| { |
| "cases": [ |
| { |
| "name": "reducesumsquare-axis0-f32-1024x512", |
| "preset": "smoke", |
| "vars": { "rows": 1024, "cols": 512 }, |
| "attrs": { "axes": [0], "keepdims": 0 }, |
| "inputs": { "x": { "shape": [1024, 512], "dtype": "float32", "dist": "normal", "seed": 118, "scale": 0.2 } }, |
| "outputs": { "y": { "shape": [512], "dtype": "float32" } }, |
| "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * 4" }] } |
| }, |
| { |
| "name": "1024x1024_axis1", |
| "attrs": { "axes": [1], "keepdims": 0 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1024, 1024] } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1024] } } |
| }, |
| { |
| "name": "reducesumsquare-lastaxis-f32-4096x4096", |
| "preset": "smoke", |
| "vars": { "rows": 4096, "cols": 4096 }, |
| "attrs": { "axes": [1], "keepdims": 0 }, |
| "inputs": { "x": { "shape": [4096, 4096], "dtype": "float32", "dist": "normal", "seed": 117, "scale": 0.2 } }, |
| "outputs": { "y": { "shape": [4096], "dtype": "float32" } }, |
| "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * 4" }] } |
| }, |
| { |
| "name": "reducesumsquare-axis0-splitk-f32-8192x512", |
| "vars": { "rows": 8192, "cols": 512 }, |
| "attrs": { "axes": [0], "keepdims": 0 }, |
| "inputs": { "x": { "shape": [8192, 512], "dtype": "float32", "dist": "normal", "seed": 211, "scale": 0.2 } }, |
| "outputs": { "y": { "shape": [512], "dtype": "float32" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * 4" }] } |
| }, |
| { |
| "name": "reducesumsquare-lastaxis-odd-f32-4096x4095", |
| "vars": { "rows": 4096, "cols": 4095 }, |
| "attrs": { "axes": [1], "keepdims": 0 }, |
| "inputs": { "x": { "shape": [4096, 4095], "dtype": "float32", "dist": "normal", "seed": 212, "scale": 0.2 } }, |
| "outputs": { "y": { "shape": [4096], "dtype": "float32" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * 4" }] } |
| }, |
| { |
| "name": "reducesumsquare-rank4-axis1-f32-32x256x28x28", |
| "vars": { "n": 32, "c": 256, "h": 28, "w": 28 }, |
| "attrs": { "axes": [1], "keepdims": 0 }, |
| "inputs": { "x": { "shape": [32, 256, 28, 28], "dtype": "float32", "dist": "normal", "seed": 213, "scale": 0.2 } }, |
| "outputs": { "y": { "shape": [32, 28, 28], "dtype": "float32" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "args.n * args.c * args.h * args.w * 4" }] } |
| }, |
| { |
| "name": "reducesumsquare-fullreduce-serial-numel-not-mul4-r3", |
| "preset": "stress", |
| "attrs": { "keepdims": 0 }, |
| "inputs": { "x": { "shape": [1001, 1001, 1], "dtype": "float32", "dist": "normal", "seed": 701, "scale": 1 } }, |
| "outputs": { "y": { "shape": [], "dtype": "float32" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "1001 * 1001 * 1 * 4" }] } |
| }, |
| { |
| "name": "reducesumsquare-int32-lastaxis-serial-4096x1024", |
| "preset": "stress", |
| "attrs": { "axes": [1], "keepdims": 0 }, |
| "inputs": { |
| "x": { "shape": [4096, 1024], "dtype": "int32", "dist": "randint", "seed": 702, "min": -10, "max": 10 } |
| }, |
| "outputs": { "y": { "shape": [4096], "dtype": "int32" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "4096 * 1024 * 4" }] } |
| }, |
| { |
| "name": "reducesumsquare-rank3-spatial-axes12-f32-128x256x256-pathology", |
| "preset": "stress", |
| "provenance": { |
| "source": "repository-authored", |
| "notes": "Rank-3 multi-axis reduction with 128 output lanes, each serially scanning a 256x256 plane." |
| }, |
| "attrs": { "axes": [1, 2], "keepdims": 1 }, |
| "inputs": { "x": { "shape": [128, 256, 256], "dtype": "float32", "dist": "normal", "seed": 305, "scale": 0.2 } }, |
| "outputs": { "y": { "shape": [128, 1, 1], "dtype": "float32", "dist": "empty" } }, |
| "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "128 * 256 * 256 * 4" }] } |
| }, |
| { |
| "name": "reducesumsquare-spatial-axes23-f32-2x64x256x256-pathology", |
| "preset": "stress", |
| "provenance": { |
| "source": "repository-authored", |
| "notes": "Rank-4 multi-axis reduction with 128 output lanes, each serially scanning a 256x256 plane." |
| }, |
| "attrs": { "axes": [2, 3], "keepdims": 1 }, |
| "inputs": { "x": { "shape": [2, 64, 256, 256], "dtype": "float32", "dist": "normal", "seed": 304, "scale": 0.2 } }, |
| "outputs": { "y": { "shape": [2, 64, 1, 1], "dtype": "float32", "dist": "empty" } }, |
| "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "2 * 64 * 256 * 256 * 4" }] } |
| }, |
| { |
| "name": "reducesumsquare-axis0-f32-131072x15-narrow-splitk-boundary", |
