| { |
| "cases": [ |
| { |
| "name": "nchw_32_128_16_16", |
| "attrs": { "p": 2 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [32, 128, 16, 16] } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [32, 128, 1, 1] } } |
| }, |
| { |
| "name": "nchw_16x256x64x64_p2_wg_parallel", |
| "preset": "smoke", |
| "vars": { "batch": 16, "channels": 256, "spatial": 4096 }, |
| "attrs": { "p": 2 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [16, 256, 64, 64], "dist": "normal", "seed": 203 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [16, 256, 1, 1], "dist": "empty" } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "args.batch * args.channels * args.spatial * 4" }] |
| } |
| }, |
| { |
| "name": "nchw_f16_16x256x64x64_p2_wg_parallel", |
| "preset": "smoke", |
| "vars": { "batch": 16, "channels": 256, "spatial": 4096 }, |
| "attrs": { "p": 2 }, |
| "inputs": { "x": { "dtype": "float16", "shape": [16, 256, 64, 64], "dist": "normal", "seed": 209 } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [16, 256, 1, 1], "dist": "empty" } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "args.batch * args.channels * args.spatial * 2" }] |
| } |
| }, |
| { |
| "name": "f32_p3_inner961_below_perf_floor_1x4096x31x31", |
| "preset": "stress", |
| "attrs": { "p": 3 }, |
| "vars": { "batch": 1, "channels": 4096, "spatial": 961 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 4096, 31, 31], "dist": "normal", "seed": 401, "scale": 1.5 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 4096, 1, 1], "dist": "empty" } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "args.batch * args.channels * args.spatial * 4" }] |
| } |
| } |
| ], |
| "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] } |
| } |
|
|