oh * W_out + ow] += in_val * w_c[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c64.json b/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c64.json deleted file mode 100644 index f55e9ad56257ab7f043f1261b2c6045b676cc204..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c64.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c64", - "definition": "deconv2d_depthwise_kh2_kw2_sh2_sw2_c64", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d_depthwise for deconv2d_depthwise_kh2_kw2_sh2_sw2_c64. Constexpr-baked dims; armbench_entry_deconv2d_depthwise calls inner_deconv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d_depthwise.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d_depthwise specialisation.\nnamespace deconv2d_depthwise_def {\nconstexpr int C = 64;\nconstexpr int Kh = 2;\nconstexpr int Kw = 2;\nconstexpr int Sh = 2;\nconstexpr int Sw = 2;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\n// Weight layout: (C, Kh, Kw) \u2014 one filter per channel.\nvoid inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d_depthwise.\n// ABI: armbench_entry_deconv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise transposed conv2d (deconv2d_depthwise).\n// LLM target: replace this file with an optimised inner_deconv2d_depthwise.\n// All per-definition constants live in deconv2d_depthwise_def:: (deconv2d_depthwise.h).\n// Weight layout: (C, Kh, Kw); no padding applied.\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" void inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n float* out_c = out_n + (long)c * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_c[i] = bias[c];\n }\n }\n\n // Scatter-accumulate from each input element (depthwise: channels independent)\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_c[ih * W + iw];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_c[oh * W_out + ow] += in_val * w_c[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c128.json b/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c128.json deleted file mode 100644 index 426df15255f4ae20c057439199f3edd1b92ab00f..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c128.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c128", - "definition": "deconv2d_depthwise_kh3_kw3_sh1_sw1_c128", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d_depthwise for deconv2d_depthwise_kh3_kw3_sh1_sw1_c128. Constexpr-baked dims; armbench_entry_deconv2d_depthwise calls inner_deconv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d_depthwise.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d_depthwise specialisation.\nnamespace deconv2d_depthwise_def {\nconstexpr int C = 128;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\n// Weight layout: (C, Kh, Kw) \u2014 one filter per channel.\nvoid inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d_depthwise.\n// ABI: armbench_entry_deconv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise transposed conv2d (deconv2d_depthwise).\n// LLM target: replace this file with an optimised inner_deconv2d_depthwise.\n// All per-definition constants live in deconv2d_depthwise_def:: (deconv2d_depthwise.h).\n// Weight layout: (C, Kh, Kw); no padding applied.\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" void inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n float* out_c = out_n + (long)c * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_c[i] = bias[c];\n }\n }\n\n // Scatter-accumulate from each input element (depthwise: channels independent)\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_c[ih * W + iw];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_c[oh * W_out + ow] += in_val * w_c[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048.json b/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048.json deleted file mode 100644 index 15ddb15e193b8a1d9d11e6daae3fb4e974e698a9..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048", - "definition": "deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d_depthwise for deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048. Constexpr-baked dims; armbench_entry_deconv2d_depthwise calls inner_deconv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d_depthwise.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d_depthwise specialisation.\nnamespace deconv2d_depthwise_def {\nconstexpr int C = 2048;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\n// Weight layout: (C, Kh, Kw) \u2014 one filter per channel.\nvoid inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d_depthwise.\n// ABI: armbench_entry_deconv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise transposed conv2d (deconv2d_depthwise).\n// LLM target: replace this file with an optimised inner_deconv2d_depthwise.\n// All per-definition constants live in deconv2d_depthwise_def:: (deconv2d_depthwise.h).\n// Weight layout: (C, Kh, Kw); no padding applied.\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" void inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n float* out_c = out_n + (long)c * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_c[i] = bias[c];\n }\n }\n\n // Scatter-accumulate from each input element (depthwise: channels independent)\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_c[ih * W + iw];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_c[oh * W_out + ow] += in_val * w_c[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c32.json b/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c32.json deleted file mode 100644 index e1e0f10b36836b9ade80042f1e558f1f35086118..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c32.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c32", - "definition": "deconv2d_depthwise_kh3_kw3_sh1_sw1_c32", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d_depthwise for deconv2d_depthwise_kh3_kw3_sh1_sw1_c32. Constexpr-baked dims; armbench_entry_deconv2d_depthwise calls inner_deconv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d_depthwise.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d_depthwise specialisation.\nnamespace deconv2d_depthwise_def {\nconstexpr int C = 32;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\n// Weight layout: (C, Kh, Kw) \u2014 one filter per channel.\nvoid inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d_depthwise.\n// ABI: armbench_entry_deconv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise transposed conv2d (deconv2d_depthwise).\n// LLM target: replace this file with an optimised inner_deconv2d_depthwise.\n// All per-definition constants live in deconv2d_depthwise_def:: (deconv2d_depthwise.h).\n// Weight layout: (C, Kh, Kw); no padding applied.\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" void inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n float* out_c = out_n + (long)c * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_c[i] = bias[c];\n }\n }\n\n // Scatter-accumulate from each input element (depthwise: channels independent)\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_c[ih * W + iw];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_c[oh * W_out + ow] += in_val * w_c[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c512.json b/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c512.json deleted file mode 100644 index c6858e0406bd606aecdac9edf35d194a4a2a7199..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c512.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c512", - "definition": "deconv2d_depthwise_kh3_kw3_sh1_sw1_c512", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d_depthwise for deconv2d_depthwise_kh3_kw3_sh1_sw1_c512. Constexpr-baked dims; armbench_entry_deconv2d_depthwise calls inner_deconv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d_depthwise.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d_depthwise specialisation.\nnamespace deconv2d_depthwise_def {\nconstexpr int C = 512;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\n// Weight layout: (C, Kh, Kw) \u2014 one filter per channel.\nvoid inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d_depthwise.\n// ABI: armbench_entry_deconv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise transposed conv2d (deconv2d_depthwise).\n// LLM target: replace this file with an optimised inner_deconv2d_depthwise.\n// All per-definition constants live in deconv2d_depthwise_def:: (deconv2d_depthwise.h).\n// Weight layout: (C, Kh, Kw); no padding applied.\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" void inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n float* out_c = out_n + (long)c * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_c[i] = bias[c];\n }\n }\n\n // Scatter-accumulate from each input element (depthwise: channels independent)\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_c[ih * W + iw];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_c[oh * W_out + ow] += in_val * w_c[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c64.json b/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c64.json deleted file mode 100644 index ecc8f36aa7316bf634c089a007aa3fe1a614b6e6..0000000000000000000000000000000000000000 --- a/solutions/ncnn/reference-scalar/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c64.json +++ /dev/null @@ -1,36 +0,0 @@ -{ - "name": "reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c64", - "definition": "deconv2d_depthwise_kh3_kw3_sh1_sw1_c64", - "dataset": "ncnn", - "author": "reference-scalar", - "description": "Scalar raw-float* deconv2d_depthwise for deconv2d_depthwise_kh3_kw3_sh1_sw1_c64. Constexpr-baked dims; armbench_entry_deconv2d_depthwise calls inner_deconv2d_depthwise. Ground-truth correctness baseline.", - "spec": { - "language": "cpp", - "target_hardware": [ - "graviton3", - "aarch64-sve" - ], - "entry_point": "deconv2d_depthwise.cpp::armbench_entry_deconv2d_depthwise", - "dependencies": [], - "isa_features": [], - "compile_flags": [ - "-O2", - "-std=c++14" - ], - "link_flags": [] - }, - "sources": [ - { - "path": "deconv2d_depthwise.h", - "content": "#pragma once\n\n// Per-definition constants for this deconv2d_depthwise specialisation.\nnamespace deconv2d_depthwise_def {\nconstexpr int C = 64;\nconstexpr int Kh = 3;\nconstexpr int Kw = 3;\nconstexpr int Sh = 1;\nconstexpr int Sw = 1;\nconstexpr int Dh = 1;\nconstexpr int Dw = 1;\n} // namespace deconv2d_depthwise_def\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n// LLM target: implement this in kernel.cpp.\n// N, H, W are var dims; H_out/W_out are pre-computed by the binding harness.\n// Input layout: (N, C, H, W), output layout: (N, C, H_out, W_out).\n// Weight layout: (C, Kh, Kw) \u2014 one filter per channel.\nvoid inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out);\n#ifdef __cplusplus\n}\n#endif\n" - }, - { - "path": "deconv2d_depthwise.cpp", - "content": "// Binding harness: computes H_out/W_out from constexpr params, calls inner_deconv2d_depthwise.\n// ABI: armbench_entry_deconv2d_depthwise(input*, output*, weight*, bias*, N, H, W)\n// H_out = (H - 1) * Sh + Kh (no-padding formula)\n// Input layout: (N, C, H, W); output layout: (N, C, H_out, W_out).\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" int armbench_entry_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W)\n{\n const int H_out = (H - 1) * Sh + Kh;\n const int W_out = (W - 1) * Sw + Kw;\n inner_deconv2d_depthwise(input, output, weight, bias, N, H, W, H_out, W_out);\n return 0;\n}\n" - }, - { - "path": "kernel.cpp", - "content": "// Reference-scalar depthwise transposed conv2d (deconv2d_depthwise).\n// LLM target: replace this file with an optimised inner_deconv2d_depthwise.\n// All per-definition constants live in deconv2d_depthwise_def:: (deconv2d_depthwise.h).\n// Weight layout: (C, Kh, Kw); no padding applied.\n#include \"deconv2d_depthwise.h\"\nusing namespace deconv2d_depthwise_def;\n\nextern \"C\" void inner_deconv2d_depthwise(\n const float* input, float* output,\n const float* weight, const float* bias,\n int N, int H, int W, int H_out, int W_out)\n{\n // Zero-fill + bias initialisation\n for (int n = 0; n < N; ++n) {\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n float* out_c = out_n + (long)c * H_out * W_out;\n for (int i = 0; i < H_out * W_out; ++i)\n out_c[i] = bias[c];\n }\n }\n\n // Scatter-accumulate from each input element (depthwise: channels independent)\n for (int n = 0; n < N; ++n) {\n const float* in_n = input + (long)n * C * H * W;\n float* out_n = output + (long)n * C * H_out * W_out;\n for (int c = 0; c < C; ++c) {\n const float* in_c = in_n + (long)c * H * W;\n float* out_c = out_n + (long)c * H_out * W_out;\n const float* w_c = weight + (long)c * Kh * Kw;\n for (int ih = 0; ih < H; ++ih) {\n for (int iw = 0; iw < W; ++iw) {\n float in_val = in_c[ih * W + iw];\n for (int kh = 0; kh < Kh; ++kh) {\n for (int kw = 0; kw < Kw; ++kw) {\n int oh = ih * Sh + kh;\n int ow = iw * Sw + kw;\n out_c[oh * W_out + ow] += in_val * w_c[kh * Kw + kw];\n }\n }\n }\n }\n }\n }\n}\n" - } - ] -} diff --git a/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin1024_cout1024.jsonl b/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin1024_cout1024.jsonl deleted file mode 100644 index 8ccc8471792099c22659aca660aa353afbc0329f..0000000000000000000000000000000000000000 --- a/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin1024_cout1024.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv1d_kw1_sw1_dw1_p0_cin1024_cout1024","workload":{"axes":{"N":1,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"fab88bdbd2105fbf9a54d8d9aaf86592","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin1024_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:22.859731+00:00","log":"","correctness":{"max_relative_error":0.0035176701570680626,"max_absolute_error":0.00003814697265625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2632760,"p5_ns":2658788,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":7337100,"instructions":15300565,"ipc":2.0854,"cache_misses":2258442.24,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin1024_cout1024","workload":{"axes":{"N":1,"W":128},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"577a3b637e825bb4ba0126e678657ed1","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin1024_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:22.859731+00:00","log":"","correctness":{"max_relative_error":0.06637168141592921,"max_absolute_error":0.00002288818359375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":30593213,"p5_ns":30613150,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":79567230,"instructions":112545914,"ipc":1.4145,"cache_misses":16526976.84,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin1024_cout1024","workload":{"axes":{"N":1,"W":512},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"1d4dd4138a2d53c8b041494f575d1db2","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin1024_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:22.859731+00:00","log":"","correctness":{"max_relative_error":0.3295774647887324,"max_absolute_error":0.00002288818359375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":126549217,"p5_ns":126736204,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":329878456,"instructions":440121816,"ipc":1.3342,"cache_misses":63416012.92,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin1024_cout1024","workload":{"axes":{"N":1,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"fab88bdbd2105fbf9a54d8d9aaf86592","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin1024_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:34:14.091879+00:00","log":"","correctness":{"max_relative_error":0.001075362917590247,"max_absolute_error":0.0000457763671875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":11546836,"p5_ns":11556121,"reference_min_ns":2632760,"speedup":0.24375960494272758,"repeat":50,"warmup":5,"cycles":30099737,"instructions":147034976,"ipc":4.8849,"cache_misses":1233262.1,"reference_cycles":7337100}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin1024_cout1024","workload":{"axes":{"N":1,"W":128},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"577a3b637e825bb4ba0126e678657ed1","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin1024_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:34:14.091879+00:00","log":"","correctness":{"max_relative_error":0.024350012133935256,"max_absolute_error":0.000057220458984375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":216217664,"p5_ns":216225919,"reference_min_ns":30593213,"speedup":0.141799398711481,"repeat":50,"warmup":5,"cycles":561125299,"instructions":1344982039,"ipc":2.3969,"cache_misses":135064633.96,"reference_cycles":79567230}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin1024_cout1024","workload":{"axes":{"N":1,"W":512},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"1d4dd4138a2d53c8b041494f575d1db2","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin1024_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:34:14.091879+00:00","log":"","correctness":{"max_relative_error":0.8033690201608759,"max_absolute_error":0.0000762939453125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":942017023,"p5_ns":942335811,"reference_min_ns":126549217,"speedup":0.1349127308735995,"repeat":32,"warmup":5,"cycles":2445124740,"instructions":5380452349,"ipc":2.2005,"cache_misses":537500218.21875,"reference_cycles":329878456}}} diff --git a/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin128_cout512.jsonl b/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin128_cout512.jsonl deleted file mode 100644 index 28abb5d4215960ef178de2479dbdb02b5d31ad84..0000000000000000000000000000000000000000 --- a/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin128_cout512.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv1d_kw1_sw1_dw1_p0_cin128_cout512","workload":{"axes":{"N":1,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"cec30cf820b457b592fa238dbd074993","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin128_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:34.198344+00:00","log":"","correctness":{"max_relative_error":0.012154696132596685,"max_absolute_error":7.62939453125e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":311197,"p5_ns":311603,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":842503,"instructions":3600591,"ipc":4.2737,"cache_misses":10212.96,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin128_cout512","workload":{"axes":{"N":1,"W":128},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"718ae81c4387579e9962f9c310c2b16e","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin128_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:34.198344+00:00","log":"","correctness":{"max_relative_error":0.006494784491241881,"max_absolute_error":7.152557373046875e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1311419,"p5_ns":1312933,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":3436189,"instructions":7930083,"ipc":2.3078,"cache_misses":508764.22,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin128_cout512","workload":{"axes":{"N":1,"W":512},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"1a0853a9c9b9554387446b61236890be","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin128_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:34.198344+00:00","log":"","correctness":{"max_relative_error":0.009389671361502348,"max_absolute_error":4.76837158203125e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":7615508,"p5_ns":7619359,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":19839001,"instructions":31031651,"ipc":1.5642,"cache_misses":4088941.24,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin128_cout512","workload":{"axes":{"N":1,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"cec30cf820b457b592fa238dbd074993","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin128_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:35:04.708190+00:00","log":"","correctness":{"max_relative_error":0.011536386395027624,"max_absolute_error":4.76837158203125e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2864870,"p5_ns":2867050,"reference_min_ns":311197,"speedup":0.11299543459834122,"repeat":50,"warmup":5,"cycles":7456080,"instructions":36994713,"ipc":4.9617,"cache_misses":3097.24,"reference_cycles":842503}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin128_cout512","workload":{"axes":{"N":1,"W":128},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"718ae81c4387579e9962f9c310c2b16e","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin128_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:35:04.708190+00:00","log":"","correctness":{"max_relative_error":0.007797150304221068,"max_absolute_error":5.7220458984375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":6558668,"p5_ns":6560229,"reference_min_ns":1311419,"speedup":0.20153265482264107,"repeat":50,"warmup":5,"cycles":17050284,"instructions":84539598,"ipc":4.9583,"cache_misses":884214.1,"reference_cycles":3436189}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin128_cout512","workload":{"axes":{"N":1,"W":512},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"1a0853a9c9b9554387446b61236890be","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin128_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:35:04.708190+00:00","log":"","correctness":{"max_relative_error":0.020738796311981823,"max_absolute_error":8.58306884765625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":53216546,"p5_ns":53242671,"reference_min_ns":7615508,"speedup":0.14343824885007198,"repeat":50,"warmup":5,"cycles":138310396,"instructions":338315429,"ipc":2.4461,"cache_misses":33814851.92,"reference_cycles":19839001}}} diff --git a/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin256_cout1024.jsonl b/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin256_cout1024.jsonl deleted file mode 100644 index 3910e060d1ffdaa2754aae331b9f7ea2da049ab9..0000000000000000000000000000000000000000 --- a/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin256_cout1024.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv1d_kw1_sw1_dw1_p0_cin256_cout1024","workload":{"axes":{"N":1,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"52c54c1b894859c882018d67dec858c1","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin256_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:37.023889+00:00","log":"","correctness":{"max_relative_error":0.06896551724137931,"max_absolute_error":0.00001239776611328125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":647458,"p5_ns":650036,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1732451,"instructions":7159651,"ipc":4.1327,"cache_misses":31065.78,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin256_cout1024","workload":{"axes":{"N":1,"W":128},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"f922cf2a09b85cd1a21756f35a10dd88","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin256_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:37.023889+00:00","log":"","correctness":{"max_relative_error":0.015898899306971057,"max_absolute_error":0.000019073486328125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":7797041,"p5_ns":7802851,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":20361805,"instructions":29662940,"ipc":1.4568,"cache_misses":4156051.82,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin256_cout1024","workload":{"axes":{"N":1,"W":512},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"2be023aa408b599aacfbc0a790eeac77","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin256_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:37.023889+00:00","log":"","correctness":{"max_relative_error":0.04918032786885246,"max_absolute_error":6.67572021484375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":29655290,"p5_ns":29669692,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":77049621,"instructions":116037251,"ipc":1.506,"cache_misses":16143959.76,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin256_cout1024","workload":{"axes":{"N":1,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"52c54c1b894859c882018d67dec858c1","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin256_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:35:08.358932+00:00","log":"","correctness":{"max_relative_error":0.05415252159381735,"max_absolute_error":0.0000133514404296875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":5710051,"p5_ns":5717562,"reference_min_ns":647458,"speedup":0.11638736827113347,"repeat":50,"warmup":5,"cycles":14885215,"instructions":73724646,"ipc":4.9529,"cache_misses":4628.66,"reference_cycles":1732451}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin256_cout1024","workload":{"axes":{"N":1,"W":128},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"f922cf2a09b85cd1a21756f35a10dd88","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin256_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:35:08.358932+00:00","log":"","correctness":{"max_relative_error":0.0021218238827456176,"max_absolute_error":9.5367431640625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":53947423,"p5_ns":53955650,"reference_min_ns":7797041,"speedup":0.14535592128777985,"repeat":50,"warmup":5,"cycles":140082391,"instructions":337143001,"ipc":2.4067,"cache_misses":33957677.14,"reference_cycles":20361805}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin256_cout1024","workload":{"axes":{"N":1,"W":512},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"2be023aa408b599aacfbc0a790eeac77","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin256_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:35:08.358932+00:00","log":"","correctness":{"max_relative_error":0.01835751142658171,"max_absolute_error":0.0000152587890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":213539132,"p5_ns":213561930,"reference_min_ns":29655290,"speedup":0.1390853988007765,"repeat":50,"warmup":5,"cycles":553973470,"instructions":1348498981,"ipc":2.4342,"cache_misses":134595381.32,"reference_cycles":77049621}}} diff --git a/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin512_cout2048.jsonl b/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin512_cout2048.jsonl deleted file mode 100644 index d56a41a8e57fb12791f4a14cf232a67b53578aab..0000000000000000000000000000000000000000 --- a/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin512_cout2048.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv1d_kw1_sw1_dw1_p0_cin512_cout2048","workload":{"axes":{"N":1,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a2f95070d3375ad6ac95f27bac9fee19","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin512_cout2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:41.476171+00:00","log":"","correctness":{"max_relative_error":0.0034673173237447026,"max_absolute_error":0.0000324249267578125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1545095,"p5_ns":1565399,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":4404261,"instructions":15531669,"ipc":3.5265,"cache_misses":117982.78,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin512_cout2048","workload":{"axes":{"N":1,"W":128},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"9c1aeac6a498507588ed42d082b38137","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin512_cout2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:41.476171+00:00","log":"","correctness":{"max_relative_error":0.02925531914893617,"max_absolute_error":0.000019073486328125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":30707474,"p5_ns":30722942,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":79795736,"instructions":114558791,"ipc":1.4357,"cache_misses":16477333.36,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin512_cout2048","workload":{"axes":{"N":1,"W":512},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a92410c8fb385fb1b3c9b3ac8f646db3","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin512_cout2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:41.476171+00:00","log":"","correctness":{"max_relative_error":0.18471337579617833,"max_absolute_error":0.0000133514404296875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":121266894,"p5_ns":121390530,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":314963411,"instructions":448089566,"ipc":1.4227,"cache_misses":64071537.7,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin512_cout2048","workload":{"axes":{"N":1,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a2f95070d3375ad6ac95f27bac9fee19","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin512_cout2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:35:23.833887+00:00","log":"","correctness":{"max_relative_error":0.0019206690638243226,"max_absolute_error":0.0000133514404296875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":11522191,"p5_ns":11527429,"reference_min_ns":1545095,"speedup":0.14678552745806528,"repeat":50,"warmup":5,"cycles":30004736,"instructions":147175161,"ipc":4.9051,"cache_misses":9586.62,"reference_cycles":4404261}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin512_cout2048","workload":{"axes":{"N":1,"W":128},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"9c1aeac6a498507588ed42d082b38137","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin512_cout2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:35:23.833887+00:00","log":"","correctness":{"max_relative_error":0.013375831131480515,"max_absolute_error":0.000034332275390625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":216175868,"p5_ns":216233358,"reference_min_ns":30707474,"speedup":0.14220930615496036,"repeat":50,"warmup":5,"cycles":561114727,"instructions":1346169327,"ipc":2.3991,"cache_misses":135305317.78,"reference_cycles":79795736}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin512_cout2048","workload":{"axes":{"N":1,"W":512},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a92410c8fb385fb1b3c9b3ac8f646db3","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin512_cout2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:35:23.833887+00:00","log":"","correctness":{"max_relative_error":0.46102624780991497,"max_absolute_error":0.0000362396240234375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":895950752,"p5_ns":896132378,"reference_min_ns":121266894,"speedup":0.13547225246952174,"repeat":34,"warmup":5,"cycles":2324929314,"instructions":5384831519,"ipc":2.3161,"cache_misses":538022155.6764706,"reference_cycles":314963411}}} diff --git a/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout256.jsonl b/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout256.jsonl deleted file mode 100644 index e4bdc3bd50962c516f00cd65dc9f67d1dbb2390f..0000000000000000000000000000000000000000 --- a/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout256.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv1d_kw1_sw1_dw1_p0_cin64_cout256","workload":{"axes":{"N":1,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a3b0a501961750c98c7eb28acf8b9e22","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin64_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:52.446754+00:00","log":"","correctness":{"max_relative_error":0.002070393374741201,"max_absolute_error":3.814697265625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":149272,"p5_ns":150862,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":453856,"instructions":1976930,"ipc":4.3559,"cache_misses":4507.98,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin64_cout256","workload":{"axes":{"N":1,"W":128},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"7ff60359cfce5a08902dad6b049c87b4","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin64_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:52.446754+00:00","log":"","correctness":{"max_relative_error":0.011064278187565859,"max_absolute_error":3.814697265625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":192394,"p5_ns":193627,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":546511,"instructions":2248662,"ipc":4.1146,"cache_misses":5436.96,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin64_cout256","workload":{"axes":{"N":1,"W":512},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"27366d29c6435d6f93bdb3323edc0f47","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin64_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:52.446754+00:00","log":"","correctness":{"max_relative_error":0.008547008547008548,"max_absolute_error":3.814697265625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2027552,"p5_ns":2030630,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":5295183,"instructions":8778362,"ipc":1.6578,"cache_misses":1063263.22,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin64_cout256","workload":{"axes":{"N":1,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a3b0a501961750c98c7eb28acf8b9e22","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin64_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:36:12.726665+00:00","log":"","correctness":{"max_relative_error":0.0008615055439635094,"max_absolute_error":3.337860107421875e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1449341,"p5_ns":1449372,"reference_min_ns":149272,"speedup":0.12009250586099778,"repeat":50,"warmup":5,"cycles":3779220,"instructions":18634655,"ipc":4.9308,"cache_misses":2711.32,"reference_cycles":453856}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin64_cout256","workload":{"axes":{"N":1,"W":128},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"7ff60359cfce5a08902dad6b049c87b4","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin64_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:36:12.726665+00:00","log":"","correctness":{"max_relative_error":0.0034505122579689147,"max_absolute_error":2.86102294921875e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1655736,"p5_ns":1655781,"reference_min_ns":192394,"speedup":0.12678242879622217,"repeat":50,"warmup":5,"cycles":4310621,"instructions":21291183,"ipc":4.9392,"cache_misses":3257.36,"reference_cycles":546511}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin64_cout256","workload":{"axes":{"N":1,"W":512},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"27366d29c6435d6f93bdb3323edc0f47","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin64_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:36:12.726665+00:00","log":"","correctness":{"max_relative_error":0.0037165859289336623,"max_absolute_error":1.9073486328125e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":13069408,"p5_ns":13078978,"reference_min_ns":2027552,"speedup":0.15574092382382018,"repeat":50,"warmup":5,"cycles":33999946,"instructions":85173053,"ipc":2.5051,"cache_misses":8339993.24,"reference_cycles":5295183}}} diff --git a/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout64.jsonl b/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout64.jsonl deleted file mode 100644 index e6a6e57045a343c69e3ed0da2914f8e385d28916..0000000000000000000000000000000000000000 --- a/traces/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout64.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv1d_kw1_sw1_dw1_p0_cin64_cout64","workload":{"axes":{"N":1,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"0dacddd55486558aa1efb586aac3e212","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin64_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:54.883416+00:00","log":"","correctness":{"max_relative_error":0.0006903143585386576,"max_absolute_error":3.814697265625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":84954,"p5_ns":85081,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":246599,"instructions":984108,"ipc":3.9907,"cache_misses":7843.06,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin64_cout64","workload":{"axes":{"N":1,"W":128},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"7bdd1ceb36ab5f41acab6bea69da6840","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin64_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:54.883416+00:00","log":"","correctness":{"max_relative_error":0.004441812259401836,"max_absolute_error":3.0994415283203125e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":48892,"p5_ns":49149,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":143331,"instructions":577014,"ipc":4.0257,"cache_misses":2215.98,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin64_cout64","workload":{"axes":{"N":1,"W":512},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"91ed2f6f0a575aa7b39b9e78451162d6","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw1_sw1_dw1_p0_cin64_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:54.883416+00:00","log":"","correctness":{"max_relative_error":0.002568662319699664,"max_absolute_error":3.814697265625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":512309,"p5_ns":512901,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1346544,"instructions":2208698,"ipc":1.6403,"cache_misses":263609.72,"reference_cycles":null}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin64_cout64","workload":{"axes":{"N":1,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"0dacddd55486558aa1efb586aac3e212","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin64_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:36:13.766964+00:00","log":"","correctness":{"max_relative_error":0.0005770733404607798,"max_absolute_error":2.86102294921875e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":723558,"p5_ns":723599,"reference_min_ns":84954,"speedup":0.13023538054393824,"repeat":50,"warmup":5,"cycles":1893487,"instructions":9322820,"ipc":4.9236,"cache_misses":6384.38,"reference_cycles":246599}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin64_cout64","workload":{"axes":{"N":1,"W":128},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"7bdd1ceb36ab5f41acab6bea69da6840","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin64_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:36:13.766964+00:00","log":"","correctness":{"max_relative_error":0.01037210877350459,"max_absolute_error":2.86102294921875e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":413913,"p5_ns":413927,"reference_min_ns":48892,"speedup":0.13257272348887758,"repeat":50,"warmup":5,"cycles":1081150,"instructions":5330013,"ipc":4.9299,"cache_misses":1115.34,"reference_cycles":143331}}} -{"definition":"conv1d_kw1_sw1_dw1_p0_cin64_cout64","workload":{"axes":{"N":1,"W":512},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"91ed2f6f0a575aa7b39b9e78451162d6","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw1_sw1_dw1_p0_cin64_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:36:13.766964+00:00","log":"","correctness":{"max_relative_error":0.0005918628107892336,"max_absolute_error":1.9073486328125e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":3251970,"p5_ns":3254145,"reference_min_ns":512309,"speedup":0.15846349215763736,"repeat":50,"warmup":5,"cycles":8497503,"instructions":21301074,"ipc":2.5067,"cache_misses":2085510.56,"reference_cycles":1346544}}} diff --git a/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin1024_cout1024.jsonl b/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin1024_cout1024.jsonl deleted file mode 100644 index 48329912613c193a9c95f933449623f208349a08..0000000000000000000000000000000000000000 --- a/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin1024_cout1024.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv1d_kw3_sw1_dw1_p1_cin1024_cout1024","workload":{"axes":{"N":1,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"615d02ab9f3b5706bbacbbc86abf53a4","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin1024_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:57.228734+00:00","log":"","correctness":{"max_relative_error":0.002910385259631491,"max_absolute_error":0.00005817413330078125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":6161087,"p5_ns":6182598,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":16834187,"instructions":37845428,"ipc":2.2481,"cache_misses":2319231.78,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin1024_cout1024","workload":{"axes":{"N":1,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"35939e65bd5851bd92830340b26df476","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin1024_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:57.228734+00:00","log":"","correctness":{"max_relative_error":0.05514382615249341,"max_absolute_error":0.0000762939453125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":27204339,"p5_ns":27216844,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":70870714,"instructions":139564142,"ipc":1.9693,"cache_misses":13127561.8,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin1024_cout1024","workload":{"axes":{"N":1,"W":256},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"82cd1753b7da537486b6f7e30dc63995","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin1024_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:57.228734+00:00","log":"","correctness":{"max_relative_error":0.21509433962264152,"max_absolute_error":0.00006103515625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":105584919,"p5_ns":105646211,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":274032196,"instructions":530168223,"ipc":1.9347,"cache_misses":62254428.62,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin1024_cout1024","workload":{"axes":{"N":1,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"615d02ab9f3b5706bbacbbc86abf53a4","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin1024_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:36:14.154052+00:00","log":"","correctness":{"max_relative_error":0.0035288975223659205,"max_absolute_error":0.00014495849609375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":32565781,"p5_ns":32581968,"reference_min_ns":6161087,"speedup":0.1988082197274801,"repeat":50,"warmup":5,"cycles":84675508,"instructions":300344116,"ipc":3.547,"cache_misses":701799.7,"reference_cycles":16834187}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin1024_cout1024","workload":{"axes":{"N":1,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"35939e65bd5851bd92830340b26df476","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin1024_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:36:14.154052+00:00","log":"","correctness":{"max_relative_error":0.07354585277397659,"max_absolute_error":0.000152587890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":154395645,"p5_ns":154409403,"reference_min_ns":27204339,"speedup":0.1768208053761486,"repeat":50,"warmup":5,"cycles":400805289,"instructions":1402927179,"ipc":3.5003,"cache_misses":67461115.02,"reference_cycles":70870714}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin1024_cout1024","workload":{"axes":{"N":1,"W":256},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"82cd1753b7da537486b6f7e30dc63995","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin1024_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:36:14.154052+00:00","log":"","correctness":{"max_relative_error":0.14531693368587853,"max_absolute_error":0.000179290771484375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":628753231,"p5_ns":628846286,"reference_min_ns":105584919,"speedup":0.16815306472517155,"repeat":48,"warmup":5,"cycles":1629659242,"instructions":5636851230,"ipc":3.4589,"cache_misses":300984989.9166667,"reference_cycles":274032196}}} diff --git a/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin128_cout256.jsonl b/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin128_cout256.jsonl deleted file mode 100644 index 353b758eaea0c96fbefe038a45757d8f3343a10a..0000000000000000000000000000000000000000 --- a/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin128_cout256.