| "preset": "stress", |
| "vars": { "rows": 131072, "cols": 15 }, |
| "attrs": { "axes": [0], "keepdims": 0 }, |
| "inputs": { "x": { "shape": [131072, 15], "dtype": "float32", "dist": "normal", "seed": 703, "scale": 0.1 } }, |
| "outputs": { "y": { "shape": [15], "dtype": "float32", "dist": "empty" } }, |
| "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * 4" }] } |
| }, |
| { |
| "name": "bench-strided_geometry_float32_rank3_odd_axis_odd_inner", |
| "attrs": { "axes": [-2], "keepdims": 1 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [3, 127, 259], "dist": "normal" } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [3, 1, 259] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "397824" }] } |
| }, |
| { |
| "name": "bench-strided_geometry_float32_rank3_even_inner", |
| "attrs": { "axes": [1], "keepdims": 0 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [7, 129, 260], "dist": "normal" } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [7, 260] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "946400" }] } |
| }, |
| { |
| "name": "bench-strided_geometry_float32_rank5_middle", |
| "attrs": { "axes": [3], "keepdims": 1 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [2, 3, 5, 17, 7], "dist": "normal" } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 5, 1, 7] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "15120" }] } |
| }, |
| { |
| "name": "bench-strided_geometry_float32_rank3_axis0", |
| "attrs": { "axes": [0], "keepdims": 0 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [17, 5, 13], "dist": "normal" } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [5, 13] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "4680" }] } |
| }, |
| { |
| "name": "bench-strided_geometry_float32_rank4_middle", |
| "attrs": { "axes": [2], "keepdims": 0 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [3, 5, 17, 4], "dist": "normal" } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [3, 5, 4] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "4320" }] } |
| }, |
| { |
| "name": "bench-strided_geometry_float32_rank4_channel", |
| "attrs": { "axes": [1], "keepdims": 1 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [2, 63, 7, 8], "dist": "normal" } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [2, 1, 7, 8] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "28672" }] } |
| }, |
| { |
| "name": "bench-strided_geometry_float16_rank3_odd_axis_odd_inner", |
| "attrs": { "axes": [-2], "keepdims": 1 }, |
| "inputs": { "x": { "dtype": "float16", "shape": [3, 127, 259], "dist": "normal" } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [3, 1, 259] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "198912" }] } |
| }, |
| { |
| "name": "bench-strided_geometry_float16_rank3_even_inner", |
| "attrs": { "axes": [1], "keepdims": 0 }, |
| "inputs": { "x": { "dtype": "float16", "shape": [7, 129, 260], "dist": "normal" } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [7, 260] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "473200" }] } |
| }, |
| { |
| "name": "bench-strided_geometry_float16_rank5_middle", |
| "attrs": { "axes": [3], "keepdims": 1 }, |
| "inputs": { "x": { "dtype": "float16", "shape": [2, 3, 5, 17, 7], "dist": "normal" } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [2, 3, 5, 1, 7] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "7560" }] } |
| }, |
| { |
| "name": "bench-strided_geometry_float16_rank3_axis0", |
| "attrs": { "axes": [0], "keepdims": 0 }, |
| "inputs": { "x": { "dtype": "float16", "shape": [17, 5, 13], "dist": "normal" } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [5, 13] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2340" }] } |
| }, |
| { |
| "name": "bench-strided_geometry_float16_rank4_middle", |
| "attrs": { "axes": [2], "keepdims": 0 }, |
| "inputs": { "x": { "dtype": "float16", "shape": [3, 5, 17, 4], "dist": "normal" } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [3, 5, 4] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2160" }] } |
| }, |
| { |
| "name": "bench-strided_geometry_float16_rank4_channel", |
| "attrs": { "axes": [1], "keepdims": 1 }, |
| "inputs": { "x": { "dtype": "float16", "shape": [2, 63, 7, 8], "dist": "normal" } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [2, 1, 7, 8] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "14336" }] } |
| } |
| ], |
| "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] } |
| } |
|
|
