{
  "op": "ai.onnx.BitCast",
  "cases": [
    {
      "name": "bitcast-f32-to-u32-33m",
      "preset": "stress",
      "provenance": { "notes": "Stress-only capacity case: input plus output occupy exactly 256 MiB of GPU storage." },
      "attrs": { "to": 12 },
      "inputs": { "input": { "dtype": "float32", "shape": [4096, 8192], "dist": "normal", "seed": 915, "scale": 1 } },
      "outputs": { "output": { "dtype": "uint32", "shape": [4096, 8192] } },
      "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "4096 * 8192 * 4 * 2" }] }
    },
    {
      "name": "bitcast-f32-to-u32-1m",
      "preset": "smoke",
      "attrs": { "to": 12 },
      "inputs": { "input": { "dtype": "float32", "shape": [1048576], "dist": "normal", "seed": 913, "scale": 1 } },
      "outputs": { "output": { "dtype": "uint32", "shape": [1048576] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.input) + numel(shapes.output)) * 4" }] }
    },
    {
      "name": "bitcast-f32-to-u32-scalar-fallback-4m",
      "preset": "edge",
      "attrs": { "to": 12 },
      "inputs": { "input": { "dtype": "float32", "shape": [4194305], "dist": "normal", "seed": 917, "scale": 1 } },
      "outputs": { "output": { "dtype": "uint32", "shape": [4194305] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "4194305 * 4 * 2" }] }
    }
  ]
}