| { |
| "cases": [ |
| { |
| "name": "4x64x128x128_g32", |
| "preset": "smoke", |
| "vars": { "dtype": "float32", "batch": 4, "channels": 64, "spatial": 16384 }, |
| "attrs": { "num_groups": 32, "epsilon": 0.00001 }, |
| "inputs": { |
| "x": { "dtype": "float32", "shape": [4, 64, 128, 128], "dist": "normal", "seed": 740, "scale": 0.5 }, |
| "scale": { "dtype": "float32", "shape": [64], "dist": "uniform", "seed": 741, "scale": 0.25, "offset": 1 }, |
| "bias": { "dtype": "float32", "shape": [64], "dist": "normal", "seed": 742, "scale": 0.1 } |
| }, |
| "outputs": { "y": { "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": "1x32x32x32_g8", |
| "attrs": { "num_groups": 8, "epsilon": 0.00001 }, |
| "inputs": { |
| "x": { "dtype": "float32", "shape": [1, 32, 32, 32] }, |
| "scale": { "dtype": "float32", "shape": [32], "data": { "kind": "constant", "value": 1.0 } }, |
| "bias": { "dtype": "float32", "shape": [32], "data": { "kind": "constant", "value": 0.0 } } |
| }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 32, 32] } } |
| }, |
| { |
| "name": "f16_vec4_4x64x128x128_g32", |
| "preset": "smoke", |
| "vars": { "dtype": "float16", "batch": 4, "channels": 64, "spatial": 16384 }, |
| "attrs": { "num_groups": 32, "epsilon": 0.00001 }, |
| "inputs": { |
| "x": { "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 }, |
| "bias": { "dtype": "float16", "shape": [64], "dist": "normal", "seed": 752, "scale": 0.1 } |
| }, |
| "outputs": { "y": { "dtype": "float16", "shape": [4, 64, 128, 128] } }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "vec4_spatial_aligned_2x128x56x56_g32", |
| "preset": "smoke", |
| "vars": { "dtype": "float32", "batch": 2, "channels": 128, "spatial": 3136 }, |
| "attrs": { "num_groups": 32, "epsilon": 0.00001 }, |
| "inputs": { |
| "x": { "dtype": "float32", "shape": [2, 128, 56, 56], "dist": "normal", "seed": 760, "scale": 0.5 }, |
| "scale": { "dtype": "float32", "shape": [128], "dist": "uniform", "seed": 761, "scale": 0.25, "offset": 1 }, |
| "bias": { "dtype": "float32", "shape": [128], "dist": "normal", "seed": 762, "scale": 0.1 } |
| }, |
| "outputs": { "y": { "dtype": "float32", "shape": [2, 128, 56, 56] } }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "scalar_spatial_unaligned_2x128x53x53_g32", |
| "preset": "smoke", |
| "vars": { "dtype": "float32", "batch": 2, "channels": 128, "spatial": 2809 }, |
| "attrs": { "num_groups": 32, "epsilon": 0.00001 }, |
| "inputs": { |
| "x": { "dtype": "float32", "shape": [2, 128, 53, 53], "dist": "normal", "seed": 770, "scale": 0.5 }, |
| "scale": { "dtype": "float32", "shape": [128], "dist": "uniform", "seed": 771, "scale": 0.25, "offset": 1 }, |
| "bias": { "dtype": "float32", "shape": [128], "dist": "normal", "seed": 772, "scale": 0.1 } |
| }, |
| "outputs": { "y": { "dtype": "float32", "shape": [2, 128, 53, 53] } }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "tiny_spatial_stem_1x32x7x7_g8", |
| "preset": "smoke", |
| "vars": { "dtype": "float32", "batch": 1, "channels": 32, "spatial": 49 }, |
| "attrs": { "num_groups": 8, "epsilon": 0.00001 }, |
| "inputs": { |
| "x": { "dtype": "float32", "shape": [1, 32, 7, 7], "dist": "normal", "seed": 780, "scale": 0.5 }, |
| "scale": { "dtype": "float32", "shape": [32], "dist": "uniform", "seed": 781, "scale": 0.25, "offset": 1 }, |
| "bias": { "dtype": "float32", "shape": [32], "dist": "normal", "seed": 782, "scale": 0.1 } |
| }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 7, 7] } }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "scalar_lowocc_layernorm_equiv_2x320x65x65_g1", |
| "preset": "stress", |
| "vars": { "dtype": "float32", "batch": 2, "channels": 320, "spatial": 4225 }, |
| "attrs": { "num_groups": 1, "epsilon": 0.00001 }, |
| "inputs": { |
| "x": { "dtype": "float32", "shape": [2, 320, 65, 65], "dist": "normal", "seed": 901, "scale": 0.5 }, |
| "scale": { "dtype": "float32", "shape": [320], "dist": "uniform", "seed": 902, "scale": 0.25, "offset": 1 }, |
| "bias": { "dtype": "float32", "shape": [320], "dist": "normal", "seed": 903, "scale": 0.1 } |
| }, |
| "outputs": { "y": { "dtype": "float32", "shape": [2, 320, 65, 65], "dist": "empty" } }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| }, |
| { |
| "name": "scalar_cliff_unaligned_spatial_8x256x63x63_g32", |
| "preset": "stress", |
| "vars": { "dtype": "float32", "batch": 8, "channels": 256, "spatial": 3969 }, |
| "attrs": { "num_groups": 32, "epsilon": 0.00001 }, |
| "inputs": { |
| "x": { "dtype": "float32", "shape": [8, 256, 63, 63], "dist": "normal", "seed": 911, "scale": 0.5 }, |
| "scale": { "dtype": "float32", "shape": [256], "dist": "uniform", "seed": 912, "scale": 0.25, "offset": 1 }, |
| "bias": { "dtype": "float32", "shape": [256], "dist": "normal", "seed": 913, "scale": 0.1 } |
| }, |
| "outputs": { "y": { "dtype": "float32", "shape": [8, 256, 63, 63], "dist": "empty" } }, |
| "bench": { |
| "metrics": [ |
| { |
| "type": "bandwidth", |
| "value": "(args.batch * args.channels * args.spatial * 2 + args.channels * 2) * dtypeBytes(args.dtype)" |
| } |
| ] |
| } |
| } |
| ] |
| } |
|
|
