| { |
| "domain": "ai.onnx", |
| "name": "ConvInteger", |
| "sinceVersion": 10, |
| "inputs": { |
| "x": { "dtype": "TX" }, |
| "w": { "dtype": "TW" }, |
| "x_zero_point": { "dtype": "TX", "optional": true }, |
| "w_zero_point": { "dtype": "TW", "optional": true } |
| }, |
| "outputs": { |
| "y": { |
| "dtype": "TY", |
| "rank": "ranks.x", |
| "shape": "[dim(shapes.x, 0), dim(shapes.w, 0), expectedOutputWidth] if ranks.x == 3 else ([dim(shapes.x, 0), dim(shapes.w, 0), expectedOutputHeight, expectedOutputWidth] if ranks.x == 4 else [dim(shapes.x, 0), dim(shapes.w, 0), expectedOutputDepth, expectedOutputHeight, expectedOutputWidth])" |
| } |
| }, |
| "attributes": { |
| "auto_pad": { "default": "NOTSET" }, |
| "group": { "default": 1 }, |
| "dilations": {}, |
| "kernel_shape": {}, |
| "pads": {}, |
| "strides": {} |
| }, |
| "attributeConstraints": { "auto_pad": { "values": ["NOTSET", "SAME_UPPER", "SAME_LOWER", "VALID"] } }, |
| "typeConstraints": { "TX": ["uint8", "int8"], "TW": ["uint8", "int8"], "TY": ["int32"] }, |
| "tunables": { "WORKGROUP_SIZE": { "default": 256 } }, |
| "derive": { |
| "inputDepth": "dim(shapes.x, 2) if ranks.x == 5 else 1", |
| "inputHeight": "dim(shapes.x, 3) if ranks.x == 5 else (dim(shapes.x, 2) if ranks.x == 4 else 1)", |
| "inputWidth": "dim(shapes.x, ranks.x - 1) if ranks.x >= 3 else 1", |
| "outputDepth": "dim(shapes.y, 2) if ranks.y == 5 else 1", |
| "outputHeight": "dim(shapes.y, 3) if ranks.y == 5 else (dim(shapes.y, 2) if ranks.y == 4 else 1)", |
| "outputWidth": "dim(shapes.y, ranks.y - 1) if ranks.y >= 3 else 1", |
| "kernelDepth": "dim(shapes.w, 2) if ranks.w == 5 else 1", |
| "kernelHeight": "dim(shapes.w, 3) if ranks.w == 5 else (dim(shapes.w, 2) if ranks.w == 4 else 1)", |
| "kernelWidth": "dim(shapes.w, ranks.w - 1) if ranks.w >= 3 else 1", |
| "spatialRank": "ranks.w - 2", |
| "kernelShapeLengthOk": "not has(attrs, \"kernel_shape\") or (attrs.kernel_shape | length) == spatialRank", |
| "stridesLengthOk": "not has(attrs, \"strides\") or (attrs.strides | length) == spatialRank", |
| "dilationsLengthOk": "not has(attrs, \"dilations\") or (attrs.dilations | length) == spatialRank", |
| "padsLengthOk": "not has(attrs, \"pads\") or (attrs.pads | length) == 2 * spatialRank", |
| "kernelD": "attrs.kernel_shape[0] if kernelShapeLengthOk and has(attrs, \"kernel_shape\") and spatialRank == 3 else 1", |
| "kernelH": "attrs.kernel_shape[spatialRank - 2] if kernelShapeLengthOk and has(attrs, \"kernel_shape\") and spatialRank >= 2 else 1", |
| "kernelW": "attrs.kernel_shape[spatialRank - 1] if kernelShapeLengthOk and has(attrs, \"kernel_shape\") and spatialRank >= 1 else 1", |
| "strideD": "attrs.strides[0] if stridesLengthOk and has(attrs, \"strides\") and spatialRank == 3 else 1", |
| "strideH": "attrs.strides[spatialRank - 2] if stridesLengthOk and has(attrs, \"strides\") and spatialRank >= 2 else 1", |
| "strideW": "attrs.strides[spatialRank - 1] if stridesLengthOk and has(attrs, \"strides\") and spatialRank >= 1 else 1", |
| "dilationD": "attrs.dilations[0] if dilationsLengthOk and has(attrs, \"dilations\") and spatialRank == 3 else 1", |
| "dilationH": "attrs.dilations[spatialRank - 2] if dilationsLengthOk and has(attrs, \"dilations\") and spatialRank >= 2 else 1", |
| "dilationW": "attrs.dilations[spatialRank - 1] if dilationsLengthOk and has(attrs, \"dilations\") and spatialRank >= 1 else 1", |
| "padFront": "attrs.pads[0] if padsLengthOk and has(attrs, \"pads\") and spatialRank == 3 else 0", |
| "padTop": "attrs.pads[spatialRank - 2] if padsLengthOk and has(attrs, \"pads\") and spatialRank >= 2 else 0", |
| "padLeft": "attrs.pads[spatialRank - 1] if padsLengthOk and has(attrs, \"pads\") and spatialRank >= 1 else 0", |
| "padBack": "attrs.pads[spatialRank] if padsLengthOk and has(attrs, \"pads\") and spatialRank == 3 else 0", |
| "padBottom": "attrs.pads[2 * spatialRank - 2] if padsLengthOk and has(attrs, \"pads\") and spatialRank >= 2 else 0", |
| "padRight": "attrs.pads[2 * spatialRank - 1] if padsLengthOk and has(attrs, \"pads\") and spatialRank >= 1 else 0", |
| "autoPadSame": "attrs.auto_pad == \"SAME_UPPER\" or attrs.auto_pad == \"SAME_LOWER\"", |
| "autoPadValid": "attrs.auto_pad == \"VALID\"", |
| "samePadDepth": "max(0, (outputDepth - 1) * strideD + (kernelDepth - 1) * dilationD + 1 - inputDepth)", |
| "samePadHeight": "max(0, (outputHeight - 1) * strideH + (kernelHeight - 1) * dilationH + 1 - inputHeight)", |
| "samePadWidth": "max(0, (outputWidth - 1) * strideW + (kernelWidth - 1) * dilationW + 1 - inputWidth)", |
| "samePadFront": "floor(samePadDepth / 2) if attrs.auto_pad == \"SAME_UPPER\" else samePadDepth - floor(samePadDepth / 2)", |
| "samePadTop": "floor(samePadHeight / 2) if attrs.auto_pad == \"SAME_UPPER\" else samePadHeight - floor(samePadHeight / 2)", |
| "samePadLeft": "floor(samePadWidth / 2) if attrs.auto_pad == \"SAME_UPPER\" else samePadWidth - floor(samePadWidth / 2)", |
| "effectivePadFront": "samePadFront if autoPadSame else (0 if autoPadValid else padFront)", |
| "effectivePadTop": "samePadTop if autoPadSame else (0 if autoPadValid else padTop)", |
| "effectivePadLeft": "samePadLeft if autoPadSame else (0 if autoPadValid else padLeft)", |
| "expectedOutputDepth": "ceil(inputDepth / strideD) if autoPadSame else floor((inputDepth + (0 if autoPadValid else padFront + padBack) - ((kernelDepth - 1) * dilationD + 1)) / strideD) + 1", |
