{ "domain": "com.microsoft", "name": "GatedAdd", "sinceVersion": 1, "inputs": { "X": { "dtype": "T" }, "Y": { "dtype": "T" }, "gate": { "dtype": "T" } }, "outputs": { "output": { "dtype": "T", "rank": "ranks.X", "shape": "shapes.X" } }, "typeConstraints": { "T": ["float32", "float16"] }, "tunables": { "WORKGROUP_SIZE": { "default": 256 } }, "derive": { "channels": "dim(shapes.X, ranks.X - 1)", "gateContract": "ranks.X >= 1 and channels > 0 and ranks.Y == ranks.X and ranks.gate == ranks.X and sameShape(shapes.Y, shapes.X) and sameShape(shapes.output, shapes.X) and dim(shapes.gate, ranks.gate - 1) == 1 and sameShape(prefix(shapes.gate, ranks.gate - 1), prefix(shapes.X, ranks.X - 1)) and f16Ok(dtypes.T)", "vec4Rows": "channels % 4 == 0 and numel(shapes.X) % 4 == 0", "scalar": "dtypes.T", "hidden": "channels if channels > 0 else 1" }, "when": ["gateContract"], "variants": [ { "id": "vec4", "priority": 30, "when": ["vec4Rows", "numel(shapes.X) > 0"], "derive": { "vec4": true, "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"" }, "passes": [ { "id": "main", "name": "GatedAdd.vec4", "shader": "gated-add.wgsl.jinja", "bindings": [ { "arg": "X", "name": "x", "elementType": "$vectorScalar" }, { "arg": "Y", "name": "y", "elementType": "$vectorScalar" }, "gate", { "arg": "output", "elementType": "$vectorScalar" }, { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.X) / 4" }] } ], "dispatch": { "x": "min(ceilDiv((numel(shapes.X) / 4), (tunables.WORKGROUP_SIZE)), 65535)", "y": "ceilDiv(ceilDiv((numel(shapes.X) / 4), (tunables.WORKGROUP_SIZE)), 65535)", "z": 1 } } ] }, { "id": "scalar", "priority": 0, "derive": { "vec4": false }, "passes": [ { "id": "main", "name": "GatedAdd.scalar", "shader": "gated-add.wgsl.jinja", "derive": { "itemsPerInvocation": 4 }, "bindings": [ { "arg": "X", "name": "x", "elementType": "$scalar" }, { "arg": "Y", "name": "y", "elementType": "$scalar" }, "gate", "output", { "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 } } ] } ] }