| { |
| "op": "ai.onnx.InstanceNormalization", |
| "cases": [ |
| { |
| "name": "nchw_4x64x128x128", |
| "preset": "smoke", |
| "vars": { "dtype": "float32", "batch": 4, "channels": 64, "spatial": 16384 }, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [4, 64, 128, 128], "dist": "normal", "seed": 750, "scale": 0.5 }, |
| "scale": { "dtype": "float32", "shape": [64], "dist": "uniform", "seed": 751, "scale": 0.25, "offset": 1 }, |
| "b": { "dtype": "float32", "shape": [64], "dist": "normal", "seed": 752, "scale": 0.1 } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [4, 64, 128, 128] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "nchw_f16_4x64x128x128", |
| "preset": "smoke", |
| "vars": { "dtype": "float16", "batch": 4, "channels": 64, "spatial": 16384 }, |
| "inputs": { |
| "input": { "dtype": "float16", "shape": [4, 64, 128, 128], "dist": "normal", "seed": 750, "scale": 0.5 }, |
| "scale": { "dtype": "float16", "shape": [64], "dist": "uniform", "seed": 751, "scale": 0.25, "offset": 1 }, |
| "b": { "dtype": "float16", "shape": [64], "dist": "normal", "seed": 752, "scale": 0.1 } |
| }, |
| "outputs": { "output": { "dtype": "float16", "shape": [4, 64, 128, 128] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "nchw_1x64x56x56", |
| "preset": "smoke", |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [1, 64, 56, 56] }, |
| "scale": { "dtype": "float32", "shape": [64] }, |
| "b": { "dtype": "float32", "shape": [64] } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 64, 56, 56] } } |
| }, |
| { |
| "name": "alignment_cliff_rank3_2x64x4095_unaligned", |
| "preset": "smoke", |
| "vars": { "dtype": "float32", "batch": 2, "channels": 64, "spatial": 4095 }, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [2, 64, 4095], "dist": "normal", "seed": 810, "scale": 0.5 }, |
| "scale": { "dtype": "float32", "shape": [64], "dist": "uniform", "seed": 811, "scale": 0.25, "offset": 1 }, |
| "b": { "dtype": "float32", "shape": [64], "dist": "normal", "seed": 812, "scale": 0.1 } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [2, 64, 4095] } }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "alignment_healthy_rank3_2x64x4096_aligned", |
| "preset": "smoke", |
| "vars": { "dtype": "float32", "batch": 2, "channels": 64, "spatial": 4096 }, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [2, 64, 4096], "dist": "normal", "seed": 810, "scale": 0.5 }, |
| "scale": { "dtype": "float32", "shape": [64], "dist": "uniform", "seed": 811, "scale": 0.25, "offset": 1 }, |
| "b": { "dtype": "float32", "shape": [64], "dist": "normal", "seed": 812, "scale": 0.1 } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [2, 64, 4096] } }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "dispatch_cliff_rows70000_yfold", |
| "preset": "smoke", |
| "vars": { "dtype": "float32", "batch": 1, "channels": 70000, "spatial": 64 }, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [1, 70000, 64], "dist": "normal", "seed": 820, "scale": 0.5 }, |
| "scale": { "dtype": "float32", "shape": [70000], "dist": "uniform", "seed": 821, "scale": 0.25, "offset": 1 }, |
| "b": { "dtype": "float32", "shape": [70000], "dist": "normal", "seed": 822, "scale": 0.1 } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 70000, 64] } }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "dispatch_healthy_rows60000_under_cap", |
| "preset": "smoke", |
| "vars": { "dtype": "float32", "batch": 1, "channels": 60000, "spatial": 64 }, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [1, 60000, 64], "dist": "normal", "seed": 820, "scale": 0.5 }, |
| "scale": { "dtype": "float32", "shape": [60000], "dist": "uniform", "seed": 821, "scale": 0.25, "offset": 1 }, |
| "b": { "dtype": "float32", "shape": [60000], "dist": "normal", "seed": 822, "scale": 0.1 } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 60000, 64] } }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "single_huge_instance_1x1x4194304_onewg", |
| "preset": "smoke", |
