Xenova's picture
Xenova HF Staff
sync 2e7068faf55e
40f0ca1 verified
Raw
History Blame
12.8 kB
{
"op": "ai.onnx.QuantizeLinear",
"tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
"cases": [
{
"name": "f32_to_u8_1m_scalar",
"preset": "smoke",
"vars": { "inputDtype": "float32", "outputDtype": "uint8", "count": 1048576 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [1024, 1024],
"data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 3.0 }
},
"y_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.02] } },
"y_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [128] } }
},
"outputs": { "y": { "dtype": "uint8", "shape": [1024, 1024] } },
"bench": {
"primary": true,
"metrics": [
{ "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
]
}
},
{
"name": "f32_to_u8_1m_no_zero_point",
"preset": "smoke",
"vars": { "inputDtype": "float32", "outputDtype": "uint8", "count": 1048576 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [1024, 1024],
"data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 3.0 }
},
"y_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.02] } }
},
"outputs": { "y": { "dtype": "uint8", "shape": [1024, 1024] } },
"bench": {
"metrics": [
{ "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
]
}
},
{
"name": "f32_to_i8_1m_plus3_scalar",
"preset": "smoke",
"vars": { "inputDtype": "float32", "outputDtype": "int8", "count": 1048579 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [1048579],
"data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 3.0 }
},
"y_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.02] } },
"y_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [0] } }
},
"outputs": { "y": { "dtype": "int8", "shape": [1048579] } },
"bench": {
"metrics": [
{ "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
]
}
},
{
"name": "f32_to_u8_1m_plus3_no_zero_point",
"preset": "smoke",
"vars": { "inputDtype": "float32", "outputDtype": "uint8", "count": 1048579 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [1048579],
"data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 3.0 }
},
"y_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.02] } }
},
"outputs": { "y": { "dtype": "uint8", "shape": [1048579] } },
"bench": {
"metrics": [
{ "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
]
}
},
{
"name": "f16_to_u8_1m_scalar",
"preset": "smoke",
"vars": { "inputDtype": "float16", "outputDtype": "uint8", "count": 1048576 },
"inputs": {
"x": { "dtype": "float16", "shape": [1024, 1024], "data": { "kind": "constant", "value": 1.25 } },
"y_scale": { "dtype": "float16", "shape": [1], "data": { "kind": "values", "values": [0.02] } },
"y_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [128] } }
},
"outputs": { "y": { "dtype": "uint8", "shape": [1024, 1024] } },
"bench": {
"metrics": [
{ "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
]
}
},
{
"name": "blocked_axis1_4096x4096_block32_with_zp",
"preset": "smoke",
"vars": { "inputDtype": "float32", "outputDtype": "uint8", "count": 16777216 },
"attrs": { "axis": 1, "block_size": 32 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [4096, 4096],
"data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 3.0 }
},
"y_scale": {
"dtype": "float32",
"shape": [4096, 128],
"data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.05, "scale": 0.01, "offset": 0.02 }
},
"y_zero_point": { "dtype": "uint8", "shape": [4096, 128], "data": { "kind": "constant", "value": 128 } }
},
"outputs": { "y": { "dtype": "uint8", "shape": [4096, 4096] } },
"bench": {
"primary": true,
"metrics": [
{ "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
]
}
},
{
"name": "pertensor_4096x4096_scalar_healthy_sibling",
"preset": "smoke",
"vars": { "inputDtype": "float32", "outputDtype": "uint8", "count": 16777216 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [4096, 4096],
"data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 3.0 }
},
"y_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.02] } },
"y_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [128] } }
},
"outputs": { "y": { "dtype": "uint8", "shape": [4096, 4096] } },
"bench": {
"metrics": [
{ "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
]
}
},
{
"name": "per_axis0_inner2049_scalar_fallback_alignment_miss",
"preset": "smoke",
"vars": { "inputDtype": "float32", "outputDtype": "uint8", "count": 1049088 },
"attrs": { "axis": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [512, 2049],
"data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 3.0 }
},
"y_scale": {
"dtype": "float32",
"shape": [512],
"data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.05, "scale": 0.01, "offset": 0.02 }
