|
| 1 | +//===- NormalizeQuantTypes.cpp - Normalize quantized types |
| 2 | +//----------------------===// |
| 3 | +// |
| 4 | +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 5 | +// See https://llvm.org/LICENSE.txt for license information. |
| 6 | +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 7 | +// |
| 8 | +//===----------------------------------------------------------------------===// |
| 9 | +// |
| 10 | +// Normalize generic quantized types to specific quantized types |
| 11 | +// |
| 12 | +//===----------------------------------------------------------------------===// |
| 13 | + |
| 14 | +#include "mlir/Dialect/Func/IR/FuncOps.h" |
| 15 | +#include "mlir/Dialect/Func/Transforms/FuncConversions.h" |
| 16 | +#include "mlir/Dialect/Quant/IR/Quant.h" |
| 17 | +#include "mlir/Dialect/Quant/IR/QuantTypes.h" |
| 18 | +#include "mlir/Dialect/Quant/Transforms/Passes.h" |
| 19 | +#include "mlir/Dialect/Tensor/IR/Tensor.h" |
| 20 | +#include "mlir/Transforms/DialectConversion.h" |
| 21 | + |
| 22 | +namespace mlir { |
| 23 | +namespace quant { |
| 24 | + |
| 25 | +#define GEN_PASS_DEF_NORMALIZEQUANTTYPES |
| 26 | +#include "mlir/Dialect/Quant/Transforms/Passes.h.inc" |
| 27 | + |
| 28 | +namespace { |
| 29 | + |
| 30 | +/// Returns true if the given sub-channel quantized type is convertible to a |
| 31 | +/// per-tensor quantized type. This is true if the sub-channel type has only |
| 32 | +/// one scale and one zero point. |
| 33 | +/// |
| 34 | +/// Assumes that `tensorType` is a tensor with element type |
| 35 | +/// `quant::UniformQuantizedSubChannelType`. |
| 36 | +static bool isConvertibleToPerTensor(TensorType tensorType) { |
| 37 | + return cast<UniformQuantizedSubChannelType>(tensorType.getElementType()) |
| 38 | + .getScales() |
| 39 | + .getType() |
| 40 | + .getNumElements() == 1; |
| 41 | +} |
| 42 | + |
| 43 | +/// Returns true if the given sub-channel quantized type is convertible to a |
| 44 | +/// per-axis quantized type. This is true if the shape of the scales tensor has |
| 45 | +/// all but one non-one value. |
| 46 | +/// |
| 47 | +/// Assumes that `tensorType` is a tensor with element type |
| 48 | +/// `quant::UniformQuantizedSubChannelType`. |
| 49 | +static bool isConvertibleToPerAxis(TensorType tensorType) { |
| 50 | + auto shape = cast<UniformQuantizedSubChannelType>(tensorType.getElementType()) |
| 51 | + .getScales() |
| 52 | + .getType() |
| 53 | + .getShape(); |
| 54 | + return llvm::count_if(shape, [](int64_t dim) { return dim != 1; }) == 1; |
| 55 | +} |
| 56 | + |
| 57 | +/// This class defines a type converter that converts sub-channel quantized |
| 58 | +/// types to per-tensor or per-axis quantized types whenever possible. |
| 59 | +class NormalizedQuantTypesConverter : public TypeConverter { |
| 60 | + |
| 61 | + static Type convertType(Type type) { |
| 62 | + auto tensorType = dyn_cast<TensorType>(type); |
| 63 | + if (!tensorType) { |
| 64 | + return type; |
| 65 | + } |
| 66 | + |
| 67 | + auto subChannelType = |
| 68 | + dyn_cast<UniformQuantizedSubChannelType>(tensorType.getElementType()); |
| 69 | + if (!subChannelType) { |
| 70 | + return type; |
| 71 | + } |
| 72 | + |
| 73 | + if (isConvertibleToPerTensor(tensorType)) { |
| 74 | + double scale = |
| 75 | + subChannelType. getScales(). getValues<APFloat>()[ 0]. convertToDouble(); |
| 76 | + int64_t zeroPoint = |
| 77 | + subChannelType. getZeroPoints(). getValues<APInt>()[ 0]. getSExtValue(); |
| 78 | + auto perTensorType = UniformQuantizedType::get( |
| 79 | + subChannelType.getFlags(), subChannelType.getStorageType(), |
| 80 | + subChannelType.getExpressedType(), scale, zeroPoint, |
| 81 | + subChannelType.getStorageTypeMin(), |
| 82 | + subChannelType.getStorageTypeMax()); |
| 83 | + return tensorType.clone(perTensorType); |
| 84 | + } |
| 85 | + |
| 86 | + if (isConvertibleToPerAxis(tensorType)) { |
| 87 | + auto shape = subChannelType.getScales().getType().getShape(); |
| 88 | + auto quantizedDimItr = |
| 89 | + llvm::find_if(shape, [](int64_t dim) { return dim != 1; }); |
