{ "domain": "com.microsoft", "name": "QuickGelu", "sinceVersion": 1, "description": "Computes `x * sigmoid(alpha * x)` elementwise, a fast approximation of GELU activation. The output has the same shape as the input. This WebGPU package implements float16 and float32; the schema-allowed double and bfloat16 types are not supported.", "inputs": [{ "role": "X", "dtype": "T", "description": "Input tensor of any shape." }], "outputs": [ { "role": "Y", "dtype": "T", "rank": "ranks.X", "shape": "shapes.X", "description": "Output tensor; same shape as the input." } ], "attributes": { "alpha": 1.702 }, "attributeDescriptions": { "alpha": "Scalar multiplier applied to `x` inside the sigmoid; defaults to 1.702, which approximates GELU." }, "typeConstraints": { "T": ["float32", "float16"] }, "args": { "X": { "kind": "tensor", "semantic": "X", "role": "input" }, "Y": { "kind": "tensor", "semantic": "Y", "role": "output" } }, "tunables": { "WORKGROUP_SIZE": 256 }, "derive": { "deviceWorkgroupCap": "min(device.limits.maxComputeInvocationsPerWorkgroup, device.limits.maxComputeWorkgroupSizeX)", "workgroupOk": "tunables.WORKGROUP_SIZE > 0 and tunables.WORKGROUP_SIZE <= deviceWorkgroupCap", "baseOk": "workgroupOk and numel(shapes.X) == numel(shapes.Y) and f16Ok(dtypes.T)", "vec4Ok": "numel(shapes.X) > 0 and numel(shapes.X) % 4 == 0" }, "bindingSets": { "scalarTail": [ { "name": "x", "arg": "X", "semantic": "X", "buffer": { "type": "read-only-storage" }, "elementType": "$scalar" }, { "name": "y", "arg": "Y", "semantic": "Y", "buffer": { "type": "storage" }, "elementType": "$scalar" }, { "name": "params", "semantic": "kernel.params", "buffer": { "type": "uniform" }, "struct": { "name": "Params", "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.X)" }] } } ] }, "variants": [ { "id": "vec4", "priority": 20, "when": ["baseOk", "vec4Ok"], "constants": { "scalar": "dtypes.T", "usesF16": "dtypes.T == \"f16\"", "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"", "vec4": true, "vec4Tail": false }, "passes": [ { "id": "main", "name": "QuickGelu.vec4", "source": { "shader": "quick-gelu.wgsl.jinja", "inputs": { "alpha": "attrs.alpha" } }, "bindings": [ { "name": "x", "arg": "X", "semantic": "X", "buffer": { "type": "read-only-storage" }, "elementType": "$vectorScalar" }, { "name": "y", "arg": "Y", "semantic": "Y", "buffer": { "type": "storage" }, "elementType": "$vectorScalar" }, { "name": "params", "semantic": "kernel.params", "buffer": { "type": "uniform" }, "struct": { "name": "Params", "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.X) / 4" }] } } ], "dispatch": { "threads": "numel(shapes.X) / 4", "workgroupSize": "tunables.WORKGROUP_SIZE" } } ] }, { "id": "vec4_tail", "priority": 10, "when": ["baseOk", "numel(shapes.X) > 0"], "constants": { "scalar": "dtypes.T", "usesF16": "dtypes.T == \"f16\"", "vec4": false, "vec4Tail": true }, "passes": [ { "id": "main", "name": "QuickGelu.vec4Tail", "source": { "shader": "quick-gelu.wgsl.jinja", "inputs": { "alpha": "attrs.alpha" } }, "bindings": "scalarTail", "dispatch": { "threads": "ceilDiv(numel(shapes.X), 4)", "workgroupSize": "tunables.WORKGROUP_SIZE" } } ] }, { "id": "scalar", "priority": 0, "when": ["baseOk", "true"], "constants": { "scalar": "dtypes.T", "usesF16": "dtypes.T == \"f16\"", "vec4": false, "vec4Tail": false }, "passes": [ { "id": "main", "name": "QuickGelu.scalar", "source": { "shader": "quick-gelu.wgsl.jinja", "inputs": { "alpha": "attrs.alpha" } }, "bindings": "scalarTail", "dispatch": { "threads": "numel(shapes.X)", "workgroupSize": "tunables.WORKGROUP_SIZE" } } ] } ] }