| "expectedOutputHeight": "ceil(inputHeight / strideH) if autoPadSame else floor((inputHeight + (0 if autoPadValid else padTop + padBottom) - ((kernelHeight - 1) * dilationH + 1)) / strideH) + 1", |
| "expectedOutputWidth": "ceil(inputWidth / strideW) if autoPadSame else floor((inputWidth + (0 if autoPadValid else padLeft + padRight) - ((kernelWidth - 1) * dilationW + 1)) / strideW) + 1", |
| "spatialAttributeLengthsOk": "kernelShapeLengthOk and stridesLengthOk and dilationsLengthOk and padsLengthOk", |
| "kernelShapeMatchesWeights": "not has(attrs, \"kernel_shape\") or (kernelW == kernelWidth and (spatialRank < 2 or kernelH == kernelHeight) and (spatialRank < 3 or kernelD == kernelDepth))", |
| "kernelExtentsOk": "kernelWidth >= 1 and (spatialRank < 2 or kernelHeight >= 1) and (spatialRank < 3 or kernelDepth >= 1)", |
| "stridesValuesOk": "not has(attrs, \"strides\") or (strideD >= 1 and floor(strideD) == strideD and strideH >= 1 and floor(strideH) == strideH and strideW >= 1 and floor(strideW) == strideW)", |
| "dilationsValuesOk": "not has(attrs, \"dilations\") or (dilationD >= 1 and floor(dilationD) == dilationD and dilationH >= 1 and floor(dilationH) == dilationH and dilationW >= 1 and floor(dilationW) == dilationW)", |
| "padsValuesOk": "padFront >= 0 and floor(padFront) == padFront and padTop >= 0 and floor(padTop) == padTop and padLeft >= 0 and floor(padLeft) == padLeft and padBack >= 0 and floor(padBack) == padBack and padBottom >= 0 and floor(padBottom) == padBottom and padRight >= 0 and floor(padRight) == padRight", |
| "explicitPadsOk": "attrs.auto_pad == \"NOTSET\" or not has(attrs, \"pads\")", |
| "spatialAttributesOk": "spatialRank >= 1 and spatialRank <= 3 and spatialAttributeLengthsOk and kernelShapeMatchesWeights and kernelExtentsOk and stridesValuesOk and dilationsValuesOk and padsValuesOk and explicitPadsOk", |
| "packedFeature": "device.wgslLanguageFeatures.has(\"packed_4x8_integer_dot_product\")", |
| "rank4TensorOk": "spatialAttributesOk and (ranks.x == 4 and ranks.w == 4 and ranks.y == 4)", |
| "outputBatchChannelsOk": "dim(shapes.y, 0) == dim(shapes.x, 0) and dim(shapes.y, 1) == dim(shapes.w, 0)", |
| "groupChannelsOk": "attrs.group >= 1 and dim(shapes.w, 1) * attrs.group == dim(shapes.x, 1) and dim(shapes.w, 0) % attrs.group == 0", |
| "output2dShapeOk": "dim(shapes.y, 2) == expectedOutputHeight and dim(shapes.y, 3) == expectedOutputWidth", |
| "aScalar": "\"u32\" if tensorDtypes.w == \"uint8\" else \"i32\"", |
| "bScalar": "\"u32\" if tensorDtypes.x == \"uint8\" else \"i32\"", |
| "wScalar": "\"u32\" if tensorDtypes.w == \"uint8\" else \"i32\"", |
| "xScalar": "\"u32\" if tensorDtypes.x == \"uint8\" else \"i32\"", |
| "aUnsigned": "tensorDtypes.w == \"uint8\"", |
| "bUnsigned": "tensorDtypes.x == \"uint8\"", |
| "wUnsigned": "tensorDtypes.w == \"uint8\"", |
| "xUnsigned": "tensorDtypes.x == \"uint8\"", |
| "bZeroPerColumn": false, |
| "rank5TensorOk": "spatialAttributesOk and (ranks.x == 5 and ranks.w == 5 and ranks.y == 5)", |
| "scalarXZeroOk": "ranks.x_zero_point == 0 or (ranks.x_zero_point == 1 and dim(shapes.x_zero_point, 0) == 1)", |
| "scalarWZeroOk": "ranks.w_zero_point == 0 or (ranks.w_zero_point == 1 and dim(shapes.w_zero_point, 0) == 1)", |
| "scalarXZeroRequiredOk": "present.x_zero_point and scalarXZeroOk", |
| "scalarWZeroRequiredOk": "present.w_zero_point and scalarWZeroOk", |
| "ungroupedChannelsOk": "attrs.group == 1 and dim(shapes.w, 1) == dim(shapes.x, 1)", |
| "output3dShapeOk": "dim(shapes.y, 2) == expectedOutputDepth and dim(shapes.y, 3) == expectedOutputHeight and dim(shapes.y, 4) == expectedOutputWidth", |
| "rank3TensorOk": "spatialAttributesOk and (ranks.x == 3 and ranks.w == 3 and ranks.y == 3)", |
| "nchwTensorOk": "rank4TensorOk or rank3TensorOk", |
| "output1dShapeOk": "dim(shapes.y, 2) == expectedOutputWidth", |
| "outputSpatialShapeOk": "output2dShapeOk if rank4TensorOk else output1dShapeOk", |
| "zeroPointsOmitted": "not present.x_zero_point and not present.w_zero_point", |
| "pointwiseTileM": 8, |
| "pointwiseTileN": 64, |
| "pointwiseKtWords": 32, |
| "pointwiseResourcesOk": "pointwiseTileN <= min(device.limits.maxComputeWorkgroupSizeX, device.limits.maxComputeInvocationsPerWorkgroup) and 4 * pointwiseTileM * (pointwiseKtWords + 1) <= device.limits.maxComputeWorkgroupStorageSize", |
| "pointwiseDispatchOk": "ceilDiv(outputHeight * outputWidth, pointwiseTileN) <= min(device.limits.maxComputeWorkgroupsPerDimension, 65535) and ceilDiv(dim(shapes.w, 0), pointwiseTileM) <= min(device.limits.maxComputeWorkgroupsPerDimension, 65535) and dim(shapes.y, 0) <= min(device.limits.maxComputeWorkgroupsPerDimension, 65535)", |
| "aVec4": "\"vec4<u32>\" if tensorDtypes.w == \"uint8\" else \"vec4<i32>\"" |
| }, |
| "when": ["outputBatchChannelsOk"], |
| "bindings": { |
| "x": { "elementType": "$bScalar" }, |
| "x_zero_point": { "elementType": "$bScalar", "length": 1 }, |
| "cols": { "scratch": "cols3d", "elementType": "$bScalar" }, |
| "params": { |
| "struct": [ |
| { "name": "outCount", "type": "u32", "value": "dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.y, 4)" }, |
| { "name": "inChannels", "type": "u32", "value": "dim(shapes.x, 1)" }, |
| { "name": "inD", "type": "u32", "value": "dim(shapes.x, 2)" }, |
| { "name": "inH", "type": "u32", "value": "dim(shapes.x, 3)" }, |
