{ "domain": "com.microsoft", "name": "BiasSplitGelu", "sinceVersion": 1, "inputs": { "X": { "dtype": "T", "rank": 3 }, "bias": { "dtype": "T", "rank": 1 } }, "outputs": { "Y": { "dtype": "T", "rank": 3, "shape": "[dim(shapes.X, 0), dim(shapes.X, 1), dim(shapes.X, 2) / 2]" } }, "typeConstraints": { "T": ["float32", "float16"] }, "tunables": { "WORKGROUP_SIZE": { "default": 64 } }, "when": ["ranks.X == 3", "ranks.Y == 3", "dim(shapes.Y, 0) == dim(shapes.X, 0)", "dim(shapes.Y, 1) == dim(shapes.X, 1)", "dim(shapes.X, 2) % 2 == 0", "dim(shapes.Y, 2) == dim(shapes.X, 2) / 2", "f16Ok(dtypes.T)", "ranks.bias == 1", "dim(shapes.bias, 0) == dim(shapes.X, 2)"], "bindings": { "x": { "arg": "X", "elementType": "$scalar" }, "bias": { "elementType": "$scalar", "length": "$twice" }, "y_y": { "arg": "Y", "name": "y", "elementType": "$scalar" }, "params_main": { "name": "params", "struct": [ { "name": "count", "type": "u32", "value": "(numel(shapes.Y) / (dim(shapes.X, ranks.X - 1) / 2)) * ceilDiv(dim(shapes.X, ranks.X - 1) / 2, 4)" } ] }, "bias_main": { "name": "bias", "elementType": "$scalar" }, "params_count": { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.Y)" }] } }, "variants": [ { "id": "vec4", "priority": 30, "when": ["numel(shapes.Y) > 0", "numel(shapes.Y) % 4 == 0", "dim(shapes.Y, 2) % 4 == 0"], "derive": { "vec4": true, "vec4Tail": false, "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"", "hidden": "dim(shapes.X, ranks.X - 1) / 2 if dim(shapes.X, ranks.X - 1) > 0 else 1", "twice": "dim(shapes.X, ranks.X - 1) if dim(shapes.X, ranks.X - 1) > 0 else 2" }, "passes": [ { "id": "main", "name": "BiasSplitGelu.Vec4", "shader": "bias-split-gelu.wgsl.jinja", "bindings": [ "x", "bias", { "arg": "Y", "name": "y", "elementType": "$vectorScalar" }, { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.Y) / 4" }] } ], "dispatch": { "x": "min(ceilDiv((numel(shapes.Y) / 4), (tunables.WORKGROUP_SIZE)), 65535)", "y": "ceilDiv(ceilDiv((numel(shapes.Y) / 4), (tunables.WORKGROUP_SIZE)), 65535)", "z": 1 } } ] }, { "id": "vec4_tail", "priority": 20, "when": ["numel(shapes.Y) > 0"], "derive": { "vec4": false, "vec4Tail": true, "hidden": "dim(shapes.X, ranks.X - 1) / 2 if dim(shapes.X, ranks.X - 1) > 0 else 1", "twice": "dim(shapes.X, ranks.X - 1) if dim(shapes.X, ranks.X - 1) > 0 else 2" }, "passes": [ { "id": "main", "name": "BiasSplitGelu.Vec4Tail", "shader": "bias-split-gelu.wgsl.jinja", "bindings": ["x", "bias", "y_y", "params_main"], "dispatch": { "x": "min(ceilDiv(((numel(shapes.Y) / (dim(shapes.X, ranks.X - 1) / 2)) * ceilDiv(dim(shapes.X, ranks.X - 1) / 2, 4)), (tunables.WORKGROUP_SIZE)), 65535)", "y": "ceilDiv(ceilDiv(((numel(shapes.Y) / (dim(shapes.X, ranks.X - 1) / 2)) * ceilDiv(dim(shapes.X, ranks.X - 1) / 2, 4)), (tunables.WORKGROUP_SIZE)), 65535)", "z": 1 } } ] }, { "id": "scalar", "priority": 0, "when": ["true"], "derive": { "vec4": false, "vec4Tail": false, "hidden": "dim(shapes.X, ranks.X - 1) / 2 if dim(shapes.X, ranks.X - 1) > 0 else 1", "twice": "dim(shapes.X, ranks.X - 1) if dim(shapes.X, ranks.X - 1) > 0 else 2" }, "passes": [ { "id": "main", "name": "BiasSplitGelu.Scalar", "shader": "bias-split-gelu.wgsl.jinja", "bindings": ["x", "bias_main", "y_y", "params_count"], "dispatch": { "x": "min(ceilDiv((numel(shapes.Y)), (tunables.WORKGROUP_SIZE)), 65535)", "y": "ceilDiv(ceilDiv((numel(shapes.Y)), (tunables.WORKGROUP_SIZE)), 65535)", "z": 1 } } ] } ], "derive": { "scalar": "dtypes.T" } }