| { |
| "op": "com.microsoft.Gelu", |
| "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] }, |
| "cases": [ |
| { |
| "name": "gelu-f32-1m", |
| "preset": "smoke", |
| "vars": { "dtype": "float32", "count": 1048576 }, |
| "inputs": { "X": { "shape": [1048576], "dtype": "float32", "dist": "normal", "seed": 650, "scale": 0.5 } }, |
| "outputs": { "Y": { "shape": [1048576], "dtype": "float32" } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] |
| } |
| }, |
| { |
| "name": "gelu-f32-8m", |
| "preset": "smoke", |
| "vars": { "dtype": "float32", "count": 8388608 }, |
| "inputs": { "X": { "shape": [8388608], "dtype": "float32", "seed": 7001, "dist": "normal", "scale": 0.5 } }, |
| "outputs": { "Y": { "shape": [8388608], "dtype": "float32" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] } |
| }, |
| { |
| "name": "gelu-f16-8m", |
| "preset": "smoke", |
| "vars": { "dtype": "float16", "count": 8388608 }, |
| "inputs": { "X": { "shape": [8388608], "dtype": "float16", "seed": 7002, "dist": "normal", "scale": 0.5 } }, |
| "outputs": { "Y": { "shape": [8388608], "dtype": "float16" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] } |
| }, |
| { |
| "name": "gelu-bert-base-b8-s512-i3072", |
| "preset": "model", |
| "provenance": { |
| "notes": "BERT class defaults (intermediate_size 3072): one FFN activation tensor for a batch of 8 full-length sequences." |
| }, |
| "vars": { "dtype": "float32", "count": 12582912 }, |
| "inputs": { "X": { "shape": [12582912], "dtype": "float32", "dist": "normal", "seed": 5900, "scale": 0.5 } }, |
| "outputs": { "Y": { "shape": [12582912], "dtype": "float32" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] } |
| }, |
| { |
| "name": "gelu-whisper-encoder-s1500-i1536", |
| "preset": "model", |
| "provenance": { |
| "notes": "Whisper class defaults (encoder_ffn_dim 1536, max_source_positions 1500): one encoder FFN activation for a full 30-second window." |
| }, |
| "vars": { "dtype": "float32", "count": 2304000 }, |
| "inputs": { "X": { "shape": [2304000], "dtype": "float32", "dist": "normal", "seed": 6000, "scale": 0.5 } }, |
| "outputs": { "Y": { "shape": [2304000], "dtype": "float32" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] } |
| }, |
| { |
| "name": "gelu-clip-text-b64-s77-i2048", |
| "preset": "model", |
| "provenance": { |
| "notes": "CLIP text class defaults (intermediate_size 2048, max_position_embeddings 77) at a 64-caption batch." |
| }, |
| "vars": { "dtype": "float32", "count": 10092544 }, |
| "inputs": { "X": { "shape": [10092544], "dtype": "float32", "dist": "normal", "seed": 6100, "scale": 0.5 } }, |
| "outputs": { "Y": { "shape": [10092544], "dtype": "float32" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] } |
| } |
| ] |
| } |
|
|