| { "name": "inW", "type": "u32", "value": "dim(shapes.x, 4)" }, |
| { |
| "name": "kRows", |
| "type": "u32", |
| "value": "dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3) * dim(shapes.w, 4)" |
| }, |
| { "name": "outH", "type": "u32", "value": "dim(shapes.y, 3)" }, |
| { "name": "outW", "type": "u32", "value": "dim(shapes.y, 4)" } |
| ] |
| }, |
| "cols_im2col": { "name": "cols", "buffer": "storage", "elementType": "$bScalar" }, |
| "params_im2col": { |
| "name": "params", |
| "struct": [ |
| { "name": "outCount", "type": "u32", "value": "dim(shapes.y, 2) * dim(shapes.y, 3)" }, |
| { "name": "inChannels", "type": "u32", "value": "dim(shapes.x, 1)" }, |
| { "name": "inH", "type": "u32", "value": "dim(shapes.x, 2)" }, |
| { "name": "inW", "type": "u32", "value": "dim(shapes.x, 3)" }, |
| { "name": "kRows", "type": "u32", "value": "dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }, |
| { "name": "outW", "type": "u32", "value": "dim(shapes.y, 3)" } |
| ] |
| }, |
| "a": { "arg": "w", "elementType": "$aVec4" }, |
| "b": { "scratch": "cols3d", "buffer": "read-only-storage", "elementType": "$bScalar" }, |
| "a_zero_point": { "arg": "w_zero_point", "elementType": "$aScalar", "length": 1 }, |
| "b_zero_point": { "arg": "x_zero_point", "elementType": "$bScalar", "length": 1 }, |
| "y": { "elementType": "i32" }, |
| "params_main": { |
| "name": "params", |
| "struct": [ |
| { "name": "M", "type": "u32", "value": "dim(shapes.w, 0)" }, |
| { "name": "N", "type": "u32", "value": "dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.y, 4)" }, |
| { |
| "name": "K", |
| "type": "u32", |
| "value": "dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3) * dim(shapes.w, 4)" |
| }, |
| { "name": "aBatchStride4", "type": "u32", "value": 0 }, |
| { |
| "name": "bBatchStride", |
| "type": "u32", |
| "value": "dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3) * dim(shapes.w, 4) * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.y, 4)" |
| }, |
| { |
| "name": "yBatchStride", |
| "type": "u32", |
| "value": "dim(shapes.w, 0) * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.y, 4)" |
| } |
| ] |
| }, |
| "b_main": { "scratch": "cols", "name": "b", "buffer": "read-only-storage", "elementType": "$bScalar" }, |
| "params__uniform": { |
| "name": "params", |
| "struct": [ |
| { "name": "M", "type": "u32", "value": "dim(shapes.w, 0)" }, |
| { "name": "N", "type": "u32", "value": "dim(shapes.y, 2) * dim(shapes.y, 3)" }, |
| { "name": "K", "type": "u32", "value": "dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }, |
| { "name": "aBatchStride4", "type": "u32", "value": 0 }, |
| { |
| "name": "bBatchStride", |
| "type": "u32", |
| "value": "(dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)) * (dim(shapes.y, 2) * dim(shapes.y, 3))" |
| }, |
| { "name": "yBatchStride", "type": "u32", "value": "dim(shapes.w, 0) * dim(shapes.y, 2) * dim(shapes.y, 3)" } |
| ] |
| }, |
| "a_w": { "arg": "w", "name": "a", "elementType": "$aScalar" }, |
| "b_x": { "arg": "x", "name": "b", "elementType": "$bScalar" }, |
| "params_quant_dp4a_matmul": { |
| "name": "params", |
| "struct": [ |
| { "name": "M", "type": "u32", "value": "dim(shapes.w, 0)" }, |
| { "name": "N", "type": "u32", "value": "outputHeight * outputWidth" }, |
| { "name": "K", "type": "u32", "value": "dim(shapes.x, 1)" }, |
| { "name": "aBatchStride4", "type": "u32", "value": 0 }, |
| { "name": "bBatchStride", "type": "u32", "value": "dim(shapes.x, 1) * inputHeight * inputWidth" }, |
| { "name": "yBatchStride", "type": "u32", "value": "dim(shapes.w, 0) * outputHeight * outputWidth" } |
| ] |
| }, |
| "x_main": { "name": "x", "elementType": "$xScalar" }, |
| "w": { "elementType": "$wScalar" }, |
| "x_zero_point_main": { "name": "x_zero_point", "elementType": "$xScalar", "length": 1 }, |
| "w_zero_point": { "elementType": "$wScalar" }, |
| "params_nchw2d_x_zero_point_only": { |
| "name": "params", |
| "struct": [ |
| { "name": "inChannels", "type": "u32", "value": "dim(shapes.x, 1)" }, |
| { "name": "inH", "type": "u32", "value": "inputHeight" }, |
| { "name": "inW", "type": "u32", "value": "inputWidth" }, |
| { "name": "outChannels", "type": "u32", "value": "dim(shapes.w, 0)" }, |
| { "name": "weightInChannels", "type": "u32", "value": "dim(shapes.w, 1)" }, |
| { "name": "inChannelsPerGroup", "type": "u32", "value": "dim(shapes.x, 1) / attrs.group" }, |
| { "name": "outChannelsPerGroup", "type": "u32", "value": "dim(shapes.w, 0) / attrs.group" }, |
| { "name": "kernelH", "type": "u32", "value": "kernelHeight" }, |
| { "name": "kernelW", "type": "u32", "value": "kernelWidth" }, |
| { "name": "outH", "type": "u32", "value": "outputHeight" }, |
| { "name": "outW", "type": "u32", "value": "outputWidth" }, |
| { "name": "strideH", "type": "u32", "value": "strideH" }, |
| { "name": "strideW", "type": "u32", "value": "strideW" }, |
| { "name": "dilationH", "type": "u32", "value": "dilationH" }, |
| { "name": "dilationW", "type": "u32", "value": "dilationW" }, |
| { "name": "padH", "type": "i32", "value": "effectivePadTop" }, |
| { "name": "padW", "type": "i32", "value": "effectivePadLeft" }, |
| { "name": "count", "type": "u32", "value": "numel(shapes.y)" } |
| ] |
| }, |
| "w_zero_point_main": { "name": "w_zero_point", "elementType": "$wScalar", "length": 1 }, |
| "params_ncdhw3d_x_zero_point_only": { |
