{
  "op": "ai.onnx.BitwiseNot",
  "cases": [
    {
      "name": "u32_33m",
      "preset": "stress",
      "provenance": { "notes": "Stress-only capacity case: input plus output occupy exactly 256 MiB of GPU storage." },
      "inputs": { "x": { "dtype": "uint32", "shape": [33554432], "data": { "kind": "constant", "value": 305419896 } } },
      "outputs": { "y": { "dtype": "uint32", "shape": [33554432] } },
      "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "33554432 * 4 * 2" }] }
    },
    {
      "name": "u32_1m",
      "preset": "smoke",
      "inputs": { "x": { "dtype": "uint32", "shape": [1048576], "data": { "kind": "constant", "value": 305419896 } } },
      "outputs": { "y": { "dtype": "uint32", "shape": [1048576] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] }
    },
    {
      "name": "u32_33m_scalar_fallback",
      "preset": "stress",
      "inputs": {
        "x": { "dtype": "uint32", "shape": [33554433], "dist": "uniform", "seed": 701, "min": 0, "max": 4294967295 }
      },
      "outputs": { "y": { "dtype": "uint32", "shape": [33554433] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "33554433 * 4 * 2", "name": "scalar-fallback BW" }] }
    }
  ]
}