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv1d_kw3_sw1_dw1_p1_cin128_cout256","workload":{"axes":{"N":1,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"3a1f65212bc45cbbae4c74095c7f12e4","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:07.429825+00:00","log":"","correctness":{"max_relative_error":0.00705371676614216,"max_absolute_error":0.000019073486328125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":415654,"p5_ns":417634,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1131452,"instructions":4077867,"ipc":3.6041,"cache_misses":14127.46,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin128_cout256","workload":{"axes":{"N":1,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"34962fa830d65cbe976a4e5be3d6556e","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:07.429825+00:00","log":"","correctness":{"max_relative_error":0.01291409320606401,"max_absolute_error":0.000019073486328125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":469498,"p5_ns":470301,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1286014,"instructions":4613326,"ipc":3.5873,"cache_misses":17834.94,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin128_cout256","workload":{"axes":{"N":1,"W":256},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"2cc9269a5e57520f933c1afc630f96b6","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:07.429825+00:00","log":"","correctness":{"max_relative_error":0.012522361359570662,"max_absolute_error":7.62939453125e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1903515,"p5_ns":1905016,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":5035302,"instructions":17480577,"ipc":3.4716,"cache_misses":77676.28,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin128_cout256","workload":{"axes":{"N":1,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"3a1f65212bc45cbbae4c74095c7f12e4","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:36:58.860659+00:00","log":"","correctness":{"max_relative_error":0.0013131855626609824,"max_absolute_error":0.000011444091796875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":3922822,"p5_ns":3923161,"reference_min_ns":415654,"speedup":0.11092988580845091,"repeat":50,"warmup":5,"cycles":10199704,"instructions":38500906,"ipc":3.7747,"cache_misses":2353.3,"reference_cycles":1131452}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin128_cout256","workload":{"axes":{"N":1,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"34962fa830d65cbe976a4e5be3d6556e","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:36:58.860659+00:00","log":"","correctness":{"max_relative_error":0.0012522316728663671,"max_absolute_error":0.00001049041748046875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":4487942,"p5_ns":4488073,"reference_min_ns":469498,"speedup":0.11019629379451665,"repeat":50,"warmup":5,"cycles":11670211,"instructions":44036292,"ipc":3.7734,"cache_misses":3145.54,"reference_cycles":1286014}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin128_cout256","workload":{"axes":{"N":1,"W":256},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"2cc9269a5e57520f933c1afc630f96b6","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:36:58.860659+00:00","log":"","correctness":{"max_relative_error":0.0026394597894881674,"max_absolute_error":0.000026702880859375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":18413124,"p5_ns":18415466,"reference_min_ns":1903515,"speedup":0.10528032042438451,"repeat":50,"warmup":5,"cycles":47827571,"instructions":176900036,"ipc":3.6987,"cache_misses":9400549.68,"reference_cycles":5035302}}} diff --git a/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin256_cout512.jsonl b/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin256_cout512.jsonl deleted file mode 100644 index b17229461988b049005fd6015864cc646cdbdf71..0000000000000000000000000000000000000000 --- a/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin256_cout512.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv1d_kw3_sw1_dw1_p1_cin256_cout512","workload":{"axes":{"N":1,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"85f2a76a3cc359869aea5aa4d8002466","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:09.865781+00:00","log":"","correctness":{"max_relative_error":0.1206896551724138,"max_absolute_error":0.0000247955322265625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":969850,"p5_ns":975754,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":2577719,"instructions":8487730,"ipc":3.2927,"cache_misses":61164.52,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin256_cout512","workload":{"axes":{"N":1,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"bde237e6b8eb515aba62da434b784405","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:09.865781+00:00","log":"","correctness":{"max_relative_error":0.16167664670658682,"max_absolute_error":0.000034332275390625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2456887,"p5_ns":2460976,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":6472263,"instructions":17852653,"ipc":2.7583,"cache_misses":734417.06,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin256_cout512","workload":{"axes":{"N":1,"W":256},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"57b9e77611a05d8991eacb6e1b3b8146","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:09.865781+00:00","log":"","correctness":{"max_relative_error":0.058333333333333334,"max_absolute_error":0.0000171661376953125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":7943465,"p5_ns":7956839,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":20730077,"instructions":67781629,"ipc":3.2697,"cache_misses":942282.0,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin256_cout512","workload":{"axes":{"N":1,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"85f2a76a3cc359869aea5aa4d8002466","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:00.508381+00:00","log":"","correctness":{"max_relative_error":0.053579067361765896,"max_absolute_error":0.0000476837158203125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":8061420,"p5_ns":8083025,"reference_min_ns":969850,"speedup":0.12207837168155004,"repeat":50,"warmup":5,"cycles":21115280,"instructions":76286539,"ipc":3.6129,"cache_misses":9409.32,"reference_cycles":2577719}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin256_cout512","workload":{"axes":{"N":1,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"bde237e6b8eb515aba62da434b784405","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:00.508381+00:00","log":"","correctness":{"max_relative_error":0.03689410729322605,"max_absolute_error":0.0000400543212890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":18837455,"p5_ns":18845435,"reference_min_ns":2456887,"speedup":0.13209051522513987,"repeat":50,"warmup":5,"cycles":48998696,"instructions":175694211,"ipc":3.5857,"cache_misses":2213622.9,"reference_cycles":6472263}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin256_cout512","workload":{"axes":{"N":1,"W":256},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"57b9e77611a05d8991eacb6e1b3b8146","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:00.508381+00:00","log":"","correctness":{"max_relative_error":0.042153549194335935,"max_absolute_error":0.000049591064453125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":75999910,"p5_ns":76000555,"reference_min_ns":7943465,"speedup":0.10504957069648195,"repeat":50,"warmup":5,"cycles":197336142,"instructions":705878564,"ipc":3.577,"cache_misses":37892795.7,"reference_cycles":20730077}}} diff --git a/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin3_cout64.jsonl b/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin3_cout64.jsonl deleted file mode 100644 index 0d0ababe58bae54766a7b7401be2bd259464f41e..0000000000000000000000000000000000000000 --- a/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin3_cout64.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv1d_kw3_sw1_dw1_p1_cin3_cout64","workload":{"axes":{"N":1,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"d591e2c5227e595786b88b4cd9db2c97","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin3_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:12.798149+00:00","log":"","correctness":{"max_relative_error":0.0005372492836676218,"max_absolute_error":4.76837158203125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":26366,"p5_ns":26425,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":73696,"instructions":367437,"ipc":4.9859,"cache_misses":1649.0,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin3_cout64","workload":{"axes":{"N":1,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"d1fc701a7c145577a8d42a7f7874ea5d","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin3_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:12.798149+00:00","log":"","correctness":{"max_relative_error":0.00010871455905374847,"max_absolute_error":4.76837158203125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":9646,"p5_ns":9672,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":28324,"instructions":124703,"ipc":4.4027,"cache_misses":330.38,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin3_cout64","workload":{"axes":{"N":1,"W":256},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"9e733654ce5751d7a59aa4a9dad47cd6","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin3_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:12.798149+00:00","log":"","correctness":{"max_relative_error":0.0002683919086902475,"max_absolute_error":4.76837158203125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":29638,"p5_ns":29705,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":81975,"instructions":416065,"ipc":5.0755,"cache_misses":1887.82,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin3_cout64","workload":{"axes":{"N":1,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"d591e2c5227e595786b88b4cd9db2c97","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin3_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:06.426666+00:00","log":"","correctness":{"max_relative_error":0.00025699473383775074,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":42916,"p5_ns":42938,"reference_min_ns":26366,"speedup":0.6369741652765413,"repeat":50,"warmup":5,"cycles":115697,"instructions":825576,"ipc":7.1357,"cache_misses":51.28,"reference_cycles":73696}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin3_cout64","workload":{"axes":{"N":1,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"d1fc701a7c145577a8d42a7f7874ea5d","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin3_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:06.426666+00:00","log":"","correctness":{"max_relative_error":0.0001269752076448078,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":13038,"p5_ns":13056,"reference_min_ns":9646,"speedup":0.7738374952188405,"repeat":50,"warmup":5,"cycles":36602,"instructions":241791,"ipc":6.6059,"cache_misses":56.82,"reference_cycles":28324}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin3_cout64","workload":{"axes":{"N":1,"W":256},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"9e733654ce5751d7a59aa4a9dad47cd6","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin3_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:06.426666+00:00","log":"","correctness":{"max_relative_error":0.00015483299666628998,"max_absolute_error":4.76837158203125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":48743,"p5_ns":48748,"reference_min_ns":29638,"speedup":0.6127002160053217,"repeat":50,"warmup":5,"cycles":133793,"instructions":944358,"ipc":7.0584,"cache_misses":70.74,"reference_cycles":81975}}} diff --git a/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin512_cout1024.jsonl b/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin512_cout1024.jsonl deleted file mode 100644 index aa37d57b9f975360f3769477c7a0b3a88bf8303b..0000000000000000000000000000000000000000 --- a/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin512_cout1024.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv1d_kw3_sw1_dw1_p1_cin512_cout1024","workload":{"axes":{"N":1,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"f4bcd3a68a365dd19c3ac9d0e84947b3","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:15.089369+00:00","log":"","correctness":{"max_relative_error":0.22470433639947437,"max_absolute_error":0.0000457763671875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2968043,"p5_ns":2998638,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":8331261,"instructions":19051141,"ipc":2.2867,"cache_misses":953585.16,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin512_cout1024","workload":{"axes":{"N":1,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"4dfe1875b6595cea954357afd2c7d59a","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:15.089369+00:00","log":"","correctness":{"max_relative_error":0.0043812396753573225,"max_absolute_error":0.0000324249267578125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":12266913,"p5_ns":12277387,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":32157275,"instructions":70295306,"ipc":2.186,"cache_misses":6199626.28,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin512_cout1024","workload":{"axes":{"N":1,"W":256},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b278cb338d3a52a1952f71cb776d8c3f","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:15.089369+00:00","log":"","correctness":{"max_relative_error":0.04941176470588235,"max_absolute_error":0.00003814697265625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":47584428,"p5_ns":47602853,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":123569841,"instructions":267067979,"ipc":2.1613,"cache_misses":31989128.9,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin512_cout1024","workload":{"axes":{"N":1,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"f4bcd3a68a365dd19c3ac9d0e84947b3","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:06.575499+00:00","log":"","correctness":{"max_relative_error":0.06493354626930026,"max_absolute_error":0.000072479248046875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":16354799,"p5_ns":16379328,"reference_min_ns":2968043,"speedup":0.19519949402952164,"repeat":50,"warmup":5,"cycles":42680751,"instructions":149752984,"ipc":3.5087,"cache_misses":35414.78,"reference_cycles":8331261}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin512_cout1024","workload":{"axes":{"N":1,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"4dfe1875b6595cea954357afd2c7d59a","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:06.575499+00:00","log":"","correctness":{"max_relative_error":0.005720985511447291,"max_absolute_error":0.000095367431640625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":77137417,"p5_ns":77148122,"reference_min_ns":12266913,"speedup":0.16055299757102573,"repeat":50,"warmup":5,"cycles":200290717,"instructions":701370090,"ipc":3.5018,"cache_misses":33230196.54,"reference_cycles":32157275}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin512_cout1024","workload":{"axes":{"N":1,"W":256},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b278cb338d3a52a1952f71cb776d8c3f","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:06.575499+00:00","log":"","correctness":{"max_relative_error":0.2600893985523897,"max_absolute_error":0.00009918212890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":311768003,"p5_ns":311792428,"reference_min_ns":47584428,"speedup":0.1527993133610199,"repeat":50,"warmup":5,"cycles":808706782,"instructions":2819592209,"ipc":3.4865,"cache_misses":150645143.92,"reference_cycles":123569841}}} diff --git a/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin64_cout128.jsonl b/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin64_cout128.jsonl deleted file mode 100644 index bd38bfe358769c54c01d449198e846902e89473e..0000000000000000000000000000000000000000 --- a/traces/conv1d/conv1d_kw3_sw1_dw1_p1_cin64_cout128.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv1d_kw3_sw1_dw1_p1_cin64_cout128","workload":{"axes":{"N":1,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"3fdfae5415725eccab553e5efd0fd198","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:20.972928+00:00","log":"","correctness":{"max_relative_error":0.0016554848828090965,"max_absolute_error":8.58306884765625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":188531,"p5_ns":191984,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":560720,"instructions":2089491,"ipc":3.7264,"cache_misses":5629.48,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin64_cout128","workload":{"axes":{"N":1,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"ce3f7d5d325c5d7c93b2a47d7dea01ba","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:20.972928+00:00","log":"","correctness":{"max_relative_error":0.0011299435028248588,"max_absolute_error":8.58306884765625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":111952,"p5_ns":113099,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":336159,"instructions":1238127,"ipc":3.6832,"cache_misses":3387.46,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin64_cout128","workload":{"axes":{"N":1,"W":256},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"7982c9a637f6534ba09decd16dd7209e","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv1d_kw3_sw1_dw1_p1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:20.972928+00:00","log":"","correctness":{"max_relative_error":0.003380009657170449,"max_absolute_error":9.5367431640625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":460485,"p5_ns":461298,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1258103,"instructions":4644306,"ipc":3.6915,"cache_misses":21035.74,"reference_cycles":null}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin64_cout128","workload":{"axes":{"N":1,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"3fdfae5415725eccab553e5efd0fd198","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:29.483757+00:00","log":"","correctness":{"max_relative_error":0.0007641572907497065,"max_absolute_error":6.67572021484375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1840123,"p5_ns":1840499,"reference_min_ns":188531,"speedup":0.11704896008202946,"repeat":50,"warmup":5,"cycles":4790474,"instructions":19413040,"ipc":4.0524,"cache_misses":1840.08,"reference_cycles":560720}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin64_cout128","workload":{"axes":{"N":1,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"ce3f7d5d325c5d7c93b2a47d7dea01ba","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:29.483757+00:00","log":"","correctness":{"max_relative_error":0.00022735084887613583,"max_absolute_error":5.7220458984375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1049453,"p5_ns":1049660,"reference_min_ns":111952,"speedup":0.12303538449038033,"repeat":50,"warmup":5,"cycles":2732214,"instructions":11069771,"ipc":4.0516,"cache_misses":1039.7,"reference_cycles":336159}}} -{"definition":"conv1d_kw3_sw1_dw1_p1_cin64_cout128","workload":{"axes":{"N":1,"W":256},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"7982c9a637f6534ba09decd16dd7209e","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv1d_kw3_sw1_dw1_p1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:29.483757+00:00","log":"","correctness":{"max_relative_error":0.019453202402575958,"max_absolute_error":7.152557373046875e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":4304372,"p5_ns":4307664,"reference_min_ns":460485,"speedup":0.11221374236442648,"repeat":50,"warmup":5,"cycles":11211666,"instructions":44446755,"ipc":3.9643,"cache_misses":759935.22,"reference_cycles":1258103}}} diff --git a/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024.jsonl b/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024.jsonl deleted file mode 100644 index 18cf89772a35cd8497a633e26f37b2e5b8aceaaf..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"46b33064e41459439fe192157458db95","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:12.588168+00:00","log":"","correctness":{"max_relative_error":0.056987185319851245,"max_absolute_error":0.0000762939453125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":8962392,"p5_ns":8987228,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":23461282,"instructions":59635350,"ipc":2.5419,"cache_misses":110686.44,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"882b121891315376bd9c3827d12370e4","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:12.588168+00:00","log":"","correctness":{"max_relative_error":0.021315848473275063,"max_absolute_error":0.000057220458984375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2445569,"p5_ns":2451327,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":6571634,"instructions":15974326,"ipc":2.4308,"cache_misses":47004.98,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"77361d3aaef253c59d648320dd79e609","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:12.588168+00:00","log":"","correctness":{"max_relative_error":0.24309313627099802,"max_absolute_error":0.000072479248046875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":36650405,"p5_ns":36676296,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":95459131,"instructions":235144251,"ipc":2.4633,"cache_misses":371355.9,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"46b33064e41459439fe192157458db95","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:44.626440+00:00","log":"","correctness":{"max_relative_error":0.056987185319851245,"max_absolute_error":0.0000762939453125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":159656289,"p5_ns":159666985,"reference_min_ns":8962392,"speedup":0.05665216061030853,"repeat":50,"warmup":5,"cycles":414128636,"instructions":2058980589,"ipc":4.9718,"cache_misses":76333110.3,"reference_cycles":23461282}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"882b121891315376bd9c3827d12370e4","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:44.626440+00:00","log":"","correctness":{"max_relative_error":0.021315848473275063,"max_absolute_error":0.000057220458984375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":39942257,"p5_ns":39955054,"reference_min_ns":2445569,"speedup":0.06333021823563557,"repeat":50,"warmup":5,"cycles":103767746,"instructions":514801511,"ipc":4.9611,"cache_misses":28969106.44,"reference_cycles":6571634}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"77361d3aaef253c59d648320dd79e609","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:44.626440+00:00","log":"","correctness":{"max_relative_error":0.24309313627099802,"max_absolute_error":0.000072479248046875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":710502294,"p5_ns":713127038,"reference_min_ns":36650405,"speedup":0.05138009023745748,"repeat":42,"warmup":5,"cycles":1857901194,"instructions":8236125151,"ipc":4.433,"cache_misses":278363263.1904762,"reference_cycles":95459131}}} diff --git a/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512.jsonl b/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512.jsonl deleted file mode 100644 index c7fb435f321ce9062317493079867d13849d1ea3..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"d3af4ea9ebca51fc9664b5f13d11c5b5","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:18.730712+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2120555,"p5_ns":2134431,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":5625066,"instructions":15346110,"ipc":2.7282,"cache_misses":50084.6,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"90d0153e72ba5f539e78f061bcde1002","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:18.730712+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":9174588,"p5_ns":9177520,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":24018106,"instructions":60989809,"ipc":2.5393,"cache_misses":204253.92,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512","workload":{"axes":{"N":1,"H":112,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"ee4b3270a4ed549b8c4ac9e46a5377b6","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:18.730712+00:00","log":"","correctness":{"max_relative_error":3.953834295272827,"max_absolute_error":6.67572021484375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":36259449,"p5_ns":36298141,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":94804944,"instructions":243625302,"ipc":2.5698,"cache_misses":888622.72,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"d3af4ea9ebca51fc9664b5f13d11c5b5","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:38:30.474852+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":33512140,"p5_ns":33513650,"reference_min_ns":2120555,"speedup":0.06461115624586135,"repeat":50,"warmup":5,"cycles":87060290,"instructions":519043754,"ipc":5.9619,"cache_misses":3240484.38,"reference_cycles":5625066}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"90d0153e72ba5f539e78f061bcde1002","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:38:30.474852+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":150268680,"p5_ns":150334096,"reference_min_ns":9174588,"speedup":0.06157707304891727,"repeat":50,"warmup":5,"cycles":390049491,"instructions":2075926163,"ipc":5.3222,"cache_misses":12966966.92,"reference_cycles":24018106}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512","workload":{"axes":{"N":1,"H":112,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"ee4b3270a4ed549b8c4ac9e46a5377b6","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:38:30.474852+00:00","log":"","correctness":{"max_relative_error":3.953834295272827,"max_absolute_error":6.67572021484375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1557573582,"p5_ns":1558123573,"reference_min_ns":36259449,"speedup":0.02332928407789455,"repeat":20,"warmup":5,"cycles":4063774254,"instructions":8310185676,"ipc":2.0449,"cache_misses":829741285.55,"reference_cycles":94804944}}} diff --git a/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024.jsonl b/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024.jsonl deleted file mode 100644 index c6304c170fba2179365bff77839a24b6896594c1..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"d285be818f9a5f05a0cd622e3d25e1a1","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:23.799712+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2108771,"p5_ns":2117647,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":5607313,"instructions":15221865,"ipc":2.7146,"cache_misses":36277.56,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"931d27e9dadb56f0af154539d71cfe33","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:23.799712+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":8421758,"p5_ns":8447847,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":22126696,"instructions":59738483,"ipc":2.6998,"cache_misses":112304.16,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"76b82d9b58f650d99a586c8feae88d3d","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:23.799712+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":34335577,"p5_ns":34359949,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":89348365,"instructions":237943471,"ipc":2.6631,"cache_misses":424371.5,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"d285be818f9a5f05a0cd622e3d25e1a1","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:39:21.808369+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":36677591,"p5_ns":36683477,"reference_min_ns":2108771,"speedup":0.05886915340783083,"repeat":50,"warmup":5,"cycles":95250444,"instructions":516662867,"ipc":5.4243,"cache_misses":3233881.6,"reference_cycles":5607313}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"931d27e9dadb56f0af154539d71cfe33","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:39:21.808369+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":147256713,"p5_ns":147270275,"reference_min_ns":8421758,"speedup":0.05793594559724517,"repeat":50,"warmup":5,"cycles":381916542,"instructions":2066268231,"ipc":5.4103,"cache_misses":12835310.08,"reference_cycles":22126696}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"76b82d9b58f650d99a586c8feae88d3d","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:39:21.808369+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":679200722,"p5_ns":679752888,"reference_min_ns":34335577,"speedup":0.05037724959305615,"repeat":44,"warmup":5,"cycles":1773585611,"instructions":8265098852,"ipc":4.6601,"cache_misses":113447662.68181819,"reference_cycles":89348365}}} diff --git a/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048.jsonl b/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048.jsonl deleted file mode 100644 index 20467044657def694a80a3ee900a8e828e27e5ba..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"2ad67410f0d0552f8f6ae305f25cefe6","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:28.671372+00:00","log":"","correctness":{"max_relative_error":0.00031587725987711185,"max_absolute_error":0.0000247955322265625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2323694,"p5_ns":2334590,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":6362293,"instructions":16087466,"ipc":2.5286,"cache_misses":55251.06,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"ba88bbd3da315b109da895c94572c808","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:28.671372+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":8322725,"p5_ns":8336107,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":21842698,"instructions":59936596,"ipc":2.744,"cache_misses":95139.0,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"1fc676a886ba582789004ddf0cbb2e2e","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:28.671372+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":34506189,"p5_ns":34517268,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":89770228,"instructions":235797748,"ipc":2.6267,"cache_misses":290494.72,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"2ad67410f0d0552f8f6ae305f25cefe6","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:40:06.576518+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":39269701,"p5_ns":39282045,"reference_min_ns":2323694,"speedup":0.06234415536349545,"repeat":50,"warmup":5,"cycles":102051154,"instructions":515489054,"ipc":5.0513,"cache_misses":3336718.8,"reference_cycles":6362293}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"ba88bbd3da315b109da895c94572c808","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:40:06.576518+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":155172042,"p5_ns":155190757,"reference_min_ns":8322725,"speedup":0.05423155618066421,"repeat":50,"warmup":5,"cycles":402767310,"instructions":2061522572,"ipc":5.1184,"cache_misses":13212428.18,"reference_cycles":21842698}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"1fc676a886ba582789004ddf0cbb2e2e","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:40:06.576518+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":654216226,"p5_ns":655778011,"reference_min_ns":34506189,"speedup":0.052702658218401324,"repeat":46,"warmup":5,"cycles":1703333969,"instructions":8245645899,"ipc":4.8409,"cache_misses":50754901.52173913,"reference_cycles":89770228}}} diff --git a/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256.jsonl b/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256.jsonl deleted file mode 100644 index 58a0f9fc54b7524ae41b03f16898ba6113270016..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"d604a2a8c81a52d1a6e45f2c3543d265","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:33.553008+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2254193,"p5_ns":2265519,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":6056384,"instructions":16025326,"ipc":2.646,"cache_misses":86342.56,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"751bd86036f055c2857b01ef4f11ee36","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:33.553008+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":543445,"p5_ns":547530,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1482784,"instructions":4055047,"ipc":2.7348,"cache_misses":19861.98,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256","workload":{"axes":{"N":1,"H":112,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"406890d52a18522aaeaa73fa1509f726","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:33.553008+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":9545950,"p5_ns":9594429,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":25474586,"instructions":63925901,"ipc":2.5094,"cache_misses":386245.22,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"d604a2a8c81a52d1a6e45f2c3543d265","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:40:51.909666+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":33807478,"p5_ns":33814373,"reference_min_ns":2254193,"speedup":0.06903288394513875,"repeat":50,"warmup":5,"cycles":87731870,"instructions":523858711,"ipc":5.9711,"cache_misses":3270124.28,"reference_cycles":6056384}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"751bd86036f055c2857b01ef4f11ee36","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:40:51.909666+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":7808865,"p5_ns":7809683,"reference_min_ns":543445,"speedup":0.07305324164694565,"repeat":50,"warmup":5,"cycles":20297306,"instructions":131007916,"ipc":6.4544,"cache_misses":815639.02,"reference_cycles":1482784}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256","workload":{"axes":{"N":1,"H":112,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"406890d52a18522aaeaa73fa1509f726","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:40:51.909666+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":155778255,"p5_ns":156134302,"reference_min_ns":9545950,"speedup":0.062414103873329595,"repeat":50,"warmup":5,"cycles":408154318,"instructions":2095241587,"ipc":5.1335,"cache_misses":16973155.32,"reference_cycles":25474586}}} diff --git a/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64.jsonl b/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64.jsonl deleted file mode 100644 index a580c94f8ba578576826f7605aa9aa9da67e2654..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64","workload":{"axes":{"N":1,"H":224,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"e6104e9e54285584a3e57d02298e419b","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:36.581064+00:00","log":"","correctness":{"max_relative_error":0.11361932754516602,"max_absolute_error":3.814697265625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":10632027,"p5_ns":10690056,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":28382110,"instructions":65759373,"ipc":2.3169,"cache_misses":751859.5,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"6841135778555a6db3b20ef4cd79cba4","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:36.581064+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":590143,"p5_ns":591460,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1604667,"instructions":4147660,"ipc":2.5847,"cache_misses":28731.46,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"6a664f25918c57fcac9147077f7e80fa","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:36.581064+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":29435,"p5_ns":31119,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":106961,"instructions":297883,"ipc":2.785,"cache_misses":2695.86,"reference_cycles":null}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64","workload":{"axes":{"N":1,"H":224,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"e6104e9e54285584a3e57d02298e419b","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:41:03.267531+00:00","log":"","correctness":{"max_relative_error":0.11361932754516602,"max_absolute_error":3.814697265625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":363572051,"p5_ns":363919911,"reference_min_ns":10632027,"speedup":0.030004426190719618,"repeat":50,"warmup":5,"cycles":945930771,"instructions":2096622295,"ipc":2.2165,"cache_misses":205339026.62,"reference_cycles":28382110}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"6841135778555a6db3b20ef4cd79cba4","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:41:03.267531+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":8467799,"p5_ns":8472924,"reference_min_ns":590143,"speedup":0.07285781496158983,"repeat":50,"warmup":5,"cycles":22024638,"instructions":130971137,"ipc":5.9466,"cache_misses":817759.86,"reference_cycles":1604667}}} -{"definition":"conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":0},"pad_left":{"type":"scalar","value":0},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"6a664f25918c57fcac9147077f7e80fa","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:41:03.267531+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":487340,"p5_ns":487471,"reference_min_ns":29435,"speedup":0.08394483676586233,"repeat":50,"warmup":5,"cycles":1274182,"instructions":8203277,"ipc":6.4381,"cache_misses":2769.56,"reference_cycles":106961}}} diff --git a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256.jsonl b/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256.jsonl deleted file mode 100644 index 975465c915ac922b161adf877327c6512f11e0cb..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"c6d116cfeff357a29d437f1f30f6fe1c","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:39.585319+00:00","log":"","correctness":{"max_relative_error":0.42028985507246375,"max_absolute_error":0.000026702880859375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":23206520,"p5_ns":23214373,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":60287789,"instructions":150471840,"ipc":2.4959,"cache_misses":11499457.46,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256","workload":{"axes":{"N":1,"H":40,"W":40},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"16f1968aecf05c6b9a34dc071ef248ac","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:39.585319+00:00","log":"","correctness":{"max_relative_error":0.10810810810810811,"max_absolute_error":0.000031948089599609375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":40908143,"p5_ns":40925312,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":106165806,"instructions":305915670,"ipc":2.8815,"cache_misses":11113787.36,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"a95305e76c5a57d09c094f3cc98493fb","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:39.585319+00:00","log":"","correctness":{"max_relative_error":0.21951219512195122,"max_absolute_error":0.0000286102294921875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":81293104,"p5_ns":81322415,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":211085218,"instructions":598504530,"ipc":2.8354,"cache_misses":23804385.74,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"c6d116cfeff357a29d437f1f30f6fe1c","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:41:24.451816+00:00","log":"","correctness":{"max_relative_error":0.08438254093778307,"max_absolute_error":0.000072479248046875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":175589351,"p5_ns":175661031,"reference_min_ns":23206520,"speedup":0.13200075093205932,"repeat":50,"warmup":5,"cycles":456723076,"instructions":1726465982,"ipc":3.7801,"cache_misses":229639.16,"reference_cycles":60287789}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256","workload":{"axes":{"N":1,"H":40,"W":40},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"16f1968aecf05c6b9a34dc071ef248ac","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:41:24.451816+00:00","log":"","correctness":{"max_relative_error":0.4362402220030089,"max_absolute_error":0.00008392333984375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":361028285,"p5_ns":361030977,"reference_min_ns":40908143,"speedup":0.11335373077117075,"repeat":50,"warmup":5,"cycles":936588547,"instructions":3540414497,"ipc":3.7801,"cache_misses":2475529.56,"reference_cycles":106165806}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"a95305e76c5a57d09c094f3cc98493fb","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:41:24.451816+00:00","log":"","correctness":{"max_relative_error":0.1631699283566095,"max_absolute_error":0.00008392333984375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":717199933,"p5_ns":717599822,"reference_min_ns":81293104,"speedup":0.11336186407847428,"repeat":42,"warmup":5,"cycles":1862047874,"instructions":6961628440,"ipc":3.7387,"cache_misses":52193079.571428575,"reference_cycles":211085218}}} diff --git a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512.jsonl b/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512.jsonl deleted file mode 100644 index 98e349ca35a75afc1b9bf05e41d89011169e2d0b..