| { |
| "cases": [ |
| { |
| "name": "axis0_small_inner_large_dim", |
| "preset": "smoke", |
| "args": { "axis": 0 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [4096, 64], "dist": "uniform", "seed": 33 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [4096, 64], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "rows_1024_cols_1024_axis1", |
| "args": { "axis": 1 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1024, 1024] } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1024, 1024] } } |
| }, |
| { |
| "name": "rows_256_cols_4096_axis1", |
| "preset": "smoke", |
| "args": { "axis": 1 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [256, 4096], "dist": "uniform", "seed": 31 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [256, 4096], "dist": "empty" } }, |
| "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "2 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "axis0_rows_16_inner_65536", |
| "preset": "smoke", |
| "args": { "axis": 0 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [16, 65536], "dist": "uniform", "seed": 32 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [16, 65536], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "axis0_rows_64_longaxis_65536", |
| "preset": "smoke", |
| "args": { "axis": 0 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [65536, 64], "dist": "uniform", "seed": 34 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [65536, 64], "dist": "empty" } }, |
| "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "2 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "axis0_rows_512_longaxis_4096_contrast", |
| "preset": "smoke", |
| "args": { "axis": 0 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [4096, 512], "dist": "uniform", "seed": 35 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [4096, 512], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "PATHO_thread_row_serial_per_row_dim128_axis0", |
| "preset": "smoke", |
| "args": { "axis": 0 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [128, 2048], "dist": "uniform", "seed": 41 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [128, 2048], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "HEALTHY_wg_strided_dim4096_axis0", |
| "preset": "smoke", |
| "args": { "axis": 0 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [4096, 2048], "dist": "uniform", "seed": 42 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [4096, 2048], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "PATHO_splitk_1d_inner1_long_axis0", |
| "preset": "smoke", |
| "args": { "axis": 0 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [262144], "dist": "uniform", "seed": 43 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [262144], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "PATHO_f16_mirror_primary_rows256_cols4096_axis1", |
| "preset": "smoke", |
| "args": { "axis": 1 }, |
| "inputs": { "x": { "dtype": "float16", "shape": [256, 4096], "dist": "uniform", "seed": 44 } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [256, 4096], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 2 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "PATHO_f16_strided_axis0_rows64_longaxis_65536", |
| "preset": "smoke", |
| "args": { "axis": 0 }, |
| "inputs": { "x": { "dtype": "float16", "shape": [65536, 64], "dist": "uniform", "seed": 45 } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [65536, 64], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 2 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "cumsum-f16-sequence-axis0-s32768-h128-ssm-state", |
| "preset": "stress", |
| "args": { "axis": 0 }, |
| "inputs": { "x": { "dtype": "float16", "shape": [32768, 128], "dist": "uniform", "seed": 46 } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [32768, 128], "dist": "empty" } }, |
| "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "2 * 2 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "PATHO_wg_strided_fewrows_longaxis_dim6144_axis0", |
| "preset": "stress", |
| "args": { "axis": 0 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [6144, 64], "dist": "normal", "seed": 51, "scale": 2 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [6144, 64], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 4 * numel(shapes.x)" }] } |
| }, |
| { |
| "name": "PATHO_splitk_scalar_demotion_f16_longaxis_axis0", |
| "preset": "stress", |
| "args": { "axis": 0 }, |
| "inputs": { "x": { "dtype": "float16", "shape": [16384, 64], "dist": "normal", "seed": 52, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [16384, 64], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 2 * numel(shapes.x)" }] } |
| } |
| ] |
| } |
|
|