| { |
| "op": "ai.onnx.MeanVarianceNormalization", |
| "cases": [ |
| { |
| "name": "1x32x32x32_default_axes", |
| "preset": "smoke", |
| "attrs": { "axes": [0, 2, 3] }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 32, 32, 32] } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 32, 32] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "3 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "8x64x128x128_default_axes", |
| "preset": "smoke", |
| "attrs": { "axes": [0, 2, 3] }, |
| "inputs": { "x": { "dtype": "float32", "shape": [8, 64, 128, 128] } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [8, 64, 128, 128] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "3 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "16x64x64x64_spatial_axes", |
| "preset": "smoke", |
| "attrs": { "axes": [2, 3] }, |
| "inputs": { "x": { "dtype": "float32", "shape": [16, 64, 64, 64] } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [16, 64, 64, 64] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "3 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "f16_vec4_8x64x128x128_default_axes", |
| "preset": "smoke", |
| "vars": { "dtype": "float16" }, |
| "attrs": { "axes": [0, 2, 3] }, |
| "inputs": { "x": { "dtype": "float16", "shape": [8, 64, 128, 128], "dist": "normal", "seed": 900, "scale": 0.5 } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [8, 64, 128, 128] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "3 * dtypeBytes(args.dtype) * numel(shapes.x)" }] |
| } |
| }, |
| { |
| "name": "f16_all_axes_flat_split_1x1x256x256", |
| "preset": "smoke", |
| "attrs": { "axes": [0, 1, 2, 3] }, |
| "inputs": { "x": { "dtype": "float16", "shape": [1, 1, 256, 256], "dist": "normal", "seed": 906, "scale": 0.5 } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [1, 1, 256, 256] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "3 * 2 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "vec4_align_cliff_8x64x128x130_unaligned_W", |
| "preset": "smoke", |
| "attrs": { "axes": [0, 2, 3] }, |
| "inputs": { "x": { "dtype": "float32", "shape": [8, 64, 128, 130], "dist": "normal", "seed": 901, "scale": 0.5 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [8, 64, 128, 130] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "3 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "channel_axis_scalar_strided_8x64x128x128", |
| "preset": "smoke", |
| "attrs": { "axes": [1] }, |
| "inputs": { "x": { "dtype": "float32", "shape": [8, 64, 128, 128], "dist": "normal", "seed": 902, "scale": 0.5 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [8, 64, 128, 128] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "3 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "channel_axis_wg_pow2_cliff_8x65x128x128", |
| "preset": "stress", |
| "attrs": { "axes": [1] }, |
| "inputs": { "x": { "dtype": "float32", "shape": [8, 65, 128, 128], "dist": "normal", "seed": 905, "scale": 0.5 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [8, 65, 128, 128] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "3 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "rank3_scalar_8x4096x1024_reduce_hidden", |
| "preset": "stress", |
| "provenance": { "notes": "Stress-only capacity case: input plus output occupy exactly 256 MiB of GPU storage." }, |
| "attrs": { "axes": [2] }, |
| "inputs": { "x": { "dtype": "float32", "shape": [8, 4096, 1024], "dist": "normal", "seed": 903, "scale": 0.5 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [8, 4096, 1024] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "3 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "rank3_rows1_decode_hidden768", |
| "preset": "smoke", |
| "attrs": { "axes": [2] }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 1, 768], "dist": "normal", "seed": 904, "scale": 0.5 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 768] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "3 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "stress_all_axes_single_workgroup_rank4_3p9m", |
| "preset": "stress", |
| "attrs": { "axes": [0, 1, 2, 3] }, |
| "inputs": { "x": { "dtype": "float32", "shape": [64, 64, 64, 15], "dist": "normal", "seed": 950, "scale": 0.5 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [64, 64, 64, 15] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "3 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "stress_tiny_reduce_launchbound_wg2_4m_rows", |
| "preset": "stress", |
| "attrs": { "axes": [2] }, |
| "inputs": { "x": { "dtype": "float32", "shape": [4194304, 1, 2], "dist": "normal", "seed": 951, "scale": 0.5 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [4194304, 1, 2] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "3 * 4 * numel(shapes.x)" }] } |
| } |
| ] |
| } |
|
|