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"9016e988b6fe5ebb83894c247f5f6ce6","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:50.087657+00:00","log":"","correctness":{"max_relative_error":0.10909090909090909,"max_absolute_error":0.000041961669921875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":28218085,"p5_ns":28235331,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":73426530,"instructions":151691752,"ipc":2.0659,"cache_misses":17958804.24,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512","workload":{"axes":{"N":1,"H":20,"W":20},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"a3653f2e63885182a71b720d78a58ee0","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:50.087657+00:00","log":"","correctness":{"max_relative_error":0.051771117166212535,"max_absolute_error":0.000049591064453125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":60937490,"p5_ns":60957737,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":158149350,"instructions":305691735,"ipc":1.9329,"cache_misses":47985378.18,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"66cd580dc7e055ebb8582f75728bf521","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:09:50.087657+00:00","log":"","correctness":{"max_relative_error":0.0538627269929209,"max_absolute_error":0.000049591064453125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":120359380,"p5_ns":120388638,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":312333862,"instructions":595458423,"ipc":1.9065,"cache_misses":96600604.08,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"9016e988b6fe5ebb83894c247f5f6ce6","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:42:29.280700+00:00","log":"","correctness":{"max_relative_error":0.10714152943004261,"max_absolute_error":0.000118255615234375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":177776227,"p5_ns":177815198,"reference_min_ns":28218085,"speedup":0.1590944221743497,"repeat":50,"warmup":5,"cycles":461527997,"instructions":1695816185,"ipc":3.6744,"cache_misses":3851875.82,"reference_cycles":73426530}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512","workload":{"axes":{"N":1,"H":20,"W":20},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"a3653f2e63885182a71b720d78a58ee0","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:42:29.280700+00:00","log":"","correctness":{"max_relative_error":0.07098970751021798,"max_absolute_error":0.0001678466796875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":363073931,"p5_ns":363135950,"reference_min_ns":60937490,"speedup":0.16783125641485613,"repeat":50,"warmup":5,"cycles":942311661,"instructions":3494336990,"ipc":3.7083,"cache_misses":9880948.3,"reference_cycles":158149350}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"66cd580dc7e055ebb8582f75728bf521","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:42:29.280700+00:00","log":"","correctness":{"max_relative_error":0.1439349574768134,"max_absolute_error":0.000141143798828125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":713246367,"p5_ns":713261864,"reference_min_ns":120359380,"speedup":0.1688442503407981,"repeat":43,"warmup":5,"cycles":1849834160,"instructions":6893153817,"ipc":3.7264,"cache_misses":17412478.93023256,"reference_cycles":312333862}}} diff --git a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64.jsonl b/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64.jsonl deleted file mode 100644 index c49ca00e19ad51ec5a4643a6268962ce0f3488a5..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64","workload":{"axes":{"N":1,"H":224,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"4932ca9d15cd5ab58f30158b896aa9bb","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:10:04.183406+00:00","log":"","correctness":{"max_relative_error":0.45029530396964446,"max_absolute_error":1.9073486328125e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":10127593,"p5_ns":10184487,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":26735845,"instructions":142346666,"ipc":5.3242,"cache_misses":369020.82,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64","workload":{"axes":{"N":1,"H":300,"W":300},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"1828114d919e55168d17e337ca88badb","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:10:04.183406+00:00","log":"","correctness":{"max_relative_error":4.467085001240753,"max_absolute_error":2.384185791015625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":18297372,"p5_ns":18317785,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":48032767,"instructions":255267218,"ipc":5.3144,"cache_misses":676535.78,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64","workload":{"axes":{"N":1,"H":416,"W":416},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"7df9d89e21dc5806ac5f63972b145252","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:10:04.183406+00:00","log":"","correctness":{"max_relative_error":0.3692890572765768,"max_absolute_error":2.384185791015625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":51815680,"p5_ns":51918743,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":136870952,"instructions":552427612,"ipc":4.0361,"cache_misses":1702899.9,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64","workload":{"axes":{"N":1,"H":224,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"4932ca9d15cd5ab58f30158b896aa9bb","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:43:34.782259+00:00","log":"","correctness":{"max_relative_error":0.7176699615686031,"max_absolute_error":1.9073486328125e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":42330458,"p5_ns":42345649,"reference_min_ns":10127593,"speedup":0.24314215070500947,"repeat":50,"warmup":5,"cycles":109959729,"instructions":734875690,"ipc":6.6831,"cache_misses":43757.22,"reference_cycles":26735845}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64","workload":{"axes":{"N":1,"H":300,"W":300},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"1828114d919e55168d17e337ca88badb","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:43:34.782259+00:00","log":"","correctness":{"max_relative_error":18.3755500106055,"max_absolute_error":2.384185791015625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":75909422,"p5_ns":75951311,"reference_min_ns":18297372,"speedup":0.2438285711089098,"repeat":50,"warmup":5,"cycles":196994006,"instructions":1318503844,"ipc":6.6931,"cache_misses":88289.1,"reference_cycles":48032767}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64","workload":{"axes":{"N":1,"H":416,"W":416},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"7df9d89e21dc5806ac5f63972b145252","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:43:34.782259+00:00","log":"","correctness":{"max_relative_error":0.3772814174089783,"max_absolute_error":2.384185791015625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":145879204,"p5_ns":145895744,"reference_min_ns":51815680,"speedup":0.36173164594732937,"repeat":50,"warmup":5,"cycles":378377047,"instructions":2535862438,"ipc":6.7019,"cache_misses":137421.34,"reference_cycles":136870952}}} diff --git a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024.jsonl b/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024.jsonl deleted file mode 100644 index 4c0ca09815615b40074d95fcc6472f3fb510ed99..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"8038056ca85a5fccb9709a05a86f884b","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:10:11.233194+00:00","log":"","correctness":{"max_relative_error":0.06493506493506493,"max_absolute_error":0.00008392333984375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":31252618,"p5_ns":31270999,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":81573743,"instructions":161780212,"ipc":1.9832,"cache_misses":16886529.92,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"b868d1dad8205e339c0e8567d69b1dab","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:10:11.233194+00:00","log":"","correctness":{"max_relative_error":0.140625,"max_absolute_error":0.00009918212890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":123440751,"p5_ns":123533463,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":320404413,"instructions":603152186,"ipc":1.8825,"cache_misses":90397987.56,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024","workload":{"axes":{"N":1,"H":20,"W":20},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"5afd92a395e35213a46c0117ae3c9ccd","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:10:11.233194+00:00","log":"","correctness":{"max_relative_error":0.0672225117248567,"max_absolute_error":0.000102996826171875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":243304477,"p5_ns":243406636,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":630712975,"instructions":1215715715,"ipc":1.9275,"cache_misses":182990314.24,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"8038056ca85a5fccb9709a05a86f884b","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:43:49.923308+00:00","log":"","correctness":{"max_relative_error":0.017922247023809523,"max_absolute_error":0.000274658203125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":178409070,"p5_ns":178434616,"reference_min_ns":31252618,"speedup":0.17614674632274477,"repeat":50,"warmup":5,"cycles":463101049,"instructions":1435457282,"ipc":3.0997,"cache_misses":3422625.56,"reference_cycles":81573743}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"b868d1dad8205e339c0e8567d69b1dab","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:43:49.923308+00:00","log":"","correctness":{"max_relative_error":0.19594780461084424,"max_absolute_error":0.000274658203125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":712837249,"p5_ns":713002835,"reference_min_ns":123440751,"speedup":0.17318049918681067,"repeat":43,"warmup":5,"cycles":1850118313,"instructions":5954828770,"ipc":3.2186,"cache_misses":15401234.860465117,"reference_cycles":320404413}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024","workload":{"axes":{"N":1,"H":20,"W":20},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"5afd92a395e35213a46c0117ae3c9ccd","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:43:49.923308+00:00","log":"","correctness":{"max_relative_error":0.10976582855653987,"max_absolute_error":0.00038909912109375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1458359456,"p5_ns":1458397903,"reference_min_ns":243304477,"speedup":0.16680775962520414,"repeat":21,"warmup":5,"cycles":3781076950,"instructions":12286916315,"ipc":3.2496,"cache_misses":72792156.90476191,"reference_cycles":630712975}}} diff --git a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128.jsonl b/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128.jsonl deleted file mode 100644 index 9d709539f49a125da6666a9e6ceea449ff71bd04..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"fe4e4809c710583bac68a0587e367d56","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:10:36.043899+00:00","log":"","correctness":{"max_relative_error":0.29119031607262946,"max_absolute_error":0.0000152587890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":16642984,"p5_ns":16652800,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":43285312,"instructions":152756570,"ipc":3.5291,"cache_misses":432230.7,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":80,"W":80},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"0c8a5295baa95c8ab9151706747aeb26","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:10:36.043899+00:00","log":"","correctness":{"max_relative_error":0.09278350515463918,"max_absolute_error":0.0000152587890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":36751474,"p5_ns":36798001,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":97100533,"instructions":311329573,"ipc":3.2063,"cache_misses":4066476.88,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":112,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"55596a1b8f9356198de62647e2df1dd0","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:10:36.043899+00:00","log":"","correctness":{"max_relative_error":0.6470588235294118,"max_absolute_error":0.0000152587890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":71956377,"p5_ns":72105491,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":189272639,"instructions":609772661,"ipc":3.2217,"cache_misses":2963010.86,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"fe4e4809c710583bac68a0587e367d56","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:45:14.571125+00:00","log":"","correctness":{"max_relative_error":0.4462053289971419,"max_absolute_error":0.0000400543212890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":172857109,"p5_ns":172940246,"reference_min_ns":16642984,"speedup":0.09644149899056546,"repeat":50,"warmup":5,"cycles":448824546,"instructions":1643903255,"ipc":3.6627,"cache_misses":802532.28,"reference_cycles":43285312}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":80,"W":80},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"0c8a5295baa95c8ab9151706747aeb26","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:45:14.571125+00:00","log":"","correctness":{"max_relative_error":0.15546668927694104,"max_absolute_error":0.00003814697265625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":359665600,"p5_ns":359901151,"reference_min_ns":36751474,"speedup":0.10396695885226212,"repeat":50,"warmup":5,"cycles":933955692,"instructions":3363485346,"ipc":3.6013,"cache_misses":40834157.34,"reference_cycles":97100533}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":112,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":1}},"uuid":"55596a1b8f9356198de62647e2df1dd0","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:45:14.571125+00:00","log":"","correctness":{"max_relative_error":0.9619907187480553,"max_absolute_error":0.00005340576171875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":711188584,"p5_ns":711412996,"reference_min_ns":71956377,"speedup":0.1025733257986986,"repeat":43,"warmup":5,"cycles":1845242294,"instructions":6603643969,"ipc":3.5787,"cache_misses":80464166.02325581,"reference_cycles":189272639}}} diff --git a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256.jsonl b/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256.jsonl deleted file mode 100644 index 2696fa32fe2257076865a34bb52a0803f5804629..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"43c5652744f65cc983f879b4edd5d7a2","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:10:45.495671+00:00","log":"","correctness":{"max_relative_error":2.0,"max_absolute_error":0.00002288818359375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":24696111,"p5_ns":24805348,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":65339863,"instructions":153726664,"ipc":2.3527,"cache_misses":12009170.1,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256","workload":{"axes":{"N":1,"H":64,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"3b6996fa809f59299e60079a141bb0e6","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:10:45.495671+00:00","log":"","correctness":{"max_relative_error":0.25,"max_absolute_error":0.000030517578125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":112684598,"p5_ns":112709167,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":292808493,"instructions":791552510,"ipc":2.7033,"cache_misses":40821451.92,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"bd4a9ef2bf55585aa51996f03681fe2b","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:10:45.495671+00:00","log":"","correctness":{"max_relative_error":1.4210526315789473,"max_absolute_error":0.000026702880859375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":94019262,"p5_ns":94297309,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":245302742,"instructions":606871321,"ipc":2.474,"cache_misses":46602719.08,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"43c5652744f65cc983f879b4edd5d7a2","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:46:19.589696+00:00","log":"","correctness":{"max_relative_error":1.5216149886449177,"max_absolute_error":0.0000705718994140625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":176262537,"p5_ns":176303340,"reference_min_ns":24696111,"speedup":0.14277045655454762,"repeat":50,"warmup":5,"cycles":457656749,"instructions":1725012432,"ipc":3.7692,"cache_misses":666490.58,"reference_cycles":65339863}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256","workload":{"axes":{"N":1,"H":64,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"3b6996fa809f59299e60079a141bb0e6","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:46:19.589696+00:00","log":"","correctness":{"max_relative_error":0.5055518430822036,"max_absolute_error":0.000095367431640625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":954433602,"p5_ns":955802101,"reference_min_ns":112684598,"speedup":0.11783502739297788,"repeat":32,"warmup":5,"cycles":2484901981,"instructions":9174262353,"ipc":3.692,"cache_misses":314519722.6875,"reference_cycles":292808493}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"bd4a9ef2bf55585aa51996f03681fe2b","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:46:19.589696+00:00","log":"","correctness":{"max_relative_error":0.46335575315687394,"max_absolute_error":0.0000743865966796875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":714534984,"p5_ns":714729053,"reference_min_ns":94019262,"speedup":0.13231947890748622,"repeat":42,"warmup":5,"cycles":1853867201,"instructions":7010008999,"ipc":3.7813,"cache_misses":20756705.35714286,"reference_cycles":245302742}}} diff --git a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256.jsonl b/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256.jsonl deleted file mode 100644 index 881f6436d83b96c9a947ec1538df5f9a72ebb3d3..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"234b7212a9735ef8ad346f2fb7155d5e","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:11:00.881831+00:00","log":"","correctness":{"max_relative_error":0.08,"max_absolute_error":0.000041961669921875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":15396853,"p5_ns":15416456,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":40657512,"instructions":79661070,"ipc":1.9593,"cache_misses":9071734.54,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256","workload":{"axes":{"N":1,"H":64,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"609adf66789f5e7d8855263f5a7d17c3","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:11:00.881831+00:00","log":"","correctness":{"max_relative_error":1.0,"max_absolute_error":0.00005340576171875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":268034939,"p5_ns":268266006,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":696347349,"instructions":1563069446,"ipc":2.2447,"cache_misses":130629642.38,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"f7240c8fff585c0e97c5ccb575b2e5d6","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:11:00.881831+00:00","log":"","correctness":{"max_relative_error":0.1945945945945946,"max_absolute_error":0.000049591064453125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":56259491,"p5_ns":56313210,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":147723351,"instructions":303607064,"ipc":2.0552,"cache_misses":37633410.8,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"234b7212a9735ef8ad346f2fb7155d5e","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:47:40.496776+00:00","log":"","correctness":{"max_relative_error":0.05069300640060241,"max_absolute_error":0.0001220703125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":88944749,"p5_ns":88949631,"reference_min_ns":15396853,"speedup":0.17605715769554584,"repeat":50,"warmup":5,"cycles":230933593,"instructions":834191803,"ipc":3.6123,"cache_misses":2615400.08,"reference_cycles":40657512}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256","workload":{"axes":{"N":1,"H":64,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"609adf66789f5e7d8855263f5a7d17c3","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:47:40.496776+00:00","log":"","correctness":{"max_relative_error":0.08176711396935418,"max_absolute_error":0.000156402587890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1958117403,"p5_ns":1958117403,"reference_min_ns":268034939,"speedup":0.13707665590623452,"repeat":16,"warmup":5,"cycles":5079984950,"instructions":18313853012,"ipc":3.6051,"cache_misses":622125287.75,"reference_cycles":696347349}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"f7240c8fff585c0e97c5ccb575b2e5d6","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:47:40.496776+00:00","log":"","correctness":{"max_relative_error":0.1207786766258446,"max_absolute_error":0.0001239776611328125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":356730191,"p5_ns":356744193,"reference_min_ns":56259491,"speedup":0.15966479225908173,"repeat":50,"warmup":5,"cycles":925209302,"instructions":3443266750,"ipc":3.7216,"cache_misses":10198278.6,"reference_cycles":147723351}}} diff --git a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512.jsonl b/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512.jsonl deleted file mode 100644 index ad699bb6dcfdfc2ddb606179b23be6c7555fffda..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"9697ab41be58545aaea7ee9706884e7c","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:11:22.448238+00:00","log":"","correctness":{"max_relative_error":0.006767451996971852,"max_absolute_error":0.00006771087646484375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":17375371,"p5_ns":17428795,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":46808694,"instructions":92220204,"ipc":1.9702,"cache_misses":6038820.18,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"2a97ba3c3d9a5ed49ff35f566b40b890","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:11:22.448238+00:00","log":"","correctness":{"max_relative_error":7.0,"max_absolute_error":0.000087738037109375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":61246704,"p5_ns":61288509,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":159224304,"instructions":314165820,"ipc":1.9731,"cache_misses":35670911.62,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"1901245960675de7b3b96791312f725c","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:11:22.448238+00:00","log":"","correctness":{"max_relative_error":0.07535641547861507,"max_absolute_error":0.000095367431640625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":245900898,"p5_ns":246030728,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":638652740,"instructions":1204122345,"ipc":1.8854,"cache_misses":180501976.68,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"9697ab41be58545aaea7ee9706884e7c","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:48:48.843751+00:00","log":"","correctness":{"max_relative_error":0.002074138703248923,"max_absolute_error":0.00019073486328125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":89174059,"p5_ns":89196050,"reference_min_ns":17375371,"speedup":0.20218077289875167,"repeat":50,"warmup":5,"cycles":231519018,"instructions":680831298,"ipc":2.9407,"cache_misses":2771865.12,"reference_cycles":46808694}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"2a97ba3c3d9a5ed49ff35f566b40b890","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:48:48.843751+00:00","log":"","correctness":{"max_relative_error":17.841352462768555,"max_absolute_error":0.0002593994140625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":356364440,"p5_ns":356425771,"reference_min_ns":61246704,"speedup":0.1721557871574201,"repeat":50,"warmup":5,"cycles":924884993,"instructions":2922321829,"ipc":3.1597,"cache_misses":15041119.94,"reference_cycles":159224304}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"1901245960675de7b3b96791312f725c","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:48:48.843751+00:00","log":"","correctness":{"max_relative_error":0.11386148730038546,"max_absolute_error":0.0003204345703125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1442312011,"p5_ns":1442926926,"reference_min_ns":245900898,"speedup":0.17075581645082177,"repeat":21,"warmup":5,"cycles":3740152185,"instructions":12115819101,"ipc":3.2394,"cache_misses":48110939.952380955,"reference_cycles":638652740}}} diff --git a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128.jsonl b/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128.jsonl deleted file mode 100644 index 24391fe0fd435e8c689fb069eafbcc9fed24b781..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"276c7d862158547796a480c8198faea6","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:11:43.071258+00:00","log":"","correctness":{"max_relative_error":0.0821917808219178,"max_absolute_error":0.0000133514404296875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":17478810,"p5_ns":17618116,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":46611049,"instructions":156597222,"ipc":3.3597,"cache_misses":578255.46,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128","workload":{"axes":{"N":1,"H":64,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"c8c1892ee91c56cd9261e095b28bfdea","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:11:43.071258+00:00","log":"","correctness":{"max_relative_error":0.22916666666666666,"max_absolute_error":0.000015020370483398438,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":22130674,"p5_ns":22305800,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":59540500,"instructions":204274909,"ipc":3.4309,"cache_misses":822480.82,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128","workload":{"axes":{"N":1,"H":32,"W":32},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"204259e7ba7659b99caddd1acd638b61","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:11:43.071258+00:00","log":"","correctness":{"max_relative_error":0.01982651796778191,"max_absolute_error":0.00001430511474609375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":5527678,"p5_ns":5560999,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":14720246,"instructions":51680508,"ipc":3.5108,"cache_misses":201228.34,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"276c7d862158547796a480c8198faea6","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:49:53.072899+00:00","log":"","correctness":{"max_relative_error":0.06786103682084517,"max_absolute_error":0.00003814697265625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":172486629,"p5_ns":172701432,"reference_min_ns":17478810,"speedup":0.10388827498761298,"repeat":50,"warmup":5,"cycles":448665155,"instructions":1656412220,"ipc":3.6919,"cache_misses":2084230.5,"reference_cycles":46611049}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128","workload":{"axes":{"N":1,"H":64,"W":64},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"c8c1892ee91c56cd9261e095b28bfdea","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:49:53.072899+00:00","log":"","correctness":{"max_relative_error":1.188565731048584,"max_absolute_error":0.00003814697265625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":230178609,"p5_ns":230392588,"reference_min_ns":22130674,"speedup":0.09962360630229905,"repeat":50,"warmup":5,"cycles":597654534,"instructions":2168032573,"ipc":3.6276,"cache_misses":82520400.86,"reference_cycles":59540500}}} -{"definition":"conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128","workload":{"axes":{"N":1,"H":32,"W":32},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"204259e7ba7659b99caddd1acd638b61","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:49:53.072899+00:00","log":"","correctness":{"max_relative_error":0.024835333652803594,"max_absolute_error":0.0000324249267578125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":56976090,"p5_ns":56989499,"reference_min_ns":5527678,"speedup":0.09940659307532279,"repeat":50,"warmup":5,"cycles":148081184,"instructions":534153191,"ipc":3.6072,"cache_misses":6644030.32,"reference_cycles":14720246}}} diff --git a/traces/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256.jsonl b/traces/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256.jsonl deleted file mode 100644 index 4dba16dc176a9af285f902de567f9b9ce81e30c8..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"3967ac3bf8c75474915b83340931a6d1","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:11:47.967846+00:00","log":"","correctness":{"max_relative_error":0.2174143935694839,"max_absolute_error":0.000080108642578125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":14064656,"p5_ns":14079350,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":36588450,"instructions":102320583,"ipc":2.7965,"cache_misses":50446.38,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"ecca723c219853bc988feb7398ba4679","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:11:47.967846+00:00","log":"","correctness":{"max_relative_error":0.011376780451866405,"max_absolute_error":0.000064849853515625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":4087660,"p5_ns":4106433,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":10740711,"instructions":29793043,"ipc":2.7738,"cache_misses":15674.32,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256","workload":{"axes":{"N":1,"H":112,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"d24edb3d10c1569abda4ed25bfe098db","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:11:47.967846+00:00","log":"","correctness":{"max_relative_error":0.16226658922560672,"max_absolute_error":0.0000705718994140625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":58966909,"p5_ns":58973440,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":153562841,"instructions":402493590,"ipc":2.621,"cache_misses":998766.98,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"3967ac3bf8c75474915b83340931a6d1","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:50:19.073983+00:00","log":"","correctness":{"max_relative_error":0.2174143935694839,"max_absolute_error":0.000080108642578125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":181858827,"p5_ns":181901182,"reference_min_ns":14064656,"speedup":0.0776264519519558,"repeat":50,"warmup":5,"cycles":471339976,"instructions":2078336700,"ipc":4.4094,"cache_misses":14665998.36,"reference_cycles":36588450}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"ecca723c219853bc988feb7398ba4679","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:50:19.073983+00:00","log":"","correctness":{"max_relative_error":0.011376780451866405,"max_absolute_error":0.000064849853515625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":44644137,"p5_ns":44650685,"reference_min_ns":4087660,"speedup":0.09264989175732694,"repeat":50,"warmup":5,"cycles":115927939,"instructions":516034589,"ipc":4.4513,"cache_misses":168413.18,"reference_cycles":10740711}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256","workload":{"axes":{"N":1,"H":112,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"d24edb3d10c1569abda4ed25bfe098db","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:50:19.073983+00:00","log":"","correctness":{"max_relative_error":0.16226658922560672,"max_absolute_error":0.0000705718994140625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":777684701,"p5_ns":777755749,"reference_min_ns":58966909,"speedup":0.07593540167509122,"repeat":39,"warmup":5,"cycles":2022282593,"instructions":8342358510,"ipc":4.1252,"cache_misses":328671857.5897436,"reference_cycles":153562841}}} diff --git a/traces/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512.jsonl b/traces/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512.jsonl deleted file mode 100644 index f5b870bb1c2bfd969c703c80824aa5a8bb443150..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"444bbc8eec2e5364afce7aa74342ed2d","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:11:54.653052+00:00","log":"","correctness":{"max_relative_error":0.04525315890919475,"max_absolute_error":0.000171661376953125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":16385960,"p5_ns":16413776,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":42640430,"instructions":118327674,"ipc":2.775,"cache_misses":35636.82,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"a559b13ebc97587786f71ee44de4bc32","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:11:54.653052+00:00","log":"","correctness":{"max_relative_error":0.025774818978658538,"max_absolute_error":0.00016021728515625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":5484726,"p5_ns":5493407,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":14379959,"instructions":38819403,"ipc":2.6995,"cache_misses":26700.14,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"d627a5a78bc25c4fbadb24438974c88d","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:11:54.653052+00:00","log":"","correctness":{"max_relative_error":3.116056341873972,"max_absolute_error":0.00016021728515625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":55950400,"p5_ns":55959717,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":145187653,"instructions":407228796,"ipc":2.8048,"cache_misses":153340.16,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"444bbc8eec2e5364afce7aa74342ed2d","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:51:07.165279+00:00","log":"","correctness":{"max_relative_error":0.04525315890919475,"max_absolute_error":0.000171661376953125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":179665575,"p5_ns":179691878,"reference_min_ns":16385960,"speedup":0.09153912781775991,"repeat":50,"warmup":5,"cycles":465816433,"instructions":2058622915,"ipc":4.4194,"cache_misses":4985260.98,"reference_cycles":42640430}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"a559b13ebc97587786f71ee44de4bc32","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:51:07.165279+00:00","log":"","correctness":{"max_relative_error":0.03341950966047394,"max_absolute_error":0.00016021728515625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":44835027,"p5_ns":44849238,"reference_min_ns":5484726,"speedup":0.12353716396367355,"repeat":50,"warmup":5,"cycles":116401887,"instructions":507592485,"ipc":4.3607,"cache_misses":964108.54,"reference_cycles":14379959}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"d627a5a78bc25c4fbadb24438974c88d","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:51:07.165279+00:00","log":"","correctness":{"max_relative_error":3.116056341873972,"max_absolute_error":0.00016021728515625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":756837895,"p5_ns":761801649,"reference_min_ns":55950400,"speedup":0.07317398044024105,"repeat":40,"warmup":5,"cycles":1984143163,"instructions":8292218362,"ipc":4.1792,"cache_misses":120687804.75,"reference_cycles":145187653}}} diff --git a/traces/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024.jsonl b/traces/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024.jsonl deleted file mode 100644 index 744833f6c971154084397ca1fc8f5f7283fb92e8..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"57dc8b5b69225a3f8ec9b9781adad2cd","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:01.369399+00:00","log":"","correctness":{"max_relative_error":0.102874755859375,"max_absolute_error":0.00020599365234375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":21804434,"p5_ns":21806173,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":56704900,"instructions":154495523,"ipc":2.7246,"cache_misses":88274.82,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"6359a7d0c2015fbc89a53a1179242bd9","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:01.369399+00:00","log":"","correctness":{"max_relative_error":0.002528246472310969,"max_absolute_error":0.00018310546875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":6198449,"p5_ns":6214892,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":16334204,"instructions":48621326,"ipc":2.9767,"cache_misses":33743.42,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"4e2093be1531531180466644db45791d","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:01.369399+00:00","log":"","correctness":{"max_relative_error":0.17307692307692307,"max_absolute_error":0.00026702880859375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":65088820,"p5_ns":65111286,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":168856471,"instructions":471710367,"ipc":2.7936,"cache_misses":105387.36,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"57dc8b5b69225a3f8ec9b9781adad2cd","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:51:55.257571+00:00","log":"","correctness":{"max_relative_error":0.102874755859375,"max_absolute_error":0.00020599365234375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":179241331,"p5_ns":179273891,"reference_min_ns":21804434,"speedup":0.1219433736295838,"repeat":50,"warmup":5,"cycles":465010097,"instructions":1847617229,"ipc":3.9733,"cache_misses":6276414.5,"reference_cycles":56704900}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"6359a7d0c2015fbc89a53a1179242bd9","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:51:55.257571+00:00","log":"","correctness":{"max_relative_error":0.002528246472310969,"max_absolute_error":0.00018310546875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":58684596,"p5_ns":58700842,"reference_min_ns":6198449,"speedup":0.10722549053260437,"repeat":50,"warmup":5,"cycles":152335083,"instructions":561923060,"ipc":3.6887,"cache_misses":1642233.68,"reference_cycles":16334204}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"4e2093be1531531180466644db45791d","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:51:55.257571+00:00","log":"","correctness":{"max_relative_error":0.11427955781882591,"max_absolute_error":0.00026702880859375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":728142610,"p5_ns":728660693,"reference_min_ns":65088820,"speedup":0.08942202986119763,"repeat":42,"warmup":5,"cycles":1888309528,"instructions":7504257358,"ipc":3.9741,"cache_misses":21982458.95238095,"reference_cycles":168856471}}} diff --git a/traces/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128.jsonl b/traces/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128.jsonl deleted file mode 100644 index e13882e277afca1d6e9840888c50eb396c20f049..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":112,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"a4aa78333cf056838a68700ae3aba650","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:09.088676+00:00","log":"","correctness":{"max_relative_error":0.1742967547792377,"max_absolute_error":0.000041961669921875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":14643979,"p5_ns":14666630,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":38320837,"instructions":101362657,"ipc":2.6451,"cache_misses":258749.48,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"c08c26b564285fbab074242c591ca462","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:09.088676+00:00","log":"","correctness":{"max_relative_error":0.10544743257410386,"max_absolute_error":0.00003528594970703125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":3535299,"p5_ns":3552608,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":9295912,"instructions":25848322,"ipc":2.7806,"cache_misses":17042.02,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":224,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"52f36d38efcb5616b0ab1fae6b74da15","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:09.088676+00:00","log":"","correctness":{"max_relative_error":0.1050962632702243,"max_absolute_error":0.000041961669921875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":59339350,"p5_ns":59585713,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":155021891,"instructions":402208456,"ipc":2.5945,"cache_misses":603759.16,"reference_cycles":null}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":112,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"a4aa78333cf056838a68700ae3aba650","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:52:43.909647+00:00","log":"","correctness":{"max_relative_error":0.1742967547792377,"max_absolute_error":0.000041961669921875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":183870523,"p5_ns":184144612,"reference_min_ns":14643979,"speedup":0.07992141435572941,"repeat":50,"warmup":5,"cycles":479481467,"instructions":2070392839,"ipc":4.318,"cache_misses":26106372.18,"reference_cycles":38320837}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"c08c26b564285fbab074242c591ca462","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:52:43.909647+00:00","log":"","correctness":{"max_relative_error":0.10544743257410386,"max_absolute_error":0.00003528594970703125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":44151159,"p5_ns":44165136,"reference_min_ns":3535299,"speedup":0.0811259742528889,"repeat":50,"warmup":5,"cycles":114586137,"instructions":515836093,"ipc":4.5017,"cache_misses":644049.68,"reference_cycles":9295912}}} -{"definition":"conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":224,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":1},"pad_left":{"type":"scalar","value":1},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"52f36d38efcb5616b0ab1fae6b74da15","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:52:43.909647+00:00","log":"","correctness":{"max_relative_error":0.1050962632702243,"max_absolute_error":0.000041961669921875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":779511243,"p5_ns":780473563,"reference_min_ns":59339350,"speedup":0.07656019504403702,"repeat":39,"warmup":5,"cycles":2024836678,"instructions":8295980815,"ipc":4.0971,"cache_misses":335068093.7948718,"reference_cycles":155021891}}} diff --git a/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128.jsonl b/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128.jsonl deleted file mode 100644 index debf22ad278b4e6c9421935b51a622dfbd741712..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"e0fc0a5d014f5201ac5d7f0fdde3eb82","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:15.837928+00:00","log":"","correctness":{"max_relative_error":0.004597893780814808,"max_absolute_error":0.000102996826171875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":5811424,"p5_ns":5827181,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":15168579,"instructions":52137251,"ipc":3.4372,"cache_misses":10016.7,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"280264e3f3da5baaad622ea522103d6c","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:15.837928+00:00","log":"","correctness":{"max_relative_error":0.035098522167487683,"max_absolute_error":0.00014495849609375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":19853512,"p5_ns":19855964,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":51568057,"instructions":161349260,"ipc":3.1289,"cache_misses":40278.5,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"f5af06f9d26c51bda23cdd491e025129","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:15.837928+00:00","log":"","correctness":{"max_relative_error":0.06222222222222222,"max_absolute_error":0.00014495849609375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":78977866,"p5_ns":78986342,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":204818644,"instructions":631042855,"ipc":3.081,"cache_misses":261812.48,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"e0fc0a5d014f5201ac5d7f0fdde3eb82","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:53:32.206277+00:00","log":"","correctness":{"max_relative_error":0.008115539952761123,"max_absolute_error":0.00016021728515625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":62037222,"p5_ns":62041403,"reference_min_ns":5811424,"speedup":0.09416507595345204,"repeat":50,"warmup":5,"cycles":161084976,"instructions":739545552,"ipc":4.591,"cache_misses":1069410.62,"reference_cycles":15168579}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"280264e3f3da5baaad622ea522103d6c","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:53:32.206277+00:00","log":"","correctness":{"max_relative_error":0.028971417546891226,"max_absolute_error":0.00020599365234375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":248171731,"p5_ns":248204118,"reference_min_ns":19853512,"speedup":0.08005370242569917,"repeat":50,"warmup":5,"cycles":644168295,"instructions":3046755366,"ipc":4.7298,"cache_misses":25721121.04,"reference_cycles":51568057}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"f5af06f9d26c51bda23cdd491e025129","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:53:32.206277+00:00","log":"","correctness":{"max_relative_error":0.19769958827806555,"max_absolute_error":0.000213623046875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":997068367,"p5_ns":997135699,"reference_min_ns":78977866,"speedup":0.0792275229321487,"repeat":31,"warmup":5,"cycles":2585195604,"instructions":12367865187,"ipc":4.7841,"cache_misses":237639652.4516129,"reference_cycles":204818644}}} diff --git a/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256.jsonl b/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256.jsonl deleted file mode 100644 index 2706238240a073f1289232b9b1b0a2b026ac165c..