|
| 1 | +//===- DecomposeGenericByUnfoldingPermutation.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 | +#include "mlir/Dialect/Affine/IR/AffineOps.h" |
| 10 | +#include "mlir/Dialect/Linalg/IR/Linalg.h" |
| 11 | +#include "mlir/Dialect/Linalg/Transforms/Transforms.h" |
| 12 | +#include <map> |
| 13 | +#include <optional> |
| 14 | +#include <utility> |
| 15 | + |
| 16 | +using namespace mlir; |
| 17 | +using namespace mlir::linalg; |
| 18 | + |
| 19 | +namespace { |
| 20 | + |
| 21 | +/// This pattern decomposes the input operand(s) of a linalg.generic that has |
| 22 | +/// a `transpose`, `broadcast`, or a mixture of two, into explicit transpose |
| 23 | +/// and broadcast. Having them folded into the linalg.generic is a good |
| 24 | +/// optimization but sometimes we may want to unwrap, i.e., `unfold` them as |
| 25 | +/// explicit transpose and broadcast. This rewrite pattern helps do it for |
| 26 | +/// each input operand. This is useful for instance when trying to recognize |
| 27 | +/// named ops. |
| 28 | +/// |
| 29 | +/// The transpose, broadcast, or mixture of both, are expressed in the affine |
| 30 | +/// map of the operand. Technically it is essentially `projected permutation`. |
| 31 | +/// |
| 32 | +/// Example |
| 33 | +/// |
| 34 | +/// ```mlir |
| 35 | +/// |
| 36 | +/// #projection = affine_map<(d0, d1, d2, d3, d4) -> (d2, d3, d1)> |
| 37 | +/// #identity = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, d3, d4)> |
| 38 | +/// ... |
| 39 | +/// %res = linalg.generic |
| 40 | +/// { indexing_maps = [#projection, #identity, #identity], |
| 41 | +/// iterator_types = ["parallel", "parallel", "parallel", |
| 42 | +/// "parallel", "parallel"]} |
| 43 | +/// ins(%x, %y : tensor<7x8x9xf32>, tensor<5x9x7x8x10xf32>) |
| 44 | +/// outs(%z : tensor<5x9x7x8x10xf32>) { |
| 45 | +/// ^bb0(%in: f32, %in_1: f32, %out: f32): |
| 46 | +/// %div = arith.divf %in, %in_1 : f32 |
| 47 | +/// linalg.yield %div : f32 |
| 48 | +/// } -> tensor<5x9x7x8x10xf32> |
| 49 | +/// ``` |
| 50 | +/// |
| 51 | +/// In the above IR operand `%x` map is a projected-permutation. This can be |
| 52 | +/// unfolded as: |
| 53 | +/// |
| 54 | +/// ```mlir |
| 55 | +/// ... |
| 56 | +/// %x_trans = linalg.transpose |
| 57 | +/// ins(%x : tensor<7x8x9xf32>) |
| 58 | +/// outs(%e1 : tensor<9x7x8xf32>) permutation = [2, 0, 1] |
| 59 | +/// ... |
| 60 | +/// %x_trans_bc = linalg.broadcast |
| 61 | +/// ins(%x_trans : tensor<9x7x8xf32>) |
| 62 | +/// outs(%e2 : tensor<5x9x7x8x10xf32>) dimensions = [0, 4] |
| 63 | +/// %2 = linalg.div |
| 64 | +/// ins(%x_trans_bc, %y : |
| 65 | +/// tensor<5x9x7x8x10xf32>, tensor<5x9x7x8x10xf32>) |
| 66 | +/// outs(%arg2 : tensor<5x9x7x8x10xf32>) -> tensor<5x9x7x8x10xf32> |
| 67 | +/// |
| 68 | +/// Note that linalg.generic has been 'specialized' to linalg.div. |
| 69 | +/// |
| 70 | +/// To unfold it, it is more optimal to transpose first and then do the |
| 71 | +/// broadcast. However, if transpose is done first, the permutation map needs |
| 72 | +/// to be expressed in terms of reduced dimension as broadcast hasn't happened |
| 73 | +/// yet. Also, the broadcast dimensions in a linalg.generic come from other |
| 74 | +/// operands (those not broadcasted along that particular dimension). We work |
| 75 | +/// this out by computing the convex-polyhedron shape of the linalg.generic |
| 76 | +/// iteration space from shapes of all the operands, both inputs and outputs. |
| 77 | +/// |
| 78 | +struct DecomposeProjectedPermutation : public OpRewritePattern<GenericOp> { |
| 79 | + using OpRewritePattern<GenericOp>::OpRewritePattern; |
| 80 | + |
| 81 | + LogicalResult matchAndRewrite(GenericOp genericOp, |
| 82 | + PatternRewriter &rewriter) const override; |
| 83 | +}; |
| 84 | + |
| 85 | +/// For the given `map`, determine what dimensions are transposed and what |
