| { |
| "domain": "com.microsoft", |
| "name": "FastGelu", |
| "sinceVersion": 1, |
| "description": "Applies the GELU (Gaussian Error Linear Unit) activation using a tanh approximation: `Y = 0.5 * X * (1 + tanh(0.797885 * X + 0.035677 * X^3))`. An optional `bias` is added to `X` before the activation is computed. This WebGPU package implements float16 and float32; the schema-allowed double and bfloat16 types are not supported.", |
| "inputs": [ |
| { "role": "X", "dtype": "T", "description": "Values transformed by FastGelu after adding the optional `bias`." }, |
| { |
| "role": "bias", |
| "dtype": "T", |
| "rank": 1, |
| "optional": true, |
| "description": "Optional 1-D bias added to `X` along the last dimension before the GELU activation." |
| } |
| ], |
| "outputs": [ |
| { |
| "role": "Y", |
| "dtype": "T", |
| "rank": "ranks.X", |
| "shape": "shapes.X", |
| "description": "Output tensor after applying the GELU activation; same shape as `X`." |
| } |
| ], |
| "typeConstraints": { "T": ["float32", "float16"] }, |
| "args": { |
| "X": { "kind": "tensor", "semantic": "X", "role": "input" }, |
| "bias": { "kind": "tensor", "semantic": "bias", "role": "input", "required": false }, |
| "Y": { "kind": "tensor", "semantic": "Y", "role": "output" } |
| }, |
| "tunables": { "WORKGROUP_SIZE": 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) <= device.limits.maxComputeWorkgroupsPerDimension * device.limits.maxComputeWorkgroupsPerDimension * 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" |
| }, |
| "constants": { "scalar": "dtypes.T", "usesF16": "dtypes.T == \"f16\"", "approximate": "\"tanh\"" }, |
| "bindingSets": { |
| "vec4Bias": [ |
| { |
| "name": "x", |
| "arg": "X", |
| "semantic": "X", |
| "buffer": { "type": "read-only-storage" }, |
| "elementType": "$vectorScalar" |
| }, |
| { |
| "name": "bias", |
| "arg": "bias", |
| "semantic": "bias", |
| "buffer": { "type": "read-only-storage" }, |
| "elementType": "$scalar", |
| "length": "$hidden" |
| }, |
| { "name": "y", "arg": "Y", "semantic": "Y", "buffer": { "type": "storage" }, "elementType": "$vectorScalar" }, |
| { |
| "name": "params", |
| "semantic": "kernel.params", |
| "buffer": { "type": "uniform" }, |
| "struct": { "name": "Params", "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.X) / 4" }] } |
| } |
| ], |
| "vec4": [ |
| { |
| "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" }] } |
| } |
| ], |
| "scalarBias": [ |
| { "name": "x", "arg": "X", "semantic": "X", "buffer": { "type": "read-only-storage" }, "elementType": "$scalar" }, |
| { |
| "name": "bias", |
| "arg": "bias", |
| "semantic": "bias", |
| "buffer": { "type": "read-only-storage" }, |
| "elementType": "$scalar" |
| }, |
| { "name": "y", "arg": "Y", "semantic": "Y", "buffer": { "type": "storage" }, "elementType": "$scalar" }, |
| { |
| "name": "params", |
| "semantic": "kernel.params", |
| "buffer": { "type": "uniform" }, |
| "struct": { "name": "Params", "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.X)" }] } |
| } |
| ], |
| "scalar": [ |
| { "name": "x", "arg": "X", "semantic": "X", "buffer": { "type": "read-only-storage" }, "elementType": "$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_bias", |
| "priority": 30, |
| "when": ["baseOk", "biasOk", "vec4Ok"], |
| "constants": { |
| "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": "vec4Bias", |
| "dispatch": { "threads": "numel(shapes.X) / 4", "workgroupSize": "workgroupSize" } |
| } |
| ] |
| }, |
| { |
| "id": "vec4_no_bias", |
| "priority": 25, |
| "when": ["baseOk", "noBiasOk", "vec4Ok"], |
| "constants": { "vec4": true, "vec4Tail": false, "hasBias": false, "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"" }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "FastGelu.vec4", |
| "shader": "elementwise-bias-gelu.wgsl.jinja", |
| "bindings": "vec4", |
| "dispatch": { "threads": "numel(shapes.X) / 4", "workgroupSize": "workgroupSize" } |
| } |
| ] |
| }, |
| { |
| "id": "vec4_tail_bias", |
| "priority": 20, |
| "when": ["baseOk", "biasOk", "numel(shapes.X) > 0"], |
| "constants": { |
| "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": "scalarBias", |
| "dispatch": { "threads": "ceilDiv(numel(shapes.X), 4)", "workgroupSize": "workgroupSize" } |
| } |
| ] |
| }, |
| { |
| "id": "vec4_tail_no_bias", |
| "priority": 15, |
| "when": ["baseOk", "noBiasOk", "numel(shapes.X) > 0"], |
| "constants": { "vec4": false, "vec4Tail": true, "hasBias": false }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "FastGelu.vec4Tail", |
| "shader": "elementwise-bias-gelu.wgsl.jinja", |
| "bindings": "scalar", |
| "dispatch": { "threads": "ceilDiv(numel(shapes.X), 4)", "workgroupSize": "workgroupSize" } |
| } |
| ] |
| }, |
| { |
| "id": "scalar_bias", |
| "priority": 10, |
| "when": ["baseOk", "biasOk", "true"], |
| "constants": { |
| "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": "scalarBias", |
| "dispatch": { "threads": "numel(shapes.X)", "workgroupSize": "workgroupSize" } |
| } |
| ] |
| }, |
| { |
| "id": "scalar_no_bias", |
| "priority": 0, |
| "when": ["baseOk", "noBiasOk", "true"], |
| "constants": { "vec4": false, "vec4Tail": false, "hasBias": false }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "FastGelu.scalar", |
| "shader": "elementwise-bias-gelu.wgsl.jinja", |
| "bindings": "scalar", |
| "dispatch": { "threads": "numel(shapes.X)", "workgroupSize": "workgroupSize" } |
| } |
| ] |
| } |
| ] |
| } |
|
|