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"9181464a38255af69c799b55a88c0daf","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:24.030675+00:00","log":"","correctness":{"max_relative_error":0.05314086294416244,"max_absolute_error":0.00034332275390625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":23251314,"p5_ns":23274097,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":60444474,"instructions":208042546,"ipc":3.4419,"cache_misses":26258.96,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"948b6e106dbd5ee28e984de515cc1057","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:24.030675+00:00","log":"","correctness":{"max_relative_error":0.01444500760263558,"max_absolute_error":0.00018310546875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":7555609,"p5_ns":7557432,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":19662161,"instructions":78165437,"ipc":3.9754,"cache_misses":42423.02,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"7fa001a54070516fac7b2090f275ec5d","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:24.030675+00:00","log":"","correctness":{"max_relative_error":0.05293524535066099,"max_absolute_error":0.00035858154296875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":79448189,"p5_ns":79463018,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":205990647,"instructions":644350642,"ipc":3.1281,"cache_misses":88053.0,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"9181464a38255af69c799b55a88c0daf","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:54:26.710126+00:00","log":"","correctness":{"max_relative_error":0.03764450215847693,"max_absolute_error":0.00034332275390625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":248239935,"p5_ns":248257027,"reference_min_ns":23251314,"speedup":0.09381908454484647,"repeat":50,"warmup":5,"cycles":644266295,"instructions":2957240382,"ipc":4.5901,"cache_misses":7098054.08,"reference_cycles":60444474}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"948b6e106dbd5ee28e984de515cc1057","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:54:26.710126+00:00","log":"","correctness":{"max_relative_error":0.04549240264402771,"max_absolute_error":0.0002288818359375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":62134233,"p5_ns":62145354,"reference_min_ns":7555609,"speedup":0.12187823168841053,"repeat":50,"warmup":5,"cycles":161326274,"instructions":696765114,"ipc":4.319,"cache_misses":1001452.92,"reference_cycles":19662161}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"7fa001a54070516fac7b2090f275ec5d","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:54:26.710126+00:00","log":"","correctness":{"max_relative_error":0.1111762509382671,"max_absolute_error":0.000396728515625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":995967779,"p5_ns":996029809,"reference_min_ns":79448189,"speedup":0.07973481598796076,"repeat":31,"warmup":5,"cycles":2583446697,"instructions":12183547488,"ipc":4.716,"cache_misses":286561670.41935486,"reference_cycles":205990647}}} diff --git a/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64.jsonl b/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64.jsonl deleted file mode 100644 index ff7b65a24c218ea8c16d6f2b2c3c15f708404082..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64","workload":{"axes":{"N":1,"H":32,"W":32},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"2a88d188621e5806a1787c31d5ec1340","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:32.543582+00:00","log":"","correctness":{"max_relative_error":0.01304945054945055,"max_absolute_error":0.000030517578125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":3245977,"p5_ns":3261757,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":8506026,"instructions":26340019,"ipc":3.0966,"cache_misses":10075.36,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64","workload":{"axes":{"N":1,"H":112,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"8a49b5357f68530bb7a559d9d0b5c813","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:32.543582+00:00","log":"","correctness":{"max_relative_error":1.5,"max_absolute_error":0.000041961669921875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":40045751,"p5_ns":40065752,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":103986190,"instructions":313316830,"ipc":3.0131,"cache_misses":455449.6,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"8e3d6a0468ee5dcfbb8cff10cedaeb8b","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:32.543582+00:00","log":"","correctness":{"max_relative_error":0.11400651465798045,"max_absolute_error":0.000034332275390625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":9897023,"p5_ns":9900553,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":25736928,"instructions":79229920,"ipc":3.0785,"cache_misses":35905.08,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64","workload":{"axes":{"N":1,"H":32,"W":32},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"2a88d188621e5806a1787c31d5ec1340","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:55:21.184246+00:00","log":"","correctness":{"max_relative_error":0.017126397772149724,"max_absolute_error":0.0000457763671875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":40680467,"p5_ns":40684611,"reference_min_ns":3245977,"speedup":0.08050633160108328,"repeat":50,"warmup":5,"cycles":105656609,"instructions":500129951,"ipc":4.7335,"cache_misses":12708953.82,"reference_cycles":8506026}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64","workload":{"axes":{"N":1,"H":112,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"8a49b5357f68530bb7a559d9d0b5c813","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:55:21.184246+00:00","log":"","correctness":{"max_relative_error":2.374734115600586,"max_absolute_error":0.000057220458984375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":497542365,"p5_ns":497640025,"reference_min_ns":40045751,"speedup":0.08058720548215687,"repeat":50,"warmup":5,"cycles":1290356073,"instructions":6239861782,"ipc":4.8358,"cache_misses":5860076.3,"reference_cycles":103986190}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"8e3d6a0468ee5dcfbb8cff10cedaeb8b","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:55:21.184246+00:00","log":"","correctness":{"max_relative_error":0.2411084832637758,"max_absolute_error":0.000041961669921875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":124063896,"p5_ns":124071394,"reference_min_ns":9897023,"speedup":0.07989348300482958,"repeat":50,"warmup":5,"cycles":322140518,"instructions":1548573858,"ipc":4.8071,"cache_misses":484060.48,"reference_cycles":25736928}}} diff --git a/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32.jsonl b/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32.jsonl deleted file mode 100644 index eb88140d7ec2f3df2b77c443c9803ca277cf83b1..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32","workload":{"axes":{"N":1,"H":224,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"87e7aa8330035a74935427a1b75f4418","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:37.836503+00:00","log":"","correctness":{"max_relative_error":0.304510710201484,"max_absolute_error":6.67572021484375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":7879622,"p5_ns":7895524,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":20683096,"instructions":60704670,"ipc":2.935,"cache_misses":137992.24,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32","workload":{"axes":{"N":1,"H":300,"W":300},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"f743b4ceac355b03beca86485e0039ab","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:37.836503+00:00","log":"","correctness":{"max_relative_error":0.39854300653425406,"max_absolute_error":6.67572021484375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":14294527,"p5_ns":14324079,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":37447951,"instructions":108714624,"ipc":2.9031,"cache_misses":226236.54,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32","workload":{"axes":{"N":1,"H":416,"W":416},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"86e89ceefd595fbaade0fe508e9a144c","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:37.836503+00:00","log":"","correctness":{"max_relative_error":99.61821903787103,"max_absolute_error":6.198883056640625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":27593885,"p5_ns":27680086,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":72392499,"instructions":208773379,"ipc":2.8839,"cache_misses":398762.9,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32","workload":{"axes":{"N":1,"H":224,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"87e7aa8330035a74935427a1b75f4418","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:55:58.481014+00:00","log":"","correctness":{"max_relative_error":0.29045746281829915,"max_absolute_error":6.67572021484375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":81102030,"p5_ns":81121298,"reference_min_ns":7879622,"speedup":0.09820784407855648,"repeat":50,"warmup":5,"cycles":210605336,"instructions":1198993982,"ipc":5.6931,"cache_misses":126720.82,"reference_cycles":20683096}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32","workload":{"axes":{"N":1,"H":300,"W":300},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"f743b4ceac355b03beca86485e0039ab","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:55:58.481014+00:00","log":"","correctness":{"max_relative_error":0.7186666811372425,"max_absolute_error":7.62939453125e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":145503797,"p5_ns":145570630,"reference_min_ns":14294527,"speedup":0.0991133275916448,"repeat":50,"warmup":5,"cycles":377829621,"instructions":2152533299,"ipc":5.6971,"cache_misses":305382.06,"reference_cycles":37447951}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32","workload":{"axes":{"N":1,"H":416,"W":416},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"86e89ceefd595fbaade0fe508e9a144c","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:55:58.481014+00:00","log":"","correctness":{"max_relative_error":33.53940634595701,"max_absolute_error":6.67572021484375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":279828058,"p5_ns":279870745,"reference_min_ns":27593885,"speedup":0.09966293010166716,"repeat":50,"warmup":5,"cycles":726373376,"instructions":4142019622,"ipc":5.7023,"cache_misses":461131.6,"reference_cycles":72392499}}} diff --git a/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512.jsonl b/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512.jsonl deleted file mode 100644 index 4bf73dba64606adabf19700329d445da4fc2a628..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"dde30152afe35a56b216aad5fe01c0c0","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:43.092514+00:00","log":"","correctness":{"max_relative_error":0.0043819188191881915,"max_absolute_error":0.000335693359375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":30096091,"p5_ns":30108287,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":78239162,"instructions":311986815,"ipc":3.9876,"cache_misses":134722.66,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"558026cffc19544d8fe7a91c58635900","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:43.092514+00:00","log":"","correctness":{"max_relative_error":0.08618654073199528,"max_absolute_error":0.0005035400390625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":92884307,"p5_ns":92901929,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":241034871,"instructions":831242786,"ipc":3.4486,"cache_misses":83273.64,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"a579f9c2aaa65ecf95218723d97b36dd","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:12:43.092514+00:00","log":"","correctness":{"max_relative_error":1.6139914772727273,"max_absolute_error":0.000579833984375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":317811732,"p5_ns":317947231,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":824115878,"instructions":2575405414,"ipc":3.1251,"cache_misses":608564.76,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"dde30152afe35a56b216aad5fe01c0c0","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:56:27.135397+00:00","log":"","correctness":{"max_relative_error":0.009817538666109318,"max_absolute_error":0.00054931640625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":248372014,"p5_ns":248390355,"reference_min_ns":30096091,"speedup":0.12139036910364306,"repeat":50,"warmup":5,"cycles":644525283,"instructions":2786552389,"ipc":4.3234,"cache_misses":6221422.9,"reference_cycles":78239162}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"558026cffc19544d8fe7a91c58635900","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:56:27.135397+00:00","log":"","correctness":{"max_relative_error":0.05963026675029516,"max_absolute_error":0.00063323974609375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":992946347,"p5_ns":992967466,"reference_min_ns":92884307,"speedup":0.09356855964093579,"repeat":31,"warmup":5,"cycles":2576024168,"instructions":11827141812,"ipc":4.5912,"cache_misses":8112958.806451613,"reference_cycles":241034871}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"a579f9c2aaa65ecf95218723d97b36dd","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:56:27.135397+00:00","log":"","correctness":{"max_relative_error":2.1889894658868965,"max_absolute_error":0.00084686279296875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":4036013836,"p5_ns":4036013836,"reference_min_ns":317811732,"speedup":0.07880804960285079,"repeat":8,"warmup":5,"cycles":10457255092,"instructions":48727534967,"ipc":4.6597,"cache_misses":1135731972.75,"reference_cycles":824115878}}} diff --git a/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128.jsonl b/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128.jsonl deleted file mode 100644 index 10070041ed8e361cab94240734233b164c89d536..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"29a4240dc9a5526990b53de998985581","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:10.403835+00:00","log":"","correctness":{"max_relative_error":0.01636904761904762,"max_absolute_error":0.0000762939453125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":9948393,"p5_ns":9953028,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":25871548,"instructions":80771640,"ipc":3.122,"cache_misses":23004.04,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"da88be470bb4531fb6d8f4e538fd8a7a","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:10.403835+00:00","log":"","correctness":{"max_relative_error":0.0735042735042735,"max_absolute_error":0.0000820159912109375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":39515075,"p5_ns":39531683,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":102520277,"instructions":315842276,"ipc":3.0808,"cache_misses":142503.82,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"453ca9c5bb34524983db573cebd4ec38","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:10.403835+00:00","log":"","correctness":{"max_relative_error":0.03765060240963856,"max_absolute_error":0.000080108642578125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2908730,"p5_ns":2914722,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":7612318,"instructions":26107074,"ipc":3.4296,"cache_misses":7048.7,"reference_cycles":null}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"29a4240dc9a5526990b53de998985581","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:58:14.716169+00:00","log":"","correctness":{"max_relative_error":0.012041172602051608,"max_absolute_error":0.00009918212890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":123990010,"p5_ns":123991637,"reference_min_ns":9948393,"speedup":0.08036628281549305,"repeat":50,"warmup":5,"cycles":321920426,"instructions":1524259320,"ipc":4.7349,"cache_misses":454071.44,"reference_cycles":25871548}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"da88be470bb4531fb6d8f4e538fd8a7a","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:58:14.716169+00:00","log":"","correctness":{"max_relative_error":0.06057320243677856,"max_absolute_error":0.0001049041748046875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":497237469,"p5_ns":497247878,"reference_min_ns":39515075,"speedup":0.0794645612647471,"repeat":50,"warmup":5,"cycles":1290138338,"instructions":6187368471,"ipc":4.7959,"cache_misses":20528875.74,"reference_cycles":102520277}}} -{"definition":"conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":2},"pad_left":{"type":"scalar","value":2},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"453ca9c5bb34524983db573cebd4ec38","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:58:14.716169+00:00","log":"","correctness":{"max_relative_error":0.037143102575231485,"max_absolute_error":0.000087738037109375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":31000229,"p5_ns":31005514,"reference_min_ns":2908730,"speedup":0.09454304963116462,"repeat":50,"warmup":5,"cycles":80516950,"instructions":369997984,"ipc":4.5953,"cache_misses":74709.82,"reference_cycles":7612318}}} diff --git a/traces/conv2d/conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64.jsonl b/traces/conv2d/conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64.jsonl deleted file mode 100644 index 58b85e0aee7b5b32ab417a1744f75f05766f153d..0000000000000000000000000000000000000000 --- a/traces/conv2d/conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64","workload":{"axes":{"N":1,"H":224,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":3},"pad_left":{"type":"scalar","value":3},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"507902e2a1bd5519a852f6d67bd21b9a","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:15.658557+00:00","log":"","correctness":{"max_relative_error":0.13022088575777613,"max_absolute_error":0.0000133514404296875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":12114948,"p5_ns":12140837,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":31616050,"instructions":73175666,"ipc":2.3145,"cache_misses":213218.26,"reference_cycles":null}}} -{"definition":"conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64","workload":{"axes":{"N":1,"H":448,"W":448},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":3},"pad_left":{"type":"scalar","value":3},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"e6b3bd13646d5cf8acd98028da45bb89","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:15.658557+00:00","log":"","correctness":{"max_relative_error":11.169481850532822,"max_absolute_error":0.000011444091796875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":48642956,"p5_ns":48760216,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":127174981,"instructions":291152621,"ipc":2.2894,"cache_misses":526970.26,"reference_cycles":null}}} -{"definition":"conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64","workload":{"axes":{"N":1,"H":320,"W":320},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":3},"pad_left":{"type":"scalar","value":3},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"09aa974271d251c08962a4f569694bf7","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:13:15.658557+00:00","log":"","correctness":{"max_relative_error":0.04177757508518205,"max_absolute_error":0.00001239776611328125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":24769824,"p5_ns":24786788,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":64620915,"instructions":148862872,"ipc":2.3036,"cache_misses":346401.16,"reference_cycles":null}}} -{"definition":"conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64","workload":{"axes":{"N":1,"H":224,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":3},"pad_left":{"type":"scalar","value":3},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"507902e2a1bd5519a852f6d67bd21b9a","tags":{"from":"tests/ncnn/candidate/convolution.cpp"},"description":null},"solution":"reference-scalar_conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:58:51.437866+00:00","log":"","correctness":{"max_relative_error":0.2862794431708115,"max_absolute_error":0.0000133514404296875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":89880980,"p5_ns":89910182,"reference_min_ns":12114948,"speedup":0.1354330436664597,"repeat":50,"warmup":5,"cycles":233444137,"instructions":1293678643,"ipc":5.5417,"cache_misses":459757.48,"reference_cycles":31616050}}} -{"definition":"conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64","workload":{"axes":{"N":1,"H":448,"W":448},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":3},"pad_left":{"type":"scalar","value":3},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"e6b3bd13646d5cf8acd98028da45bb89","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:58:51.437866+00:00","log":"","correctness":{"max_relative_error":12.348656532789855,"max_absolute_error":0.0000133514404296875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":363994839,"p5_ns":364050126,"reference_min_ns":48642956,"speedup":0.13457290622380627,"repeat":50,"warmup":5,"cycles":945026637,"instructions":5191429362,"ipc":5.4934,"cache_misses":1376111.98,"reference_cycles":127174981}}} -{"definition":"conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64","workload":{"axes":{"N":1,"H":320,"W":320},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"pad_top":{"type":"scalar","value":3},"pad_left":{"type":"scalar","value":3},"activation_type":{"type":"scalar","value":0},"with_bias":{"type":"scalar","value":0}},"uuid":"09aa974271d251c08962a4f569694bf7","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:58:51.437866+00:00","log":"","correctness":{"max_relative_error":0.07939028374074272,"max_absolute_error":0.000011444091796875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":184480900,"p5_ns":184516653,"reference_min_ns":24769824,"speedup":0.1349238534127835,"repeat":50,"warmup":5,"cycles":478943592,"instructions":2645277041,"ipc":5.5231,"cache_misses":906369.78,"reference_cycles":64620915}}} diff --git a/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024.jsonl b/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024.jsonl deleted file mode 100644 index adb4a656289d3538773baaefb3bc47865f35e82b..0000000000000000000000000000000000000000 --- a/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"356687f67442549a958d0fb9ceae1ddd","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:23.298873+00:00","log":"","correctness":{"max_relative_error":0.0006138148059172062,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":110763,"p5_ns":111245,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":307334,"instructions":1475823,"ipc":4.802,"cache_misses":10689.34,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"f8455e9a94b2550b8c529b8745ef514e","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:23.298873+00:00","log":"","correctness":{"max_relative_error":0.0494475084587488,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":3057590,"p5_ns":3066931,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":8806179,"instructions":22379418,"ipc":2.5413,"cache_misses":215058.02,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"c8de676aa2bc562dbb92d24b6bbc98e0","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:23.298873+00:00","log":"","correctness":{"max_relative_error":0.09109174689830554,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":780779,"p5_ns":805683,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":2221962,"instructions":6970018,"ipc":3.1369,"cache_misses":55074.84,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"356687f67442549a958d0fb9ceae1ddd","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:30.026840+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":233789,"p5_ns":235445,"reference_min_ns":110763,"speedup":0.49309613428443994,"repeat":50,"warmup":5,"cycles":623274,"instructions":3607818,"ipc":5.7885,"cache_misses":514.82,"reference_cycles":307334}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"f8455e9a94b2550b8c529b8745ef514e","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:30.026840+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":16036395,"p5_ns":16040034,"reference_min_ns":3057590,"speedup":0.20964440356284444,"repeat":50,"warmup":5,"cycles":42005314,"instructions":247782892,"ipc":5.8988,"cache_misses":16150.94,"reference_cycles":8806179}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"c8de676aa2bc562dbb92d24b6bbc98e0","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:30.026840+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":4020163,"p5_ns":4045858,"reference_min_ns":780779,"speedup":0.21072976148814554,"repeat":50,"warmup":5,"cycles":10544130,"instructions":61266045,"ipc":5.8104,"cache_misses":3476.46,"reference_cycles":2221962}}} diff --git a/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128.jsonl b/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128.jsonl deleted file mode 100644 index bfa241e24eb85be58ae7eff16185f5ba3672752c..0000000000000000000000000000000000000000 --- a/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"2253a4b54d7158d98d03d77c8faccbb8","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:25.926851+00:00","log":"","correctness":{"max_relative_error":0.010683180087101908,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":354011,"p5_ns":356884,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":976840,"instructions":2822410,"ipc":2.8893,"cache_misses":15098.08,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"c56f2c7545b95fe39056ddd416b5fdd7","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:25.926851+00:00","log":"","correctness":{"max_relative_error":0.00928617652244472,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":91023,"p5_ns":91421,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":269649,"instructions":896595,"ipc":3.325,"cache_misses":5712.38,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"8bfd4f907cfe5518adcf3b18998f246d","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:25.926851+00:00","log":"","correctness":{"max_relative_error":0.0004626310497011229,"max_absolute_error":4.76837158203125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":32596,"p5_ns":32776,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":93350,"instructions":360453,"ipc":3.8613,"cache_misses":2290.9,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"2253a4b54d7158d98d03d77c8faccbb8","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:31.381283+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1995847,"p5_ns":2001232,"reference_min_ns":354011,"speedup":0.18714252461557176,"repeat":50,"warmup":5,"cycles":5219765,"instructions":30982226,"ipc":5.9356,"cache_misses":1866.5,"reference_cycles":976840}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"c56f2c7545b95fe39056ddd416b5fdd7","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:31.381283+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":483381,"p5_ns":483837,"reference_min_ns":91023,"speedup":0.21122153316147285,"repeat":50,"warmup":5,"cycles":1276617,"instructions":7667831,"ipc":6.0064,"cache_misses":626.68,"reference_cycles":269649}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"8bfd4f907cfe5518adcf3b18998f246d","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:31.381283+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":119633,"p5_ns":119672,"reference_min_ns":32596,"speedup":0.29444048422606467,"repeat":50,"warmup":5,"cycles":317042,"instructions":1884269,"ipc":5.9433,"cache_misses":438.8,"reference_cycles":93350}}} diff --git a/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048.jsonl b/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048.jsonl deleted file mode 100644 index 1d4f7a1d4037fa16cb4353de97c671cf16e3980b..0000000000000000000000000000000000000000 --- a/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"3dae9d35a0335188af578afb54ef8151","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:28.237252+00:00","log":"","correctness":{"max_relative_error":0.002994697687163859,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":493188,"p5_ns":493935,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1346563,"instructions":5331000,"ipc":3.959,"cache_misses":24402.24,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"d3b09cb41d385638873b847f731b9c48","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:28.237252+00:00","log":"","correctness":{"max_relative_error":0.3085977756662302,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":6311657,"p5_ns":6398710,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":17973981,"instructions":44733601,"ipc":2.4888,"cache_misses":430491.84,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"fb87b8cfce2253f8812665d43d231cca","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:28.237252+00:00","log":"","correctness":{"max_relative_error":0.007799331705269452,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1631534,"p5_ns":1651688,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":4857009,"instructions":13912333,"ipc":2.8644,"cache_misses":103813.8,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"3dae9d35a0335188af578afb54ef8151","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:31.684118+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1935834,"p5_ns":1938970,"reference_min_ns":493188,"speedup":0.2638206973082731,"repeat":50,"warmup":5,"cycles":5104084,"instructions":29979651,"ipc":5.8737,"cache_misses":1890.68,"reference_cycles":1346563}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"d3b09cb41d385638873b847f731b9c48","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:31.684118+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":31979426,"p5_ns":32009446,"reference_min_ns":6311657,"speedup":0.21293287709843475,"repeat":50,"warmup":5,"cycles":84411488,"instructions":495554608,"ipc":5.8707,"cache_misses":30175.26,"reference_cycles":17973981}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"fb87b8cfce2253f8812665d43d231cca","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:31.684118+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":7800332,"p5_ns":7803794,"reference_min_ns":1631534,"speedup":0.23935031073102486,"repeat":50,"warmup":5,"cycles":20292470,"instructions":122518017,"ipc":6.0376,"cache_misses":5924.66,"reference_cycles":4857009}}} diff --git a/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256.jsonl b/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256.jsonl deleted file mode 100644 index f3222a51f213ad855088f906d255b6887772a5b8..0000000000000000000000000000000000000000 --- a/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"bdf7c3b7b2f15227b1c3dd8bee2a65af","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:31.223019+00:00","log":"","correctness":{"max_relative_error":0.0009555858204238001,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":180972,"p5_ns":181787,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":543333,"instructions":1764571,"ipc":3.2477,"cache_misses":6545.56,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"3105af14ead75832a36cb5d634172a75","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:31.223019+00:00","log":"","correctness":{"max_relative_error":0.003234383739047061,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":768003,"p5_ns":779656,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":2142075,"instructions":5616895,"ipc":2.6222,"cache_misses":42179.14,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"f1171a1af3eb5282b8991051b712e1d6","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:31.223019+00:00","log":"","correctness":{"max_relative_error":0.00048386489437813007,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":63833,"p5_ns":64080,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":181272,"instructions":692194,"ipc":3.8185,"cache_misses":3599.24,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"bdf7c3b7b2f15227b1c3dd8bee2a65af","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:34.331741+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":972246,"p5_ns":976388,"reference_min_ns":180972,"speedup":0.21212286840787115,"repeat":50,"warmup":5,"cycles":2561407,"instructions":15327111,"ipc":5.9839,"cache_misses":885.98,"reference_cycles":543333}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"3105af14ead75832a36cb5d634172a75","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:34.331741+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":4078854,"p5_ns":4084238,"reference_min_ns":768003,"speedup":0.20123941099230386,"repeat":50,"warmup":5,"cycles":10644411,"instructions":61954101,"ipc":5.8203,"cache_misses":4216.0,"reference_cycles":2142075}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"f1171a1af3eb5282b8991051b712e1d6","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:34.331741+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":236337,"p5_ns":236382,"reference_min_ns":63833,"speedup":0.2889275672780827,"repeat":50,"warmup":5,"cycles":627396,"instructions":3757181,"ipc":5.9885,"cache_misses":504.0,"reference_cycles":181272}}} diff --git a/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32.jsonl b/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32.jsonl deleted file mode 100644 index 36132afcf2a30a585d8a82cb7f6db4d4a77bb579..0000000000000000000000000000000000000000 --- a/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32","workload":{"axes":{"N":1,"H":224,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"55851b01f1c851e1b4195dace8193993","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:33.603518+00:00","log":"","correctness":{"max_relative_error":0.024685582170641727,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1413194,"p5_ns":1435993,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":4144020,"instructions":9398751,"ipc":2.268,"cache_misses":135733.5,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b22b06e0a17a530a99c1427a5b5b136e","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:33.603518+00:00","log":"","correctness":{"max_relative_error":0.009957783911210686,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":82988,"p5_ns":83206,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":255317,"instructions":727558,"ipc":2.8496,"cache_misses":5703.46,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a19d9a59ec1952ea93ecde2fa4348f68","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:33.603518+00:00","log":"","correctness":{"max_relative_error":0.0011883573779413827,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":24890,"p5_ns":24989,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":73669,"instructions":246094,"ipc":3.3405,"cache_misses":1647.56,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32","workload":{"axes":{"N":1,"H":224,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"55851b01f1c851e1b4195dace8193993","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:34.799292+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":7906957,"p5_ns":7909250,"reference_min_ns":1413194,"speedup":0.20105962373988406,"repeat":50,"warmup":5,"cycles":20610901,"instructions":124951607,"ipc":6.0624,"cache_misses":7145.84,"reference_cycles":4144020}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b22b06e0a17a530a99c1427a5b5b136e","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:34.799292+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":495440,"p5_ns":495766,"reference_min_ns":82988,"speedup":0.19589785140529895,"repeat":50,"warmup":5,"cycles":1303317,"instructions":7753976,"ipc":5.9494,"cache_misses":559.32,"reference_cycles":255317}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a19d9a59ec1952ea93ecde2fa4348f68","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:34.799292+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":124568,"p5_ns":125033,"reference_min_ns":24890,"speedup":0.22279043745180166,"repeat":50,"warmup":5,"cycles":330665,"instructions":1925373,"ipc":5.8227,"cache_misses":446.74,"reference_cycles":73669}}} diff --git a/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512.jsonl b/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512.jsonl deleted file mode 100644 index b4703ee44181000f8c4725b454783fa655e6c24a..0000000000000000000000000000000000000000 --- a/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"0dab312563a150818b5583d92d5bd3b3","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:36.016048+00:00","log":"","correctness":{"max_relative_error":0.0003722374924504813,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":123971,"p5_ns":124454,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":348488,"instructions":1354591,"ipc":3.8871,"cache_misses":6446.04,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"3182ffdfc7f055f190e0b6688beed430","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:36.016048+00:00","log":"","correctness":{"max_relative_error":0.012886767862965026,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1504726,"p5_ns":1527878,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":4406302,"instructions":11204445,"ipc":2.5428,"cache_misses":115223.12,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"87dfa303cfcc5e84b08ae06f7e14ff2e","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:36.016048+00:00","log":"","correctness":{"max_relative_error":0.07757174598123945,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":396163,"p5_ns":397415,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1080360,"instructions":3499570,"ipc":3.2393,"cache_misses":20993.26,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"0dab312563a150818b5583d92d5bd3b3","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:35.461389+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":476555,"p5_ns":477187,"reference_min_ns":123971,"speedup":0.27622788911822804,"repeat":50,"warmup":5,"cycles":1261596,"instructions":7505409,"ipc":5.9491,"cache_misses":612.22,"reference_cycles":348488}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"3182ffdfc7f055f190e0b6688beed430","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:35.461389+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":7980879,"p5_ns":7987558,"reference_min_ns":1504726,"speedup":0.2113098557925864,"repeat":50,"warmup":5,"cycles":20852326,"instructions":123898299,"ipc":5.9417,"cache_misses":7429.38,"reference_cycles":4406302}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"87dfa303cfcc5e84b08ae06f7e14ff2e","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:35.461389+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1993293,"p5_ns":2000013,"reference_min_ns":396163,"speedup":0.20640107751312264,"repeat":50,"warmup":5,"cycles":5234275,"instructions":30638180,"ipc":5.8534,"cache_misses":1594.82,"reference_cycles":1080360}}} diff --git a/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64.jsonl b/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64.jsonl deleted file mode 100644 index c997fdd29f0cb738c70e0535bd8f6ab79c5c2210..0000000000000000000000000000000000000000 --- a/traces/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64","workload":{"axes":{"N":1,"H":112,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"acc8c2d1d5c3519abd9b0f1cfe9c59e7","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:38.455576+00:00","log":"","correctness":{"max_relative_error":0.06394452059652589,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":748574,"p5_ns":762388,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":2079630,"instructions":4986527,"ipc":2.3978,"cache_misses":46542.3,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"add76356814d51218f9e3382d153b348","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:38.455576+00:00","log":"","correctness":{"max_relative_error":0.008634770792412141,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":164797,"p5_ns":166075,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":501059,"instructions":1426103,"ipc":2.8462,"cache_misses":10740.46,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"cbb6ade47dc757cf9ada527b2d5683b6","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:38.455576+00:00","log":"","correctness":{"max_relative_error":0.0015531802663193571,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":47705,"p5_ns":48015,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":137396,"instructions":462874,"ipc":3.3689,"cache_misses":2252.48,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64","workload":{"axes":{"N":1,"H":112,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"acc8c2d1d5c3519abd9b0f1cfe9c59e7","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:36.239297+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":3962570,"p5_ns":3967532,"reference_min_ns":748574,"speedup":0.2011331031045679,"repeat":50,"warmup":5,"cycles":10339571,"instructions":62304191,"ipc":6.0258,"cache_misses":3522.88,"reference_cycles":2079630}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"add76356814d51218f9e3382d153b348","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:36.239297+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":992528,"p5_ns":994224,"reference_min_ns":164797,"speedup":0.19247721660779069,"repeat":50,"warmup":5,"cycles":2603212,"instructions":15496448,"ipc":5.9528,"cache_misses":1152.52,"reference_cycles":501059}}} -{"definition":"conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"cbb6ade47dc757cf9ada527b2d5683b6","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:36.239297+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":243892,"p5_ns":244978,"reference_min_ns":47705,"speedup":0.21067933028396538,"repeat":50,"warmup":5,"cycles":652157,"instructions":3839625,"ipc":5.8876,"cache_misses":509.06,"reference_cycles":137396}}} diff --git a/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024.jsonl b/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024.jsonl deleted file mode 100644 index 65d521d03aa54851ce05d12fd966e25a7ac197dc..0000000000000000000000000000000000000000 --- a/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"6e8f3e53ef305ff8941cb19fcfdba43e","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:40.821687+00:00","log":"","correctness":{"max_relative_error":0.017882561832541204,"max_absolute_error":1.9073486328125e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":440538,"p5_ns":442289,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1217669,"instructions":4477767,"ipc":3.6773,"cache_misses":15322.62,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"684ebc734e645351931cb7dffcc7a31c","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:40.821687+00:00","log":"","correctness":{"max_relative_error":0.043050820388281105,"max_absolute_error":1.9073486328125e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1505542,"p5_ns":1531656,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":4145347,"instructions":11789463,"ipc":2.844,"cache_misses":56794.0,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"c170cae0a7ab5ee7ae3ddcd32a5ca89a","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:40.821687+00:00","log":"","correctness":{"max_relative_error":0.3144782919849114,"max_absolute_error":2.384185791015625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":5947028,"p5_ns":6007136,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":16274325,"instructions":41014048,"ipc":2.5202,"cache_misses":214055.74,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"6e8f3e53ef305ff8941cb19fcfdba43e","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:36.712094+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2819645,"p5_ns":2822886,"reference_min_ns":440538,"speedup":0.16479302365206175,"repeat":50,"warmup":5,"cycles":7389081,"instructions":43486733,"ipc":5.8853,"cache_misses":2579.58,"reference_cycles":1217669}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"684ebc734e645351931cb7dffcc7a31c","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:36.712094+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":11877240,"p5_ns":11885011,"reference_min_ns":1505542,"speedup":0.13419752325453257,"repeat":50,"warmup":5,"cycles":30889892,"instructions":181980288,"ipc":5.8913,"cache_misses":8760.18,"reference_cycles":4145347}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"c170cae0a7ab5ee7ae3ddcd32a5ca89a","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:36.712094+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":48145639,"p5_ns":48151419,"reference_min_ns":5947028,"speedup":0.13020655296156186,"repeat":50,"warmup":5,"cycles":124988525,"instructions":744508522,"ipc":5.9566,"cache_misses":35922.64,"reference_cycles":16274325}}} diff --git a/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128.jsonl b/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128.jsonl deleted file mode 100644 index 7ca4ad199e00140cf440673946323d5b4eca0158..0000000000000000000000000000000000000000 --- a/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"388655f1308e585190af07c49b18a223","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:43.654773+00:00","log":"","correctness":{"max_relative_error":0.00606831376517616,"max_absolute_error":1.430511474609375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":174858,"p5_ns":175478,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":514093,"instructions":1499309,"ipc":2.9164,"cache_misses":3639.42,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"791d0759f55a5470817453ad5e10d862","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:43.654773+00:00","log":"","correctness":{"max_relative_error":0.0005000661599558705,"max_absolute_error":1.430511474609375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":55501,"p5_ns":55704,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":160393,"instructions":585316,"ipc":3.6493,"cache_misses":2716.9,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"d55d25e09ea55038aa79b467a9ef22b8","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:43.654773+00:00","log":"","correctness":{"max_relative_error":0.12048851036477583,"max_absolute_error":2.384185791015625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":725113,"p5_ns":733171,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1965484,"instructions":5152476,"ipc":2.6215,"cache_misses":23836.56,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"388655f1308e585190af07c49b18a223","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:40.441023+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1476433,"p5_ns":1477895,"reference_min_ns":174858,"speedup":0.13316670647130102,"repeat":50,"warmup":5,"cycles":3860522,"instructions":22759489,"ipc":5.8954,"cache_misses":1892.54,"reference_cycles":514093}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"791d0759f55a5470817453ad5e10d862","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:40.441023+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":350472,"p5_ns":350639,"reference_min_ns":55501,"speedup":0.17399147795071618,"repeat":50,"warmup":5,"cycles":921844,"instructions":5445715,"ipc":5.9074,"cache_misses":738.44,"reference_cycles":160393}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"d55d25e09ea55038aa79b467a9ef22b8","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:40.441023+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":5998625,"p5_ns":6002324,"reference_min_ns":725113,"speedup":0.1257834169197768,"repeat":50,"warmup":5,"cycles":15625939,"instructions":93076468,"ipc":5.9565,"cache_misses":4956.24,"reference_cycles":1965484}}} diff --git a/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256.jsonl b/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256.jsonl deleted file mode 100644 index 7ac290b2babe69708a7305cffcbad5092aa208c4..0000000000000000000000000000000000000000 --- a/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"27369bbd6ef35cb882b2cb402d35a437","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:46.001479+00:00","log":"","correctness":{"max_relative_error":0.000836304658318373,"max_absolute_error":1.430511474609375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":110417,"p5_ns":110649,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":318292,"instructions":1141355,"ipc":3.5859,"cache_misses":4481.84,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"aec67bf96654549384120f8d8f1a304b","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:46.001479+00:00","log":"","correctness":{"max_relative_error":0.0034947509352950706,"max_absolute_error":1.9073486328125e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":365426,"p5_ns":376577,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1008250,"instructions":2969264,"ipc":2.945,"cache_misses":5939.84,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"255f34a0bd8154c18a77a899e96356fb","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:46.001479+00:00","log":"","correctness":{"max_relative_error":0.014796670592849304,"max_absolute_error":2.86102294921875e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1465503,"p5_ns":1487268,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":4027791,"instructions":10275278,"ipc":2.5511,"cache_misses":47538.4,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"27369bbd6ef35cb882b2cb402d35a437","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:41.041237+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":700614,"p5_ns":702608,"reference_min_ns":110417,"speedup":0.17242477692017663,"repeat":50,"warmup":5,"cycles":1845976,"instructions":10880070,"ipc":5.8939,"cache_misses":902.14,"reference_cycles":318292}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"aec67bf96654549384120f8d8f1a304b","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:41.041237+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2954934,"p5_ns":2959781,"reference_min_ns":365426,"speedup":0.13062613872761525,"repeat":50,"warmup":5,"cycles":7718593,"instructions":45504165,"ipc":5.8954,"cache_misses":3402.84,"reference_cycles":1008250}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"255f34a0bd8154c18a77a899e96356fb","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:41.041237+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":12017010,"p5_ns":12019333,"reference_min_ns":1465503,"speedup":0.12902018181514163,"repeat":50,"warmup":5,"cycles":31218302,"instructions":186136911,"ipc":5.9624,"cache_misses":7818.62,"reference_cycles":4027791}}} diff --git a/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32.jsonl b/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32.jsonl deleted file mode 100644 index 8b5a78ec341fe79642f8aeb5e1f3ee41d049b245..0000000000000000000000000000000000000000 --- a/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32","workload":{"axes":{"N":1,"H":224,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"011d9c161d78506d8b82b8f78ae0879c","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:48.419940+00:00","log":"","correctness":{"max_relative_error":0.04527594245517467,"max_absolute_error":2.1457672119140625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2826843,"p5_ns":2846841,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":7977753,"instructions":18506687,"ipc":2.3198,"cache_misses":133855.7,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"9cde362cff975aceba771d7f0e8fcb31","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:48.419940+00:00","log":"","correctness":{"max_relative_error":0.007540134939582841,"max_absolute_error":1.430511474609375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":46068,"p5_ns":46109,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":134678,"instructions":396752,"ipc":2.9459,"cache_misses":1820.54,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"1b022377a1985784a2f40ddf4b10bedb","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:48.419940+00:00","log":"","correctness":{"max_relative_error":0.00014716195208430577,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":16342,"p5_ns":16380,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":48858,"instructions":168115,"ipc":3.4409,"cache_misses":1267.0,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32","workload":{"axes":{"N":1,"H":224,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"011d9c161d78506d8b82b8f78ae0879c","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:42.099670+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":24296030,"p5_ns":24300785,"reference_min_ns":2826843,"speedup":0.12649814482142455,"repeat":50,"warmup":5,"cycles":63066166,"instructions":378594754,"ipc":6.0031,"cache_misses":14561.76,"reference_cycles":7977753}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"9cde362cff975aceba771d7f0e8fcb31","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:42.099670+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":367723,"p5_ns":368063,"reference_min_ns":46068,"speedup":0.13924294056343153,"repeat":50,"warmup":5,"cycles":967216,"instructions":5697752,"ipc":5.8909,"cache_misses":736.74,"reference_cycles":134678}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"1b022377a1985784a2f40ddf4b10bedb","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:42.099670+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":87886,"p5_ns":87977,"reference_min_ns":16342,"speedup":0.20889753511341044,"repeat":50,"warmup":5,"cycles":233885,"instructions":1369510,"ipc":5.8555,"cache_misses":193.32,"reference_cycles":48858}}} diff --git a/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512.jsonl b/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512.jsonl deleted file mode 100644 index 6713813aa0eb57beb3da0cde06b899384e58e025..0000000000000000000000000000000000000000 --- a/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"4ff49086ff80584a86ff0f316e900fa2","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:50.914851+00:00","log":"","correctness":{"max_relative_error":0.001956270099070383,"max_absolute_error":1.1920928955078125e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":97176,"p5_ns":97263,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":265328,"instructions":1151277,"ipc":4.3391,"cache_misses":6617.44,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a5f5a3b81d2e592988ed251a554c0305","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:50.914851+00:00","log":"","correctness":{"max_relative_error":0.10881398676664746,"max_absolute_error":2.384185791015625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":760297,"p5_ns":769836,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":2035960,"instructions":5909427,"ipc":2.9025,"cache_misses":13225.26,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"10c62583c34f530c91179b8038ef486c","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:50.914851+00:00","log":"","correctness":{"max_relative_error":0.0015223093658437012,"max_absolute_error":1.430511474609375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":217940,"p5_ns":218848,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":618863,"instructions":2253767,"ipc":3.6418,"cache_misses":8356.32,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"4ff49086ff80584a86ff0f316e900fa2","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:43.655659+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":325463,"p5_ns":326488,"reference_min_ns":97176,"speedup":0.3055376745889265,"repeat":50,"warmup":5,"cycles":868397,"instructions":4975458,"ipc":5.7295,"cache_misses":634.0,"reference_cycles":265328}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a5f5a3b81d2e592988ed251a554c0305","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:43.655659+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":5937736,"p5_ns":5942095,"reference_min_ns":760297,"speedup":0.13174210419956914,"repeat":50,"warmup":5,"cycles":15454133,"instructions":90996504,"ipc":5.8882,"cache_misses":4010.4,"reference_cycles":2035960}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"10c62583c34f530c91179b8038ef486c","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:43.655659+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1408037,"p5_ns":1411776,"reference_min_ns":217940,"speedup":0.16783271591211105,"repeat":50,"warmup":5,"cycles":3687380,"instructions":21748721,"ipc":5.8982,"cache_misses":1507.92,"reference_cycles":618863}}} diff --git a/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64.jsonl b/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64.jsonl deleted file mode 100644 index 74149c1f697cbcb4eff13c9451fc8b133642a482..0000000000000000000000000000000000000000 --- a/traces/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"f907513b7dd056049de530275012b29d","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:53.291483+00:00","log":"","correctness":{"max_relative_error":0.19763890872252962,"max_absolute_error":1.9073486328125e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":362393,"p5_ns":364858,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":973266,"instructions":2590644,"ipc":2.6618,"cache_misses":12256.6,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"c1f95fba5f35520fa98162b977892a84","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:53.291483+00:00","log":"","correctness":{"max_relative_error":0.002522303296006571,"max_absolute_error":1.430511474609375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":87284,"p5_ns":87760,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":252786,"instructions":764390,"ipc":3.0239,"cache_misses":2120.32,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"db7190846c6250ccbb68ac952dbfacd7","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:53.291483+00:00","log":"","correctness":{"max_relative_error":0.019365052964772953,"max_absolute_error":1.430511474609375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":30437,"p5_ns":30522,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":88352,"instructions":307295,"ipc":3.4781,"cache_misses":1843.12,"reference_cycles":null}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"f907513b7dd056049de530275012b29d","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:44.250367+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":3002698,"p5_ns":3006781,"reference_min_ns":362393,"speedup":0.12444959212660366,"repeat":50,"warmup":5,"cycles":7820564,"instructions":46542357,"ipc":5.9513,"cache_misses":3705.08,"reference_cycles":973266}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"c1f95fba5f35520fa98162b977892a84","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:44.250367+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":737528,"p5_ns":737738,"reference_min_ns":87284,"speedup":0.1306784021819539,"repeat":50,"warmup":5,"cycles":1934413,"instructions":11384260,"ipc":5.8851,"cache_misses":1274.68,"reference_cycles":252786}}} -{"definition":"conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"db7190846c6250ccbb68ac952dbfacd7","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:37:44.250367+00:00","log":"","correctness":{"max_relative_error":0.0,"max_absolute_error":0.0,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":175293,"p5_ns":175373,"reference_min_ns":30437,"speedup":0.19105902909178193,"repeat":50,"warmup":5,"cycles":462433,"instructions":2728313,"ipc":5.8999,"cache_misses":414.34,"reference_cycles":88352}}} diff --git a/traces/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin128_cout256.jsonl b/traces/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin128_cout256.jsonl deleted file mode 100644 index 373c61ae1145111cbdbba36aaf7032ba948364ad..0000000000000000000000000000000000000000 --- a/traces/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin128_cout256.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin128_cout256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"1e10a2446b915b48bbeaebd551683f65","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh1_sw1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:55.625262+00:00","log":"","correctness":{"max_relative_error":1.0641057539959342,"max_absolute_error":0.0000762939453125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":249054634,"p5_ns":249064753,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":646516889,"instructions":259761929,"ipc":0.4018,"cache_misses":159650.66,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin128_cout256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b7800d870b1e523293dc20bdc2d71a64","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh1_sw1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:55.625262+00:00","log":"","correctness":{"max_relative_error":0.007769730544344406,"max_absolute_error":0.000072479248046875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":59009281,"p5_ns":59012439,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":153273408,"instructions":107632359,"ipc":0.7022,"cache_misses":45146.66,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin128_cout256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b4d85782dfbc580e8685167a2413f29e","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh1_sw1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:14:55.625262+00:00","log":"","correctness":{"max_relative_error":0.13797113946143616,"max_absolute_error":0.000091552734375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":988819906,"p5_ns":988892554,"reference_min_ns":null,"speedup":null,"repeat":31,"warmup":5,"cycles":2565702797,"instructions":979091611,"ipc":0.3816,"cache_misses":571782.9677419355,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin128_cout256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"1e10a2446b915b48bbeaebd551683f65","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh1_sw1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:32.745521+00:00","log":"","correctness":{"max_relative_error":0.9056285936005262,"max_absolute_error":0.00008392333984375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":125513880,"p5_ns":125537764,"reference_min_ns":249054634,"speedup":1.9870666913558503,"repeat":50,"warmup":5,"cycles":325362451,"instructions":1288054156,"ipc":3.9588,"cache_misses":32944339.98,"reference_cycles":646516889}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin128_cout256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b7800d870b1e523293dc20bdc2d71a64","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh1_sw1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:32.745521+00:00","log":"","correctness":{"max_relative_error":0.020784363285388314,"max_absolute_error":0.000072479248046875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":38816775,"p5_ns":38835449,"reference_min_ns":59009281,"speedup":1.5193628813815154,"repeat":50,"warmup":5,"cycles":100880053,"instructions":322102384,"ipc":3.1929,"cache_misses":8905670.9,"reference_cycles":153273408}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin128_cout256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b4d85782dfbc580e8685167a2413f29e","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh1_sw1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:32.745521+00:00","log":"","correctness":{"max_relative_error":0.14932004238696808,"max_absolute_error":0.000095367431640625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":526496043,"p5_ns":526692813,"reference_min_ns":988819906,"speedup":1.8745134601032203,"repeat":50,"warmup":5,"cycles":1368729994,"instructions":5152163582,"ipc":3.7642,"cache_misses":154957205.4,"reference_cycles":2565702797}}} diff --git a/traces/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin256_cout512.jsonl b/traces/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin256_cout512.jsonl deleted file mode 100644 index 1d9ad2634de67f71c3d1e5c011bf453a1c0d6ca1..0000000000000000000000000000000000000000 --- a/traces/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin256_cout512.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin256_cout512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"c80b48c47e1d5ec3b0cf61cccbf3563a","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh1_sw1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:15:51.835216+00:00","log":"","correctness":{"max_relative_error":0.07931206888292776,"max_absolute_error":0.000152587890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":245479525,"p5_ns":245536658,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":637384443,"instructions":430297021,"ipc":0.6751,"cache_misses":136361.8,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin256_cout512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"3f04b4e83624544f809891e9096ea567","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh1_sw1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:15:51.835216+00:00","log":"","correctness":{"max_relative_error":1.3982718664271232,"max_absolute_error":0.0001583099365234375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":995884028,"p5_ns":995986359,"reference_min_ns":null,"speedup":null,"repeat":31,"warmup":5,"cycles":2584115070,"instructions":1038224497,"ipc":0.4018,"cache_misses":593101.4516129033,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin256_cout512","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b9ed91b58b72562bbbe6f29af161ab71","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh1_sw1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:15:51.835216+00:00","log":"","correctness":{"max_relative_error":0.5333004476752908,"max_absolute_error":0.000179290771484375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":3955342995,"p5_ns":3955342995,"reference_min_ns":null,"speedup":null,"repeat":8,"warmup":5,"cycles":10260432404,"instructions":3913657384,"ipc":0.3814,"cache_misses":2234725.5,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin256_cout512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"c80b48c47e1d5ec3b0cf61cccbf3563a","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh1_sw1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:00:11.787645+00:00","log":"","correctness":{"max_relative_error":0.05919327217857827,"max_absolute_error":0.000156402587890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":160341777,"p5_ns":160390884,"reference_min_ns":245479525,"speedup":1.5312137689679892,"repeat":50,"warmup":5,"cycles":416260914,"instructions":1287057166,"ipc":3.0919,"cache_misses":50495798.92,"reference_cycles":637384443}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin256_cout512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"3f04b4e83624544f809891e9096ea567","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh1_sw1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:00:11.787645+00:00","log":"","correctness":{"max_relative_error":1.3801712152612118,"max_absolute_error":0.0001621246337890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":659150647,"p5_ns":659645329,"reference_min_ns":995884028,"speedup":1.5106468801954271,"repeat":46,"warmup":5,"cycles":1710601666,"instructions":5148135151,"ipc":3.0095,"cache_misses":262267362.8478261,"reference_cycles":2584115070}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin256_cout512","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b9ed91b58b72562bbbe6f29af161ab71","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh1_sw1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:00:11.787645+00:00","log":"","correctness":{"max_relative_error":0.3206136828271028,"max_absolute_error":0.000186920166015625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2682425587,"p5_ns":2682425587,"reference_min_ns":3955342995,"speedup":1.477635102781965,"repeat":12,"warmup":5,"cycles":6943820152,"instructions":20592398641,"ipc":2.9656,"cache_misses":1654950587.4166667,"reference_cycles":10260432404}}} diff --git a/traces/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024.jsonl b/traces/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024.jsonl deleted file mode 100644 index 2780a156c111131cc005752ed9537cb66024f14d..0000000000000000000000000000000000000000 --- a/traces/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"d4bc1434edde5e678092cc989f06ea1f","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:17:40.259366+00:00","log":"","correctness":{"max_relative_error":0.01945721677253531,"max_absolute_error":0.000244140625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":204642824,"p5_ns":204693440,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":531770446,"instructions":804443454,"ipc":1.5128,"cache_misses":168037.34,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"3f55d9b5825054b482ed0a68a8b07a15","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:17:40.259366+00:00","log":"","correctness":{"max_relative_error":0.10767061060125177,"max_absolute_error":0.00037384033203125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":3986532219,"p5_ns":3986532219,"reference_min_ns":null,"speedup":null,"repeat":8,"warmup":5,"cycles":10341829140,"instructions":4151398736,"ipc":0.4014,"cache_misses":2390631.875,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"9b18355b1a5f587fb9ec22fdc3564a96","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:17:40.259366+00:00","log":"","correctness":{"max_relative_error":0.2309723681436133,"max_absolute_error":0.0003204345703125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":972170312,"p5_ns":972304589,"reference_min_ns":null,"speedup":null,"repeat":31,"warmup":5,"cycles":2523252600,"instructions":1720551131,"ipc":0.6819,"cache_misses":512172.8387096774,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"d4bc1434edde5e678092cc989f06ea1f","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:01:43.772189+00:00","log":"","correctness":{"max_relative_error":0.03773470223362642,"max_absolute_error":0.000244140625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":144143810,"p5_ns":144250406,"reference_min_ns":204642824,"speedup":1.4209786770059094,"repeat":50,"warmup":5,"cycles":374228308,"instructions":1260662178,"ipc":3.3687,"cache_misses":46979531.58,"reference_cycles":531770446}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"3f55d9b5825054b482ed0a68a8b07a15","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:01:43.772189+00:00","log":"","correctness":{"max_relative_error":0.1137196771419059,"max_absolute_error":0.00040435791015625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2922702093,"p5_ns":2922702093,"reference_min_ns":3986532219,"speedup":1.364696801205778,"repeat":11,"warmup":5,"cycles":7578114883,"instructions":20173565362,"ipc":2.6621,"cache_misses":1019687843.0,"reference_cycles":10341829140}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"9b18355b1a5f587fb9ec22fdc3564a96","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:01:43.772189+00:00","log":"","correctness":{"max_relative_error":0.2637705423489123,"max_absolute_error":0.00032806396484375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":656112992,"p5_ns":656397666,"reference_min_ns":972170312,"speedup":1.4834136703375758,"repeat":46,"warmup":5,"cycles":1700977044,"instructions":5042937655,"ipc":2.9647,"cache_misses":187066628.7173913,"reference_cycles":2523252600}}} diff --git a/traces/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin64_cout128.jsonl b/traces/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin64_cout128.jsonl deleted file mode 100644 index e01bcc546b8addf464f9d713df59f2dbfbd70f52..0000000000000000000000000000000000000000 --- a/traces/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin64_cout128.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin64_cout128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b6f7a251d2d2532195a006e4b5bf52e9","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh1_sw1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:19:26.088187+00:00","log":"","correctness":{"max_relative_error":1.6513323397249788,"max_absolute_error":0.0000400543212890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":247283589,"p5_ns":247304662,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":641782446,"instructions":245117629,"ipc":0.3819,"cache_misses":148202.24,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin64_cout128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"ab31f9e56d8a5a6d93b51a3db4f7b926","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh1_sw1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:19:26.088187+00:00","log":"","correctness":{"max_relative_error":0.2932703483618047,"max_absolute_error":0.0000457763671875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":62278292,"p5_ns":62283099,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":161711824,"instructions":65046345,"ipc":0.4022,"cache_misses":42002.5,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin64_cout128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"498a15f55c405bcdb487e8d2b2753f6c","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh1_sw1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:19:26.088187+00:00","log":"","correctness":{"max_relative_error":0.004498488472573649,"max_absolute_error":0.00003814697265625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":15213527,"p5_ns":15217253,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":39545656,"instructions":26957923,"ipc":0.6817,"cache_misses":9823.78,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin64_cout128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b6f7a251d2d2532195a006e4b5bf52e9","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh1_sw1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:03:16.207654+00:00","log":"","correctness":{"max_relative_error":0.676909369391364,"max_absolute_error":0.000041961669921875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":99691104,"p5_ns":99788867,"reference_min_ns":247283589,"speedup":2.477563236860219,"repeat":50,"warmup":5,"cycles":259037766,"instructions":1290361308,"ipc":4.9814,"cache_misses":10741051.2,"reference_cycles":641782446}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin64_cout128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"ab31f9e56d8a5a6d93b51a3db4f7b926","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh1_sw1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:03:16.207654+00:00","log":"","correctness":{"max_relative_error":0.18681649386027213,"max_absolute_error":0.0000476837158203125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":26475053,"p5_ns":26477351,"reference_min_ns":62278292,"speedup":2.349605020609484,"repeat":50,"warmup":5,"cycles":68825110,"instructions":322633021,"ipc":4.6877,"cache_misses":4290543.66,"reference_cycles":161711824}}} -{"definition":"deconv2d_kh3_kw3_sh1_sw1_cin64_cout128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"498a15f55c405bcdb487e8d2b2753f6c","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh1_sw1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:03:16.207654+00:00","log":"","correctness":{"max_relative_error":0.004498488472573649,"max_absolute_error":0.00003814697265625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":8836281,"p5_ns":8836869,"reference_min_ns":15213527,"speedup":1.7213016611536887,"repeat":50,"warmup":5,"cycles":22974274,"instructions":80697945,"ipc":3.5125,"cache_misses":1745301.94,"reference_cycles":39545656}}} diff --git a/traces/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin128_cout256.jsonl b/traces/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin128_cout256.jsonl deleted file mode 100644 index e786a29e413d5e813575d107dd94c40c54c6de2d..0000000000000000000000000000000000000000 --- a/traces/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin128_cout256.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin128_cout256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"757cdaaa84715203a02fb77704f20e52","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh2_sw2_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:19:46.574729+00:00","log":"","correctness":{"max_relative_error":0.38183048266719505,"max_absolute_error":0.000030517578125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":20961092,"p5_ns":20965651,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":54467403,"instructions":91486825,"ipc":1.6797,"cache_misses":28253.06,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin128_cout256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"397d7a1f2c54514f91797479cc497e99","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh2_sw2_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:19:46.574729+00:00","log":"","correctness":{"max_relative_error":9.569357872009277,"max_absolute_error":0.0000362396240234375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":376211130,"p5_ns":376243299,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":976165421,"instructions":844171153,"ipc":0.8648,"cache_misses":3757078.14,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin128_cout256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"cd31a62ce3de56cba0188e1ac64fe8f1","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh2_sw2_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:19:46.574729+00:00","log":"","correctness":{"max_relative_error":0.21390967223321505,"max_absolute_error":0.000030517578125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":95974369,"p5_ns":96008601,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":249201754,"instructions":224628684,"ipc":0.9014,"cache_misses":119516.6,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin128_cout256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"757cdaaa84715203a02fb77704f20e52","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh2_sw2_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:03:23.961213+00:00","log":"","correctness":{"max_relative_error":0.38183048266719505,"max_absolute_error":0.000030517578125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":32933476,"p5_ns":32946128,"reference_min_ns":20961092,"speedup":0.6367178622990999,"repeat":50,"warmup":5,"cycles":85544016,"instructions":322157840,"ipc":3.766,"cache_misses":9265643.2,"reference_cycles":54467403}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin128_cout256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"397d7a1f2c54514f91797479cc497e99","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh2_sw2_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:03:23.961213+00:00","log":"","correctness":{"max_relative_error":9.569357872009277,"max_absolute_error":0.00003814697265625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":821019175,"p5_ns":823109484,"reference_min_ns":376211130,"speedup":0.4490880454494894,"repeat":36,"warmup":5,"cycles":2173661559,"instructions":5155720239,"ipc":2.3719,"cache_misses":184255417.58333334,"reference_cycles":976165421}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin128_cout256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"cd31a62ce3de56cba0188e1ac64fe8f1","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh2_sw2_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:03:23.961213+00:00","log":"","correctness":{"max_relative_error":0.21390967223321505,"max_absolute_error":0.000030517578125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":137428765,"p5_ns":137509032,"reference_min_ns":95974369,"speedup":0.6924075262290085,"repeat":50,"warmup":5,"cycles":359906189,"instructions":1288487409,"ipc":3.5801,"cache_misses":45529880.08,"reference_cycles":249201754}}} diff --git a/traces/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin256_cout512.jsonl b/traces/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin256_cout512.jsonl deleted file mode 100644 index 519a6a27240c5de5c66c28b1a50c8ed0b92ffd2b..0000000000000000000000000000000000000000 --- a/traces/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin256_cout512.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin256_cout512","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"2c450cd625e356a6a7901d93f4023df1","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh2_sw2_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:20:16.581440+00:00","log":"","correctness":{"max_relative_error":0.03795766378660637,"max_absolute_error":0.0000591278076171875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":20790696,"p5_ns":20795708,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":54118304,"instructions":172490982,"ipc":3.1873,"cache_misses":31360.34,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin256_cout512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"6ed1ded2d1b450e28303b76f1c208064","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh2_sw2_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:20:16.581440+00:00","log":"","correctness":{"max_relative_error":0.033878062611484634,"max_absolute_error":0.000072479248046875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":83724366,"p5_ns":83793266,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":217659387,"instructions":365191965,"ipc":1.6778,"cache_misses":89485.74,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin256_cout512","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"00136af5b6be5dedbab1cc642549cd54","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh2_sw2_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:20:16.581440+00:00","log":"","correctness":{"max_relative_error":3.998158645629883,"max_absolute_error":0.000080108642578125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1503793754,"p5_ns":1503909714,"reference_min_ns":null,"speedup":null,"repeat":20,"warmup":5,"cycles":3900646671,"instructions":3366758847,"ipc":0.8631,"cache_misses":14168266.55,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin256_cout512","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"2c450cd625e356a6a7901d93f4023df1","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh2_sw2_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:04:09.094721+00:00","log":"","correctness":{"max_relative_error":0.03734354812661499,"max_absolute_error":0.0000591278076171875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":38661196,"p5_ns":38689623,"reference_min_ns":20790696,"speedup":0.5383000716890557,"repeat":50,"warmup":5,"cycles":100535569,"instructions":321856448,"ipc":3.2014,"cache_misses":10466540.1,"reference_cycles":54118304}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin256_cout512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"6ed1ded2d1b450e28303b76f1c208064","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh2_sw2_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:04:09.094721+00:00","log":"","correctness":{"max_relative_error":0.033878062611484634,"max_absolute_error":0.000064849853515625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":162871280,"p5_ns":162918953,"reference_min_ns":83724366,"speedup":0.5150338998605453,"repeat":50,"warmup":5,"cycles":422611768,"instructions":1287270690,"ipc":3.046,"cache_misses":54843885.48,"reference_cycles":217659387}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin256_cout512","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"00136af5b6be5dedbab1cc642549cd54","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh2_sw2_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:04:09.094721+00:00","log":"","correctness":{"max_relative_error":3.998158645629883,"max_absolute_error":0.0000762939453125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":3166376204,"p5_ns":3166376204,"reference_min_ns":1503793754,"speedup":0.4750401972351678,"repeat":10,"warmup":5,"cycles":8211192850,"instructions":20598425716,"ipc":2.5086,"cache_misses":1615468872.5,"reference_cycles":3900646671}}} diff --git a/traces/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024.jsonl b/traces/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024.jsonl deleted file mode 100644 index 90f4531799be1089168de721e6d695c87c4225e6..0000000000000000000000000000000000000000 --- a/traces/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"c793189925d45c92b5df3139c412f90f","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:21:04.043937+00:00","log":"","correctness":{"max_relative_error":0.32074916762793626,"max_absolute_error":0.00009918212890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":83324276,"p5_ns":83354280,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":216690614,"instructions":689466130,"ipc":3.1818,"cache_misses":118503.0,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"89d19f1d5cfb5e14b73cd4020af05f5b","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:21:04.043937+00:00","log":"","correctness":{"max_relative_error":1.7788882785373263,"max_absolute_error":0.00011444091796875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":334532404,"p5_ns":334671947,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":868767342,"instructions":1459314340,"ipc":1.6798,"cache_misses":318503.54,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"f012cbf24a565fa7a174a4a9da91198e","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:21:04.043937+00:00","log":"","correctness":{"max_relative_error":46.54415007560484,"max_absolute_error":0.0001430511474609375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":6034790627,"p5_ns":6034790627,"reference_min_ns":null,"speedup":null,"repeat":5,"warmup":5,"cycles":15654673943,"instructions":13523016828,"ipc":0.8638,"cache_misses":56107273.6,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"c793189925d45c92b5df3139c412f90f","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:05:11.642613+00:00","log":"","correctness":{"max_relative_error":0.32074916762793626,"max_absolute_error":0.00009918212890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":165444598,"p5_ns":165581115,"reference_min_ns":83324276,"speedup":0.5045385957555532,"repeat":50,"warmup":5,"cycles":429482731,"instructions":1260936471,"ipc":2.9359,"cache_misses":49602334.4,"reference_cycles":216690614}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"89d19f1d5cfb5e14b73cd4020af05f5b","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:05:11.642613+00:00","log":"","correctness":{"max_relative_error":1.7788882785373263,"max_absolute_error":0.00011444091796875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":754787175,"p5_ns":755111268,"reference_min_ns":334532404,"speedup":0.4366279579153149,"repeat":40,"warmup":5,"cycles":1989719912,"instructions":5044130283,"ipc":2.5351,"cache_misses":250889369.9,"reference_cycles":868767342}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"f012cbf24a565fa7a174a4a9da91198e","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:05:11.642613+00:00","log":"","correctness":{"max_relative_error":46.54415007560484,"max_absolute_error":0.0001468658447265625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":18916750179,"p5_ns":18916750179,"reference_min_ns":6034790627,"speedup":0.319158092795948,"repeat":2,"warmup":5,"cycles":49049904409,"instructions":80751169768,"ipc":1.6463,"cache_misses":7438731045.5,"reference_cycles":15654673943}}} diff --git a/traces/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin64_cout128.jsonl b/traces/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin64_cout128.jsonl deleted file mode 100644 index a0bab1447478757043f342290645892be2b44762..0000000000000000000000000000000000000000 --- a/traces/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin64_cout128.