| 86 | +/// dimensions are broadcasted. |
| 87 | +/// Returns : |
| 88 | +/// transpose-permutation, broadcast-dimensions` (empty if not needed) |
| 89 | +/// |
| 90 | +std::pair<SmallVector<int64_t>, SmallVector<int64_t>> |
| 91 | +computeTransposeBroadcast(AffineMap &map) { |
| 92 | + assert(map.isProjectedPermutation(false) && "not a projection"); |
| 93 | + |
| 94 | + // As the map is a projection it likely operates on a smaller set of |
| 95 | + // dimensions as far as the transpose is concerned (rest are broadcast). |
| 96 | + int64_t minorSize = map.getNumResults(); |
| 97 | + |
| 98 | + SmallVector<int64_t> minorResult; |
| 99 | + for (int64_t i = 0; i < minorSize; ++i) { |
| 100 | + auto expr = cast<AffineDimExpr>(map. getResults()[i]); |
| 101 | + minorResult.push_back(expr.getPosition()); |
| 102 | + } |
| 103 | + |
| 104 | + // If dims are not monotonically increasing then transpose is present. |
| 105 | + SmallVector<int64_t> sortedResMap(minorResult); |
| 106 | + std::sort(sortedResMap.begin(), sortedResMap.end()); |
| 107 | + bool hasTranspose = !std::equal(minorResult.begin(), minorResult.end(), |
| 108 | + sortedResMap.begin(), sortedResMap.end()); |
| 109 | + |
| 110 | + // Walk the sorted map result to determine which dimensions are broadcasted. |
| 111 | + SmallVector<int64_t> broadcast; |
| 112 | + for (int64_t i = 0, j = 0; i < map.getNumInputs(); ++i) { |
| 113 | + if (j < minorSize && sortedResMap[j] == i) { |
| 114 | + j++; |
| 115 | + continue; |
| 116 | + } |
| 117 | + broadcast.push_back(i); |
| 118 | + } |
| 119 | + |
| 120 | + SmallVector<int64_t> permutation; |
| 121 | + if (hasTranspose) { |
| 122 | + // Consider an operand `x : tensor<7x8x9>` of a genericOp that has |
| 123 | + // affine map `affine_map<(d0, d1, d2, d3, d4) -> (d2, d3, d1)>` |
| 124 | + // `x`s access is both transposed and broadcast. But when specifying |
| 125 | + // the `linalg.transpose(x : tensor<7x8x9>)` the dimensions need to be |
| 126 | + // specified as `affine_map<(d0,d1,d2) -> (d1, d2, d0)` instead of |
| 127 | + // refering to d3, d4. Therefore, re-base the transpose dimensions so |
| 128 | + // that they start from d0. |
| 129 | + permutation.resize(minorSize); |
| 130 | + std::map<int64_t, int64_t> minorMap; |
| 131 | + for (int64_t i = 0; i < minorSize; ++i) |
| 132 | + minorMap.insert({sortedResMap[i], i}); |
| 133 | + |
| 134 | + // Re-map the dimensions. |
| 135 | + SmallVector<int64_t> remappedResult(minorSize); |
| 136 | + for (int64_t i = 0; i < minorSize; ++i) |
| 137 | + remappedResult[i] = minorMap[minorResult[i]]; |
| 138 | + |
| 139 | + /// Calculate the permutation for the transpose. |
| 140 | + for (unsigned i = 0; i < minorSize; ++i) { |
| 141 | + permutation[remappedResult[i]] = i; |
| 142 | + } |
| 143 | + } |
| 144 | + return {permutation, broadcast}; |
| 145 | +} |
| 146 | + |
| 147 | +LogicalResult DecomposeProjectedPermutation::matchAndRewrite( |
| 148 | + GenericOp op, PatternRewriter &rewriter) const { |
| 149 | + if (!op.hasPureTensorSemantics() || op.isSingleInputOutput() || |
| 150 | + op.isSingleYieldOp() || !op.isAllParallelLoops()) |
| 151 | + return failure(); |
| 152 | + |
| 153 | + // If the map of an operand is not a `projected permutation` then |
| 154 | + // it cannot be decomposed to mere transpose and broadcast. |
| 155 | + // The requirement that all maps be `projected permutation` may be |
| 156 | + // over-restrictive but since we need to determine shape of the |
| 157 | + // iteration space as well, reject if any map violates assumption. |
| 158 | + for (auto &opOperand : op->getOpOperands()) { |
| 159 | + auto map = op.getMatchingIndexingMap(&opOperand); |
| 160 | + if (!map.isProjectedPermutation(false)) |
| 161 | + return failure(); |
| 162 | + } |
| 163 | + |
