| { |
| "op": "ai.onnx.ReduceLogSumExp", |
| "cases": [ |
| { |
| "name": "1024x1024_axis1", |
| "attrs": { "axes": [1], "keepdims": 0 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1024, 1024] } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1024] } } |
| }, |
| { |
| "name": "reducelogsumexp-fullreduce-r3-256x256x256-keepdims", |
| "preset": "smoke", |
| "vars": { "rows": 65536, "cols": 256 }, |
| "attrs": { "keepdims": 1 }, |
| "inputs": { "x": { "shape": [256, 256, 256], "dtype": "float32", "dist": "normal", "seed": 811, "scale": 0.5 } }, |
| "outputs": { "y": { "shape": [1, 1, 1], "dtype": "float32" } }, |
| "bench": { |
| "metrics": [ |
| { "type": "bandwidth", "name": "stable max-plus-exp traffic", "value": "2 * args.rows * args.cols * 4" } |
| ] |
| } |
| }, |
| { |
| "name": "reducelogsumexp-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", "name": "stable max-plus-exp traffic", "value": "2 * args.rows * args.cols * 4" } |
| ] |
| } |
| }, |
| { |
| "name": "reducelogsumexp-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", "name": "stable max-plus-exp traffic", "value": "2 * args.rows * args.cols * 4" } |
| ] |
| } |
| }, |
| { |
| "name": "axis0-large-65536x512", |
| "preset": "smoke", |
| "vars": { "rows": 65536, "cols": 512 }, |
| "attrs": { "axes": [0], "keepdims": 0 }, |
| "inputs": { "x": { "shape": [65536, 512], "dtype": "float32", "dist": "normal", "scale": 0.2, "seed": 311 } }, |
| "outputs": { "y": { "shape": [512], "dtype": "float32" } }, |
| "bench": { |
| "primary": true, |
| "metrics": [ |
| { "type": "bandwidth", "name": "stable max-plus-exp traffic", "value": "2 * args.rows * args.cols * 4" } |
| ] |
| } |
| }, |
| { |
| "name": "axis0-tall-16384x1024", |
| "preset": "smoke", |
| "vars": { "rows": 16384, "cols": 1024 }, |
| "attrs": { "axes": [0], "keepdims": 0 }, |
| "inputs": { "x": { "shape": [16384, 1024], "dtype": "float32", "dist": "normal", "scale": 0.2, "seed": 312 } }, |
| "outputs": { "y": { "shape": [1024], "dtype": "float32" } }, |
| "bench": { |
| "primary": true, |
| "metrics": [ |
| { "type": "bandwidth", "name": "stable max-plus-exp traffic", "value": "2 * args.rows * args.cols * 4" } |
| ] |
| } |
| }, |
| { |
| "name": "fullreduce-single-lane-r3-101x101x101-numel-not-mult4", |
| "preset": "stress", |
| "vars": { "numel": 1030301 }, |
| "attrs": { "keepdims": 1 }, |
| "inputs": { "x": { "shape": [101, 101, 101], "dtype": "float32", "dist": "normal", "seed": 733, "scale": 0.3 } }, |
| "outputs": { "y": { "shape": [1, 1, 1], "dtype": "float32", "dist": "empty" } }, |
| "bench": { |
| "metrics": [{ "type": "bandwidth", "name": "stable max-plus-exp traffic", "value": "2 * numel(shapes.x) * 4" }] |
| } |
| }, |
| { |
| "name": "axis0-few-lane-starvation-131072x8", |
| "preset": "stress", |
| "vars": { "rows": 131072, "cols": 8 }, |
| "attrs": { "axes": [0], "keepdims": 0 }, |
| "inputs": { "x": { "shape": [131072, 8], "dtype": "float32", "dist": "normal", "seed": 941, "scale": 0.2 } }, |
| "outputs": { "y": { "shape": [8], "dtype": "float32", "dist": "empty" } }, |
| "bench": { |
| "primary": true, |
| "metrics": [ |
| { "type": "bandwidth", "name": "stable max-plus-exp traffic", "value": "2 * args.rows * args.cols * 4" } |
| ] |
| } |
| }, |
| { |
| "name": "reducelogsumexp-spatial-axes12-f32-16x256x256-low-lane", |
| "preset": "stress", |
| "provenance": { |
| "source": "authored for variant coverage", |
| "notes": "Low-lane rank-3 contiguous-suffix reduction that verifies the vec4 subgroup reducer and no-subgroup workgroup-tree fallback." |
| }, |
| "vars": { "batch": 16, "height": 256, "width": 256 }, |
| "attrs": { "axes": [1, 2], "keepdims": 1 }, |
| "inputs": { "x": { "shape": [16, 256, 256], "dtype": "float32", "dist": "normal", "seed": 910, "scale": 0.2 } }, |
| "outputs": { "y": { "shape": [16, 1, 1], "dtype": "float32", "dist": "empty" } }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "name": "stable max-plus-exp traffic", |
| "value": "2 * args.batch * args.height * args.width * 4" |
| } |
| ] |
| } |
| } |
| ], |
| "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] } |
| } |
|
|