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin64_cout128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"66eb34df79a9580482245ff6dcd9e904","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh2_sw2_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:22:37.243320+00:00","log":"","correctness":{"max_relative_error":0.031782547515463444,"max_absolute_error":0.0000133514404296875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":24028552,"p5_ns":24035690,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":62405336,"instructions":56493479,"ipc":0.9053,"cache_misses":37046.14,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin64_cout128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"cc0b953d48495c6e903ebf005f51d9d3","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh2_sw2_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:22:37.243320+00:00","log":"","correctness":{"max_relative_error":0.2205037995586245,"max_absolute_error":0.0000152587890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":5256913,"p5_ns":5258566,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":13710888,"instructions":22973514,"ipc":1.6756,"cache_misses":8393.4,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin64_cout128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a3436efc560d52b6833375a176619fd7","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh3_kw3_sh2_sw2_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:22:37.243320+00:00","log":"","correctness":{"max_relative_error":0.3361523376768501,"max_absolute_error":0.0000152587890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":94197781,"p5_ns":94208033,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":244386667,"instructions":212292492,"ipc":0.8687,"cache_misses":948183.56,"reference_cycles":null}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin64_cout128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"66eb34df79a9580482245ff6dcd9e904","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh2_sw2_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:08:28.091666+00:00","log":"","correctness":{"max_relative_error":0.031782547515463444,"max_absolute_error":0.0000133514404296875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":27617525,"p5_ns":27739042,"reference_min_ns":24028552,"speedup":0.8603299455453574,"repeat":50,"warmup":5,"cycles":72536515,"instructions":322812850,"ipc":4.4503,"cache_misses":4500994.68,"reference_cycles":62405336}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin64_cout128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"cc0b953d48495c6e903ebf005f51d9d3","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh2_sw2_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:08:28.091666+00:00","log":"","correctness":{"max_relative_error":0.2205037995586245,"max_absolute_error":0.0000133514404296875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":6799042,"p5_ns":6800103,"reference_min_ns":5256913,"speedup":0.773960484848348,"repeat":50,"warmup":5,"cycles":17715230,"instructions":80728027,"ipc":4.557,"cache_misses":1213303.0,"reference_cycles":13710888}}} -{"definition":"deconv2d_kh3_kw3_sh2_sw2_cin64_cout128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a3436efc560d52b6833375a176619fd7","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh3_kw3_sh2_sw2_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:08:28.091666+00:00","log":"","correctness":{"max_relative_error":0.3361523376768501,"max_absolute_error":0.0000171661376953125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":174633917,"p5_ns":176899552,"reference_min_ns":94197781,"speedup":0.5183733995931,"repeat":50,"warmup":5,"cycles":471449089,"instructions":1291604058,"ipc":2.7396,"cache_misses":18145448.08,"reference_cycles":244386667}}} diff --git a/traces/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin128_cout256.jsonl b/traces/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin128_cout256.jsonl deleted file mode 100644 index 6d1dc3093dfda1ba0a0fa1c435dda7b30dc78b6e..0000000000000000000000000000000000000000 --- a/traces/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin128_cout256.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin128_cout256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"1127aa3e0c165ab59166fbed66e348b7","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh1_sw1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:22:46.515271+00:00","log":"","correctness":{"max_relative_error":0.20255186898367747,"max_absolute_error":0.0001373291015625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":504509538,"p5_ns":504621774,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1309648126,"instructions":407812655,"ipc":0.3114,"cache_misses":253762.28,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin128_cout256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a212111de9cf54e49315e02fc5709525","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh1_sw1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:22:46.515271+00:00","log":"","correctness":{"max_relative_error":0.05294349895364817,"max_absolute_error":0.000118255615234375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":115773420,"p5_ns":115794752,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":300679461,"instructions":166126858,"ipc":0.5525,"cache_misses":60813.48,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin128_cout256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b6cbdf3552d25b4082b3ab88e0031132","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh1_sw1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:22:46.515271+00:00","log":"","correctness":{"max_relative_error":0.42641246623280943,"max_absolute_error":0.000152587890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1999022287,"p5_ns":1999022287,"reference_min_ns":null,"speedup":null,"repeat":16,"warmup":5,"cycles":5187313289,"instructions":1557462678,"ipc":0.3002,"cache_misses":1138049.625,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin128_cout256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"1127aa3e0c165ab59166fbed66e348b7","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh1_sw1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:08:40.438919+00:00","log":"","correctness":{"max_relative_error":0.548657717023577,"max_absolute_error":0.0001373291015625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":216407025,"p5_ns":216439283,"reference_min_ns":504509538,"speedup":2.3347219147370533,"repeat":50,"warmup":5,"cycles":560943947,"instructions":2189314327,"ipc":3.9029,"cache_misses":38455191.9,"reference_cycles":1309648126}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin128_cout256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a212111de9cf54e49315e02fc5709525","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh1_sw1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:08:40.438919+00:00","log":"","correctness":{"max_relative_error":0.06018694754185407,"max_absolute_error":0.0001125335693359375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":49962830,"p5_ns":49967756,"reference_min_ns":115773420,"speedup":2.31679712440487,"repeat":50,"warmup":5,"cycles":129782387,"instructions":547344429,"ipc":4.2174,"cache_misses":6939189.12,"reference_cycles":300679461}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin128_cout256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b6cbdf3552d25b4082b3ab88e0031132","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh1_sw1_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:08:40.438919+00:00","log":"","correctness":{"max_relative_error":0.27853801255141825,"max_absolute_error":0.00016021728515625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":909298428,"p5_ns":909357546,"reference_min_ns":1999022287,"speedup":2.2003361211494696,"repeat":33,"warmup":5,"cycles":2357509491,"instructions":8757272903,"ipc":3.7146,"cache_misses":303678639.3333333,"reference_cycles":5187313289}}} diff --git a/traces/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin256_cout512.jsonl b/traces/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin256_cout512.jsonl deleted file mode 100644 index 901a7f6404706118bc05ba89bb55884cf9bf44b8..0000000000000000000000000000000000000000 --- a/traces/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin256_cout512.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin256_cout512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"281906993858518cb1aaca5ab9e280a0","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh1_sw1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:24:07.592600+00:00","log":"","correctness":{"max_relative_error":0.07076172585098164,"max_absolute_error":0.00023651123046875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":463135275,"p5_ns":463209813,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1202241354,"instructions":664167779,"ipc":0.5524,"cache_misses":217251.3,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin256_cout512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"6cdad3b9d35c515ca57413fcd9f351c4","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh1_sw1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:24:07.592600+00:00","log":"","correctness":{"max_relative_error":0.1798017381926442,"max_absolute_error":0.0003204345703125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2019009492,"p5_ns":2019009492,"reference_min_ns":null,"speedup":null,"repeat":15,"warmup":5,"cycles":5238146884,"instructions":1630336663,"ipc":0.3112,"cache_misses":841925.2666666667,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin256_cout512","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"6d2d53dd8df15ddfb3a8d0929224f919","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh1_sw1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:24:07.592600+00:00","log":"","correctness":{"max_relative_error":2.0336895282451923,"max_absolute_error":0.000335693359375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":7996888169,"p5_ns":7996888169,"reference_min_ns":null,"speedup":null,"repeat":4,"warmup":5,"cycles":20746273311,"instructions":6227111405,"ipc":0.3002,"cache_misses":4479051.5,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin256_cout512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"281906993858518cb1aaca5ab9e280a0","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh1_sw1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:09:31.141485+00:00","log":"","correctness":{"max_relative_error":0.06781188931557898,"max_absolute_error":0.00026702880859375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":225817676,"p5_ns":225965331,"reference_min_ns":463135275,"speedup":2.0527552519883914,"repeat":50,"warmup":5,"cycles":585672039,"instructions":2187440279,"ipc":3.7349,"cache_misses":43844232.08,"reference_cycles":1202241354}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin256_cout512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"6cdad3b9d35c515ca57413fcd9f351c4","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh1_sw1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:09:31.141485+00:00","log":"","correctness":{"max_relative_error":0.23343103177657162,"max_absolute_error":0.00032806396484375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1087389983,"p5_ns":1087502509,"reference_min_ns":2019009492,"speedup":1.8596477503239885,"repeat":28,"warmup":5,"cycles":2816741441,"instructions":8750752525,"ipc":3.1067,"cache_misses":177143824.42857143,"reference_cycles":5238146884}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin256_cout512","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"6d2d53dd8df15ddfb3a8d0929224f919","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh1_sw1_cin256_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:09:31.141485+00:00","log":"","correctness":{"max_relative_error":2.015570537860577,"max_absolute_error":0.00032806396484375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":4206549765,"p5_ns":4206549765,"reference_min_ns":7996888169,"speedup":1.9047285681053845,"repeat":8,"warmup":5,"cycles":10891984117,"instructions":35001358545,"ipc":3.2135,"cache_misses":2234112878.25,"reference_cycles":20746273311}}} diff --git a/traces/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024.jsonl b/traces/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024.jsonl deleted file mode 100644 index cf2fd92d66515ff3173d3005ac2be8662981e56e..0000000000000000000000000000000000000000 --- a/traces/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"6280b53d3fe952eeb170df8162906061","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:26:38.433821+00:00","log":"","correctness":{"max_relative_error":0.7414608898944305,"max_absolute_error":0.00042724609375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":397177282,"p5_ns":397222528,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1032216916,"instructions":1325581475,"ipc":1.2842,"cache_misses":944576.96,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"0471154089685f21b99250b1388fc051","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:26:38.433821+00:00","log":"","correctness":{"max_relative_error":5.424152271167652,"max_absolute_error":0.000640869140625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":8088857387,"p5_ns":8088857387,"reference_min_ns":null,"speedup":null,"repeat":4,"warmup":5,"cycles":20982406914,"instructions":6567065442,"ipc":0.313,"cache_misses":3475909.75,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"86533bd15aee5ff199928247341d140f","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:26:38.433821+00:00","log":"","correctness":{"max_relative_error":0.07859461690751911,"max_absolute_error":0.000457763671875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1878858616,"p5_ns":1878858616,"reference_min_ns":null,"speedup":null,"repeat":16,"warmup":5,"cycles":4874618498,"instructions":2749810358,"ipc":0.5641,"cache_misses":1501044.0,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"6280b53d3fe952eeb170df8162906061","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:11:20.128700+00:00","log":"","correctness":{"max_relative_error":0.8218857760856293,"max_absolute_error":0.00045013427734375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":208944282,"p5_ns":209018413,"reference_min_ns":397177282,"speedup":1.9041677939772026,"repeat":50,"warmup":5,"cycles":542082961,"instructions":2160698565,"ipc":3.9859,"cache_misses":49237764.72,"reference_cycles":1032216916}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"0471154089685f21b99250b1388fc051","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:11:20.128700+00:00","log":"","correctness":{"max_relative_error":5.58311297442462,"max_absolute_error":0.00067138671875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":5074039348,"p5_ns":5074039348,"reference_min_ns":8088857387,"speedup":1.5961120203084072,"repeat":6,"warmup":5,"cycles":13145948810,"instructions":34581476080,"ipc":2.6306,"cache_misses":2220510745.6666665,"reference_cycles":20982406914}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"86533bd15aee5ff199928247341d140f","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:11:20.128700+00:00","log":"","correctness":{"max_relative_error":0.1446709884358569,"max_absolute_error":0.00049591064453125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":957342669,"p5_ns":957655054,"reference_min_ns":1878858616,"speedup":1.9654249524959426,"repeat":32,"warmup":5,"cycles":2480185515,"instructions":8643244974,"ipc":3.4849,"cache_misses":209787815.59375,"reference_cycles":4874618498}}} diff --git a/traces/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin64_cout128.jsonl b/traces/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin64_cout128.jsonl deleted file mode 100644 index 5a79679c875bed32cb1ab0bac2c63bc9c2ccdc37..0000000000000000000000000000000000000000 --- a/traces/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin64_cout128.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin64_cout128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b8291a0b9c45578f9ae1e64330530844","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh1_sw1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:29:05.664784+00:00","log":"","correctness":{"max_relative_error":1.9902560442862427,"max_absolute_error":0.0000972747802734375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":499791266,"p5_ns":499823104,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1297026301,"instructions":389721399,"ipc":0.3005,"cache_misses":281169.14,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin64_cout128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a374addb3be755c895fd22b974d73bc6","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh1_sw1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:29:05.664784+00:00","log":"","correctness":{"max_relative_error":0.018516928639254013,"max_absolute_error":0.000091552734375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":126139691,"p5_ns":126142156,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":327532305,"instructions":102065293,"ipc":0.3116,"cache_misses":69733.78,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin64_cout128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"5fdc5d99e43a5bcb980207e9b1ad02b7","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh1_sw1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:29:05.664784+00:00","log":"","correctness":{"max_relative_error":0.01153113556556043,"max_absolute_error":0.000057220458984375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":28967684,"p5_ns":28972171,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":75248265,"instructions":41579192,"ipc":0.5526,"cache_misses":16663.02,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin64_cout128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b8291a0b9c45578f9ae1e64330530844","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh1_sw1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:13:09.795555+00:00","log":"","correctness":{"max_relative_error":1.8009069318602071,"max_absolute_error":0.0000972747802734375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":180211031,"p5_ns":180530600,"reference_min_ns":499791266,"speedup":2.7580151755336124,"repeat":50,"warmup":5,"cycles":470275259,"instructions":2192968786,"ipc":4.6632,"cache_misses":20641345.98,"reference_cycles":1297026301}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin64_cout128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a374addb3be755c895fd22b974d73bc6","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh1_sw1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:13:09.795555+00:00","log":"","correctness":{"max_relative_error":0.019455815631401502,"max_absolute_error":0.00009918212890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":44808765,"p5_ns":44841940,"reference_min_ns":126139691,"speedup":2.8103491767994933,"repeat":50,"warmup":5,"cycles":116545057,"instructions":548281957,"ipc":4.7045,"cache_misses":2678918.28,"reference_cycles":327532305}}} -{"definition":"deconv2d_kh4_kw4_sh1_sw1_cin64_cout128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"5fdc5d99e43a5bcb980207e9b1ad02b7","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh1_sw1_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:13:09.795555+00:00","log":"","correctness":{"max_relative_error":0.010460313779426817,"max_absolute_error":0.000057220458984375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":10424408,"p5_ns":10434924,"reference_min_ns":28967684,"speedup":2.7692716956647074,"repeat":50,"warmup":5,"cycles":27172583,"instructions":137094713,"ipc":5.0453,"cache_misses":323545.02,"reference_cycles":75248265}}} diff --git a/traces/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin128_cout256.jsonl b/traces/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin128_cout256.jsonl deleted file mode 100644 index d7542638c2fc15dbb774927da5b6432e5aa06f1a..0000000000000000000000000000000000000000 --- a/traces/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin128_cout256.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin128_cout256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"5ff57762aad35880848ad483fc149572","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh2_sw2_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:29:44.638642+00:00","log":"","correctness":{"max_relative_error":0.17463923495450662,"max_absolute_error":0.00003814697265625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":25595512,"p5_ns":25601843,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":66705752,"instructions":145935645,"ipc":2.1878,"cache_misses":42162.84,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin128_cout256","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"15748e6d15d054d292c8942325690e9c","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh2_sw2_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:29:44.638642+00:00","log":"","correctness":{"max_relative_error":0.006924927690742433,"max_absolute_error":0.000030517578125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":7617165,"p5_ns":7632506,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":19880208,"instructions":73535025,"ipc":3.6989,"cache_misses":11388.92,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin128_cout256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b3d7f1a4e8585a46ae8fbbec63913dfd","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh2_sw2_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:29:44.638642+00:00","log":"","correctness":{"max_relative_error":0.04972899797925799,"max_absolute_error":0.0000362396240234375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":92735607,"p5_ns":92737861,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":240792185,"instructions":314821383,"ipc":1.3074,"cache_misses":105494.36,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin128_cout256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"5ff57762aad35880848ad483fc149572","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh2_sw2_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:13:23.255816+00:00","log":"","correctness":{"max_relative_error":0.21477536152947943,"max_absolute_error":0.00003814697265625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":56594451,"p5_ns":56605524,"reference_min_ns":25595512,"speedup":0.4543189350668994,"repeat":50,"warmup":5,"cycles":146825824,"instructions":547541035,"ipc":3.7292,"cache_misses":15542572.28,"reference_cycles":66705752}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin128_cout256","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"15748e6d15d054d292c8942325690e9c","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh2_sw2_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:13:23.255816+00:00","log":"","correctness":{"max_relative_error":0.006879078364727525,"max_absolute_error":0.000030517578125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":14098192,"p5_ns":14126658,"reference_min_ns":7617165,"speedup":0.5396054005792609,"repeat":50,"warmup":5,"cycles":36842122,"instructions":136933726,"ipc":3.7168,"cache_misses":3051678.3,"reference_cycles":19880208}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin128_cout256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b3d7f1a4e8585a46ae8fbbec63913dfd","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh2_sw2_cin128_cout256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:13:23.255816+00:00","log":"","correctness":{"max_relative_error":0.03749047987876084,"max_absolute_error":0.00003814697265625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":228719728,"p5_ns":229339416,"reference_min_ns":92735607,"speedup":0.40001870812659346,"repeat":50,"warmup":5,"cycles":601952309,"instructions":2189869140,"ipc":3.6379,"cache_misses":84999182.62,"reference_cycles":240792185}}} diff --git a/traces/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin32_cout64.jsonl b/traces/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin32_cout64.jsonl deleted file mode 100644 index 5994be25cfb51e8bbb6661017683e0b12eff3888..0000000000000000000000000000000000000000 --- a/traces/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin32_cout64.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin32_cout64","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"044007b2e13158ec990f463b57d0061d","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh2_sw2_cin32_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:29:54.074188+00:00","log":"","correctness":{"max_relative_error":0.05619021815550396,"max_absolute_error":9.5367431640625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":23762761,"p5_ns":23785070,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":61765577,"instructions":75188860,"ipc":1.2173,"cache_misses":294242.14,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin32_cout64","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"15bc416b1e1e59eea11239db250d6cae","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh2_sw2_cin32_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:29:54.074188+00:00","log":"","correctness":{"max_relative_error":0.057945404928126366,"max_absolute_error":9.059906005859375e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":5845817,"p5_ns":5851786,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":15231684,"instructions":19945313,"ipc":1.3095,"cache_misses":9573.1,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin32_cout64","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"1f34a863a180506db9e5909d6e67dae2","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh2_sw2_cin32_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:29:54.074188+00:00","log":"","correctness":{"max_relative_error":0.012502523165049849,"max_absolute_error":7.62939453125e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1602601,"p5_ns":1604635,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":4186898,"instructions":9209242,"ipc":2.1995,"cache_misses":2949.14,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin32_cout64","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"044007b2e13158ec990f463b57d0061d","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh2_sw2_cin32_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:13:40.521217+00:00","log":"","correctness":{"max_relative_error":0.05619021815550396,"max_absolute_error":0.000011444091796875,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":48731642,"p5_ns":48886239,"reference_min_ns":23762761,"speedup":0.46277592667887624,"repeat":50,"warmup":5,"cycles":133467567,"instructions":550652782,"ipc":4.1257,"cache_misses":6602947.78,"reference_cycles":61765577}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin32_cout64","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"15bc416b1e1e59eea11239db250d6cae","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh2_sw2_cin32_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:13:40.521217+00:00","log":"","correctness":{"max_relative_error":0.06650423515968992,"max_absolute_error":9.5367431640625e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":10987239,"p5_ns":11036570,"reference_min_ns":5845817,"speedup":0.5298546153409451,"repeat":50,"warmup":5,"cycles":28746912,"instructions":137695556,"ipc":4.7899,"cache_misses":1274872.42,"reference_cycles":15231684}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin32_cout64","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"1f34a863a180506db9e5909d6e67dae2","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh2_sw2_cin32_cout64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:13:40.521217+00:00","log":"","correctness":{"max_relative_error":0.023483491281542708,"max_absolute_error":7.152557373046875e-6,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2709946,"p5_ns":2714062,"reference_min_ns":1602601,"speedup":0.5905534256529851,"repeat":50,"warmup":5,"cycles":7089787,"instructions":34449787,"ipc":4.8591,"cache_misses":207673.42,"reference_cycles":4186898}}} diff --git a/traces/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin512_cout512.jsonl b/traces/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin512_cout512.jsonl deleted file mode 100644 index bf7ea08cad2e8a4e23ee0147d0710a10a53ce468..0000000000000000000000000000000000000000 --- a/traces/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin512_cout512.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin512_cout512","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b3693969f85550cbb0492e8f94b24747","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh2_sw2_cin512_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:29:58.118669+00:00","log":"","correctness":{"max_relative_error":0.06423373472814564,"max_absolute_error":0.0001373291015625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":59518411,"p5_ns":59582452,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":155095511,"instructions":587380684,"ipc":3.7872,"cache_misses":92882.5,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin512_cout512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"140cb7f0f21e5ccd9c00e5864aad30f1","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh2_sw2_cin512_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:29:58.118669+00:00","log":"","correctness":{"max_relative_error":4.778336361028995,"max_absolute_error":0.000141143798828125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":204207719,"p5_ns":204287894,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":530347196,"instructions":1165131539,"ipc":2.1969,"cache_misses":239409.86,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin512_cout512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"d0af6900a2665b82a7cd797a2d9c54c9","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh2_sw2_cin512_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:29:58.118669+00:00","log":"","correctness":{"max_relative_error":0.2287376863046389,"max_absolute_error":0.000152587890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":742547859,"p5_ns":742728723,"reference_min_ns":null,"speedup":null,"repeat":41,"warmup":5,"cycles":1926545832,"instructions":2510546510,"ipc":1.3031,"cache_misses":615840.6097560975,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin512_cout512","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b3693969f85550cbb0492e8f94b24747","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh2_sw2_cin512_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:13:44.382565+00:00","log":"","correctness":{"max_relative_error":0.06423373472814564,"max_absolute_error":0.0001373291015625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":112284767,"p5_ns":112324867,"reference_min_ns":59518411,"speedup":0.5319156719904381,"repeat":50,"warmup":5,"cycles":291579134,"instructions":1081126273,"ipc":3.7078,"cache_misses":11782680.66,"reference_cycles":155095511}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin512_cout512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"140cb7f0f21e5ccd9c00e5864aad30f1","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh2_sw2_cin512_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:13:44.382565+00:00","log":"","correctness":{"max_relative_error":4.412686591825233,"max_absolute_error":0.000141143798828125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":495325062,"p5_ns":495352317,"reference_min_ns":204207719,"speedup":0.4129813159840571,"repeat":50,"warmup":5,"cycles":1284191743,"instructions":4324243115,"ipc":3.3673,"cache_misses":128363493.86,"reference_cycles":530347196}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin512_cout512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"d0af6900a2665b82a7cd797a2d9c54c9","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh2_sw2_cin512_cout512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:13:44.382565+00:00","log":"","correctness":{"max_relative_error":0.4159487259298253,"max_absolute_error":0.000156402587890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2509946480,"p5_ns":2509946480,"reference_min_ns":742547859,"speedup":0.2941028264626929,"repeat":12,"warmup":5,"cycles":6550585913,"instructions":17299865109,"ipc":2.641,"cache_misses":995996175.1666666,"reference_cycles":1926545832}}} diff --git a/traces/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin64_cout128.jsonl b/traces/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin64_cout128.jsonl deleted file mode 100644 index 888b7a95b04fecd6bab8a03f08509105c639635c..0000000000000000000000000000000000000000 --- a/traces/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin64_cout128.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin64_cout128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"081d1b988b495ddcb2d502a7ede5b7ce","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh2_sw2_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:30:50.752055+00:00","log":"","correctness":{"max_relative_error":0.07078972165407639,"max_absolute_error":0.000019073486328125,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":23266429,"p5_ns":23269470,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":60435370,"instructions":79048762,"ipc":1.308,"cache_misses":35315.84,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin64_cout128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"9df9693d73905250bae823389d6dbf40","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh2_sw2_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:30:50.752055+00:00","log":"","correctness":{"max_relative_error":0.049958992214649944,"max_absolute_error":0.0000152587890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":6414974,"p5_ns":6416118,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":16700970,"instructions":36590568,"ipc":2.1909,"cache_misses":8778.22,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin64_cout128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"713dfd524ac65093a33658a678099203","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_kh4_kw4_sh2_sw2_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:30:50.752055+00:00","log":"","correctness":{"max_relative_error":1.0484463393898396,"max_absolute_error":0.0000247955322265625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":95017089,"p5_ns":95034014,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":246495171,"instructions":298188442,"ipc":1.2097,"cache_misses":1172980.46,"reference_cycles":null}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin64_cout128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"081d1b988b495ddcb2d502a7ede5b7ce","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh2_sw2_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:15:04.709490+00:00","log":"","correctness":{"max_relative_error":0.07078972165407639,"max_absolute_error":0.0000209808349609375,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":45176083,"p5_ns":45272889,"reference_min_ns":23266429,"speedup":0.512519606606487,"repeat":50,"warmup":5,"cycles":117918162,"instructions":548504223,"ipc":4.6516,"cache_misses":4412124.42,"reference_cycles":60435370}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin64_cout128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"9df9693d73905250bae823389d6dbf40","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh2_sw2_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:15:04.709490+00:00","log":"","correctness":{"max_relative_error":0.04641905242521828,"max_absolute_error":0.0000152587890625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":11082160,"p5_ns":11084223,"reference_min_ns":6414974,"speedup":0.5781038340180985,"repeat":50,"warmup":5,"cycles":28889222,"instructions":137171671,"ipc":4.7482,"cache_misses":1106032.2,"reference_cycles":16700970}}} -{"definition":"deconv2d_kh4_kw4_sh2_sw2_cin64_cout128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"713dfd524ac65093a33658a678099203","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_kh4_kw4_sh2_sw2_cin64_cout128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T03:15:04.709490+00:00","log":"","correctness":{"max_relative_error":1.0484463393898396,"max_absolute_error":0.0000247955322265625,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":254810990,"p5_ns":258127750,"reference_min_ns":95017089,"speedup":0.3595547881181913,"repeat":50,"warmup":5,"cycles":685556636,"instructions":2194319868,"ipc":3.2008,"cache_misses":27500078.98,"reference_cycles":246495171}}} diff --git a/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024.jsonl b/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024.jsonl deleted file mode 100644 index 78f65856ee29da2d095934eb29e2cdc03d2acdea..0000000000000000000000000000000000000000 --- a/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"0dcbf44bb78f59c6a3bd009cbac9c31b","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:00.100807+00:00","log":"","correctness":{"max_relative_error":0.0009769215303308825,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":8092813,"p5_ns":8107321,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":21167157,"instructions":69612945,"ipc":3.2887,"cache_misses":30419.84,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"59aefe54d11e59b4b1ddc44557dbc71e","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:00.100807+00:00","log":"","correctness":{"max_relative_error":0.011475245157877604,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":32603299,"p5_ns":32630543,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":85010406,"instructions":278120173,"ipc":3.2716,"cache_misses":140803.08,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"46c3e8394950585db030c655c26ded31","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:00.100807+00:00","log":"","correctness":{"max_relative_error":0.07719504833221436,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":149719749,"p5_ns":150190603,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":390324668,"instructions":1183905073,"ipc":3.0331,"cache_misses":929553.32,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"0dcbf44bb78f59c6a3bd009cbac9c31b","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:27.484963+00:00","log":"","correctness":{"max_relative_error":0.0009769215303308825,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":389379,"p5_ns":389565,"reference_min_ns":8092813,"speedup":20.395965161286593,"repeat":50,"warmup":5,"cycles":1037811,"instructions":6022102,"ipc":5.8027,"cache_misses":1268.14,"reference_cycles":21167157}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"59aefe54d11e59b4b1ddc44557dbc71e","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:27.484963+00:00","log":"","correctness":{"max_relative_error":0.011475245157877604,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1561654,"p5_ns":1568897,"reference_min_ns":32603299,"speedup":20.519050571820074,"repeat":50,"warmup":5,"cycles":4142999,"instructions":23649546,"ipc":5.7083,"cache_misses":3333.98,"reference_cycles":85010406}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"46c3e8394950585db030c655c26ded31","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:27.484963+00:00","log":"","correctness":{"max_relative_error":0.07719504833221436,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":6351304,"p5_ns":6364729,"reference_min_ns":149719749,"speedup":23.351106173036854,"repeat":50,"warmup":5,"cycles":16715468,"instructions":93845664,"ipc":5.6143,"cache_misses":256000.52,"reference_cycles":390324668}}} diff --git a/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c128.jsonl b/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c128.jsonl deleted file mode 100644 index 6e367f2cc96d14dc0de3926ad1d116c795d3198d..0000000000000000000000000000000000000000 --- a/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c128.