| "name": "params", |
| "struct": [ |
| { "name": "inChannels", "type": "u32", "value": "dim(shapes.x, 1)" }, |
| { "name": "inD", "type": "u32", "value": "dim(shapes.x, 2)" }, |
| { "name": "inH", "type": "u32", "value": "dim(shapes.x, 3)" }, |
| { "name": "inW", "type": "u32", "value": "dim(shapes.x, 4)" }, |
| { "name": "outChannels", "type": "u32", "value": "dim(shapes.w, 0)" }, |
| { "name": "weightInChannels", "type": "u32", "value": "dim(shapes.w, 1)" }, |
| { "name": "inChannelsPerGroup", "type": "u32", "value": "dim(shapes.x, 1) / attrs.group" }, |
| { "name": "outChannelsPerGroup", "type": "u32", "value": "dim(shapes.w, 0) / attrs.group" }, |
| { "name": "kernelD", "type": "u32", "value": "dim(shapes.w, 2)" }, |
| { "name": "kernelH", "type": "u32", "value": "dim(shapes.w, 3)" }, |
| { "name": "kernelW", "type": "u32", "value": "dim(shapes.w, 4)" }, |
| { "name": "outD", "type": "u32", "value": "dim(shapes.y, 2)" }, |
| { "name": "outH", "type": "u32", "value": "dim(shapes.y, 3)" }, |
| { "name": "outW", "type": "u32", "value": "dim(shapes.y, 4)" }, |
| { "name": "strideD", "type": "u32", "value": "strideD" }, |
| { "name": "strideH", "type": "u32", "value": "strideH" }, |
| { "name": "strideW", "type": "u32", "value": "strideW" }, |
| { "name": "dilationD", "type": "u32", "value": "dilationD" }, |
| { "name": "dilationH", "type": "u32", "value": "dilationH" }, |
| { "name": "dilationW", "type": "u32", "value": "dilationW" }, |
| { "name": "padD", "type": "i32", "value": "effectivePadFront" }, |
| { "name": "padH", "type": "i32", "value": "effectivePadTop" }, |
| { "name": "padW", "type": "i32", "value": "effectivePadLeft" }, |
| { "name": "count", "type": "u32", "value": "numel(shapes.y)" } |
| ] |
| }, |
| "a_zero_point_per_row": { "arg": "w_zero_point", "name": "a_zero_point", "elementType": "$aScalar" } |
| }, |
| "variants": [ |
| { |
| "id": "dp4a_im2col_ncdhw3d_scalar_zero_points", |
| "priority": 20, |
| "when": ["packedFeature", "rank5TensorOk", "scalarXZeroRequiredOk", "scalarWZeroRequiredOk", "ungroupedChannelsOk", "dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3) * dim(shapes.w, 4) % 4 == 0", "dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3) * dim(shapes.w, 4) >= 32", "dim(shapes.w, 0) >= 8", "dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.y, 4) >= 256", "output3dShapeOk", "dim(shapes.x, 0) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3) * dim(shapes.w, 4) * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.y, 4) * 4 <= device.limits.maxStorageBufferBindingSize", "dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3) * dim(shapes.w, 4) <= min(device.limits.maxComputeWorkgroupsPerDimension, 65535)", "ceilDiv(dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.y, 4), tunables.WORKGROUP_SIZE) <= min(device.limits.maxComputeWorkgroupsPerDimension, 65535)", "ceilDiv(dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.y, 4), 64) <= min(device.limits.maxComputeWorkgroupsPerDimension, 65535)", "ceilDiv(dim(shapes.w, 0), 8) <= min(device.limits.maxComputeWorkgroupsPerDimension, 65535)", "dim(shapes.x, 0) <= min(device.limits.maxComputeWorkgroupsPerDimension, 65535)"], |
| "derive": { |
| "hasAZero": true, |
| "hasBZero": true, |
| "aPacked": true, |
| "aZeroPerRow": false, |
| "tileM": 8, |
| "tileN": 64, |
| "ktWords": 32, |
| "hasBias": false |
| }, |
| "intermediates": [ |
| { |
| "id": "cols3d", |
| "dtype": "int32", |
| "shape": "[dim(shapes.x, 0) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3) * dim(shapes.w, 4) * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.y, 4)]" |
| } |
| ], |
| "passes": [ |
| { |
| "id": "im2col", |
| "name": "ConvInteger.Im2Col3d", |
| "shader": "conv-int-im2col-spatial.wgsl.jinja", |
| "derive": { |
| "spatialRankSpec": 3, |
| "kernelDSpec": "dim(shapes.w, 2)", |
| "kernelHSpec": "dim(shapes.w, 3)", |
| "kernelWSpec": "dim(shapes.w, 4)", |
| "strideDSpec": "strideD", |
| "strideHSpec": "strideH", |
| "strideWSpec": "strideW", |
| "dilationDSpec": "dilationD", |
| "dilationHSpec": "dilationH", |
| "dilationWSpec": "dilationW", |
| "padFrontSpec": "effectivePadFront", |
| "padTopSpec": "effectivePadTop", |
| "padLeftSpec": "effectivePadLeft" |
| }, |
| "bindings": ["x", "x_zero_point", "cols", "params"], |
| "dispatch": { |
| "x": "ceil(dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.y, 4) / tunables.WORKGROUP_SIZE)", |
| "y": "dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3) * dim(shapes.w, 4)", |
| "z": "dim(shapes.x, 0)" |
| } |
| }, |
| { |
| "id": "main", |
| "name": "ConvInteger.Im2Col3dDp4a", |
| "shader": "quant-dp4a-matmul.wgsl.jinja", |
| "bindings": ["a", "b", "a_zero_point", "b_zero_point", "y", "params_main"], |
| "dispatch": { |
| "x": "ceil(dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.y, 4) / 64)", |
| "y": "ceil(dim(shapes.w, 0) / 8)", |
| "z": "dim(shapes.y, 0)" |
| } |
| } |
| ] |
| }, |
| { |
| "id": "dp4a_im2col_nchw_scalar_zero_points", |
| "priority": 20, |
| "when": ["packedFeature", "rank4TensorOk", "scalarXZeroRequiredOk", "scalarWZeroRequiredOk", "ungroupedChannelsOk", "(dim(shapes.w, 2) > 1 or dim(shapes.w, 3) > 1)", "dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3) % 4 == 0", "dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3) >= 64", "numel(shapes.y) >= 8192", "dim(shapes.y, 2) * dim(shapes.y, 3) <= 4194240", "dim(shapes.w, 0) <= 524280", "dim(shapes.y, 0) <= min(device.limits.maxComputeWorkgroupsPerDimension, 65535)"], |
