| { |
| "op": "com.microsoft.GatedAdd", |
| "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] }, |
| "cases": [ |
| { |
| "name": "gatedadd-f32-t2048-h2048", |
| "preset": "smoke", |
| "vars": { "dtype": "float32" }, |
| "inputs": { |
| "X": { "shape": [2048, 2048], "dtype": "float32", "dist": "normal", "seed": 7101, "scale": 2 }, |
| "Y": { "shape": [2048, 2048], "dtype": "float32", "dist": "normal", "seed": 7102, "scale": 2 }, |
| "gate": { "shape": [2048, 1], "dtype": "float32", "dist": "normal", "seed": 7103, "scale": 1 } |
| }, |
| "outputs": { "output": { "shape": [2048, 2048], "dtype": "float32" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2048 * 2048 * 3 * dtypeBytes(args.dtype)" }] } |
| }, |
| { |
| "name": "gatedadd-f16-t2048-h2048", |
| "preset": "model", |
| "vars": { "dtype": "float16" }, |
| "inputs": { |
| "X": { "shape": [2048, 2048], "dtype": "float16", "dist": "normal", "seed": 7104, "scale": 2 }, |
| "Y": { "shape": [2048, 2048], "dtype": "float16", "dist": "normal", "seed": 7105, "scale": 2 }, |
| "gate": { "shape": [2048, 1], "dtype": "float16", "dist": "normal", "seed": 7106, "scale": 1 } |
| }, |
| "outputs": { "output": { "shape": [2048, 2048], "dtype": "float16" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2048 * 2048 * 3 * dtypeBytes(args.dtype)" }] } |
| }, |
| { |
| "name": "gatedadd-f32-channels-odd-h1023", |
| "preset": "stress", |
| "vars": { "dtype": "float32" }, |
| "inputs": { |
| "X": { "shape": [4096, 1023], "dtype": "float32", "dist": "normal", "seed": 7107, "scale": 2 }, |
| "Y": { "shape": [4096, 1023], "dtype": "float32", "dist": "normal", "seed": 7108, "scale": 2 }, |
| "gate": { "shape": [4096, 1], "dtype": "float32", "dist": "normal", "seed": 7109, "scale": 1 } |
| }, |
| "outputs": { "output": { "shape": [4096, 1023], "dtype": "float32" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "4096 * 1023 * 3 * dtypeBytes(args.dtype)" }] } |
| } |
| ] |
| } |
|
|