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"804046907b3858dca08617f94d7c6729","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:13.678266+00:00","log":"","correctness":{"max_relative_error":0.00337376314051011,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":4048316,"p5_ns":4055961,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":10615614,"instructions":34782283,"ipc":3.2765,"cache_misses":7978.16,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"8254a10d459951308de23684f9e5a924","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:13.678266+00:00","log":"","correctness":{"max_relative_error":0.0005283665012668919,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1007772,"p5_ns":1009160,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":2637026,"instructions":8718377,"ipc":3.3061,"cache_misses":1621.4,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"fd9b08900a38550ba0e42cab3ec8855d","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:13.678266+00:00","log":"","correctness":{"max_relative_error":0.02917783400591682,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":16207754,"p5_ns":16273049,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":42466062,"instructions":139038556,"ipc":3.2741,"cache_misses":73034.42,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"804046907b3858dca08617f94d7c6729","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:28.292483+00:00","log":"","correctness":{"max_relative_error":0.00337376314051011,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":193828,"p5_ns":194845,"reference_min_ns":4048316,"speedup":20.41275805310279,"repeat":50,"warmup":5,"cycles":520048,"instructions":2966121,"ipc":5.7036,"cache_misses":905.6,"reference_cycles":10615614}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"8254a10d459951308de23684f9e5a924","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:28.292483+00:00","log":"","correctness":{"max_relative_error":0.0005283665012668919,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":49558,"p5_ns":49578,"reference_min_ns":1007772,"speedup":19.750784556042394,"repeat":50,"warmup":5,"cycles":133515,"instructions":762430,"ipc":5.7105,"cache_misses":434.14,"reference_cycles":2637026}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"fd9b08900a38550ba0e42cab3ec8855d","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:28.292483+00:00","log":"","correctness":{"max_relative_error":0.02917783400591682,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":787053,"p5_ns":788505,"reference_min_ns":16207754,"speedup":20.351778633395348,"repeat":50,"warmup":5,"cycles":2086602,"instructions":11740779,"ipc":5.6267,"cache_misses":30731.78,"reference_cycles":42466062}}} diff --git a/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c256.jsonl b/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c256.jsonl deleted file mode 100644 index 80e97c0ef287fc7921de9b3b167080654b8e8e5c..0000000000000000000000000000000000000000 --- a/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c256.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"d432881dc24254099a5d4c349eeac2e4","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:17.167055+00:00","log":"","correctness":{"max_relative_error":0.0012561285007766812,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2008723,"p5_ns":2010880,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":5261913,"instructions":17417088,"ipc":3.31,"cache_misses":2962.0,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"1b7ab7d018a9567396c51468cb8c17e5","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:17.167055+00:00","log":"","correctness":{"max_relative_error":0.2827104330062866,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":8112152,"p5_ns":8128739,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":21222385,"instructions":69547507,"ipc":3.2771,"cache_misses":26164.18,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a96dbf5e240d51b0bee1c7a164395aef","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:17.167055+00:00","log":"","correctness":{"max_relative_error":0.05534350872039795,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":32366149,"p5_ns":32431681,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":84531522,"instructions":278057388,"ipc":3.2894,"cache_misses":144588.9,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c256","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"d432881dc24254099a5d4c349eeac2e4","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:28.519982+00:00","log":"","correctness":{"max_relative_error":0.0012561285007766812,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":98687,"p5_ns":98724,"reference_min_ns":2008723,"speedup":19.838607881283085,"repeat":50,"warmup":5,"cycles":265236,"instructions":1513637,"ipc":5.7068,"cache_misses":531.84,"reference_cycles":5261913}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c256","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"1b7ab7d018a9567396c51468cb8c17e5","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:28.519982+00:00","log":"","correctness":{"max_relative_error":0.2827104330062866,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":387811,"p5_ns":389012,"reference_min_ns":8112152,"speedup":20.55482591864032,"repeat":50,"warmup":5,"cycles":1032477,"instructions":5920681,"ipc":5.7344,"cache_misses":1313.28,"reference_cycles":21222385}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c256","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"a96dbf5e240d51b0bee1c7a164395aef","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c256","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:28.519982+00:00","log":"","correctness":{"max_relative_error":0.05534350872039795,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1768353,"p5_ns":1774965,"reference_min_ns":32366149,"speedup":16.7720974307784,"repeat":50,"warmup":5,"cycles":5040009,"instructions":23475986,"ipc":4.6579,"cache_misses":75791.02,"reference_cycles":84531522}}} diff --git a/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c32.jsonl b/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c32.jsonl deleted file mode 100644 index 820bdd544159aef408e70adabdf8066009b21826..0000000000000000000000000000000000000000 --- a/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c32.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c32","workload":{"axes":{"N":1,"H":112,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"841062afb6c956a29f615060215c02bc","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:21.880743+00:00","log":"","correctness":{"max_relative_error":0.004878666089928668,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":16301818,"p5_ns":16329455,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":42528974,"instructions":139030624,"ipc":3.2691,"cache_misses":74870.02,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c32","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"84b088cc1a53505c91cd3ac3e053efdc","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:21.880743+00:00","log":"","correctness":{"max_relative_error":0.0008380690731676707,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1015807,"p5_ns":1017934,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":2656017,"instructions":8710069,"ipc":3.2794,"cache_misses":1664.0,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c32","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b6000adadce75cdebc308dfde3c20771","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:21.880743+00:00","log":"","correctness":{"max_relative_error":0.0004557483601120283,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":250474,"p5_ns":250490,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":659294,"instructions":2193950,"ipc":3.3277,"cache_misses":883.66,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c32","workload":{"axes":{"N":1,"H":112,"W":112},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"841062afb6c956a29f615060215c02bc","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:28.846920+00:00","log":"","correctness":{"max_relative_error":0.004878666089928668,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":955576,"p5_ns":977482,"reference_min_ns":16301818,"speedup":15.464446006400465,"repeat":50,"warmup":5,"cycles":2750113,"instructions":11700067,"ipc":4.2544,"cache_misses":44816.82,"reference_cycles":42528974}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c32","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"84b088cc1a53505c91cd3ac3e053efdc","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:28.846920+00:00","log":"","correctness":{"max_relative_error":0.0008380690731676707,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":48685,"p5_ns":48695,"reference_min_ns":1015807,"speedup":20.42335906742126,"repeat":50,"warmup":5,"cycles":130048,"instructions":749595,"ipc":5.764,"cache_misses":415.9,"reference_cycles":2656017}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c32","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"b6000adadce75cdebc308dfde3c20771","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:28.846920+00:00","log":"","correctness":{"max_relative_error":0.0004557483601120283,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":13031,"p5_ns":13043,"reference_min_ns":250474,"speedup":17.932165587771312,"repeat":50,"warmup":5,"cycles":36766,"instructions":198884,"ipc":5.4094,"cache_misses":398.08,"reference_cycles":659294}}} diff --git a/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c512.jsonl b/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c512.jsonl deleted file mode 100644 index 5153f176d2bc9c49e080f9f2b31298138d762b6f..0000000000000000000000000000000000000000 --- a/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c512.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c512","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"92ea00f8c83d50eba7904087c4466d04","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:25.156759+00:00","log":"","correctness":{"max_relative_error":0.0002657306728078358,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":997410,"p5_ns":998865,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":2598471,"instructions":8749777,"ipc":3.3673,"cache_misses":1721.36,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"88937b0830cc52278bdaf72fc6f14e80","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:25.156759+00:00","log":"","correctness":{"max_relative_error":0.3891468048095703,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":16201720,"p5_ns":16254372,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":42347607,"instructions":139071704,"ipc":3.2841,"cache_misses":78213.84,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"1df1501cab9c53219b8f8fab2261e690","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:25.156759+00:00","log":"","correctness":{"max_relative_error":0.00211126955947435,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":4017477,"p5_ns":4027181,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":10526899,"instructions":34814661,"ipc":3.3072,"cache_misses":8408.3,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c512","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"92ea00f8c83d50eba7904087c4466d04","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:29.084363+00:00","log":"","correctness":{"max_relative_error":0.0002657306728078358,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":49172,"p5_ns":49197,"reference_min_ns":997410,"speedup":19.847019285850678,"repeat":50,"warmup":5,"cycles":130925,"instructions":799708,"ipc":6.1081,"cache_misses":409.84,"reference_cycles":2598471}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"88937b0830cc52278bdaf72fc6f14e80","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:29.084363+00:00","log":"","correctness":{"max_relative_error":0.3891468048095703,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":779904,"p5_ns":785657,"reference_min_ns":16201720,"speedup":20.34715945290676,"repeat":50,"warmup":5,"cycles":2081254,"instructions":11830530,"ipc":5.6843,"cache_misses":2065.82,"reference_cycles":42347607}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"1df1501cab9c53219b8f8fab2261e690","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:29.084363+00:00","log":"","correctness":{"max_relative_error":0.00211126955947435,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":194425,"p5_ns":194693,"reference_min_ns":4017477,"speedup":20.304246438477175,"repeat":50,"warmup":5,"cycles":518458,"instructions":3016453,"ipc":5.8181,"cache_misses":876.2,"reference_cycles":10526899}}} diff --git a/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c64.jsonl b/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c64.jsonl deleted file mode 100644 index 58ffaca45af7e36f005de30b4fa4bd2fbd79b035..0000000000000000000000000000000000000000 --- a/traces/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c64.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c64","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"df7119dd1a2358c4a82a212336455e32","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:28.655034+00:00","log":"","correctness":{"max_relative_error":0.0037953123754384567,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":8163575,"p5_ns":8189300,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":21349594,"instructions":69528753,"ipc":3.2567,"cache_misses":27161.04,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c64","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"23a01e83413a57af9f5971b2ef24b5bb","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:28.655034+00:00","log":"","correctness":{"max_relative_error":0.0005732094369283536,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":2024495,"p5_ns":2026691,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":5289621,"instructions":17400786,"ipc":3.2896,"cache_misses":3000.02,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c64","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"5ec8dfa7bd225537bddc52c1b1e6dc66","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh2_kw2_sh2_sw2_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:28.655034+00:00","log":"","correctness":{"max_relative_error":0.00011918364816164561,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":500049,"p5_ns":500559,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1314142,"instructions":4368968,"ipc":3.3246,"cache_misses":1142.16,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c64","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"df7119dd1a2358c4a82a212336455e32","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:29.317138+00:00","log":"","correctness":{"max_relative_error":0.0037953123754384567,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":394635,"p5_ns":395493,"reference_min_ns":8163575,"speedup":20.201253731625734,"repeat":50,"warmup":5,"cycles":1056845,"instructions":5876095,"ipc":5.56,"cache_misses":16651.68,"reference_cycles":21349594}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c64","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"23a01e83413a57af9f5971b2ef24b5bb","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:29.317138+00:00","log":"","correctness":{"max_relative_error":0.0005732094369283536,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":96716,"p5_ns":96922,"reference_min_ns":2024495,"speedup":20.274282012855352,"repeat":50,"warmup":5,"cycles":260903,"instructions":1488412,"ipc":5.7049,"cache_misses":3735.3,"reference_cycles":5289621}}} -{"definition":"deconv2d_depthwise_kh2_kw2_sh2_sw2_c64","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"5ec8dfa7bd225537bddc52c1b1e6dc66","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh2_kw2_sh2_sw2_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:29.317138+00:00","log":"","correctness":{"max_relative_error":0.00011918364816164561,"max_absolute_error":1.1920928955078125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":25439,"p5_ns":25444,"reference_min_ns":500049,"speedup":18.98527860847455,"repeat":50,"warmup":5,"cycles":69219,"instructions":386761,"ipc":5.5875,"cache_misses":406.78,"reference_cycles":1314142}}} diff --git a/traces/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c128.jsonl b/traces/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c128.jsonl deleted file mode 100644 index 9163cfc3b864749a97a11d8e5089162cc46a6bd7..0000000000000000000000000000000000000000 --- a/traces/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c128.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"4ba29780f29a5e91914dae8b36dc0bd3","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:31.521290+00:00","log":"","correctness":{"max_relative_error":0.0008753921046401515,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":3268964,"p5_ns":3274735,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":8529110,"instructions":24523490,"ipc":2.8753,"cache_misses":9600.68,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"dc3f4d9557ea5826813c4313c30ba029","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:31.521290+00:00","log":"","correctness":{"max_relative_error":0.0006004609128376008,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":853557,"p5_ns":857049,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":2253329,"instructions":6577117,"ipc":2.9188,"cache_misses":2944.28,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"e6b38f97acd7523081b87ed332a21f9e","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:31.521290+00:00","log":"","correctness":{"max_relative_error":0.015288619307784346,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":12152379,"p5_ns":12156642,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":31644102,"instructions":94680129,"ipc":2.992,"cache_misses":46130.66,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c128","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"4ba29780f29a5e91914dae8b36dc0bd3","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:29.507090+00:00","log":"","correctness":{"max_relative_error":0.0008753921046401515,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":385988,"p5_ns":386022,"reference_min_ns":3268964,"speedup":8.405888086539418,"repeat":50,"warmup":5,"cycles":1014659,"instructions":5982923,"ipc":5.8965,"cache_misses":671.42,"reference_cycles":8529110}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c128","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"dc3f4d9557ea5826813c4313c30ba029","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:29.507090+00:00","log":"","correctness":{"max_relative_error":0.0006004609128376008,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":96685,"p5_ns":96698,"reference_min_ns":853557,"speedup":8.719132470447114,"repeat":50,"warmup":5,"cycles":258435,"instructions":1536730,"ipc":5.9463,"cache_misses":505.24,"reference_cycles":2253329}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c128","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"e6b38f97acd7523081b87ed332a21f9e","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c128","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:29.507090+00:00","log":"","correctness":{"max_relative_error":0.015288619307784346,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1864963,"p5_ns":1868469,"reference_min_ns":12152379,"speedup":6.497611451726069,"repeat":50,"warmup":5,"cycles":4870113,"instructions":27956748,"ipc":5.7405,"cache_misses":2021.68,"reference_cycles":31644102}}} diff --git a/traces/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048.jsonl b/traces/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048.jsonl deleted file mode 100644 index 11aaaa9f73ef7eecabc0445e67dcb7b957313059..0000000000000000000000000000000000000000 --- a/traces/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"f58a8438eaa2511e96daf4814339b033","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:34.733207+00:00","log":"","correctness":{"max_relative_error":0.012588644927402711,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":3786712,"p5_ns":3795327,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":10044819,"instructions":29953128,"ipc":2.9819,"cache_misses":20831.62,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"69ec210dc1c25b908213f6d32d96ade8","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:34.733207+00:00","log":"","correctness":{"max_relative_error":0.018513315092257367,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":51130344,"p5_ns":51131974,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":133000396,"instructions":392020041,"ipc":2.9475,"cache_misses":202353.2,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"e71f316e44305e0797f3617beaef5b7b","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:34.733207+00:00","log":"","correctness":{"max_relative_error":0.029899966332220262,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":13749541,"p5_ns":13763279,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":36030716,"instructions":104941737,"ipc":2.9126,"cache_misses":51902.4,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048","workload":{"axes":{"N":1,"H":7,"W":7},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"f58a8438eaa2511e96daf4814339b033","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:29.973251+00:00","log":"","correctness":{"max_relative_error":0.012588644927402711,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":396383,"p5_ns":396551,"reference_min_ns":3786712,"speedup":9.572994805043805,"repeat":50,"warmup":5,"cycles":1049287,"instructions":6412743,"ipc":6.1115,"cache_misses":745.46,"reference_cycles":10044819}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"69ec210dc1c25b908213f6d32d96ade8","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:29.973251+00:00","log":"","correctness":{"max_relative_error":0.018513315092257367,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":6166764,"p5_ns":6175665,"reference_min_ns":51130344,"speedup":8.270358328064363,"repeat":50,"warmup":5,"cycles":16081576,"instructions":95558753,"ipc":5.9421,"cache_misses":5239.86,"reference_cycles":133000396}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"e71f316e44305e0797f3617beaef5b7b","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:29.973251+00:00","log":"","correctness":{"max_relative_error":0.029899966332220262,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1533409,"p5_ns":1534454,"reference_min_ns":13749541,"speedup":8.995020757039656,"repeat":50,"warmup":5,"cycles":4005629,"instructions":24363032,"ipc":6.0822,"cache_misses":1204.48,"reference_cycles":36030716}}} diff --git a/traces/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c32.jsonl b/traces/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c32.jsonl deleted file mode 100644 index d56d84d8f9251d580a815c6642cc769cab9e4d4e..0000000000000000000000000000000000000000 --- a/traces/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c32.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c32","workload":{"axes":{"N":1,"H":224,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"ed78fc8321d557658c11ba75ccbeadfe","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:40.896949+00:00","log":"","correctness":{"max_relative_error":704.918790006559,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":48054636,"p5_ns":48101678,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":127871124,"instructions":368886474,"ipc":2.8848,"cache_misses":180217.8,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c32","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"58eec703e91c5c9db79875ff13e25553","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:40.896949+00:00","log":"","correctness":{"max_relative_error":0.00026129042332952596,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":798317,"p5_ns":799765,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":2090675,"instructions":6144654,"ipc":2.9391,"cache_misses":2774.78,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c32","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"00d4ec44c2da523c83e6aa8ea3e90f37","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:40.896949+00:00","log":"","correctness":{"max_relative_error":0.0001343470982142857,"max_absolute_error":4.76837158203125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":214545,"p5_ns":214829,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":566372,"instructions":1659231,"ipc":2.9296,"cache_misses":950.8,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c32","workload":{"axes":{"N":1,"H":224,"W":224},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"ed78fc8321d557658c11ba75ccbeadfe","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:30.770946+00:00","log":"","correctness":{"max_relative_error":704.918790006559,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":6525449,"p5_ns":6528044,"reference_min_ns":48054636,"speedup":7.521976395896132,"repeat":50,"warmup":5,"cycles":16999671,"instructions":98428368,"ipc":5.79,"cache_misses":5816.46,"reference_cycles":127871124}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c32","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"58eec703e91c5c9db79875ff13e25553","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:30.770946+00:00","log":"","correctness":{"max_relative_error":0.00026129042332952596,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":97293,"p5_ns":97337,"reference_min_ns":798317,"speedup":8.134257512032091,"repeat":50,"warmup":5,"cycles":257021,"instructions":1504002,"ipc":5.8517,"cache_misses":442.26,"reference_cycles":2090675}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c32","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"00d4ec44c2da523c83e6aa8ea3e90f37","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c32","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:30.770946+00:00","log":"","correctness":{"max_relative_error":0.0001343470982142857,"max_absolute_error":4.76837158203125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":24964,"p5_ns":25019,"reference_min_ns":214545,"speedup":8.354185411903533,"repeat":50,"warmup":5,"cycles":67795,"instructions":391472,"ipc":5.7744,"cache_misses":348.26,"reference_cycles":566372}}} diff --git a/traces/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c512.jsonl b/traces/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c512.jsonl deleted file mode 100644 index 5f480413204881707d6a502b10f315777f5f75a6..0000000000000000000000000000000000000000 --- a/traces/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c512.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"01d3d28db36455eba8e43f1a509d38f0","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:46.008644+00:00","log":"","correctness":{"max_relative_error":0.002594402897014819,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":3426208,"p5_ns":3426634,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":8925176,"instructions":26249602,"ipc":2.9411,"cache_misses":10693.22,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"3d9f21bc00145659b3c3346afd3dd9ad","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:46.008644+00:00","log":"","correctness":{"max_relative_error":0.015220226545479302,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":12769516,"p5_ns":12770387,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":33181094,"instructions":98020605,"ipc":2.9541,"cache_misses":48569.04,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c512","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"74f2a6991a295553888b63da9d138ab7","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:46.008644+00:00","log":"","correctness":{"max_relative_error":0.03227555124383224,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":48543059,"p5_ns":48575125,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":126261315,"instructions":378667971,"ipc":2.9991,"cache_misses":190588.04,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c512","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"01d3d28db36455eba8e43f1a509d38f0","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:31.485424+00:00","log":"","correctness":{"max_relative_error":0.002594402897014819,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":383206,"p5_ns":383419,"reference_min_ns":3426208,"speedup":8.805682883273972,"repeat":50,"warmup":5,"cycles":1013570,"instructions":6101249,"ipc":6.0196,"cache_misses":609.42,"reference_cycles":8925176}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c512","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"3d9f21bc00145659b3c3346afd3dd9ad","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:31.485424+00:00","log":"","correctness":{"max_relative_error":0.015220226545479302,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1541792,"p5_ns":1543075,"reference_min_ns":12769516,"speedup":8.243340455132664,"repeat":50,"warmup":5,"cycles":4025200,"instructions":23902030,"ipc":5.9381,"cache_misses":1671.58,"reference_cycles":33181094}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c512","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"74f2a6991a295553888b63da9d138ab7","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c512","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:31.485424+00:00","log":"","correctness":{"max_relative_error":0.03227555124383224,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":7482061,"p5_ns":7488214,"reference_min_ns":48543059,"speedup":6.482146713078614,"repeat":50,"warmup":5,"cycles":19478318,"instructions":111792815,"ipc":5.7393,"cache_misses":7081.48,"reference_cycles":126261315}}} diff --git a/traces/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c64.jsonl b/traces/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c64.jsonl deleted file mode 100644 index aef9081e25a546aa77cad0ad1a5ad81fbee75d17..0000000000000000000000000000000000000000 --- a/traces/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c64.jsonl +++ /dev/null @@ -1,6 +0,0 @@ -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c64","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"5b2f0118d2a75f6189bc23ec7b513450","tags":{"from":"gen_definitions"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:51.954965+00:00","log":"","correctness":{"max_relative_error":0.002521626895779361,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":6187839,"p5_ns":6193282,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":16322493,"instructions":47351320,"ipc":2.901,"cache_misses":21222.12,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c64","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"9ac0c99c519256eda187e71998d681f2","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:51.954965+00:00","log":"","correctness":{"max_relative_error":0.005900182029753353,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":1633036,"p5_ns":1635618,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":4261344,"instructions":12269654,"ipc":2.8793,"cache_misses":4944.68,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c64","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"2b63b0d83ec351888b793dd003ca8c80","tags":{"from":"gen_workload"},"description":null},"solution":"baseline-ncnn-arm_deconv2d_depthwise_kh3_kw3_sh1_sw1_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:31:51.954965+00:00","log":"","correctness":{"max_relative_error":0.00041428830481224127,"max_absolute_error":4.76837158203125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":427948,"p5_ns":428470,"reference_min_ns":null,"speedup":null,"repeat":50,"warmup":5,"cycles":1125803,"instructions":3298416,"ipc":2.9298,"cache_misses":1749.2,"reference_cycles":null}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c64","workload":{"axes":{"N":1,"H":56,"W":56},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"5b2f0118d2a75f6189bc23ec7b513450","tags":{"from":"gen_definitions"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:32.357799+00:00","log":"","correctness":{"max_relative_error":0.002521626895779361,"max_absolute_error":7.152557373046875e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":930962,"p5_ns":933643,"reference_min_ns":6187839,"speedup":6.697456996337477,"repeat":50,"warmup":5,"cycles":2437118,"instructions":13983729,"ipc":5.7378,"cache_misses":1147.02,"reference_cycles":16322493}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c64","workload":{"axes":{"N":1,"H":28,"W":28},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"9ac0c99c519256eda187e71998d681f2","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:32.357799+00:00","log":"","correctness":{"max_relative_error":0.005900182029753353,"max_absolute_error":9.5367431640625e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":192589,"p5_ns":192644,"reference_min_ns":1633036,"speedup":8.384973200160562,"repeat":50,"warmup":5,"cycles":508212,"instructions":2997402,"ipc":5.8979,"cache_misses":510.92,"reference_cycles":4261344}}} -{"definition":"deconv2d_depthwise_kh3_kw3_sh1_sw1_c64","workload":{"axes":{"N":1,"H":14,"W":14},"inputs":{"input":{"type":"random"},"weight":{"type":"random"},"bias":{"type":"random"}},"uuid":"2b63b0d83ec351888b793dd003ca8c80","tags":{"from":"gen_workload"},"description":null},"solution":"reference-scalar_deconv2d_depthwise_kh3_kw3_sh1_sw1_c64","evaluation":{"status":"PASSED","environment":{"hardware":"aarch64-ip-172-31-59-130","cpu_pinned":0,"libs":{"python":"3.12.3","numpy":"2.4.6"}},"timestamp":"2026-06-15T02:59:32.357799+00:00","log":"","correctness":{"max_relative_error":0.00041428830481224127,"max_absolute_error":4.76837158203125e-7,"matched_ratio":1.0,"extra":null},"performance":{"min_ns":48702,"p5_ns":48726,"reference_min_ns":427948,"speedup":8.621821774292366,"repeat":50,"warmup":5,"cycles":130576,"instructions":772223,"ipc":5.914,"cache_misses":419.32,"reference_cycles":1125803}}} diff --git a/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024.jsonl b/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024.jsonl deleted file mode 100644 index 255c38737269e55f348c803055cf3830c759d48a..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c1024_c1024.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "46b33064e41459439fe192157458db95", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 7, "W": 7}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "882b121891315376bd9c3827d12370e4", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "77361d3aaef253c59d648320dd79e609", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512.jsonl b/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512.jsonl deleted file mode 100644 index 67b6ddc4c0d85507a64c287c343afc057d280720..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c128_c512.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "d3af4ea9ebca51fc9664b5f13d11c5b5", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "90d0153e72ba5f539e78f061bcde1002", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 112, "W": 112}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "ee4b3270a4ed549b8c4ac9e46a5377b6", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024.jsonl b/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024.jsonl deleted file mode 100644 index 6d316a75a60ef14d6011e732cc2784ba8bdc1a9c..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c256_c1024.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "d285be818f9a5f05a0cd622e3d25e1a1", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "931d27e9dadb56f0af154539d71cfe33", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "76b82d9b58f650d99a586c8feae88d3d", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048.jsonl b/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048.jsonl deleted file mode 100644 index 683299be768121460d29fe98cb952ed310f7e755..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c512_c2048.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 7, "W": 7}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "2ad67410f0d0552f8f6ae305f25cefe6", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "ba88bbd3da315b109da895c94572c808", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "1fc676a886ba582789004ddf0cbb2e2e", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256.jsonl b/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256.jsonl deleted file mode 100644 index 218f7ed411d2ea8e13dae5d5f3c6c168564d8108..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c256.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "d604a2a8c81a52d1a6e45f2c3543d265", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "751bd86036f055c2857b01ef4f11ee36", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 112, "W": 112}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "406890d52a18522aaeaa73fa1509f726", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64.jsonl b/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64.jsonl deleted file mode 100644 index 9dff76dfced2ac814d6367917204d18df2e7c1ce..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh1_kw1_sh1_sw1_dh1_dw1_c64_c64.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 224, "W": 224}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "e6104e9e54285584a3e57d02298e419b", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "6841135778555a6db3b20ef4cd79cba4", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 0}, "pad_left": {"type": "scalar", "value": 0}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "6a664f25918c57fcac9147077f7e80fa", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256.jsonl b/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256.jsonl deleted file mode 100644 index e0430446860c014e89e3f73c2510bc5bf9e11fa6..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c128_c256.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 1}}, "uuid": "c6d116cfeff357a29d437f1f30f6fe1c", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 40, "W": 40}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 1}}, "uuid": "16f1968aecf05c6b9a34dc071ef248ac", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 1}}, "uuid": "a95305e76c5a57d09c094f3cc98493fb", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512.jsonl b/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512.jsonl deleted file mode 100644 index 78d615aecbf9986ce7166e123d11bd4abc2d3594..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c256_c512.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 1}}, "uuid": "9016e988b6fe5ebb83894c247f5f6ce6", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 20, "W": 20}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 1}}, "uuid": "a3653f2e63885182a71b720d78a58ee0", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 1}}, "uuid": "66cd580dc7e055ebb8582f75728bf521", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64.jsonl b/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64.jsonl deleted file mode 100644 index 9bff6c99e4eda377a9c95aa099893ac447e54c90..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c3_c64.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 224, "W": 224}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 1}}, "uuid": "4932ca9d15cd5ab58f30158b896aa9bb", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 300, "W": 300}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 1}}, "uuid": "1828114d919e55168d17e337ca88badb", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 416, "W": 416}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 1}}, "uuid": "7df9d89e21dc5806ac5f63972b145252", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024.jsonl b/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024.jsonl deleted file mode 100644 index e1f32e017bd33491f46332e47d7d45c3dac8606b..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c512_c1024.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 7, "W": 7}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 1}}, "uuid": "8038056ca85a5fccb9709a05a86f884b", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 1}}, "uuid": "b868d1dad8205e339c0e8567d69b1dab", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 20, "W": 20}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 1}}, "uuid": "5afd92a395e35213a46c0117ae3c9ccd", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128.jsonl b/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128.jsonl deleted file mode 100644 index c0e63c05e03effec919b418aa4bbe682c13df940..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh1_dw1_c64_c128.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 1}}, "uuid": "fe4e4809c710583bac68a0587e367d56", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 80, "W": 80}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 1}}, "uuid": "0c8a5295baa95c8ab9151706747aeb26", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 112, "W": 112}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 1}}, "uuid": "55596a1b8f9356198de62647e2df1dd0", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256.jsonl b/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256.jsonl deleted file mode 100644 index f022c3a64f41ae0853e49cf8be8e604f928ffc37..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c128_c256.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "43c5652744f65cc983f879b4edd5d7a2", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 64, "W": 64}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "3b6996fa809f59299e60079a141bb0e6", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "bd4a9ef2bf55585aa51996f03681fe2b", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256.jsonl b/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256.jsonl deleted file mode 100644 index 4e08d64a9613fdbcf8a55fd5ff2a1f8278b1ca24..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c256_c256.