| "derive": { "hasAZero": true, "hasBZero": true, "tileM": 8, "tileN": 64, "ktWords": 32 }, |
| "intermediates": [ |
| { |
| "id": "cols", |
| "dtype": "int32", |
| "shape": "[dim(shapes.x, 0) * (dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)) * (dim(shapes.y, 2) * dim(shapes.y, 3))]" |
| } |
| ], |
| "passes": [ |
| { |
| "id": "im2col", |
| "name": "ConvInteger.Im2Col", |
| "shader": "conv-int-im2col-spatial.wgsl.jinja", |
| "derive": { |
| "spatialRankSpec": 2, |
| "kernelHSpec": "dim(shapes.w, 2)", |
| "kernelWSpec": "dim(shapes.w, 3)", |
| "strideHSpec": "strideH", |
| "strideWSpec": "strideW", |
| "dilationHSpec": "dilationH", |
| "dilationWSpec": "dilationW", |
| "padTopSpec": "effectivePadTop", |
| "padLeftSpec": "effectivePadLeft" |
| }, |
| "bindings": ["x", "x_zero_point", "cols_im2col", "params_im2col"], |
| "dispatch": { |
| "x": "ceil(dim(shapes.y, 2) * dim(shapes.y, 3) / tunables.WORKGROUP_SIZE)", |
| "y": "dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)", |
| "z": "dim(shapes.x, 0)" |
| } |
| }, |
| { |
| "id": "main", |
| "name": "ConvInteger.Im2ColDp4a", |
| "shader": "quant-dp4a-matmul.wgsl.jinja", |
| "bindings": ["a", "b_main", "a_zero_point", "b_zero_point", "y", "params__uniform"], |
| "dispatch": { |
| "x": "ceil(dim(shapes.y, 2) * dim(shapes.y, 3) / 64)", |
| "y": "ceil(dim(shapes.w, 0) / 8)", |
| "z": "dim(shapes.y, 0)" |
| } |
| } |
| ] |
| }, |
| { |
| "id": "dp4a_pointwise_1x1_tail", |
| "priority": 21, |
| "when": ["nchwTensorOk", "ungroupedChannelsOk", "kernelHeight == 1", "kernelWidth == 1", "strideH == 1", "strideW == 1", "padTop == 0 and padBottom == 0", "padLeft == 0 and padRight == 0", "outputHeight == inputHeight", "outputWidth == inputWidth", "scalarXZeroRequiredOk", "scalarWZeroRequiredOk", "dim(shapes.x, 1) >= 32", "dim(shapes.x, 1) % 4 != 0", "pointwiseResourcesOk", "pointwiseDispatchOk"], |
| "derive": { |
| "hasAZero": true, |
| "hasBZero": true, |
| "tileM": "pointwiseTileM", |
| "tileN": "pointwiseTileN", |
| "ktWords": "pointwiseKtWords", |
| "aPacked": false |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger.Dp4aPointwiseTail", |
| "shader": "quant-dp4a-matmul.wgsl.jinja", |
| "bindings": ["a_w", "b_x", "a_zero_point", "b_zero_point", "y", "params_quant_dp4a_matmul"], |
| "dispatch": { |
| "x": "ceilDiv(outputHeight * outputWidth, pointwiseTileN)", |
| "y": "ceilDiv(dim(shapes.w, 0), pointwiseTileM)", |
| "z": "dim(shapes.y, 0)" |
| } |
| } |
| ] |
| }, |
| { |
| "id": "dp4a_pointwise_1x1_tail_x_zero_point_only", |
| "priority": 21, |
| "when": ["nchwTensorOk", "ungroupedChannelsOk", "kernelHeight == 1", "kernelWidth == 1", "strideH == 1", "strideW == 1", "padTop == 0 and padBottom == 0", "padLeft == 0 and padRight == 0", "outputHeight == inputHeight", "outputWidth == inputWidth", "scalarXZeroRequiredOk", "not present.w_zero_point", "dim(shapes.x, 1) >= 32", "dim(shapes.x, 1) % 4 != 0", "pointwiseResourcesOk", "pointwiseDispatchOk"], |
| "derive": { |
| "hasAZero": false, |
| "hasBZero": true, |
| "tileM": "pointwiseTileM", |
| "tileN": "pointwiseTileN", |
| "ktWords": "pointwiseKtWords", |
| "aPacked": false |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger.Dp4aPointwiseTail", |
| "shader": "quant-dp4a-matmul.wgsl.jinja", |
| "bindings": ["a_w", "b_x", "b_zero_point", "y", "params_quant_dp4a_matmul"], |
| "dispatch": { |
| "x": "ceilDiv(outputHeight * outputWidth, pointwiseTileN)", |
| "y": "ceilDiv(dim(shapes.w, 0), pointwiseTileM)", |
| "z": "dim(shapes.y, 0)" |
| } |
| } |
| ] |
| }, |
| { |
| "id": "dp4a_pointwise_1x1_tail_w_zero_point_only", |
| "priority": 21, |
| "when": ["nchwTensorOk", "ungroupedChannelsOk", "kernelHeight == 1", "kernelWidth == 1", "strideH == 1", "strideW == 1", "padTop == 0 and padBottom == 0", "padLeft == 0 and padRight == 0", "outputHeight == inputHeight", "outputWidth == inputWidth", "not present.x_zero_point", "scalarWZeroRequiredOk", "dim(shapes.x, 1) >= 32", "dim(shapes.x, 1) % 4 != 0", "pointwiseResourcesOk", "pointwiseDispatchOk"], |
| "derive": { |
| "hasAZero": true, |
| "hasBZero": false, |
| "tileM": "pointwiseTileM", |
| "tileN": "pointwiseTileN", |
| "ktWords": "pointwiseKtWords", |
| "aPacked": false |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger.Dp4aPointwiseTail", |
| "shader": "quant-dp4a-matmul.wgsl.jinja", |
| "bindings": ["a_w", "b_x", "a_zero_point", "y", "params_quant_dp4a_matmul"], |
| "dispatch": { |
| "x": "ceilDiv(outputHeight * outputWidth, pointwiseTileN)", |
| "y": "ceilDiv(dim(shapes.w, 0), pointwiseTileM)", |
| "z": "dim(shapes.y, 0)" |
| } |
| } |
| ] |
| }, |
| { |
| "id": "dp4a_pointwise_1x1_tail_no_zero_points", |
| "priority": 21, |
| "when": ["nchwTensorOk", "ungroupedChannelsOk", "kernelHeight == 1", "kernelWidth == 1", "strideH == 1", "strideW == 1", "padTop == 0 and padBottom == 0", "padLeft == 0 and padRight == 0", "outputHeight == inputHeight", "outputWidth == inputWidth", "not present.x_zero_point", "not present.w_zero_point", "dim(shapes.x, 1) >= 32", "dim(shapes.x, 1) % 4 != 0", "pointwiseResourcesOk", "pointwiseDispatchOk"], |
| "derive": { |
| "hasAZero": false, |
| "hasBZero": false, |
| "tileM": "pointwiseTileM", |
| "tileN": "pointwiseTileN", |
