|
| 1 | +//===- DistributionUtils.cpp - Distribution tools for GPUOps --------------===// |
| 2 | +// |
| 3 | +// Part of the MLIR 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 | +// This file implements distribution utility methods. |
| 10 | +// |
| 11 | +//===----------------------------------------------------------------------===// |
| 12 | + |
| 13 | +#include "mlir/Dialect/GPU/Utils/DistributionUtils.h" |
| 14 | +#include "mlir/Dialect/Affine/IR/AffineOps.h" |
| 15 | +#include "mlir/Dialect/Arith/IR/Arith.h" |
| 16 | +#include "mlir/IR/Value.h" |
| 17 | + |
| 18 | +#include <numeric> |
| 19 | + |
| 20 | +using namespace mlir; |
| 21 | +using namespace mlir::gpu; |
| 22 | + |
| 23 | +WarpExecuteOnLane0Op mlir::gpu::moveRegionToNewWarpOpAndReplaceReturns( |
| 24 | + RewriterBase &rewriter, WarpExecuteOnLane0Op warpOp, |
| 25 | + ValueRange newYieldedValues, TypeRange newReturnTypes) { |
| 26 | + // Create a new op before the existing one, with the extra operands. |
| 27 | + OpBuilder::InsertionGuard g(rewriter); |
| 28 | + rewriter.setInsertionPoint(warpOp); |
| 29 | + auto newWarpOp = rewriter.create<WarpExecuteOnLane0Op>( |
| 30 | + warpOp.getLoc(), newReturnTypes, warpOp.getLaneid(), warpOp.getWarpSize(), |
| 31 | + warpOp.getArgs(), warpOp.getBody()->getArgumentTypes()); |
| 32 | + |
| 33 | + Region &opBody = warpOp.getBodyRegion(); |
| 34 | + Region &newOpBody = newWarpOp.getBodyRegion(); |
| 35 | + Block &newOpFirstBlock = newOpBody.front(); |
| 36 | + rewriter.inlineRegionBefore(opBody, newOpBody, newOpBody.begin()); |
| 37 | + rewriter.eraseBlock(&newOpFirstBlock); |
| 38 | + assert(newWarpOp.getWarpRegion().hasOneBlock() && |
| 39 | + "expected WarpOp with single block"); |
| 40 | + |
| 41 | + auto yield = |
| 42 | + cast<gpu::YieldOp>(newOpBody.getBlocks().begin()->getTerminator()); |
| 43 | + |
| 44 | + rewriter.modifyOpInPlace( |
| 45 | + yield, [&]() { yield.getValuesMutable().assign(newYieldedValues); }); |
| 46 | + return newWarpOp; |
| 47 | +} |
| 48 | + |
| 49 | +WarpExecuteOnLane0Op mlir::gpu::moveRegionToNewWarpOpAndAppendReturns( |
| 50 | + RewriterBase &rewriter, WarpExecuteOnLane0Op warpOp, |
| 51 | + ValueRange newYieldedValues, TypeRange newReturnTypes, |
| 52 | + llvm::SmallVector<size_t> &indices) { |
| 53 | + SmallVector<Type> types(warpOp.getResultTypes().begin(), |
| 54 | + warpOp.getResultTypes().end()); |
| 55 | + auto yield = cast<gpu::YieldOp>( |
| 56 | + warpOp.getBodyRegion().getBlocks().begin()->getTerminator()); |
| 57 | + llvm::SmallSetVector<Value, 32> yieldValues(yield.getOperands().begin(), |
| 58 | + yield.getOperands().end()); |
| 59 | + for (auto newRet : llvm::zip(newYieldedValues, newReturnTypes)) { |
| 60 | + if (yieldValues.insert(std::get<0>(newRet))) { |
| 61 | + types.push_back(std::get<1>(newRet)); |
| 62 | + indices.push_back(yieldValues.size() - 1); |
| 63 | + } else { |
| 64 | + // If the value already exit the region don't create a new output. |
| 65 | + for (auto [idx, yieldOperand] : |
| 66 | + llvm::enumerate(yieldValues.getArrayRef())) { |
| 67 | + if (yieldOperand == std::get<0>(newRet)) { |
| 68 | + indices.push_back(idx); |
| 69 | + break; |
| 70 | + } |
| 71 | + } |
| 72 | + } |
| 73 | + } |
| 74 | + yieldValues.insert(newYieldedValues.begin(), newYieldedValues.end()); |
