{
  "op": "ai.onnx.Sum",
  "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
  "cases": [
    {
      "name": "sum-f32-4m",
      "preset": "smoke",
      "vars": { "dtype": "float32", "count": 4194304 },
      "inputs": {
        "a": { "dtype": "float32", "shape": [4194304], "dist": "normal", "seed": 702, "scale": 2 },
        "b": { "dtype": "float32", "shape": [4194304], "dist": "normal", "seed": 703, "scale": 2 }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4194304] } },
      "bench": {
        "primary": true,
        "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 3" }]
      }
    },
    {
      "name": "sum-f32-broadcast-scalar-4m-unaligned",
      "preset": "stress",
      "vars": { "dtype": "float32", "count": 4194302 },
      "inputs": {
        "a": { "dtype": "float32", "shape": [4194302], "dist": "normal", "seed": 711, "scale": 2 },
        "b": { "dtype": "float32", "shape": [4194302], "dist": "normal", "seed": 712, "scale": 2 }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [4194302] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 3" }] }
    },
    {
      "name": "sum-f32-four-input-rank4-mixed-broadcast-pathology",
      "preset": "stress",
      "vars": { "dtype": "float32", "count": 2097152 },
      "inputs": {
        "a": { "dtype": "float32", "shape": [8, 64, 64, 64], "dist": "normal", "seed": 713, "scale": 2 },
        "b": { "dtype": "float32", "shape": [1, 64, 1, 1], "dist": "normal", "seed": 714, "scale": 0.25 },
        "c": { "dtype": "float32", "shape": [8, 1, 64, 64], "dist": "normal", "seed": 715, "scale": 0.5 },
        "d": { "dtype": "float32", "shape": [1], "data": { "kind": "constant", "value": 0.125 } }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [8, 64, 64, 64], "dist": "empty" } },
      "bench": {
        "metrics": [
          {
            "type": "bandwidth",
            "value": "dtypeBytes(args.dtype) * (numel(shapes.a) + numel(shapes.b) + numel(shapes.c) + numel(shapes.d) + numel(shapes.y))"
          }
        ]
      }
    },
    {
      "name": "sum-f32-three-input-rank4-mixed-broadcast-pathology",
      "preset": "stress",
      "vars": { "dtype": "float32", "count": 2097152 },
      "inputs": {
        "a": { "dtype": "float32", "shape": [8, 64, 64, 64], "dist": "normal", "seed": 716, "scale": 2 },
        "b": { "dtype": "float32", "shape": [1, 64, 1, 1], "dist": "normal", "seed": 717, "scale": 0.25 },
        "c": { "dtype": "float32", "shape": [8, 1, 64, 64], "dist": "normal", "seed": 718, "scale": 0.5 }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [8, 64, 64, 64], "dist": "empty" } },
      "bench": {
        "primary": true,
        "metrics": [
          {
            "type": "bandwidth",
            "value": "dtypeBytes(args.dtype) * (numel(shapes.a) + numel(shapes.b) + numel(shapes.c) + numel(shapes.y))"
          }
        ]
      }
    }
  ]
}