| "ktWords": "pointwiseKtWords", |
| "aPacked": false |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger.Dp4aPointwiseTail", |
| "shader": "quant-dp4a-matmul.wgsl.jinja", |
| "bindings": ["a_w", "b_x", "y", "params_quant_dp4a_matmul"], |
| "dispatch": { |
| "x": "ceilDiv(outputHeight * outputWidth, pointwiseTileN)", |
| "y": "ceilDiv(dim(shapes.w, 0), pointwiseTileM)", |
| "z": "dim(shapes.y, 0)" |
| } |
| } |
| ] |
| }, |
| { |
| "id": "dp4a_pointwise_1x1", |
| "priority": 20, |
| "when": ["nchwTensorOk", "ungroupedChannelsOk", "kernelHeight == 1", "kernelWidth == 1", "strideH == 1", "strideW == 1", "padTop == 0 and padBottom == 0", "padLeft == 0 and padRight == 0", "outputHeight == inputHeight", "outputWidth == inputWidth", "scalarXZeroRequiredOk", "scalarWZeroRequiredOk", "dim(shapes.x, 1) % 4 == 0", "pointwiseResourcesOk", "pointwiseDispatchOk"], |
| "derive": { |
| "aVec4": "\"vec4<u32>\" if tensorDtypes.w == \"uint8\" else \"vec4<i32>\"", |
| "hasAZero": true, |
| "hasBZero": true, |
| "tileM": "pointwiseTileM", |
| "tileN": "pointwiseTileN", |
| "ktWords": "pointwiseKtWords" |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger.Dp4aPointwise", |
| "shader": "quant-dp4a-matmul.wgsl.jinja", |
| "bindings": ["a", "b_x", "a_zero_point", "b_zero_point", "y", "params_quant_dp4a_matmul"], |
| "dispatch": { |
| "x": "ceilDiv(outputHeight * outputWidth, pointwiseTileN)", |
| "y": "ceilDiv(dim(shapes.w, 0), pointwiseTileM)", |
| "z": "dim(shapes.y, 0)" |
| } |
| } |
| ] |
| }, |
| { |
| "id": "dp4a_pointwise_1x1_x_zero_point_only", |
| "priority": 20, |
| "when": ["nchwTensorOk", "ungroupedChannelsOk", "kernelHeight == 1", "kernelWidth == 1", "strideH == 1", "strideW == 1", "padTop == 0 and padBottom == 0", "padLeft == 0 and padRight == 0", "outputHeight == inputHeight", "outputWidth == inputWidth", "scalarXZeroRequiredOk", "not present.w_zero_point", "dim(shapes.x, 1) % 4 == 0", "pointwiseResourcesOk", "pointwiseDispatchOk"], |
| "derive": { |
| "aVec4": "\"vec4<u32>\" if tensorDtypes.w == \"uint8\" else \"vec4<i32>\"", |
| "hasAZero": false, |
| "hasBZero": true, |
| "tileM": "pointwiseTileM", |
| "tileN": "pointwiseTileN", |
| "ktWords": "pointwiseKtWords" |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger.Dp4aPointwise", |
| "shader": "quant-dp4a-matmul.wgsl.jinja", |
| "bindings": ["a", "b_x", "b_zero_point", "y", "params_quant_dp4a_matmul"], |
| "dispatch": { |
| "x": "ceilDiv(outputHeight * outputWidth, pointwiseTileN)", |
| "y": "ceilDiv(dim(shapes.w, 0), pointwiseTileM)", |
| "z": "dim(shapes.y, 0)" |
| } |
| } |
| ] |
| }, |
| { |
| "id": "dp4a_pointwise_1x1_w_zero_point_only", |
| "priority": 20, |
| "when": ["nchwTensorOk", "ungroupedChannelsOk", "kernelHeight == 1", "kernelWidth == 1", "strideH == 1", "strideW == 1", "padTop == 0 and padBottom == 0", "padLeft == 0 and padRight == 0", "outputHeight == inputHeight", "outputWidth == inputWidth", "not present.x_zero_point", "scalarWZeroRequiredOk", "dim(shapes.x, 1) % 4 == 0", "pointwiseResourcesOk", "pointwiseDispatchOk"], |
| "derive": { |
| "aVec4": "\"vec4<u32>\" if tensorDtypes.w == \"uint8\" else \"vec4<i32>\"", |
| "hasAZero": true, |
| "hasBZero": false, |
| "tileM": "pointwiseTileM", |
| "tileN": "pointwiseTileN", |
| "ktWords": "pointwiseKtWords" |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger.Dp4aPointwise", |
| "shader": "quant-dp4a-matmul.wgsl.jinja", |
| "bindings": ["a", "b_x", "a_zero_point", "y", "params_quant_dp4a_matmul"], |
| "dispatch": { |
| "x": "ceilDiv(outputHeight * outputWidth, pointwiseTileN)", |
| "y": "ceilDiv(dim(shapes.w, 0), pointwiseTileM)", |
| "z": "dim(shapes.y, 0)" |
| } |
| } |
| ] |
| }, |
| { |
| "id": "dp4a_pointwise_1x1_no_zero_points", |
| "priority": 20, |
| "when": ["nchwTensorOk", "ungroupedChannelsOk", "kernelHeight == 1", "kernelWidth == 1", "strideH == 1", "strideW == 1", "padTop == 0 and padBottom == 0", "padLeft == 0 and padRight == 0", "outputHeight == inputHeight", "outputWidth == inputWidth", "not present.x_zero_point", "not present.w_zero_point", "dim(shapes.x, 1) % 4 == 0", "pointwiseResourcesOk", "pointwiseDispatchOk"], |
| "derive": { |
| "aVec4": "\"vec4<u32>\" if tensorDtypes.w == \"uint8\" else \"vec4<i32>\"", |
| "hasAZero": false, |
| "hasBZero": false, |
| "tileM": "pointwiseTileM", |
| "tileN": "pointwiseTileN", |
| "ktWords": "pointwiseKtWords" |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger.Dp4aPointwise", |
| "shader": "quant-dp4a-matmul.wgsl.jinja", |
| "bindings": ["a", "b_x", "y", "params_quant_dp4a_matmul"], |
| "dispatch": { |
| "x": "ceilDiv(outputHeight * outputWidth, pointwiseTileN)", |
| "y": "ceilDiv(dim(shapes.w, 0), pointwiseTileM)", |
| "z": "dim(shapes.y, 0)" |
| } |
| } |
| ] |
| }, |
| { |
| "id": "dp4a_pointwise_1x1_tail_per_channel", |
| "priority": 21, |
| "when": ["nchwTensorOk", "ungroupedChannelsOk", "kernelHeight == 1", "kernelWidth == 1", "strideH == 1", "strideW == 1", "padTop == 0 and padBottom == 0", "padLeft == 0 and padRight == 0", "outputHeight == inputHeight", "outputWidth == inputWidth", "scalarXZeroRequiredOk", "present.w_zero_point", "ranks.w_zero_point == 1", "dim(shapes.w_zero_point, 0) == dim(shapes.w, 0)", "dim(shapes.x, 1) >= 32", "dim(shapes.x, 1) % 4 != 0", "pointwiseResourcesOk", "pointwiseDispatchOk"], |
| "derive": { |
| "hasAZero": true, |
| "hasBZero": true, |