| 75 | + WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndReplaceReturns( |
| 76 | + rewriter, warpOp, yieldValues.getArrayRef(), types); |
| 77 | + rewriter.replaceOp(warpOp, |
| 78 | + newWarpOp.getResults().take_front(warpOp.getNumResults())); |
| 79 | + return newWarpOp; |
| 80 | +} |
| 81 | + |
| 82 | +bool mlir::gpu::canBeHoisted(Operation *op, |
| 83 | + function_ref<bool(Value)> definedOutside) { |
| 84 | + return llvm::all_of(op->getOperands(), definedOutside) && |
| 85 | + isMemoryEffectFree(op) && op->getNumRegions() == 0; |
| 86 | +} |
| 87 | + |
| 88 | +OpOperand * |
| 89 | +mlir::gpu::getWarpResult(WarpExecuteOnLane0Op warpOp, |
| 90 | + const std::function<bool(Operation *)> &fn) { |
| 91 | + auto yield = cast<gpu::YieldOp>( |
| 92 | + warpOp.getBodyRegion().getBlocks().begin()->getTerminator()); |
| 93 | + for (OpOperand &yieldOperand : yield->getOpOperands()) { |
| 94 | + Value yieldValues = yieldOperand.get(); |
| 95 | + Operation *definedOp = yieldValues.getDefiningOp(); |
| 96 | + if (definedOp && fn(definedOp)) { |
| 97 | + if (!warpOp.getResult(yieldOperand.getOperandNumber()).use_empty()) |
| 98 | + return &yieldOperand; |
| 99 | + } |
| 100 | + } |
| 101 | + return {}; |
| 102 | +} |
| 103 | + |
| 104 | +bool mlir::gpu::delinearizeLaneId(OpBuilder &builder, Location loc, |
| 105 | + ArrayRef<int64_t> originalShape, |
| 106 | + ArrayRef<int64_t> distributedShape, |
| 107 | + int64_t warpSize, Value laneId, |
| 108 | + SmallVectorImpl<Value> &delinearizedIds) { |
| 109 | + // If the original shape and the distributed shape is the same, we don't |
| 110 | + // distribute at all--every thread is handling the whole. For such case, we |
| 111 | + // should not rely on lane IDs later. So just return an empty lane ID vector. |
| 112 | + if (originalShape == distributedShape) { |
| 113 | + delinearizedIds.clear(); |
| 114 | + return true; |
| 115 | + } |
| 116 | + |
| 117 | + SmallVector<int64_t> sizes; |
| 118 | + for (auto [large, small] : llvm::zip_equal(originalShape, distributedShape)) { |
| 119 | + if (large % small != 0) |
| 120 | + return false; |
| 121 | + sizes.push_back(large / small); |
| 122 | + } |
| 123 | + if (std::accumulate(sizes.begin(), sizes.end(), 1, |
| 124 | + std::multiplies<int64_t>()) != warpSize) |
| 125 | + return false; |
| 126 | + |
| 127 | + AffineExpr s0, s1; |
| 128 | + bindSymbols(builder.getContext(), s0, s1); |
| 129 | + |
| 130 | + int64_t usedThreads = 1; |
| 131 | + |
| 132 | + Value zero = builder. create<arith::ConstantIndexOp>(loc, 0); |
| 133 | + delinearizedIds.assign(sizes.size(), zero); |
| 134 | + |
| 135 | + for (int i = sizes.size() - 1; i >= 0; --i) { |
| 136 | + usedThreads *= sizes[i]; |
| 137 | + if (usedThreads == warpSize) { |
| 138 | + // We've used up all available threads. Don't need to perform modulo |
| 139 | + // anymore. And we can stop the calculation for further dimensions. |
| 140 | + delinearizedIds[i] = laneId; |
| 141 | + break; |
| 142 | + } |
| 143 | + delinearizedIds[i] = |
| 144 | + affine::makeComposedAffineApply(builder, loc, s0 % sizes[i], {laneId}); |
| 145 | + laneId = affine::makeComposedAffineApply( |
| 146 | + builder, loc, s0.floorDiv(usedThreads), {laneId}); |
| 147 | + } |
| 148 | + return true; |
| 149 | +} |
0 commit comments