{
  "domain": "ai.onnx",
  "name": "BitCast",
  "sinceVersion": 26,
  "description": "Reinterprets the raw bit pattern of a tensor as a different data type without any value conversion. The target type must have the same bit-width as the input type, and the output tensor has the same shape as the input.",
  "inputs": [{ "role": "input", "dtype": "T", "description": "Input tensor to be bitwise reinterpreted." }],
  "outputs": [
    {
      "role": "output",
      "dtype": "U",
      "rank": "ranks.input",
      "description": "Output tensor with the same shape as the input, reinterpreted as the target type.",
      "shape": "shapes.input"
    }
  ],
  "attributes": {},
  "attributeDescriptions": {
    "to": "Required TensorProto DataType enum integer naming the output dtype; the target type must have the same bit-width as the input type."
  },
  "attributeConstraints": { "to": { "required": true } },
  "typeConstraints": {
    "T": ["float32", "int8", "int32", "uint8", "uint32"],
    "U": ["float32", "int8", "int32", "uint8", "uint32"]
  },
  "args": {
    "input": { "kind": "tensor", "semantic": "input", "role": "input" },
    "output": { "kind": "tensor", "semantic": "output", "role": "output" }
  },
  "tunables": { "WORKGROUP_SIZE": 256 },
  "derive": {
    "deviceWorkgroupCap": "min(device.limits.maxComputeInvocationsPerWorkgroup, device.limits.maxComputeWorkgroupSizeX)",
    "bitcastShapeOk": "ranks.input >= 0 and ranks.output == ranks.input and numel(shapes.input) == numel(shapes.output)",
    "bitcastTypeOk": "attrs.to == onnxDtypeCode(logicalDtypes.U) and (((tensorDtypes.input == \"int8\" or tensorDtypes.input == \"uint8\") and (tensorDtypes.output == \"int8\" or tensorDtypes.output == \"uint8\")) or (tensorDtypes.input != \"int8\" and tensorDtypes.input != \"uint8\" and tensorDtypes.output != \"int8\" and tensorDtypes.output != \"uint8\"))",
    "bitcastWorkgroupSize": "min(tunables.WORKGROUP_SIZE, deviceWorkgroupCap)",
    "bitcastDispatchFits": "ceilDiv(ceilDiv(numel(shapes.output), bitcastWorkgroupSize), device.limits.maxComputeWorkgroupsPerDimension) <= device.limits.maxComputeWorkgroupsPerDimension",
    "bitcastBaseOk": "bitcastShapeOk and bitcastTypeOk and bitcastDispatchFits"
  },
  "constants": {
    "inScalar": "dtypes.T",
    "outScalar": "dtypes.U",
    "inputIsInt8": "tensorDtypes.input == \"int8\"",
    "inputIsUint8": "tensorDtypes.input == \"uint8\"",
    "outputIsInt8": "tensorDtypes.output == \"int8\"",
    "outputIsUint8": "tensorDtypes.output == \"uint8\""
  },
  "bindingSets": {
    "vec4Slots": [
      {
        "name": "input",
        "arg": "input",
        "semantic": "input",
        "buffer": { "type": "read-only-storage" },
        "elementType": "$inVec4"
      },
      {
        "name": "output",
        "arg": "output",
        "semantic": "output",
        "buffer": { "type": "storage" },
        "elementType": "$outVec4"
      },
      {
        "name": "params",
        "semantic": "kernel.params",
        "buffer": { "type": "uniform" },
        "struct": {
          "name": "Params",
          "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.output) / 4" }]
        }
      }
    ],
    "scalarSlots": [
      {
        "name": "input",
        "arg": "input",
        "semantic": "input",
        "buffer": { "type": "read-only-storage" },
        "elementType": "$inScalar"
      },
      {
        "name": "output",
        "arg": "output",
        "semantic": "output",
        "buffer": { "type": "storage" },
        "elementType": "$outScalar"
      },
      {
        "name": "params",
        "semantic": "kernel.params",
        "buffer": { "type": "uniform" },
        "struct": { "name": "Params", "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.output)" }] }
      }
    ]
  },
  "variants": [
    {
      "id": "slot32_vec4",
      "priority": 20,
      "when": ["numel(shapes.input) % 4 == 0", "bitcastBaseOk"],
      "constants": { "inVec4": "\"vec4<\" ~ dtypes.T ~ \">\"", "outVec4": "\"vec4<\" ~ dtypes.U ~ \">\"" },
      "passes": [
        {
          "id": "main",
          "name": "BitCast.vec4",
          "source": {
            "shader": "bitcast.wgsl.jinja",
            "inputs": { "vectorized": true, "workgroupSize": "bitcastWorkgroupSize" }
          },
          "bindings": "vec4Slots",
          "dispatch": { "threads": "numel(shapes.output) / 4", "workgroupSize": "bitcastWorkgroupSize" }
        }
      ]
    },
    {
      "id": "slot32",
      "when": ["bitcastBaseOk"],
      "passes": [
        {
          "id": "main",
          "name": "BitCast",
          "source": { "shader": "bitcast.wgsl.jinja", "inputs": { "workgroupSize": "bitcastWorkgroupSize" } },
          "bindings": "scalarSlots",
          "dispatch": { "threads": "numel(shapes.output)", "workgroupSize": "bitcastWorkgroupSize" }
        }
      ]
    }
  ]
}