| "tileM": "pointwiseTileM", |
| "tileN": "pointwiseTileN", |
| "ktWords": "pointwiseKtWords", |
| "aPacked": false, |
| "aZeroPerRow": true |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger.Dp4aPointwiseTail", |
| "shader": "quant-dp4a-matmul.wgsl.jinja", |
| "bindings": ["a_w", "b_x", "a_zero_point_per_row", "b_zero_point", "y", "params_quant_dp4a_matmul"], |
| "dispatch": { |
| "x": "ceilDiv(outputHeight * outputWidth, pointwiseTileN)", |
| "y": "ceilDiv(dim(shapes.w, 0), pointwiseTileM)", |
| "z": "dim(shapes.y, 0)" |
| } |
| } |
| ] |
| }, |
| { |
| "id": "dp4a_pointwise_1x1_tail_per_channel_w_zero_point_only", |
| "priority": 21, |
| "when": ["nchwTensorOk", "ungroupedChannelsOk", "kernelHeight == 1", "kernelWidth == 1", "strideH == 1", "strideW == 1", "padTop == 0 and padBottom == 0", "padLeft == 0 and padRight == 0", "outputHeight == inputHeight", "outputWidth == inputWidth", "not present.x_zero_point", "present.w_zero_point", "ranks.w_zero_point == 1", "dim(shapes.w_zero_point, 0) == dim(shapes.w, 0)", "dim(shapes.x, 1) >= 32", "dim(shapes.x, 1) % 4 != 0", "pointwiseResourcesOk", "pointwiseDispatchOk"], |
| "derive": { |
| "hasAZero": true, |
| "hasBZero": false, |
| "tileM": "pointwiseTileM", |
| "tileN": "pointwiseTileN", |
| "ktWords": "pointwiseKtWords", |
| "aPacked": false, |
| "aZeroPerRow": true |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger.Dp4aPointwiseTail", |
| "shader": "quant-dp4a-matmul.wgsl.jinja", |
| "bindings": ["a_w", "b_x", "a_zero_point_per_row", "y", "params_quant_dp4a_matmul"], |
| "dispatch": { |
| "x": "ceilDiv(outputHeight * outputWidth, pointwiseTileN)", |
| "y": "ceilDiv(dim(shapes.w, 0), pointwiseTileM)", |
| "z": "dim(shapes.y, 0)" |
| } |
| } |
| ] |
| }, |
| { |
| "id": "dp4a_pointwise_1x1_per_channel", |
| "priority": 20, |
| "when": ["nchwTensorOk", "ungroupedChannelsOk", "kernelHeight == 1", "kernelWidth == 1", "strideH == 1", "strideW == 1", "padTop == 0 and padBottom == 0", "padLeft == 0 and padRight == 0", "outputHeight == inputHeight", "outputWidth == inputWidth", "scalarXZeroRequiredOk", "present.w_zero_point", "ranks.w_zero_point == 1", "dim(shapes.w_zero_point, 0) == dim(shapes.w, 0)", "dim(shapes.x, 1) % 4 == 0", "pointwiseResourcesOk", "pointwiseDispatchOk"], |
| "derive": { |
| "aVec4": "\"vec4<u32>\" if tensorDtypes.w == \"uint8\" else \"vec4<i32>\"", |
| "hasAZero": true, |
| "hasBZero": true, |
| "tileM": "pointwiseTileM", |
| "tileN": "pointwiseTileN", |
| "ktWords": "pointwiseKtWords", |
| "aZeroPerRow": true |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger.Dp4aPointwise", |
| "shader": "quant-dp4a-matmul.wgsl.jinja", |
| "bindings": ["a", "b_x", "a_zero_point_per_row", "b_zero_point", "y", "params_quant_dp4a_matmul"], |
| "dispatch": { |
| "x": "ceilDiv(outputHeight * outputWidth, pointwiseTileN)", |
| "y": "ceilDiv(dim(shapes.w, 0), pointwiseTileM)", |
| "z": "dim(shapes.y, 0)" |
| } |
| } |
| ] |
| }, |
| { |
| "id": "dp4a_pointwise_1x1_per_channel_w_zero_point_only", |
| "priority": 20, |
| "when": ["nchwTensorOk", "ungroupedChannelsOk", "kernelHeight == 1", "kernelWidth == 1", "strideH == 1", "strideW == 1", "padTop == 0 and padBottom == 0", "padLeft == 0 and padRight == 0", "outputHeight == inputHeight", "outputWidth == inputWidth", "not present.x_zero_point", "present.w_zero_point", "ranks.w_zero_point == 1", "dim(shapes.w_zero_point, 0) == dim(shapes.w, 0)", "dim(shapes.x, 1) % 4 == 0", "pointwiseResourcesOk", "pointwiseDispatchOk"], |
| "derive": { |
| "aVec4": "\"vec4<u32>\" if tensorDtypes.w == \"uint8\" else \"vec4<i32>\"", |
| "hasAZero": true, |
| "hasBZero": false, |
| "tileM": "pointwiseTileM", |
| "tileN": "pointwiseTileN", |
| "ktWords": "pointwiseKtWords", |
| "aZeroPerRow": true |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger.Dp4aPointwise", |
| "shader": "quant-dp4a-matmul.wgsl.jinja", |
| "bindings": ["a", "b_x", "a_zero_point_per_row", "y", "params_quant_dp4a_matmul"], |
| "dispatch": { |
| "x": "ceilDiv(outputHeight * outputWidth, pointwiseTileN)", |
| "y": "ceilDiv(dim(shapes.w, 0), pointwiseTileM)", |
| "z": "dim(shapes.y, 0)" |
| } |
| } |
| ] |
| }, |
| { |
| "id": "nchw2d_per_channel_zero_point", |
| "priority": 5, |
| "when": ["nchwTensorOk", "scalarXZeroRequiredOk", "ranks.w_zero_point == 1", "dim(shapes.w_zero_point, 0) == dim(shapes.w, 0)", "groupChannelsOk", "outputSpatialShapeOk"], |
| "derive": { "wZeroPerChannel": true }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvIntegerPerChannel", |
| "shader": "conv-int-accumulate-spatial.wgsl.jinja", |
| "derive": { "spatialRankSpec": 2, "xZeroOmitted": "false", "wZeroOmitted": "false" }, |
| "bindings": ["x_main", "w", "x_zero_point_main", "w_zero_point", "y", "params_nchw2d_x_zero_point_only"], |
| "dispatch": { |
| "x": "min(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "y": "ceilDiv(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "nchw2d_per_channel_w_zero_point_only", |
| "priority": 5, |
| "when": ["nchwTensorOk", "not present.x_zero_point", "present.w_zero_point", "ranks.w_zero_point == 1", "dim(shapes.w_zero_point, 0) == dim(shapes.w, 0)", "groupChannelsOk", "outputSpatialShapeOk"], |
| "derive": { "wZeroPerChannel": true }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvIntegerPerChannel", |
| "shader": "conv-int-accumulate-spatial.wgsl.jinja", |
