| { |
| "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] }, |
| "cases": [ |
| { |
| "name": "castlike-f32-to-u32-1m", |
| "preset": "smoke", |
| "vars": { "inputDtype": "float32", "outputDtype": "uint32", "count": 1048576 }, |
| "inputs": { |
| "x": { "dtype": "float32", "shape": [1048576], "dist": "uniform", "seed": 912, "scale": 255 }, |
| "target": { "dtype": "uint32", "shape": [1], "data": { "kind": "values", "values": [0] } } |
| }, |
| "outputs": { "y": { "dtype": "uint32", "shape": [1048576] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [ |
| { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" } |
| ] |
| } |
| }, |
| { |
| "name": "castlike-f32-to-u32-1m-plus3", |
| "preset": "smoke", |
| "vars": { "inputDtype": "float32", "outputDtype": "uint32", "count": 1048579 }, |
| "inputs": { |
| "x": { "dtype": "float32", "shape": [1048579], "dist": "uniform", "seed": 913, "scale": 255 }, |
| "target": { "dtype": "uint32", "shape": [0], "data": { "kind": "values", "values": [] } } |
| }, |
| "outputs": { "y": { "dtype": "uint32", "shape": [1048579] } }, |
| "bench": { |
| "metrics": [ |
| { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" } |
| ] |
| } |
| }, |
| { |
| "name": "castlike-f32-to-f16-1m", |
| "preset": "smoke", |
| "vars": { "inputDtype": "float32", "outputDtype": "float16", "count": 1048576 }, |
| "inputs": { |
| "x": { "dtype": "float32", "shape": [1048576], "dist": "normal", "seed": 914 }, |
| "target": { "dtype": "float16", "shape": [0], "data": { "kind": "values", "values": [] } } |
| }, |
| "outputs": { "y": { "dtype": "float16", "shape": [1048576] } }, |
| "bench": { |
| "metrics": [ |
| { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" } |
| ] |
| } |
| }, |
| { |
| "name": "castlike-i8-to-f32-1m-plus3", |
| "preset": "smoke", |
| "vars": { "inputDtype": "int8", "outputDtype": "float32", "count": 1048579 }, |
| "inputs": { |
| "x": { "dtype": "int8", "shape": [1048579], "dist": "uniform", "seed": 915, "min": -128, "max": 128 }, |
| "target": { "dtype": "float32", "shape": [0], "data": { "kind": "values", "values": [] } } |
| }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1048579] } }, |
| "bench": { |
| "metrics": [ |
| { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" } |
| ] |
| } |
| }, |
| { |
| "name": "castlike-f16-to-f32-unaligned-1m-plus1", |
| "preset": "edge", |
| "vars": { "inputDtype": "float16", "outputDtype": "float32", "count": 1048577 }, |
| "inputs": { |
| "x": { "dtype": "float16", "shape": [1048577], "dist": "normal", "seed": 9231 }, |
| "target": { "dtype": "float32", "shape": [0], "data": { "kind": "values", "values": [] } } |
| }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1048577] } }, |
| "bench": { |
| "metrics": [ |
| { "type": "bandwidth", "value": "args.count * (dtypeBytes(args.inputDtype) + dtypeBytes(args.outputDtype))" } |
| ] |
| } |
| } |
| ] |
| } |
|
|