| "vars": { "dtype": "float32", "batch": 1, "channels": 1, "spatial": 4194304 }, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [1, 1, 4194304], "dist": "normal", "seed": 830, "scale": 0.5 }, |
| "scale": { "dtype": "float32", "shape": [1], "dist": "uniform", "seed": 831, "scale": 0.25, "offset": 1 }, |
| "b": { "dtype": "float32", "shape": [1], "dist": "normal", "seed": 832, "scale": 0.1 } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 1, 4194304] } }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "onewg_per_plane_tiny_spatial_launchbound_1x65280x3x1", |
| "preset": "stress", |
| "vars": { "dtype": "float32", "batch": 1, "channels": 65280, "spatial": 3 }, |
| "attrs": { "epsilon": 0.00001 }, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [1, 65280, 3, 1], "dist": "normal", "seed": 901, "scale": 0.5 }, |
| "scale": { "dtype": "float32", "shape": [65280], "dist": "uniform", "seed": 902, "scale": 0.25, "offset": 1 }, |
| "b": { "dtype": "float32", "shape": [65280], "dist": "normal", "seed": 903, "scale": 0.1 } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 65280, 3, 1] } }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "onewg_per_plane_tiny_spatial_launchbound_rank3_1x49152x6", |
| "preset": "stress", |
| "vars": { "dtype": "float32", "batch": 1, "channels": 49152, "spatial": 6 }, |
| "attrs": { "epsilon": 0.00001 }, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [1, 49152, 6], "dist": "normal", "seed": 911, "scale": 0.5 }, |
| "scale": { "dtype": "float32", "shape": [49152], "dist": "uniform", "seed": 912, "scale": 0.25, "offset": 1 }, |
| "b": { "dtype": "float32", "shape": [49152], "dist": "normal", "seed": 913, "scale": 0.1 } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 49152, 6] } }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "rank7_spatial_flatten_gpu_gap", |
| "vars": { "dtype": "float32", "batch": 1, "channels": 2, "spatial": 32 }, |
| "attrs": { "epsilon": 0.00001 }, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [1, 2, 2, 2, 2, 2, 2], "dist": "normal", "seed": 931, "scale": 0.5 }, |
| "scale": { "dtype": "float32", "shape": [2], "dist": "uniform", "seed": 932, "scale": 0.25, "offset": 1 }, |
| "b": { "dtype": "float32", "shape": [2], "dist": "normal", "seed": 933, "scale": 0.1 } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 2, 2, 2, 2, 2, 2] } }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "splitk-priority-cliff-c256-256x256", |
| "preset": "stress", |
| "provenance": { |
| "source": "authored for variant coverage", |
| "notes": "Realistic 16.8M-element feature map that pins the selector boundary between plane_subgroup_vec4 and plane_splitk." |
| }, |
| "vars": { "dtype": "float32", "batch": 1, "channels": 256, "spatial": 65536 }, |
| "attrs": { "epsilon": 0.00001 }, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [1, 256, 256, 256], "dist": "normal", "seed": 991, "scale": 0.5 }, |
| "scale": { "dtype": "float32", "shape": [256], "dist": "uniform", "seed": 992, "scale": 0.25, "offset": 1 }, |
| "b": { "dtype": "float32", "shape": [256], "dist": "normal", "seed": 993, "scale": 0.1 } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 256, 256, 256], "dist": "empty" } }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "splitk-priority-cliff-c32-512x512", |
| "preset": "stress", |
| "provenance": { |
| "source": "authored for variant coverage", |
| "notes": "Realistic 8.4M-element high-resolution feature map that pins the selector boundary between plane_subgroup_vec4 and plane_splitk." |
| }, |
| "vars": { "dtype": "float32", "batch": 1, "channels": 32, "spatial": 262144 }, |
| "attrs": { "epsilon": 0.00001 }, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [1, 32, 512, 512], "dist": "normal", "seed": 994, "scale": 0.5 }, |
| "scale": { "dtype": "float32", "shape": [32], "dist": "uniform", "seed": 995, "scale": 0.25, "offset": 1 }, |
| "b": { "dtype": "float32", "shape": [32], "dist": "normal", "seed": 996, "scale": 0.1 } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 32, 512, 512], "dist": "empty" } }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| } |
| ] |
| } |
|
|
