{ "domain": "com.microsoft", "name": "QuickGelu", "sinceVersion": 1, "inputs": { "X": { "dtype": "T" } }, "outputs": { "Y": { "dtype": "T", "rank": "ranks.X", "shape": "shapes.X" } }, "attributes": { "alpha": { "default": 1.702 } }, "typeConstraints": { "T": ["float32", "float16"] }, "tunables": { "WORKGROUP_SIZE": { "default": 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", "scalar": "dtypes.T", "alpha": "attrs.alpha" }, "when": ["baseOk"], "variants": [ { "id": "vec4", "priority": 20, "when": ["vec4Ok"], "derive": { "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"", "vec4": true, "vec4Tail": false }, "passes": [ { "id": "main", "name": "QuickGelu.Vec4", "shader": "quick-gelu.wgsl.jinja", "bindings": [ { "arg": "X", "name": "x", "elementType": "$vectorScalar" }, { "arg": "Y", "name": "y", "elementType": "$vectorScalar" }, { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.X) / 4" }] } ], "dispatch": { "x": "min(ceilDiv((numel(shapes.X) / 4), (tunables.WORKGROUP_SIZE)), 65535)", "y": "ceilDiv(ceilDiv((numel(shapes.X) / 4), (tunables.WORKGROUP_SIZE)), 65535)", "z": 1 } } ] }, { "id": "vec4_tail", "priority": 10, "derive": { "vec4": false, "vec4Tail": "numel(shapes.X) > 0" }, "passes": [ { "id": "main", "name": "QuickGelu.Vec4Tail", "shader": "quick-gelu.wgsl.jinja", "bindings": [ { "arg": "X", "name": "x", "elementType": "$scalar" }, { "arg": "Y", "name": "y", "elementType": "$scalar" }, { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.X)" }] } ], "dispatch": { "x": "min(ceilDiv((ceilDiv(numel(shapes.X), 4)), (tunables.WORKGROUP_SIZE)), 65535)", "y": "ceilDiv(ceilDiv((ceilDiv(numel(shapes.X), 4)), (tunables.WORKGROUP_SIZE)), 65535)", "z": 1 } } ] } ] }