|
| 1 | +//===- TosaFoldConstantClamp.cpp ------------------------------------------===// |
| 2 | +// |
| 3 | +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | +// See https://llvm.org/LICENSE.txt for license information. |
| 5 | +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | +// |
| 7 | +//===----------------------------------------------------------------------===// |
| 8 | +// |
| 9 | +// Fold TOSA Clamp operation on constant data |
| 10 | +// |
| 11 | +//===----------------------------------------------------------------------===// |
| 12 | + |
| 13 | +#include "mlir/Dialect/Tosa/IR/TosaOps.h" |
| 14 | +#include "mlir/Dialect/Tosa/Transforms/Passes.h" |
| 15 | +#include "mlir/Dialect/Tosa/Transforms/TosaFoldCommon.h" |
| 16 | +#include "mlir/IR/Matchers.h" |
| 17 | +#include "mlir/Pass/Pass.h" |
| 18 | +#include <llvm/ADT/APFloat.h> |
| 19 | +#include <llvm/ADT/APInt.h> |
| 20 | +#include <llvm/Support/Debug.h> |
| 21 | +#include <mlir/IR/BuiltinAttributes.h> |
| 22 | +#include <mlir/IR/BuiltinTypes.h> |
| 23 | +#include <mlir/Support/LogicalResult.h> |
| 24 | + |
| 25 | +using namespace mlir; |
| 26 | +using namespace mlir::tosa; |
| 27 | + |
| 28 | +namespace { |
| 29 | + |
| 30 | +struct TosaFoldConstantClamp : public OpRewritePattern<ClampOp> { |
| 31 | + |
| 32 | + using OpRewritePattern::OpRewritePattern; |
| 33 | + |
| 34 | + static void |
| 35 | + changeSemanticsLossless(APFloat &floatVal, |
| 36 | + const llvm::fltSemantics *floatSemantics) { |
| 37 | + bool losesInfo; |
| 38 | + floatVal.convert(*floatSemantics, tosaRoundingMode, &losesInfo); |
| 39 | + assert(!losesInfo); |
| 40 | + } |
| 41 | + |
| 42 | + DenseElementsAttr applyClamp(DenseElementsAttr inputValues, |
| 43 | + const APInt &lowerBound, const APInt &upperBound, |
| 44 | + TensorType resultType) const { |
| 45 | + |
| 46 | + // Determine the width for the APInt comparison |
| 47 | + auto comparisonWidth = |
| 48 | + std::max(inputValues.getElementType().getIntOrFloatBitWidth(), |
| 49 | + lowerBound.getBitWidth()); |
| 50 | + |
| 51 | + auto resultingIntType = cast<IntegerType>(resultType.getElementType()); |
| 52 | + |
| 53 | + // Ensure that the value is larger than the lower bound |
| 54 | + auto clampLower = [&lowerBound, &comparisonWidth](const APInt &val, |
| 55 | + IntegerType type) { |
| 56 | + auto clampedLower = llvm::APIntOps::smax( |
| 57 | + val.sext(comparisonWidth), lowerBound.sext(comparisonWidth)); |
| 58 | + // Make sure the output value has the correct type |
| 59 | + assert(type.getWidth() >= clampedLower.getSignificantBits()); |
| 60 | + return clampedLower.trunc(type.getWidth()); |
| 61 | + }; |
| 62 | + auto newTensor = applyElementWise<APInt, APInt, IntegerType>( |
| 63 | + inputValues, clampLower, resultingIntType); |
| 64 | + |
| 65 | + // Next, make sure the upper bound is adhered to |
| 66 | + auto clampUpper = [&upperBound, &comparisonWidth](const APInt &val, |
| 67 | + IntegerType type) { |
| 68 | + auto clampedUpper = llvm::APIntOps::smin( |
| 69 | + val.sext(comparisonWidth), upperBound.sext(comparisonWidth)); |
| 70 | + assert(type.getWidth() >= clampedUpper.getSignificantBits()); |
| 71 | + return clampedUpper.trunc(type.getWidth()); |
| 72 | + }; |
| 73 | + newTensor = applyElementWise<APInt, APInt, IntegerType>( |
| 74 | + newTensor, clampUpper, resultingIntType); |
| 75 | + |
| 76 | + return newTensor; |
| 77 | + } |
| 78 | + |
| 79 | + DenseElementsAttr applyClamp(DenseElementsAttr inputValues, |
| 80 | + APFloat lowerBound, APFloat upperBound, |
| 81 | + TensorType resultType) const { |
| 82 | + auto inputValType = cast<FloatType>(inputValues.getElementType()); |
| 83 | + auto inputWidth = inputValType.getWidth(); |
| 84 | + auto bWidth = APFloat::semanticsSizeInBits(lowerBound.getSemantics()); |
| 85 | + auto *comparisonSem = inputWidth < bWidth |
| 86 | + ? &lowerBound.getSemantics() |
| 87 | + : &inputValType.getFloatSemantics(); |
| 88 | + |
| 89 | + changeSemanticsLossless(lowerBound, comparisonSem); |
| 90 | + changeSemanticsLossless(upperBound, comparisonSem); |
| 91 | + |
