{ "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] }, "cases": [ { "name": "rank3_sequence_window", "inputs": { "data": { "dtype": "float32", "shape": [1, 128, 768] }, "starts": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 32 }, "ends": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 96 }, "axes": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 1 }, "steps": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 1 } }, "outputs": { "output": { "dtype": "float32", "shape": [1, 64, 768] } } }, { "name": "rank3_end_past_dim", "inputs": { "data": { "dtype": "float32", "shape": [1, 128, 768] }, "starts": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 120 }, "ends": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 999 }, "axes": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 1 }, "steps": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 1 } }, "outputs": { "output": { "dtype": "float32", "shape": [1, 8, 768] } } }, { "name": "slice-rows-aligned-f32-4096x4096", "preset": "smoke", "vars": { "outRows": 3584, "cols": 4096 }, "inputs": { "data": { "dtype": "float32", "shape": [4096, 4096], "dist": "normal", "seed": 131, "scale": 0.5 }, "starts": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 256 }, "ends": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 3840 }, "axes": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 0 }, "steps": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 1 } }, "outputs": { "output": { "dtype": "float32", "shape": [3584, 4096] } }, "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "args.outRows * args.cols * 4 * 2" }] } }, { "name": "slice-inner-misaligned-f32-4096x4100", "preset": "smoke", "vars": { "rows": 4096, "outCols": 4096 }, "inputs": { "data": { "dtype": "float32", "shape": [4096, 4100], "dist": "normal", "seed": 132, "scale": 0.5 }, "starts": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 2 }, "ends": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 4098 }, "axes": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 1 }, "steps": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 1 } }, "outputs": { "output": { "dtype": "float32", "shape": [4096, 4096] } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.outCols * 4 * 2" }] } }, { "name": "slice-rows-aligned-f16-4096x2048", "preset": "edge", "vars": { "outRows": 3072, "cols": 2048, "dtype": "float16" }, "inputs": { "data": { "dtype": "float16", "shape": [4096, 2048], "dist": "normal", "seed": 133, "scale": 0.5 }, "starts": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 512 }, "ends": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 3584 }, "axes": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 0 }, "steps": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 1 } }, "outputs": { "output": { "dtype": "float16", "shape": [3072, 2048] } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "args.outRows * args.cols * dtypeBytes(args.dtype) * 2" }] } }, { "name": "slice-inner-true-misaligned-f32-4096x4097", "preset": "edge", "vars": { "rows": 4096, "outCols": 4097 }, "inputs": { "data": { "dtype": "float32", "shape": [8192, 4097], "dist": "normal", "seed": 141, "scale": 0.5 }, "starts": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 2048 }, "ends": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 6144 }, "axes": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 0 }, "steps": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 1 } }, "outputs": { "output": { "dtype": "float32", "shape": [4096, 4097] } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.outCols * 4 * 2" }] } }, { "name": "slice-no-axes-no-steps-f32-4096x4096", "preset": "edge", "vars": { "outRows": 3584, "cols": 4096 }, "inputs": { "data": { "dtype": "float32", "shape": [4096, 4096], "dist": "normal", "seed": 142, "scale": 0.5 }, "starts": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 256 }, "ends": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 3840 } }, "outputs": { "output": { "dtype": "float32", "shape": [3584, 4096] } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "args.outRows * args.cols * 4 * 2" }] } }, { "name": "slice-inner-vec4-unaligned-reverse-f32-4096x4096", "preset": "stress", "vars": { "rows": 4096, "cols": 4096 }, "inputs": { "data": { "dtype": "float32", "shape": [4096, 4096], "dist": "normal", "seed": 151, "scale": 0.5 }, "starts": { "dtype": "int32", "shape": [1], "dist": "constant", "value": -1 }, "ends": { "dtype": "int32", "shape": [1], "dist": "constant", "value": -2147483648 }, "axes": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 1 }, "steps": { "dtype": "int32", "shape": [1], "dist": "constant", "value": -1 } }, "outputs": { "output": { "dtype": "float32", "shape": [4096, 4096], "dist": "empty" } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * 4 * 2" }] } }, { "name": "slice-rows-aligned-f32-2048x2048", "preset": "smoke", "vars": { "outRows": 1792, "cols": 2048 }, "inputs": { "data": { "dtype": "float32", "shape": [2048, 2048], "dist": "normal", "seed": 134, "scale": 0.5 }, "starts": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 128 }, "ends": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 1920 }, "axes": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 0 }, "steps": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 1 } }, "outputs": { "output": { "dtype": "float32", "shape": [1792, 2048] } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "args.outRows * args.cols * 4 * 2" }] } }, { "name": "slice-middle-axis-f32-24x512x197-vit-tokens-pathology", "preset": "stress", "provenance": { "notes": "Sequence-window slice on the middle axis of a rank-3 tensor whose last extent is the ViT-B/16 token count: the inner_vec4 family needs both data and output last extents % 4 == 0 and rank2_single_axis_x4 serves rank 2 only, so 197 drops the copy to generic_metadata with a per-axis divmod address chain. Batch 24 keeps the sliced output above 1M elements. Control: slice-middle-axis-f32-24x512x200-control." }, "vars": { "batch": 24, "outRows": 256, "cols": 197 }, "inputs": { "data": { "dtype": "float32", "shape": [24, 512, 197], "dist": "normal", "seed": 152, "scale": 0.5 }, "starts": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 0 }, "ends": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 256 }, "axes": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 1 }, "steps": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 1 } }, "outputs": { "output": { "dtype": "float32", "shape": [24, 256, 197], "dist": "empty" } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "args.batch * args.outRows * args.cols * 4 * 2" }] } }, { "name": "slice-middle-axis-f32-24x512x200-control", "preset": "stress", "provenance": { "notes": "Nearest multiple-of-4 last extent for slice-middle-axis-f32-24x512x197-vit-tokens-pathology (1.5% more bytes) on the inner_vec4 route." }, "vars": { "batch": 24, "outRows": 256, "cols": 200 }, "inputs": { "data": { "dtype": "float32", "shape": [24, 512, 200], "dist": "normal", "seed": 153, "scale": 0.5 }, "starts": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 0 }, "ends": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 256 }, "axes": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 1 }, "steps": { "dtype": "int32", "shape": [1], "dist": "constant", "value": 1 } }, "outputs": { "output": { "dtype": "float32", "shape": [24, 256, 200], "dist": "empty" } }, "bench": { "metrics": [{ "type": "bandwidth", "value": "args.batch * args.outRows * args.cols * 4 * 2" }] } } ] }