| "derive": { "spatialRankSpec": 2, "xZeroOmitted": "true", "wZeroOmitted": "false" }, |
| "bindings": ["x_main", "w", "w_zero_point", "y", "params_nchw2d_x_zero_point_only"], |
| "dispatch": { |
| "x": "min(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "y": "ceilDiv(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "nchw2d_scalar_zero_points", |
| "when": ["nchwTensorOk", "scalarXZeroRequiredOk", "scalarWZeroRequiredOk", "groupChannelsOk", "outputSpatialShapeOk"], |
| "derive": {}, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger", |
| "shader": "conv-int-accumulate-spatial.wgsl.jinja", |
| "derive": { "spatialRankSpec": 2, "xZeroOmitted": "false", "wZeroOmitted": "false" }, |
| "bindings": ["x_main", "w", "x_zero_point_main", "w_zero_point_main", "y", "params_nchw2d_x_zero_point_only"], |
| "dispatch": { |
| "x": "min(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "y": "ceilDiv(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "nchw2d_x_zero_point_only", |
| "when": ["nchwTensorOk", "present.x_zero_point", "not present.w_zero_point", "scalarXZeroRequiredOk", "groupChannelsOk", "outputSpatialShapeOk"], |
| "derive": {}, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger", |
| "shader": "conv-int-accumulate-spatial.wgsl.jinja", |
| "derive": { "spatialRankSpec": 2, "xZeroOmitted": "false", "wZeroOmitted": "true" }, |
| "bindings": ["x_main", "w", "x_zero_point_main", "y", "params_nchw2d_x_zero_point_only"], |
| "dispatch": { |
| "x": "min(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "y": "ceilDiv(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "nchw2d_w_zero_point_only", |
| "when": ["nchwTensorOk", "not present.x_zero_point", "present.w_zero_point", "scalarWZeroRequiredOk", "groupChannelsOk", "outputSpatialShapeOk"], |
| "derive": {}, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger", |
| "shader": "conv-int-accumulate-spatial.wgsl.jinja", |
| "derive": { "spatialRankSpec": 2, "xZeroOmitted": "true", "wZeroOmitted": "false" }, |
| "bindings": ["x_main", "w", "w_zero_point_main", "y", "params_nchw2d_x_zero_point_only"], |
| "dispatch": { |
| "x": "min(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "y": "ceilDiv(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "ncdhw3d_scalar_zero_points", |
| "when": ["rank5TensorOk", "scalarXZeroRequiredOk", "scalarWZeroRequiredOk", "groupChannelsOk", "output3dShapeOk"], |
| "derive": {}, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger3d", |
| "shader": "conv-int-accumulate-spatial.wgsl.jinja", |
| "derive": { "spatialRankSpec": 3, "xZeroOmitted": "false", "wZeroOmitted": "false" }, |
| "bindings": ["x_main", "w", "x_zero_point_main", "w_zero_point_main", "y", "params_ncdhw3d_x_zero_point_only"], |
| "dispatch": { |
| "x": "min(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "y": "ceilDiv(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "ncdhw3d_x_zero_point_only", |
| "when": ["rank5TensorOk", "scalarXZeroRequiredOk", "not present.w_zero_point", "groupChannelsOk", "output3dShapeOk"], |
| "derive": {}, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger3d", |
| "shader": "conv-int-accumulate-spatial.wgsl.jinja", |
| "derive": { "spatialRankSpec": 3, "xZeroOmitted": "false", "wZeroOmitted": "true" }, |
| "bindings": ["x_main", "w", "x_zero_point_main", "y", "params_ncdhw3d_x_zero_point_only"], |
| "dispatch": { |
| "x": "min(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "y": "ceilDiv(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "ncdhw3d_w_zero_point_only", |
| "when": ["rank5TensorOk", "not present.x_zero_point", "scalarWZeroRequiredOk", "groupChannelsOk", "output3dShapeOk"], |
| "derive": {}, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger3d", |
| "shader": "conv-int-accumulate-spatial.wgsl.jinja", |
| "derive": { "spatialRankSpec": 3, "xZeroOmitted": "true", "wZeroOmitted": "false" }, |
| "bindings": ["x_main", "w", "w_zero_point_main", "y", "params_ncdhw3d_x_zero_point_only"], |
| "dispatch": { |
| "x": "min(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "y": "ceilDiv(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "ncdhw3d_implicit_zero_points", |
| "when": ["rank5TensorOk", "zeroPointsOmitted", "groupChannelsOk", "output3dShapeOk"], |
| "derive": {}, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger3d", |
| "shader": "conv-int-accumulate-spatial.wgsl.jinja", |
| "derive": { "spatialRankSpec": 3, "xZeroOmitted": "true", "wZeroOmitted": "true" }, |
| "bindings": ["x_main", "w", "y", "params_ncdhw3d_x_zero_point_only"], |
| "dispatch": { |
| "x": "min(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "y": "ceilDiv(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "nchw_implicit_zero_points", |
| "when": ["nchwTensorOk", "zeroPointsOmitted", "groupChannelsOk", "outputSpatialShapeOk"], |
| "derive": {}, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "ConvInteger", |
| "shader": "conv-int-accumulate-spatial.wgsl.jinja", |
| "derive": { "spatialRankSpec": 2, "xZeroOmitted": "true", "wZeroOmitted": "true" }, |
| "bindings": ["x_main", "w", "y", "params_nchw2d_x_zero_point_only"], |
| "dispatch": { |
| "x": "min(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "y": "ceilDiv(ceilDiv((numel(shapes.y)), (tunables.WORKGROUP_SIZE)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| } |
| ] |
| } |
|
|
