{
  "domain": "com.microsoft",
  "name": "FastGelu",
  "sinceVersion": 1,
  "inputs": { "X": { "dtype": "T" }, "bias": { "dtype": "T", "rank": 1, "optional": true } },
  "outputs": { "Y": { "dtype": "T", "rank": "ranks.X", "shape": "shapes.X" } },
  "typeConstraints": { "T": ["float32", "float16"] },
  "tunables": { "WORKGROUP_SIZE": { "default": 256 } },
  "derive": {
    "deviceWorkgroupCap": "min(device.limits.maxComputeInvocationsPerWorkgroup, device.limits.maxComputeWorkgroupSizeX)",
    "storageBufferLimit": "min(device.limits.maxStorageBufferBindingSize, device.limits.maxBufferSize)",
    "workgroupSize": "min(tunables.WORKGROUP_SIZE, deviceWorkgroupCap)",
    "buffersFit": "numel(shapes.X) * dtypeBytes(dtypes.T) <= storageBufferLimit and numel(shapes.Y) * dtypeBytes(dtypes.T) <= storageBufferLimit and (not present.bias or numel(shapes.bias) * dtypeBytes(dtypes.T) <= storageBufferLimit)",
    "dispatchFits": "numel(shapes.X) <= min(device.limits.maxComputeWorkgroupsPerDimension, 65535) * min(device.limits.maxComputeWorkgroupsPerDimension, 65535) * workgroupSize * 4",
    "baseOk": "workgroupSize > 0 and ranks.X >= 1 and numel(shapes.X) == numel(shapes.Y) and f16Ok(dtypes.T) and buffersFit and dispatchFits",
    "biasOk": "present.bias and ranks.bias == 1 and dim(shapes.bias, 0) == dim(shapes.X, ranks.X - 1)",
    "noBiasOk": "not present.bias",
    "vec4Ok": "numel(shapes.X) > 0 and numel(shapes.X) % 4 == 0 and dim(shapes.X, ranks.X - 1) % 4 == 0",
    "scalar": "dtypes.T",
    "approximate": "\"tanh\""
  },
  "when": ["baseOk"],
  "bindings": {
    "x": { "arg": "X", "buffer": "read-only-storage", "elementType": "$vectorScalar" },
    "y": { "arg": "Y", "buffer": "storage", "elementType": "$vectorScalar" },
    "params": { "buffer": "uniform", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.X) / 4" }] },
    "x_2": { "arg": "X", "name": "x", "buffer": "read-only-storage", "elementType": "$scalar" },
    "y_2": { "arg": "Y", "name": "y", "buffer": "storage", "elementType": "$scalar" },
    "params_2": {
      "name": "params",
      "buffer": "uniform",
      "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.X)" }]
    }
  },
  "variants": [
    {
      "id": "vec4_bias",
      "priority": 30,
      "when": ["biasOk", "vec4Ok"],
      "derive": {
        "vec4": true,
        "vec4Tail": false,
        "hasBias": true,
        "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"",
        "hidden": "dim(shapes.X, ranks.X - 1) if dim(shapes.X, ranks.X - 1) > 0 else 1"
      },
      "passes": [
        {
          "id": "main",
          "name": "FastGelu.vec4Bias",
          "shader": "elementwise-bias-gelu.wgsl.jinja",
          "bindings": ["x", { "arg": "bias", "elementType": "$scalar", "length": "$hidden" }, "y", "params"],
          "dispatch": {
            "x": "min(ceilDiv((numel(shapes.X) / 4), (workgroupSize)), 65535)",
            "y": "ceilDiv(ceilDiv((numel(shapes.X) / 4), (workgroupSize)), 65535)",
            "z": 1
          }
        }
      ]
    },
    {
      "id": "vec4_no_bias",
      "priority": 25,
      "when": ["noBiasOk", "vec4Ok"],
      "derive": { "vec4": true, "vec4Tail": false, "hasBias": false, "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"" },
      "passes": [
        {
          "id": "main",
          "name": "FastGelu.vec4",
          "shader": "elementwise-bias-gelu.wgsl.jinja",
          "bindings": ["x", "y", "params"],
          "dispatch": {
            "x": "min(ceilDiv((numel(shapes.X) / 4), (workgroupSize)), 65535)",
            "y": "ceilDiv(ceilDiv((numel(shapes.X) / 4), (workgroupSize)), 65535)",
            "z": 1
          }
        }
      ]
    },
    {
      "id": "vec4_tail_bias",
      "priority": 20,
      "when": ["biasOk", "numel(shapes.X) > 0"],
      "derive": {
        "vec4": false,
        "vec4Tail": true,
        "hasBias": true,
        "hidden": "dim(shapes.X, ranks.X - 1) if dim(shapes.X, ranks.X - 1) > 0 else 1"
      },
      "passes": [
        {
          "id": "main",
          "name": "FastGelu.vec4TailBias",
          "shader": "elementwise-bias-gelu.wgsl.jinja",
          "bindings": ["x_2", "bias", "y_2", "params_2"],
          "dispatch": {
            "x": "min(ceilDiv((ceilDiv(numel(shapes.X), 4)), (workgroupSize)), 65535)",
            "y": "ceilDiv(ceilDiv((ceilDiv(numel(shapes.X), 4)), (workgroupSize)), 65535)",
            "z": 1
          }
        }
      ]
    },
    {
      "id": "vec4_tail_no_bias",
      "priority": 15,
      "when": ["noBiasOk", "numel(shapes.X) > 0"],
      "derive": { "vec4": false, "vec4Tail": true, "hasBias": false },
      "passes": [
        {
          "id": "main",
          "name": "FastGelu.vec4Tail",
          "shader": "elementwise-bias-gelu.wgsl.jinja",
          "bindings": ["x_2", "y_2", "params_2"],
          "dispatch": {
            "x": "min(ceilDiv((ceilDiv(numel(shapes.X), 4)), (workgroupSize)), 65535)",
            "y": "ceilDiv(ceilDiv((ceilDiv(numel(shapes.X), 4)), (workgroupSize)), 65535)",
            "z": 1
          }
        }
      ]
    },
    {
      "id": "scalar_bias",
      "priority": 10,
      "when": ["biasOk", "true"],
      "derive": {
        "vec4": false,
        "vec4Tail": false,
        "hasBias": true,
        "hidden": "dim(shapes.X, ranks.X - 1) if dim(shapes.X, ranks.X - 1) > 0 else 1"
      },
      "passes": [
        {
          "id": "main",
          "name": "FastGelu.scalarBias",
          "shader": "elementwise-bias-gelu.wgsl.jinja",
          "bindings": ["x_2", "bias", "y_2", "params_2"],
          "dispatch": {
            "x": "min(ceilDiv((numel(shapes.X)), (workgroupSize)), 65535)",
            "y": "ceilDiv(ceilDiv((numel(shapes.X)), (workgroupSize)), 65535)",
            "z": 1
          }
        }
      ]
    },
    {
      "id": "scalar_no_bias",
      "priority": 0,
      "when": ["noBiasOk", "true"],
      "derive": { "vec4": false, "vec4Tail": false, "hasBias": false },
      "passes": [
        {
          "id": "main",
          "name": "FastGelu.scalar",
          "shader": "elementwise-bias-gelu.wgsl.jinja",
          "bindings": ["x_2", "y_2", "params_2"],
          "dispatch": {
            "x": "min(ceilDiv((numel(shapes.X)), (workgroupSize)), 65535)",
            "y": "ceilDiv(ceilDiv((numel(shapes.X)), (workgroupSize)), 65535)",
            "z": 1
          }
        }
      ]
    }
  ]
}