},
"y_zero_point": { "dtype": "uint8", "shape": [512], "data": { "kind": "constant", "value": 128 } }
},
"outputs": { "y": { "dtype": "uint8", "shape": [512, 2049] } },
"bench": {
"metrics": [
{ "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
]
}
},
{
"name": "per_axis0_inner2048_vec4_healthy_sibling",
"preset": "smoke",
"vars": { "inputDtype": "float32", "outputDtype": "uint8", "count": 1048576 },
"attrs": { "axis": 0 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [512, 2048],
"data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 3.0 }
},
"y_scale": {
"dtype": "float32",
"shape": [512],
"data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.05, "scale": 0.01, "offset": 0.02 }
},
"y_zero_point": { "dtype": "uint8", "shape": [512], "data": { "kind": "constant", "value": 128 } }
},
"outputs": { "y": { "dtype": "uint8", "shape": [512, 2048] } },
"bench": {
"metrics": [
{ "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
]
}
},
{
"name": "f16_per_axis0_inner2048_vec4_unbenched",
"preset": "smoke",
"vars": { "inputDtype": "float16", "outputDtype": "uint8", "count": 1048576 },
"attrs": { "axis": 0 },
"inputs": {
"x": { "dtype": "float16", "shape": [512, 2048], "data": { "kind": "constant", "value": 1.25 } },
"y_scale": {
"dtype": "float16",
"shape": [512],
"data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.05, "scale": 0.01, "offset": 0.02 }
},
"y_zero_point": { "dtype": "uint8", "shape": [512], "data": { "kind": "constant", "value": 128 } }
},
"outputs": { "y": { "dtype": "uint8", "shape": [512, 2048] } },
"bench": {
"metrics": [
{ "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
]
}
},
{
"name": "per_axis1_lastdim_inner1_generic_scalar_cliff",
"preset": "stress",
"vars": { "inputDtype": "float32", "outputDtype": "uint8", "count": 4194304 },
"attrs": { "axis": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2048, 2048],
"data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 3.0 }
},
"y_scale": {
"dtype": "float32",
"shape": [2048],
"data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.05, "scale": 0.01, "offset": 0.02 }
},
"y_zero_point": { "dtype": "uint8", "shape": [2048], "data": { "kind": "constant", "value": 128 } }
},
"outputs": { "y": { "dtype": "uint8", "shape": [2048, 2048] } },
"bench": {
"metrics": [
{ "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
]
}
},
{
"name": "per_axis1_lastdim_inner1_generic_scalar_no_zero_point_cliff",
"preset": "stress",
"vars": { "inputDtype": "float32", "outputDtype": "uint8", "count": 4194304 },
"attrs": { "axis": 1 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [2048, 2048],
"data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 3.0 }
},
"y_scale": {
"dtype": "float32",
"shape": [2048],
"data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.05, "scale": 0.01, "offset": 0.02 }
}
},
"outputs": { "y": { "dtype": "uint8", "shape": [2048, 2048] } },
"bench": {
"metrics": [
{ "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
]
}
},
{
"name": "blocked-rank4-axis1-bs32-no-zp-scalar-pathology",
"preset": "stress",
"vars": { "inputDtype": "float32", "outputDtype": "uint8", "count": 4194304 },
"attrs": { "axis": 1, "block_size": 32 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [4, 512, 32, 64],
"data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 3.0 }
},
"y_scale": {
"dtype": "float32",
"shape": [4, 16, 32, 64],
"data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.05, "scale": 0.005, "offset": 0.02 }
}
},
"outputs": { "y": { "dtype": "uint8", "shape": [4, 512, 32, 64], "dist": "empty" } },
"bench": {
"primary": true,
"metrics": [
{ "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" }
]
}
},
{
"name": "blocked-rank4-axis1-bs32-with-zp-scalar-pathology",
"preset": "stress",
"vars": { "inputDtype": "float32", "outputDtype": "uint8", "count": 4194304 },
"attrs": { "axis": 1, "block_size": 32 },
"inputs": {
"x": {
"dtype": "float32",
"shape": [4, 512, 32, 64],
"data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 3.0 }
},
"y_scale": {
"dtype": "float32",
"shape": [4, 16, 32, 64],
"data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.05, "scale": 0.005, "offset": 0.02 }
},
"y_zero_point": {
"dtype": "uint8",
"shape": [4, 16, 32, 64],
"dist": "randint",
"seed": 204,
"min": 96,
"max": 160
}
},
"outputs": { "y": { "dtype": "uint8", "shape": [4, 512, 32, 64], "dist": "empty" } },
"bench": {
"primary": true,
"metrics": [
{
"type": "bandwidth",
"value": "dtypeBytes(args.inputDtype) * (numel(shapes.x) + numel(shapes.y_scale)) + dtypeBytes(args.outputDtype) * (numel(shapes.y_zero_point) + numel(shapes.y))"
}
]
}
}
]
}