{
  "op": "ai.onnx.Where",
  "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
  "cases": [
    {
      "name": "where-f32-1m",
      "preset": "smoke",
      "vars": { "dtype": "float32", "count": 1048576 },
      "inputs": {
        "condition": { "shape": [1048576], "dtype": "bool", "dist": "uniform", "seed": 870, "scale": 1 },
        "x": { "shape": [1048576], "dtype": "float32", "dist": "normal", "seed": 871, "scale": 1 },
        "y": { "shape": [1048576], "dtype": "float32", "dist": "normal", "seed": 872, "scale": 1 }
      },
      "outputs": { "output": { "shape": [1048576], "dtype": "float32" } },
      "bench": {
        "primary": true,
        "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 3 + args.count * 4" }]
      }
    },
    {
      "name": "where-f32-rank4-broadcast",
      "preset": "smoke",
      "vars": { "dtype": "float32", "count": 98304 },
      "inputs": {
        "condition": { "shape": [1, 1, 128, 1], "dtype": "bool", "dist": "linearMod", "mod": 2 },
        "x": { "shape": [2, 1, 128, 384], "dtype": "float32", "dist": "normal", "seed": 873, "scale": 1 },
        "y": { "shape": [1, 1, 1, 384], "dtype": "float32", "dist": "normal", "seed": 874, "scale": 1 }
      },
      "outputs": { "output": { "shape": [2, 1, 128, 384], "dtype": "float32" } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 3 + args.count * 4" }]
      }
    },
    {
      "name": "where-f32-zero-size",
      "preset": "edge",
      "vars": { "dtype": "float32", "count": 0 },
      "inputs": {
        "condition": { "shape": [0, 1], "dtype": "bool", "dist": "empty" },
        "x": { "shape": [0, 3], "dtype": "float32", "dist": "empty" },
        "y": { "shape": [1, 3], "dtype": "float32", "dist": "normal", "seed": 875, "scale": 1 }
      },
      "outputs": { "output": { "shape": [0, 3], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "0" }] }
    },
    {
      "name": "where-f32-1m-scalar-condition",
      "vars": { "dtype": "float32", "count": 1048576 },
      "inputs": {
        "condition": { "shape": [1], "dtype": "bool", "data": { "kind": "values", "values": [1] } },
        "x": { "shape": [1048576], "dtype": "float32", "dist": "normal", "seed": 875, "scale": 1 },
        "y": { "shape": [1048576], "dtype": "float32", "dist": "normal", "seed": 876, "scale": 1 }
      },
      "outputs": { "output": { "shape": [1048576], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 3" }] }
    },
    {
      "name": "where-f16-1m",
      "preset": "smoke",
      "vars": { "dtype": "float16", "count": 1048576 },
      "inputs": {
        "condition": { "shape": [1048576], "dtype": "bool", "dist": "uniform", "seed": 970, "scale": 1 },
        "x": { "shape": [1048576], "dtype": "float16", "dist": "normal", "seed": 971, "scale": 1 },
        "y": { "shape": [1048576], "dtype": "float16", "dist": "normal", "seed": 972, "scale": 1 }
      },
      "outputs": { "output": { "shape": [1048576], "dtype": "float16" } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 3 + args.count * 4" }]
      }
    },
    {
      "name": "where-f32-1m-odd-scalar-fallback",
      "preset": "smoke",
      "vars": { "dtype": "float32", "count": 1048575 },
      "inputs": {
        "condition": { "shape": [1048575], "dtype": "bool", "dist": "uniform", "seed": 973, "scale": 1 },
        "x": { "shape": [1048575], "dtype": "float32", "dist": "normal", "seed": 974, "scale": 1 },
        "y": { "shape": [1048575], "dtype": "float32", "dist": "normal", "seed": 975, "scale": 1 }
      },
      "outputs": { "output": { "shape": [1048575], "dtype": "float32" } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 3 + args.count * 4" }]
      }
    },
    {
      "name": "where-f32-large-rank4-broadcast-scalar",
      "preset": "smoke",
      "vars": { "dtype": "float32", "count": 1048576 },
      "inputs": {
        "condition": { "shape": [1, 1, 512, 1], "dtype": "bool", "dist": "linearMod", "mod": 2 },
        "x": { "shape": [4, 1, 512, 512], "dtype": "float32", "dist": "normal", "seed": 976, "scale": 1 },
        "y": { "shape": [1], "dtype": "float32", "data": { "kind": "values", "values": [0.0] } }
      },
      "outputs": { "output": { "shape": [4, 1, 512, 512], "dtype": "float32" } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 3 + args.count * 4" }]
      }
    },
    {
      "name": "where-f32-4m-mixed-axis-broadcast-fallback",
      "preset": "stress",
      "vars": { "dtype": "float32", "count": 4194304 },
      "inputs": {
        "condition": { "shape": [512, 1, 1], "dtype": "bool", "dist": "linearMod", "mod": 2 },
        "x": { "shape": [2, 512, 1, 8], "dtype": "float32", "dist": "normal", "seed": 9101, "scale": 2 },
        "y": { "shape": [1, 1, 512, 8], "dtype": "float32", "dist": "normal", "seed": 9102, "scale": 2 }
      },
      "outputs": { "output": { "shape": [2, 512, 512, 8], "dtype": "float32" } },
      "bench": {
        "metrics": [
          {
            "type": "bandwidth",
            "value": "(numel(shapes.x) + numel(shapes.y) + numel(shapes.output)) * dtypeBytes(args.dtype) + numel(shapes.condition) * 4"
          }
        ]
      }
    }
  ]
}