| 164 | + // Decomposing linalg.generic involves creating `tensor.empty` |
| 165 | + // which can have dynamic shapes but then we would have to work |
| 166 | + // out which operand can supply that runtime-value (tensor.dim). |
| 167 | + // Leaving it as a future TODO. |
| 168 | + if (llvm::any_of(op->getOpOperands(), [](OpOperand &oper) { |
| 169 | + auto opType = cast<RankedTensorType>(oper.get().getType()); |
| 170 | + return ShapedType::isDynamicShape(opType.getShape()); |
| 171 | + })) |
| 172 | + return failure(); |
| 173 | + |
| 174 | + auto outputShape = op.getStaticLoopRanges(); |
| 175 | + |
| 176 | + auto loc = op.getLoc(); |
| 177 | + bool isChanged = false; |
| 178 | + SmallVector<Value> newInitValues = op.getDpsInputs(); |
| 179 | + SmallVector<AffineMap> newMap = op.getIndexingMapsArray(); |
| 180 | + |
| 181 | + // Walk over each input operand and unfold if it is transposed, broadcast |
| 182 | + // or mix of two via operand's affine-map. |
| 183 | + for (int64_t i = 0; i < op.getNumDpsInputs(); ++i) { |
| 184 | + auto &map = newMap[i]; |
| 185 | + auto inputRTType = cast<RankedTensorType>(newInitValues[i].getType()); |
| 186 | + auto elType = inputRTType.getElementType(); |
| 187 | + |
| 188 | + /// Nothing to do if map is already an identity. |
| 189 | + if (map.isIdentity()) |
| 190 | + continue; |
| 191 | + |
| 192 | + auto [permutation, broadcastedDims] = computeTransposeBroadcast(map); |
| 193 | + |
| 194 | + // Does it need transpose? |
| 195 | + if (!permutation.empty()) { |
| 196 | + /// linalg.transpose permutes the dimensions of input using |
| 197 | + /// rule: dim(result, i) = dim(input, permutation[i]) |
| 198 | + SmallVector<int64_t> transposedShape(map.getNumResults()); |
| 199 | + for (int64_t i = 0; i < map.getNumResults(); ++i) |
| 200 | + transposedShape[i] = inputRTType.getShape()[permutation[i]]; |
| 201 | + |
| 202 | + Value emptyTensor = |
| 203 | + rewriter.create<tensor::EmptyOp>(loc, transposedShape, elType); |
| 204 | + |
| 205 | + auto transposeOp = rewriter.create<TransposeOp>(loc, newInitValues[i], |
| 206 | + emptyTensor, permutation); |
| 207 | + newInitValues[i] = transposeOp->getResult(0); |
| 208 | + isChanged = true; |
| 209 | + } |
| 210 | + |
| 211 | + // Does it require broadcast? |
| 212 | + if (!broadcastedDims.empty()) { |
| 213 | + assert(broadcastedDims.size() && "should have non size broadcast"); |
| 214 | + Value emptyTensor = rewriter.create<tensor::EmptyOp>( |
| 215 | + loc, outputShape, inputRTType.getElementType()); |
| 216 | + |
| 217 | + auto broadcastOp = rewriter.create<linalg::BroadcastOp>( |
| 218 | + loc, newInitValues[i], emptyTensor, broadcastedDims); |
| 219 | + |
| 220 | + newInitValues[i] = broadcastOp->getResult(0); |
| 221 | + isChanged = true; |
| 222 | + } |
| 223 | + newMap[i] = rewriter.getMultiDimIdentityMap(map.getNumDims()); |
| 224 | + } |
| 225 | + |
| 226 | + if (isChanged) { |
| 227 | + SmallVector<Value> operands = op->getOperands(); |
| 228 | + ValueRange operandsRef(operands); |
| 229 | + |
| 230 | + auto newOp = rewriter.create<linalg::GenericOp>( |
| 231 | + /*location=*/op.getLoc(), |
| 232 | + /*resultTensorTypes=*/op->getResultTypes(), |
| 233 | + /*inputs=*/newInitValues, |
| 234 | + /*outputs=*/operandsRef.drop_front(op.getNumDpsInputs()), |
| 235 | + /*indexingMaps=*/newMap, |
| 236 | + /*iteratorTypes=*/op.getIteratorTypesArray()); |
| 237 | + |
| 238 | + newOp.getRegion().takeBody(op->getRegion(0)); |
| 239 | + rewriter.replaceOp(op, newOp->getResults()); |
| 240 | + } |
| 241 | + return success(); |
| 242 | +} |
| 243 | + |
| 244 | +} // namespace |
| 245 | + |
| 246 | +void mlir::linalg::populateDecomposeProjectedPermutationPatterns( |
| 247 | + RewritePatternSet &patterns) { |
| 248 | + patterns.insert<DecomposeProjectedPermutation>(patterns.getContext()); |
| 249 | +} |
0 commit comments