{ "domain": "com.microsoft", "name": "BiasAdd", "sinceVersion": 1, "inputs": { "X": { "dtype": "T", "rank": 3 }, "bias": { "dtype": "T", "rank": 1 }, "skip": { "dtype": "T", "rank": 3 } }, "outputs": { "Y": { "dtype": "T", "rank": 3, "shape": "shapes.X" } }, "typeConstraints": { "T": ["float32", "float16"] }, "tunables": { "WORKGROUP_SIZE": { "default": 256 } }, "derive": { "scalar": "dtypes.T", "hidden": "dim(shapes.X, ranks.X - 1) if dim(shapes.X, ranks.X - 1) > 0 else 1" }, "when": ["ranks.X == 3", "ranks.skip == 3", "ranks.Y == 3", "sameShape(shapes.X, shapes.skip)", "sameShape(shapes.X, shapes.Y)", "f16Ok(dtypes.T)", "ranks.bias == 1", "dim(shapes.bias, 0) == dim(shapes.X, 2)"], "bindings": { "x": { "arg": "X", "buffer": "read-only-storage", "elementType": "$scalar" } }, "variants": [ { "id": "vec4", "priority": 30, "when": ["numel(shapes.X) > 0", "numel(shapes.X) % 4 == 0", "dim(shapes.X, 2) % 4 == 0"], "derive": { "vec4": true, "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"" }, "passes": [ { "id": "main", "name": "BiasAdd.vec4", "shader": "bias-add.wgsl.jinja", "bindings": [ "x", { "arg": "bias", "elementType": "$scalar", "length": "$hidden" }, "skip", { "arg": "Y", "name": "y", "elementType": "$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": "scalar", "priority": 0, "derive": { "vec4": false }, "passes": [ { "id": "main", "name": "BiasAdd.scalar", "shader": "bias-add.wgsl.jinja", "derive": { "itemsPerInvocation": 4 }, "bindings": [ "x", "bias", "skip", { "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 } } ] } ] }