| 92 | + auto resultingFloatType = cast<FloatType>(resultType.getElementType()); |
| 93 | + |
| 94 | + // Ensure that the value is larger than the lower bound |
| 95 | + auto clampLower = [&lowerBound, &comparisonSem](APFloat val, |
| 96 | + FloatType type) { |
| 97 | + if (val.isNaN()) { |
| 98 | + return APFloat::getNaN(type.getFloatSemantics()); |
| 99 | + } |
| 100 | + changeSemanticsLossless(val, comparisonSem); |
| 101 | + auto clampedLower = val < lowerBound ? lowerBound : val; |
| 102 | + changeSemanticsLossless(clampedLower, &type.getFloatSemantics()); |
| 103 | + return clampedLower; |
| 104 | + }; |
| 105 | + auto newTensor = applyElementWise<APFloat, APFloat, FloatType>( |
| 106 | + inputValues, clampLower, resultingFloatType); |
| 107 | + |
| 108 | + // Next, make sure the upper bound is adhered to |
| 109 | + auto clampUpper = [&upperBound, &comparisonSem](APFloat val, |
| 110 | + FloatType type) { |
| 111 | + if (val.isNaN()) { |
| 112 | + return APFloat::getNaN(type.getFloatSemantics()); |
| 113 | + } |
| 114 | + changeSemanticsLossless(val, comparisonSem); |
| 115 | + auto clampedUpper = val < upperBound ? val : upperBound; |
| 116 | + changeSemanticsLossless(clampedUpper, &type.getFloatSemantics()); |
| 117 | + return clampedUpper; |
| 118 | + }; |
| 119 | + newTensor = applyElementWise<APFloat, APFloat, FloatType>( |
| 120 | + newTensor, clampUpper, resultingFloatType); |
| 121 | + |
| 122 | + return newTensor; |
| 123 | + } |
| 124 | + |
| 125 | + LogicalResult matchAndRewrite(ClampOp clampOp, |
| 126 | + PatternRewriter &rewriter) const override { |
| 127 | + auto valsToClamp = clampOp.getInput(); |
| 128 | + auto inputElementType = valsToClamp.getType().getElementType(); |
| 129 | + |
| 130 | + // Check if the input is constant |
| 131 | + if (failed(notifyIfNoTosaDenseConstantTensor(valsToClamp, clampOp, |
| 132 | + rewriter))) { |
| 133 | + return failure(); |
| 134 | + } |
| 135 | + |
| 136 | + if (isa<IntegerType>(inputElementType) && |
| 137 | + cast<IntegerType>(inputElementType).isUnsigned()) { |
| 138 | + return rewriter.notifyMatchFailure( |
| 139 | + clampOp, "Currently, unsigned integer clamps are unsupported."); |
| 140 | + } |
| 141 | + |
| 142 | + // Extract the tensor values |
| 143 | + DenseElementsAttr inputValues; |
| 144 | + matchPattern(valsToClamp, m_Constant(&inputValues)); |
| 145 | + |
| 146 | + if (!constantUnaryOpShouldBeFolded(clampOp, inputValues)) { |
| 147 | + return rewriter.notifyMatchFailure( |
| 148 | + clampOp, |
| 149 | + "Currently, clamps will only be folded if this requires only " |
| 150 | + "little additional memory usage."); |
| 151 | + } |
| 152 | + |
| 153 | + // Apply the clamp to all values of the int/float tensor |
| 154 | + auto resultType = clampOp.getType(); |
| 155 | + DenseElementsAttr newTensor; |
| 156 | + if (isa<IntegerType>(inputElementType)) { |
| 157 | + auto lowerBoundVal = clampOp.getMinIntAttr().getValue(); |
| 158 | + auto upperBoundVal = clampOp.getMaxIntAttr().getValue(); |
| 159 | + assert(lowerBoundVal.getBitWidth() == upperBoundVal.getBitWidth()); |
| 160 | + |
| 161 | + newTensor = |
| 162 | + applyClamp(inputValues, lowerBoundVal, upperBoundVal, resultType); |
| 163 | + } else { |
| 164 | + assert(isa<FloatType>(inputElementType)); |
| 165 | + auto lowerBoundVal = clampOp.getMinFp(); |
| 166 | + auto upperBoundVal = clampOp.getMaxFp(); |
| 167 | + assert(APFloat::getSizeInBits(lowerBoundVal.getSemantics()) == |
| 168 | + APFloat::getSizeInBits(upperBoundVal.getSemantics())); |
| 169 | + |
| 170 | + newTensor = |
| 171 | + applyClamp(inputValues, lowerBoundVal, upperBoundVal, resultType); |
| 172 | + } |
| 173 | + |
| 174 | + rewriter.replaceOpWithNewOp<ConstOp>(clampOp, newTensor.getType(), |
| 175 | + newTensor); |
| 176 | + |
| 177 | + return success(); |
| 178 | + } |
| 179 | +}; |
| 180 | + |
| 181 | +} // namespace |
| 182 | + |
| 183 | +void mlir::tosa::populateTosaFoldConstantClampPatterns( |
| 184 | + MLIRContext *ctx, RewritePatternSet &patterns) { |
| 185 | + patterns.add<TosaFoldConstantClamp>(ctx); |
| 186 | +} |
0 commit comments