{
  "op": "com.microsoft.BiasSoftmax",
  "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256], "BLOCK_COLS": [1024, 2048, 4096, 8192] },
  "cases": [
    {
      "name": "biassoftmax-f32-attn-32x512x512",
      "preset": "smoke",
      "vars": { "dtype": "float32" },
      "attrs": { "axis": 2, "is_inner_broadcast": 0 },
      "inputs": {
        "data": { "shape": [32, 512, 512], "dtype": "float32", "dist": "normal", "seed": 510, "scale": 2 },
        "bias": { "shape": [1, 512, 512], "dtype": "float32", "dist": "normal", "seed": 511, "scale": 1 }
      },
      "outputs": { "output": { "shape": [32, 512, 512], "dtype": "float32" } },
      "bench": {
        "primary": true,
        "metrics": [{ "type": "bandwidth", "value": "(32 * 512 * 512 * 2 + 512 * 512) * dtypeBytes(args.dtype)" }]
      }
    },
    {
      "name": "biassoftmax-f32-suffix-64x16x256",
      "preset": "stress",
      "vars": { "dtype": "float32" },
      "attrs": { "axis": 1, "is_inner_broadcast": 0 },
      "inputs": {
        "data": { "shape": [64, 16, 256], "dtype": "float32", "dist": "normal", "seed": 512, "scale": 2 },
        "bias": { "shape": [1, 16, 256], "dtype": "float32", "dist": "normal", "seed": 513, "scale": 1 }
      },
      "outputs": { "output": { "shape": [64, 16, 256], "dtype": "float32" } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "(64 * 16 * 256 * 2 + 16 * 256) * dtypeBytes(args.dtype)" }]
      }
    },
    {
      "name": "biassoftmax-f16-attn-32x512x512",
      "preset": "model",
      "vars": { "dtype": "float16" },
      "attrs": { "axis": 2, "is_inner_broadcast": 0 },
      "inputs": {
        "data": { "shape": [32, 512, 512], "dtype": "float16", "dist": "normal", "seed": 514, "scale": 2 },
        "bias": { "shape": [1, 512, 512], "dtype": "float16", "dist": "normal", "seed": 515, "scale": 1 }
      },
      "outputs": { "output": { "shape": [32, 512, 512], "dtype": "float16" } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "(32 * 512 * 512 * 2 + 512 * 512) * dtypeBytes(args.dtype)" }]
      }
    },
    {
      "name": "biassoftmax-f32-lowocc-8x524288-axis1",
      "preset": "stress",
      "vars": { "dtype": "float32" },
      "attrs": { "axis": 1, "is_inner_broadcast": 0 },
      "inputs": {
        "data": { "shape": [8, 524288], "dtype": "float32", "dist": "normal", "seed": 530, "scale": 2 },
        "bias": { "shape": [1, 524288], "dtype": "float32", "dist": "normal", "seed": 531, "scale": 1 }
      },
      "outputs": { "output": { "shape": [8, 524288], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "(8 * 524288 * 2 + 524288) * dtypeBytes(args.dtype)" }] }
    },
    {
      "name": "biassoftmax-f32-launchbound-262144x2-axis1",
      "preset": "stress",
      "vars": { "dtype": "float32" },
      "attrs": { "axis": 1, "is_inner_broadcast": 0 },
      "inputs": {
        "data": { "shape": [262144, 2], "dtype": "float32", "dist": "normal", "seed": 532, "scale": 2 },
        "bias": { "shape": [1, 2], "dtype": "float32", "dist": "normal", "seed": 533, "scale": 1 }
      },
      "outputs": { "output": { "shape": [262144, 2], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "(262144 * 2 * 2 + 2) * dtypeBytes(args.dtype)" }] }
    }
  ]
}