| 90 | + auto scales = llvm::to_vector(llvm::map_range( |
| 91 | + subChannelType. getScales(). getValues<APFloat>(), |
| 92 | + [](APFloat scale) { return scale.convertToDouble(); })); |
| 93 | + auto zeroPoints = llvm::to_vector(llvm::map_range( |
| 94 | + subChannelType. getZeroPoints(). getValues<APInt>(), |
| 95 | + [](APInt zeroPoint) { return zeroPoint.getSExtValue(); })); |
| 96 | + auto perAxisType = UniformQuantizedPerAxisType::get( |
| 97 | + subChannelType.getFlags(), subChannelType.getStorageType(), |
| 98 | + subChannelType.getExpressedType(), scales, zeroPoints, |
| 99 | + quantizedDimItr - shape.begin(), subChannelType.getStorageTypeMin(), |
| 100 | + subChannelType.getStorageTypeMax()); |
| 101 | + return tensorType.clone(perAxisType); |
| 102 | + } |
| 103 | + return type; |
| 104 | + } |
| 105 | + |
| 106 | +public: |
| 107 | + explicit NormalizedQuantTypesConverter() { addConversion(convertType); } |
| 108 | +}; |
| 109 | + |
| 110 | +/// This class implements a conversion pattern that converts any generic |
| 111 | +/// operation with sub-channel quantized types to an equivalent operation with |
| 112 | +/// per-tensor or per-axis quantized types. |
| 113 | +class ConvertGenericOpwithSubChannelType : public ConversionPattern { |
| 114 | +public: |
| 115 | + ConvertGenericOpwithSubChannelType(TypeConverter &typeConverter, |
| 116 | + MLIRContext *context) |
| 117 | + : ConversionPattern(typeConverter, MatchAnyOpTypeTag{}, 0, context) {} |
| 118 | + |
| 119 | + LogicalResult |
| 120 | + matchAndRewrite(Operation *op, ArrayRef<Value> operands, |
| 121 | + ConversionPatternRewriter &rewriter) const final { |
| 122 | + SmallVector<Type> resultTypes; |
| 123 | + if (failed(typeConverter->convertTypes(op->getResultTypes(), resultTypes))) |
| 124 | + return failure(); |
| 125 | + |
| 126 | + auto *newOp = Operation::create( |
| 127 | + op->getLoc(), op->getName(), resultTypes, operands, op->getAttrs(), |
| 128 | + op->getPropertiesStorage(), op->getSuccessors(), op->getNumRegions()); |
| 129 | + for (auto regions : llvm::zip(op->getRegions(), newOp->getRegions())) { |
| 130 | + Region &before = std::get<0>(regions); |
| 131 | + Region &parent = std::get<1>(regions); |
| 132 | + rewriter.inlineRegionBefore(before, parent, parent.end()); |
| 133 | + if (failed(rewriter.convertRegionTypes(&parent, *typeConverter))) |
| 134 | + return failure(); |
| 135 | + } |
| 136 | + rewriter.insert(newOp); |
| 137 | + rewriter.replaceOp(op, newOp->getResults()); |
| 138 | + return success(); |
| 139 | + } |
| 140 | +}; |
| 141 | + |
| 142 | +// Conversion pass |
| 143 | +class NormalizeQuantTypes |
| 144 | + : public impl::NormalizeQuantTypesBase<NormalizeQuantTypes> { |
| 145 | +public: |
| 146 | + void runOnOperation() override { |
| 147 | + |
| 148 | + auto moduleOp = cast<ModuleOp>(getOperation()); |
| 149 | + auto *context = &getContext(); |
| 150 | + |
| 151 | + NormalizedQuantTypesConverter typeConverter; |
| 152 | + ConversionTarget target(*context); |
| 153 | + |
| 154 | + // Determine legal operations. |
| 155 | + target.addDynamicallyLegalOp<func::FuncOp>([&](func::FuncOp op) { |
| 156 | + return typeConverter.isSignatureLegal(op.getFunctionType()) && |
| 157 | + typeConverter.isLegal(&op.getBody()); |
| 158 | + }); |
| 159 | + target.markUnknownOpDynamicallyLegal([&](Operation *op) { |
| 160 | + return typeConverter.isLegal(op->getOperandTypes()) && |
| 161 | + typeConverter.isLegal(op->getResultTypes()); |
| 162 | + }); |
| 163 | + |
| 164 | + // Register conversion patterns |
| 165 | + RewritePatternSet patterns(context); |
| 166 | + populateFunctionOpInterfaceTypeConversionPattern<func::FuncOp>( |
| 167 | + patterns, typeConverter); |
| 168 | + patterns.add<ConvertGenericOpwithSubChannelType>(typeConverter, context); |
| 169 | + |
| 170 | + // Apply conversion |
| 171 | + if (failed(applyFullConversion(moduleOp, target, std::move(patterns)))) |
| 172 | + signalPassFailure(); |
| 173 | + } |
| 174 | +}; |
| 175 | + |
| 176 | +} // namespace |
| 177 | + |
| 178 | +} // namespace quant |
| 179 | +} // namespace mlir |
0 commit comments