| { |
| "op": "ai.onnx.Hardmax", |
| "cases": [ |
| { |
| "name": "rows_4096_cols_128", |
| "vars": { "rows": 4096, "cols": 128, "dtype": "float32" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [4096, 128] } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [4096, 128] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] |
| }, |
| "preset": "smoke" |
| }, |
| { |
| "name": "rows_4096_cols_512", |
| "vars": { "rows": 4096, "cols": 512, "dtype": "float32" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [4096, 512] } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [4096, 512] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] } |
| }, |
| { |
| "name": "rows_4096_cols_128_f16", |
| "vars": { "rows": 4096, "cols": 128, "dtype": "float16" }, |
| "inputs": { "x": { "dtype": "float16", "shape": [4096, 128] } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [4096, 128] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] } |
| }, |
| { |
| "name": "rows_1024_cols_1025", |
| "vars": { "rows": 1024, "cols": 1025, "dtype": "float32" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1024, 1025] } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1024, 1025] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] } |
| }, |
| { |
| "name": "rank3_axis1_scalar_fallback", |
| "vars": { "outer": 1024, "axisDim": 1024, "inner": 8, "dtype": "float32" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1024, 1024, 8] } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1024, 1024, 8] } }, |
| "bench": { |
| "metrics": [ |
| { "type": "bandwidth", "value": "args.outer * args.axisDim * args.inner * dtypeBytes(args.dtype) * 2" } |
| ] |
| } |
| }, |
| { |
| "name": "unaligned_last_axis_127_scalar_cliff", |
| "vars": { "rows": 16384, "cols": 127, "dtype": "float32" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [16384, 127] } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [16384, 127] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] } |
| }, |
| { |
| "name": "rows_1024_cols_1025_f16_subgroup", |
| "vars": { "rows": 1024, "cols": 1025, "dtype": "float16" }, |
| "inputs": { "x": { "dtype": "float16", "shape": [1024, 1025] } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [1024, 1025] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] } |
| }, |
| { |
| "name": "axis0_huge_axis_few_threads_launch_poor", |
| "preset": "stress", |
| "vars": { "outer": 1, "axisDim": 65536, "inner": 64, "dtype": "float32" }, |
| "attrs": { "axis": 0 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [65536, 64], "dist": "normal", "seed": 1301, "scale": 2 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [65536, 64], "dist": "empty" } }, |
| "bench": { |
| "metrics": [{ "type": "bandwidth", "value": "args.axisDim * args.inner * dtypeBytes(args.dtype) * 2" }] |
| } |
| }, |
| { |
| "name": "unaligned_last_axis_1023_scalar_row_cliff", |
| "preset": "edge", |
| "vars": { "rows": 16384, "cols": 1023, "dtype": "float32" }, |
| "attrs": { "axis": -1 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [16384, 1023], "dist": "normal", "seed": 2207, "scale": 2 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [16384, 1023], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype) * 2" }] } |
| } |
| ] |
| } |
|
|