| { |
| "op": "ai.onnx.Cast", |
| "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] }, |
| "cases": [ |
| { |
| "name": "cast-f32-u32-1m", |
| "preset": "smoke", |
| "vars": { "inputDtype": "float32", "outputDtype": "uint32", "count": 1048576 }, |
| "attrs": { "to": 12 }, |
| "inputs": { |
| "x": { "shape": [1048576], "dtype": "float32", "dist": "uniform", "seed": 103, "scale": 32, "signed": false } |
| }, |
| "outputs": { "y": { "shape": [1048576], "dtype": "uint32" } }, |
| "bench": { |
| "primary": true, |
| "metrics": [ |
| { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" } |
| ] |
| } |
| }, |
| { |
| "name": "cast-f32-u32-1m-plus3", |
| "preset": "smoke", |
| "vars": { "inputDtype": "float32", "outputDtype": "uint32", "count": 1048579 }, |
| "attrs": { "to": 12 }, |
| "inputs": { |
| "x": { "shape": [1048579], "dtype": "float32", "dist": "uniform", "seed": 104, "scale": 32, "signed": false } |
| }, |
| "outputs": { "y": { "shape": [1048579], "dtype": "uint32" } }, |
| "bench": { |
| "metrics": [ |
| { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" } |
| ] |
| } |
| }, |
| { |
| "name": "cast-f32-u32-16m", |
| "preset": "smoke", |
| "vars": { "inputDtype": "float32", "outputDtype": "uint32", "count": 16777216 }, |
| "attrs": { "to": 12 }, |
| "inputs": { |
| "x": { "shape": [16777216], "dtype": "float32", "dist": "uniform", "seed": 108, "scale": 32, "signed": false } |
| }, |
| "outputs": { "y": { "shape": [16777216], "dtype": "uint32" } }, |
| "bench": { |
| "metrics": [ |
| { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" } |
| ] |
| } |
| }, |
| { |
| "name": "cast-f32-f16-1m", |
| "preset": "smoke", |
| "vars": { "inputDtype": "float32", "outputDtype": "float16", "count": 1048576 }, |
| "attrs": { "to": 10 }, |
| "inputs": { "x": { "shape": [1048576], "dtype": "float32", "dist": "normal", "seed": 105 } }, |
| "outputs": { "y": { "shape": [1048576], "dtype": "float16" } }, |
| "bench": { |
| "metrics": [ |
| { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" } |
| ] |
| } |
| }, |
| { |
| "name": "cast-f16-f32-1m", |
| "preset": "smoke", |
| "vars": { "inputDtype": "float16", "outputDtype": "float32", "count": 1048576 }, |
| "attrs": { "to": 1 }, |
| "inputs": { "x": { "shape": [1048576], "dtype": "float16", "dist": "normal", "seed": 106 } }, |
| "outputs": { "y": { "shape": [1048576], "dtype": "float32" } }, |
| "bench": { |
| "metrics": [ |
| { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" } |
| ] |
| } |
| }, |
| { |
| "name": "cast-i8-f32-1m-plus3", |
| "preset": "smoke", |
| "vars": { "inputDtype": "int8", "outputDtype": "float32", "count": 1048579 }, |
| "attrs": { "to": 1 }, |
| "inputs": { |
| "x": { "shape": [1048579], "dtype": "int8", "dist": "uniform", "seed": 107, "scale": 127, "signed": true } |
| }, |
| "outputs": { "y": { "shape": [1048579], "dtype": "float32" } }, |
| "bench": { |
| "metrics": [ |
| { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" } |
| ] |
| } |
| } |
| ] |
| } |
|
|