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "234b7212a9735ef8ad346f2fb7155d5e", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 64, "W": 64}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "609adf66789f5e7d8855263f5a7d17c3", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "f7240c8fff585c0e97c5ccb575b2e5d6", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512.jsonl b/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512.jsonl deleted file mode 100644 index 8a2bffcaca4e5215b7fc6ca19407fbb005dbeb53..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c512_c512.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 7, "W": 7}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "9697ab41be58545aaea7ee9706884e7c", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "2a97ba3c3d9a5ed49ff35f566b40b890", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "1901245960675de7b3b96791312f725c", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128.jsonl b/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128.jsonl deleted file mode 100644 index d1e52eb90dd4ffe2e0250571fd374c3c4c633937..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh3_kw3_sh1_sw1_dh2_dw2_c64_c128.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "276c7d862158547796a480c8198faea6", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 64, "W": 64}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "c8c1892ee91c56cd9261e095b28bfdea", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 32, "W": 32}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "204259e7ba7659b99caddd1acd638b61", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256.jsonl b/workloads/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256.jsonl deleted file mode 100644 index 0b59af9e067b7625f49d10c2d9b392f4d6206b0b..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c128_c256.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "3967ac3bf8c75474915b83340931a6d1", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "ecca723c219853bc988feb7398ba4679", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 112, "W": 112}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "d24edb3d10c1569abda4ed25bfe098db", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512.jsonl b/workloads/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512.jsonl deleted file mode 100644 index 298c395af8d3b79d7b08f1a317f3d7014670a72b..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c256_c512.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "444bbc8eec2e5364afce7aa74342ed2d", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "a559b13ebc97587786f71ee44de4bc32", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "d627a5a78bc25c4fbadb24438974c88d", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024.jsonl b/workloads/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024.jsonl deleted file mode 100644 index 925630344b306abc88e7152ebe65489cb321d3b9..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c512_c1024.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "57dc8b5b69225a3f8ec9b9781adad2cd", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 7, "W": 7}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "6359a7d0c2015fbc89a53a1179242bd9", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "4e2093be1531531180466644db45791d", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128.jsonl b/workloads/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128.jsonl deleted file mode 100644 index 4e36635cc7fc5d8ef2b778d02d0811f08ba1c013..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh3_kw3_sh2_sw2_dh1_dw1_c64_c128.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 112, "W": 112}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "a4aa78333cf056838a68700ae3aba650", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "c08c26b564285fbab074242c591ca462", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 224, "W": 224}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 1}, "pad_left": {"type": "scalar", "value": 1}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "52f36d38efcb5616b0ab1fae6b74da15", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128.jsonl b/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128.jsonl deleted file mode 100644 index 8258224983271b740e8705fc5db99a3285715e5e..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c128_c128.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "e0fc0a5d014f5201ac5d7f0fdde3eb82", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "280264e3f3da5baaad622ea522103d6c", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "f5af06f9d26c51bda23cdd491e025129", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256.jsonl b/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256.jsonl deleted file mode 100644 index bfa09cf01b966e761a2fb21d703ff6c119dd26a4..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c256_c256.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "9181464a38255af69c799b55a88c0daf", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 7, "W": 7}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "948b6e106dbd5ee28e984de515cc1057", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "7fa001a54070516fac7b2090f275ec5d", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64.jsonl b/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64.jsonl deleted file mode 100644 index 3db9bd0002e4e76e8bb9ed7ea5adb4e8ed80d5b5..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c32_c64.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 32, "W": 32}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "2a88d188621e5806a1787c31d5ec1340", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 112, "W": 112}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "8a49b5357f68530bb7a559d9d0b5c813", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "8e3d6a0468ee5dcfbb8cff10cedaeb8b", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32.jsonl b/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32.jsonl deleted file mode 100644 index 474e768769f82de8dde2cd646204416a57a5368f..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c3_c32.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 224, "W": 224}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "87e7aa8330035a74935427a1b75f4418", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 300, "W": 300}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "f743b4ceac355b03beca86485e0039ab", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 416, "W": 416}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "86e89ceefd595fbaade0fe508e9a144c", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512.jsonl b/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512.jsonl deleted file mode 100644 index a59eb9192b1a5120d751a39ab5f8cfa3ee8c34b2..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c512_c512.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 7, "W": 7}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "dde30152afe35a56b216aad5fe01c0c0", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "558026cffc19544d8fe7a91c58635900", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "a579f9c2aaa65ecf95218723d97b36dd", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128.jsonl b/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128.jsonl deleted file mode 100644 index bad0ca22102b2295524e7a61be61925080a66dd7..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh5_kw5_sh1_sw1_dh1_dw1_c64_c128.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "29a4240dc9a5526990b53de998985581", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "da88be470bb4531fb6d8f4e538fd8a7a", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 2}, "pad_left": {"type": "scalar", "value": 2}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "453ca9c5bb34524983db573cebd4ec38", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv/conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64.jsonl b/workloads/conv/conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64.jsonl deleted file mode 100644 index 08a92533c2ea0ee1ad575e932c162dca9a91c014..0000000000000000000000000000000000000000 --- a/workloads/conv/conv2d_kh7_kw7_sh2_sw2_dh1_dw1_c3_c64.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 224, "W": 224}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 3}, "pad_left": {"type": "scalar", "value": 3}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "507902e2a1bd5519a852f6d67bd21b9a", "tags": {"from": "tests/ncnn/candidate/convolution.cpp"}} -{"axes": {"N": 1, "H": 448, "W": 448}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 3}, "pad_left": {"type": "scalar", "value": 3}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "e6b3bd13646d5cf8acd98028da45bb89", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 320, "W": 320}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "pad_top": {"type": "scalar", "value": 3}, "pad_left": {"type": "scalar", "value": 3}, "activation_type": {"type": "scalar", "value": 0}, "with_bias": {"type": "scalar", "value": 0}}, "uuid": "09aa974271d251c08962a4f569694bf7", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin1024_cout1024.jsonl b/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin1024_cout1024.jsonl deleted file mode 100644 index 859063e24991a467b91d09735df748a44f5a991f..0000000000000000000000000000000000000000 --- a/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin1024_cout1024.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "fab88bdbd2105fbf9a54d8d9aaf86592", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "W": 128}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "577a3b637e825bb4ba0126e678657ed1", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "W": 512}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "1d4dd4138a2d53c8b041494f575d1db2", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin128_cout512.jsonl b/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin128_cout512.jsonl deleted file mode 100644 index 2f3e8edc72374ece4f9c9c88dc0a74f634e61735..0000000000000000000000000000000000000000 --- a/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin128_cout512.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "cec30cf820b457b592fa238dbd074993", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "W": 128}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "718ae81c4387579e9962f9c310c2b16e", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "W": 512}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "1a0853a9c9b9554387446b61236890be", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin256_cout1024.jsonl b/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin256_cout1024.jsonl deleted file mode 100644 index 84081197ac5f18ff6091fffdd0b345326fa4732d..0000000000000000000000000000000000000000 --- a/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin256_cout1024.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "52c54c1b894859c882018d67dec858c1", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "W": 128}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "f922cf2a09b85cd1a21756f35a10dd88", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "W": 512}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "2be023aa408b599aacfbc0a790eeac77", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin512_cout2048.jsonl b/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin512_cout2048.jsonl deleted file mode 100644 index 8c449d6429b3101b1cc141d2de2e6282e1a8c95e..0000000000000000000000000000000000000000 --- a/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin512_cout2048.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "a2f95070d3375ad6ac95f27bac9fee19", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "W": 128}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "9c1aeac6a498507588ed42d082b38137", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "W": 512}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "a92410c8fb385fb1b3c9b3ac8f646db3", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout256.jsonl b/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout256.jsonl deleted file mode 100644 index ed50c991a9905d86cda370029294440f97856a2a..0000000000000000000000000000000000000000 --- a/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout256.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "W": 112}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "a3b0a501961750c98c7eb28acf8b9e22", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "W": 128}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "7ff60359cfce5a08902dad6b049c87b4", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "W": 512}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "27366d29c6435d6f93bdb3323edc0f47", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout64.jsonl b/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout64.jsonl deleted file mode 100644 index 3970e69bbb1d015f1c88b2c5df945747f6cf58b6..0000000000000000000000000000000000000000 --- a/workloads/conv1d/conv1d_kw1_sw1_dw1_p0_cin64_cout64.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "W": 224}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "0dacddd55486558aa1efb586aac3e212", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "W": 128}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "7bdd1ceb36ab5f41acab6bea69da6840", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "W": 512}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "91ed2f6f0a575aa7b39b9e78451162d6", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin1024_cout1024.jsonl b/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin1024_cout1024.jsonl deleted file mode 100644 index af3a0a3a4e04d4d20d57ff8cf9ed226569673481..0000000000000000000000000000000000000000 --- a/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin1024_cout1024.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "615d02ab9f3b5706bbacbbc86abf53a4", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "W": 64}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "35939e65bd5851bd92830340b26df476", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "W": 256}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "82cd1753b7da537486b6f7e30dc63995", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin128_cout256.jsonl b/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin128_cout256.jsonl deleted file mode 100644 index 01b6ddb8edb35c04d88a2db1003bb1c69721b6bc..0000000000000000000000000000000000000000 --- a/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin128_cout256.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "3a1f65212bc45cbbae4c74095c7f12e4", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "W": 64}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "34962fa830d65cbe976a4e5be3d6556e", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "W": 256}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "2cc9269a5e57520f933c1afc630f96b6", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin256_cout512.jsonl b/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin256_cout512.jsonl deleted file mode 100644 index d63421710944d8c0a5627349ae4f3afa95320049..0000000000000000000000000000000000000000 --- a/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin256_cout512.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "85f2a76a3cc359869aea5aa4d8002466", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "W": 64}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "bde237e6b8eb515aba62da434b784405", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "W": 256}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "57b9e77611a05d8991eacb6e1b3b8146", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin3_cout64.jsonl b/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin3_cout64.jsonl deleted file mode 100644 index 78892d85cd74cf90a4ebb7d298c4c131bcc60460..0000000000000000000000000000000000000000 --- a/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin3_cout64.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "W": 224}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "d591e2c5227e595786b88b4cd9db2c97", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "W": 64}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "d1fc701a7c145577a8d42a7f7874ea5d", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "W": 256}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "9e733654ce5751d7a59aa4a9dad47cd6", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin512_cout1024.jsonl b/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin512_cout1024.jsonl deleted file mode 100644 index f17e8953382a5252e62ed94cb0f096815618f990..0000000000000000000000000000000000000000 --- a/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin512_cout1024.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "f4bcd3a68a365dd19c3ac9d0e84947b3", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "W": 64}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "4dfe1875b6595cea954357afd2c7d59a", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "W": 256}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "b278cb338d3a52a1952f71cb776d8c3f", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin64_cout128.jsonl b/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin64_cout128.jsonl deleted file mode 100644 index 8697b02482541b55c118ea4b5fcde7416ba8065e..0000000000000000000000000000000000000000 --- a/workloads/conv1d/conv1d_kw3_sw1_dw1_p1_cin64_cout128.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "W": 112}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "3fdfae5415725eccab553e5efd0fd198", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "W": 64}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "ce3f7d5d325c5d7c93b2a47d7dea01ba", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "W": 256}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "7982c9a637f6534ba09decd16dd7209e", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024.jsonl b/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024.jsonl deleted file mode 100644 index cdd064c5c05f71152c7fbef2cfde69ec9327b6be..0000000000000000000000000000000000000000 --- a/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c1024.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 7, "W": 7}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "356687f67442549a958d0fb9ceae1ddd", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "f8455e9a94b2550b8c529b8745ef514e", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "c8de676aa2bc562dbb92d24b6bbc98e0", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128.jsonl b/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128.jsonl deleted file mode 100644 index 37238a11dd8d7c562e1b0a3d318adbb8b8ee8014..0000000000000000000000000000000000000000 --- a/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c128.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "2253a4b54d7158d98d03d77c8faccbb8", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "c56f2c7545b95fe39056ddd416b5fdd7", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "8bfd4f907cfe5518adcf3b18998f246d", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048.jsonl b/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048.jsonl deleted file mode 100644 index ef318b54bca50788daccfb198d6f4ade30f21f56..0000000000000000000000000000000000000000 --- a/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c2048.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "3dae9d35a0335188af578afb54ef8151", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "d3b09cb41d385638873b847f731b9c48", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "fb87b8cfce2253f8812665d43d231cca", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256.jsonl b/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256.jsonl deleted file mode 100644 index c91b8274dc832eaa0fc2501db9ca68942434d5ed..0000000000000000000000000000000000000000 --- a/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c256.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "bdf7c3b7b2f15227b1c3dd8bee2a65af", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "3105af14ead75832a36cb5d634172a75", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "f1171a1af3eb5282b8991051b712e1d6", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32.jsonl b/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32.jsonl deleted file mode 100644 index a224bbe52b468f8f15a59bfd902489331ba6b5fd..0000000000000000000000000000000000000000 --- a/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c32.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 224, "W": 224}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "55851b01f1c851e1b4195dace8193993", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "b22b06e0a17a530a99c1427a5b5b136e", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "a19d9a59ec1952ea93ecde2fa4348f68", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512.jsonl b/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512.jsonl deleted file mode 100644 index ce89833c9a3b37d78f93976bec8d2329a25516e9..0000000000000000000000000000000000000000 --- a/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c512.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "0dab312563a150818b5583d92d5bd3b3", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "3182ffdfc7f055f190e0b6688beed430", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "87dfa303cfcc5e84b08ae06f7e14ff2e", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64.jsonl b/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64.jsonl deleted file mode 100644 index 543cac21955c8a87906e159303285044d42af6b7..0000000000000000000000000000000000000000 --- a/workloads/conv2d_depthwise/conv2d_depthwise_kh3_kw3_sh1_sw1_dh1_dw1_p1_c64.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 112, "W": 112}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "acc8c2d1d5c3519abd9b0f1cfe9c59e7", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "add76356814d51218f9e3382d153b348", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "cbb6ade47dc757cf9ada527b2d5683b6", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024.jsonl b/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024.jsonl deleted file mode 100644 index ec1dc69c59e9e4d42f91c34248b27231626ca4bd..0000000000000000000000000000000000000000 --- a/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c1024.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "6e8f3e53ef305ff8941cb19fcfdba43e", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "684ebc734e645351931cb7dffcc7a31c", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "c170cae0a7ab5ee7ae3ddcd32a5ca89a", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128.jsonl b/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128.jsonl deleted file mode 100644 index 66ca8b3e5b3d0d15ebaa1c8e610bc2aa0de90f2d..0000000000000000000000000000000000000000 --- a/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c128.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "388655f1308e585190af07c49b18a223", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "791d0759f55a5470817453ad5e10d862", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "d55d25e09ea55038aa79b467a9ef22b8", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256.jsonl b/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256.jsonl deleted file mode 100644 index 17b82916e2e36ec7f0154b92443e7dc1c8072db6..0000000000000000000000000000000000000000 --- a/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c256.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "27369bbd6ef35cb882b2cb402d35a437", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "aec67bf96654549384120f8d8f1a304b", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "255f34a0bd8154c18a77a899e96356fb", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32.jsonl b/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32.jsonl deleted file mode 100644 index 3d1ba7416d91d1f3879cc7f032ca856043b4d2bf..0000000000000000000000000000000000000000 --- a/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c32.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 224, "W": 224}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "011d9c161d78506d8b82b8f78ae0879c", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "9cde362cff975aceba771d7f0e8fcb31", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "1b022377a1985784a2f40ddf4b10bedb", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512.jsonl b/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512.jsonl deleted file mode 100644 index 00c70a042aaf3f0c6fef00e60bf7a7979b039413..0000000000000000000000000000000000000000 --- a/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c512.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 7, "W": 7}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "4ff49086ff80584a86ff0f316e900fa2", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "a5f5a3b81d2e592988ed251a554c0305", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "10c62583c34f530c91179b8038ef486c", "tags": {"from": "gen_workload"}} diff --git a/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64.jsonl b/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64.jsonl deleted file mode 100644 index a90c4a5006f37dfadba0746d20078419d0f7171e..0000000000000000000000000000000000000000 --- a/workloads/conv2d_depthwise/conv2d_depthwise_kh5_kw5_sh1_sw1_dh1_dw1_p2_c64.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "f907513b7dd056049de530275012b29d", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "c1f95fba5f35520fa98162b977892a84", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "db7190846c6250ccbb68ac952dbfacd7", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin128_cout256.jsonl b/workloads/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin128_cout256.jsonl deleted file mode 100644 index 4d2981f122e2aa824df2abc2375c936d7d4fd9a0..0000000000000000000000000000000000000000 --- a/workloads/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin128_cout256.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "1e10a2446b915b48bbeaebd551683f65", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "b7800d870b1e523293dc20bdc2d71a64", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "b4d85782dfbc580e8685167a2413f29e", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin256_cout512.jsonl b/workloads/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin256_cout512.jsonl deleted file mode 100644 index 6b7ac2543a24e3ae1ceda6dbf48a79d4d8b7ab12..0000000000000000000000000000000000000000 --- a/workloads/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin256_cout512.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "c80b48c47e1d5ec3b0cf61cccbf3563a", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "3f04b4e83624544f809891e9096ea567", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "b9ed91b58b72562bbbe6f29af161ab71", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024.jsonl b/workloads/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024.jsonl deleted file mode 100644 index fd6801a72064565a286fe3b942c291f1ac411d2b..0000000000000000000000000000000000000000 --- a/workloads/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin512_cout1024.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 7, "W": 7}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "d4bc1434edde5e678092cc989f06ea1f", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "3f55d9b5825054b482ed0a68a8b07a15", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "9b18355b1a5f587fb9ec22fdc3564a96", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin64_cout128.jsonl b/workloads/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin64_cout128.jsonl deleted file mode 100644 index 03a94d43c2556e43f47141818543d19056d6a47b..0000000000000000000000000000000000000000 --- a/workloads/deconv2d/deconv2d_kh3_kw3_sh1_sw1_cin64_cout128.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "b6f7a251d2d2532195a006e4b5bf52e9", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "ab31f9e56d8a5a6d93b51a3db4f7b926", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "498a15f55c405bcdb487e8d2b2753f6c", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin128_cout256.jsonl b/workloads/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin128_cout256.jsonl deleted file mode 100644 index 491bb496f31d749a32b7b6a16b1eb533cddec114..0000000000000000000000000000000000000000 --- a/workloads/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin128_cout256.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "757cdaaa84715203a02fb77704f20e52", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "397d7a1f2c54514f91797479cc497e99", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "cd31a62ce3de56cba0188e1ac64fe8f1", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin256_cout512.jsonl b/workloads/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin256_cout512.jsonl deleted file mode 100644 index 68055a810dbd4ea539f805da0fb3c8b1f213f3b7..0000000000000000000000000000000000000000 --- a/workloads/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin256_cout512.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 7, "W": 7}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "2c450cd625e356a6a7901d93f4023df1", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "6ed1ded2d1b450e28303b76f1c208064", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "00136af5b6be5dedbab1cc642549cd54", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024.jsonl b/workloads/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024.jsonl deleted file mode 100644 index f7bcbf568d548c01c38404efcb61e1ead503d866..0000000000000000000000000000000000000000 --- a/workloads/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin512_cout1024.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 7, "W": 7}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "c793189925d45c92b5df3139c412f90f", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "89d19f1d5cfb5e14b73cd4020af05f5b", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "f012cbf24a565fa7a174a4a9da91198e", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin64_cout128.jsonl b/workloads/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin64_cout128.jsonl deleted file mode 100644 index 9bf3ec792ceab33eddbd190aef77fbaf4e8a7d09..0000000000000000000000000000000000000000 --- a/workloads/deconv2d/deconv2d_kh3_kw3_sh2_sw2_cin64_cout128.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "66eb34df79a9580482245ff6dcd9e904", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "cc0b953d48495c6e903ebf005f51d9d3", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "a3436efc560d52b6833375a176619fd7", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin128_cout256.jsonl b/workloads/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin128_cout256.jsonl deleted file mode 100644 index 1c0a026032b3a5fb83447723e29a864ba91b5b3c..0000000000000000000000000000000000000000 --- a/workloads/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin128_cout256.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "1127aa3e0c165ab59166fbed66e348b7", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "a212111de9cf54e49315e02fc5709525", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "b6cbdf3552d25b4082b3ab88e0031132", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin256_cout512.jsonl b/workloads/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin256_cout512.jsonl deleted file mode 100644 index 9c57efbef1e878741e42bcc2bbca5f30c81dbc6e..0000000000000000000000000000000000000000 --- a/workloads/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin256_cout512.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "281906993858518cb1aaca5ab9e280a0", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "6cdad3b9d35c515ca57413fcd9f351c4", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "6d2d53dd8df15ddfb3a8d0929224f919", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024.jsonl b/workloads/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024.jsonl deleted file mode 100644 index d53b596d9b482e19e341befede9429549ea8e746..0000000000000000000000000000000000000000 --- a/workloads/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin512_cout1024.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 7, "W": 7}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "6280b53d3fe952eeb170df8162906061", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "0471154089685f21b99250b1388fc051", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "86533bd15aee5ff199928247341d140f", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin64_cout128.jsonl b/workloads/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin64_cout128.jsonl deleted file mode 100644 index c9c18c4e15c6626ab280bbae71195b125f65a394..0000000000000000000000000000000000000000 --- a/workloads/deconv2d/deconv2d_kh4_kw4_sh1_sw1_cin64_cout128.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "b8291a0b9c45578f9ae1e64330530844", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "a374addb3be755c895fd22b974d73bc6", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "5fdc5d99e43a5bcb980207e9b1ad02b7", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin128_cout256.jsonl b/workloads/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin128_cout256.jsonl deleted file mode 100644 index f1a36bb910079d41e19f7398bb939bdf6f28c45e..0000000000000000000000000000000000000000 --- a/workloads/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin128_cout256.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "5ff57762aad35880848ad483fc149572", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 7, "W": 7}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "15748e6d15d054d292c8942325690e9c", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "b3d7f1a4e8585a46ae8fbbec63913dfd", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin32_cout64.jsonl b/workloads/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin32_cout64.jsonl deleted file mode 100644 index fc20f1a66fec5dbe2d019d3feb9fd55a659524cb..0000000000000000000000000000000000000000 --- a/workloads/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin32_cout64.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "044007b2e13158ec990f463b57d0061d", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "15bc416b1e1e59eea11239db250d6cae", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "1f34a863a180506db9e5909d6e67dae2", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin512_cout512.jsonl b/workloads/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin512_cout512.jsonl deleted file mode 100644 index 75fe3eeb078b2b0ec699ff43a41d2d631b0c328f..0000000000000000000000000000000000000000 --- a/workloads/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin512_cout512.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 7, "W": 7}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "b3693969f85550cbb0492e8f94b24747", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "140cb7f0f21e5ccd9c00e5864aad30f1", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "d0af6900a2665b82a7cd797a2d9c54c9", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin64_cout128.jsonl b/workloads/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin64_cout128.jsonl deleted file mode 100644 index e5a557a0aada0d809f7bcbda0c10b75e165252e1..0000000000000000000000000000000000000000 --- a/workloads/deconv2d/deconv2d_kh4_kw4_sh2_sw2_cin64_cout128.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "081d1b988b495ddcb2d502a7ede5b7ce", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "9df9693d73905250bae823389d6dbf40", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "713dfd524ac65093a33658a678099203", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024.jsonl b/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024.jsonl deleted file mode 100644 index cb3c58103c816ca1c27fc0dac1d6dd9b8cf0c06f..0000000000000000000000000000000000000000 --- a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c1024.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "0dcbf44bb78f59c6a3bd009cbac9c31b", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "59aefe54d11e59b4b1ddc44557dbc71e", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "46c3e8394950585db030c655c26ded31", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c128.jsonl b/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c128.jsonl deleted file mode 100644 index e4ecbea531e5cab307e3e469fa6f699ac262e5e3..0000000000000000000000000000000000000000 --- a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c128.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "804046907b3858dca08617f94d7c6729", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "8254a10d459951308de23684f9e5a924", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "fd9b08900a38550ba0e42cab3ec8855d", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c256.jsonl b/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c256.jsonl deleted file mode 100644 index a69e9542bf6c89262d7908efb3967e5bd3678ba9..0000000000000000000000000000000000000000 --- a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c256.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "d432881dc24254099a5d4c349eeac2e4", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "1b7ab7d018a9567396c51468cb8c17e5", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "a96dbf5e240d51b0bee1c7a164395aef", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c32.jsonl b/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c32.jsonl deleted file mode 100644 index 1937295d57dee183f05f56b49dc74c207923ad46..0000000000000000000000000000000000000000 --- a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c32.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 112, "W": 112}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "841062afb6c956a29f615060215c02bc", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "84b088cc1a53505c91cd3ac3e053efdc", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "b6000adadce75cdebc308dfde3c20771", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c512.jsonl b/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c512.jsonl deleted file mode 100644 index 9b28b4e39924ce358092a7fa2bf839641d978778..0000000000000000000000000000000000000000 --- a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c512.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 7, "W": 7}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "92ea00f8c83d50eba7904087c4466d04", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "88937b0830cc52278bdaf72fc6f14e80", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "1df1501cab9c53219b8f8fab2261e690", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c64.jsonl b/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c64.jsonl deleted file mode 100644 index 90abe075079fc47102e9c967422f0bb885174b19..0000000000000000000000000000000000000000 --- a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh2_kw2_sh2_sw2_c64.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "df7119dd1a2358c4a82a212336455e32", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "23a01e83413a57af9f5971b2ef24b5bb", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "5ec8dfa7bd225537bddc52c1b1e6dc66", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c128.jsonl b/workloads/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c128.jsonl deleted file mode 100644 index 707aaba89e511aea8b632a491dd4cd9011fd3ce9..0000000000000000000000000000000000000000 --- a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c128.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "4ba29780f29a5e91914dae8b36dc0bd3", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "dc3f4d9557ea5826813c4313c30ba029", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "e6b38f97acd7523081b87ed332a21f9e", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048.jsonl b/workloads/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048.jsonl deleted file mode 100644 index fcd60aadbb280624776cc49e84815c857eb66762..0000000000000000000000000000000000000000 --- a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c2048.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 7, "W": 7}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "f58a8438eaa2511e96daf4814339b033", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "69ec210dc1c25b908213f6d32d96ade8", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "e71f316e44305e0797f3617beaef5b7b", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c32.jsonl b/workloads/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c32.jsonl deleted file mode 100644 index 5f864e6e15e6c7dc6d1e84ef323f8ddc040fadf9..0000000000000000000000000000000000000000 --- a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c32.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 224, "W": 224}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "ed78fc8321d557658c11ba75ccbeadfe", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "58eec703e91c5c9db79875ff13e25553", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "00d4ec44c2da523c83e6aa8ea3e90f37", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c512.jsonl b/workloads/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c512.jsonl deleted file mode 100644 index 881cf0977dc81eb4e0679f761b4e039d6cac72ff..0000000000000000000000000000000000000000 --- a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c512.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "01d3d28db36455eba8e43f1a509d38f0", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "3d9f21bc00145659b3c3346afd3dd9ad", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "74f2a6991a295553888b63da9d138ab7", "tags": {"from": "gen_workload"}} diff --git a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c64.jsonl b/workloads/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c64.jsonl deleted file mode 100644 index 3f92cba5449181cd848bfc5e4635dc1bedc84435..0000000000000000000000000000000000000000 --- a/workloads/deconv2d_depthwise/deconv2d_depthwise_kh3_kw3_sh1_sw1_c64.jsonl +++ /dev/null @@ -1,3 +0,0 @@ -{"axes": {"N": 1, "H": 56, "W": 56}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "5b2f0118d2a75f6189bc23ec7b513450", "tags": {"from": "gen_definitions"}} -{"axes": {"N": 1, "H": 28, "W": 28}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "9ac0c99c519256eda187e71998d681f2", "tags": {"from": "gen_workload"}} -{"axes": {"N": 1, "H": 14, "W": 14}, "inputs": {"input": {"type": "random"}, "weight": {"type": "random"}, "bias": {"type": "random"}}, "uuid": "2b63b0d83ec351888b793dd003ca8c80", "tags": {"from": "gen_workload"}}