| 1 | //===- ParallelLoopMapper.cpp - Utilities for mapping parallel loops to GPU =// |
| 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 | // This file implements utilities to generate mappings for parallel loops to |
| 10 | // GPU devices. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "mlir/Dialect/GPU/Transforms/Passes.h" |
| 15 | |
| 16 | #include "mlir/Dialect/Func/IR/FuncOps.h" |
| 17 | #include "mlir/Dialect/GPU/IR/GPUDialect.h" |
| 18 | #include "mlir/Dialect/GPU/Transforms/ParallelLoopMapper.h" |
| 19 | #include "mlir/Dialect/SCF/IR/SCF.h" |
| 20 | #include "mlir/IR/AffineMap.h" |
| 21 | |
| 22 | namespace mlir { |
| 23 | #define GEN_PASS_DEF_GPUMAPPARALLELLOOPSPASS |
| 24 | #include "mlir/Dialect/GPU/Transforms/Passes.h.inc" |
| 25 | } // namespace mlir |
| 26 | |
| 27 | namespace mlir { |
| 28 | |
| 29 | using scf::ParallelOp; |
| 30 | |
| 31 | StringRef gpu::getMappingAttrName() { return "mapping" ; } |
| 32 | |
| 33 | LogicalResult |
| 34 | gpu::setMappingAttr(ParallelOp ploopOp, |
| 35 | ArrayRef<ParallelLoopDimMappingAttr> mapping) { |
| 36 | // Verify that each processor is mapped to only once. |
| 37 | llvm::DenseSet<gpu::Processor> specifiedMappings; |
| 38 | for (auto dimAttr : mapping) { |
| 39 | gpu::Processor processor = dimAttr.getProcessor(); |
| 40 | if (processor != gpu::Processor::Sequential && |
| 41 | specifiedMappings.count(V: processor)) |
| 42 | return ploopOp.emitError( |
| 43 | message: "invalid mapping multiple loops to same processor" ); |
| 44 | specifiedMappings.insert(V: processor); |
| 45 | } |
| 46 | ArrayRef<Attribute> mappingAsAttrs(mapping.data(), mapping.size()); |
| 47 | ploopOp->setDiscardableAttr( |
| 48 | name: getMappingAttrName(), |
| 49 | value: ArrayAttr::get(context: ploopOp.getContext(), value: mappingAsAttrs)); |
| 50 | return success(); |
| 51 | } |
| 52 | |
| 53 | namespace gpu { |
| 54 | namespace { |
| 55 | enum MappingLevel { MapGrid = 0, MapBlock = 1, Sequential = 2 }; |
| 56 | enum class MappingPolicy { OutermostFirst, InnermostFirst }; |
| 57 | } // namespace |
| 58 | |
| 59 | static constexpr int kNumHardwareIds = 3; |
| 60 | |
| 61 | /// Bounded increment on MappingLevel. Increments to the next |
| 62 | /// level unless Sequential was already reached. |
| 63 | static MappingLevel &operator++(MappingLevel &mappingLevel) { |
| 64 | if (mappingLevel < Sequential) { |
| 65 | mappingLevel = static_cast<MappingLevel>(mappingLevel + 1); |
| 66 | } |
| 67 | return mappingLevel; |
| 68 | } |
| 69 | |
| 70 | // Map the policy string to a typed mapping policy. |
| 71 | // TODO: Revisit this and possibly use a loop interchange pass instead. |
| 72 | static FailureOr<MappingPolicy> getMappingPolicyFromStr(StringRef policy) { |
| 73 | std::string policyCanonical = policy.trim().lower(); |
| 74 | |
| 75 | std::optional<MappingPolicy> option = |
| 76 | llvm::StringSwitch<std::optional<MappingPolicy>>(policyCanonical) |
| 77 | .Case(S: "innermost-first" , Value: MappingPolicy::InnermostFirst) |
| 78 | .Case(S: "outermost-first" , Value: MappingPolicy::OutermostFirst) |
| 79 | .Default(Value: std::nullopt); |
| 80 | |
| 81 | if (!option) |
| 82 | return failure(); |
| 83 | return *option; |
| 84 | } |
| 85 | |
| 86 | /// Computed the hardware id to use for a given mapping level. Will |
| 87 | /// assign x,y and z hardware ids for the first 3 dimensions and use |
| 88 | /// sequential after. |
| 89 | static Processor getHardwareIdForMapping(MappingLevel level, int dimension) { |
| 90 | |
| 91 | if (dimension >= kNumHardwareIds || level == Sequential) |
| 92 | return Processor::Sequential; |
| 93 | |
| 94 | switch (level) { |
| 95 | case MapGrid: |
| 96 | switch (dimension) { |
| 97 | case 0: |
| 98 | return Processor::BlockX; |
| 99 | case 1: |
| 100 | return Processor::BlockY; |
| 101 | case 2: |
| 102 | return Processor::BlockZ; |
| 103 | default: |
| 104 | return Processor::Sequential; |
| 105 | } |
| 106 | break; |
| 107 | case MapBlock: |
| 108 | switch (dimension) { |
| 109 | case 0: |
| 110 | return Processor::ThreadX; |
| 111 | case 1: |
| 112 | return Processor::ThreadY; |
| 113 | case 2: |
| 114 | return Processor::ThreadZ; |
| 115 | default: |
| 116 | return Processor::Sequential; |
| 117 | } |
| 118 | default:; |
| 119 | } |
| 120 | return Processor::Sequential; |
| 121 | } |
| 122 | |
| 123 | /// Add mapping information to the given parallel loop. Do not add |
| 124 | /// mapping information if the loop already has it. Also, don't |
| 125 | /// start a mapping at a nested loop. |
| 126 | static void |
| 127 | mapParallelOp(ParallelOp parallelOp, MappingLevel mappingLevel = MapGrid, |
| 128 | MappingPolicy mappingPolicy = MappingPolicy::OutermostFirst) { |
| 129 | // Do not try to add a mapping to already mapped loops or nested loops. |
| 130 | if (parallelOp->getDiscardableAttr(name: getMappingAttrName()) || |
| 131 | ((mappingLevel == MapGrid) && parallelOp->getParentOfType<ParallelOp>())) |
| 132 | return; |
| 133 | |
| 134 | const int numLoops = static_cast<int>(parallelOp.getNumLoops()); |
| 135 | const int loopsToMap = std::min(a: numLoops, b: kNumHardwareIds); |
| 136 | |
| 137 | MLIRContext *ctx = parallelOp.getContext(); |
| 138 | Builder b(ctx); |
| 139 | SmallVector<ParallelLoopDimMappingAttr, 4> attrs; |
| 140 | attrs.reserve(N: numLoops); |
| 141 | |
| 142 | for (int i = 0; i < numLoops; ++i) { |
| 143 | |
| 144 | // Determine the mapping to use for this loop. |
| 145 | // If the are more loops to map than HW IDs map to sequential. |
| 146 | int hwMapping = kNumHardwareIds; |
| 147 | if (i < loopsToMap) { |
| 148 | hwMapping = (mappingPolicy == MappingPolicy::OutermostFirst) |
| 149 | ? i |
| 150 | : (loopsToMap - 1 - i); |
| 151 | } |
| 152 | |
| 153 | attrs.push_back(Elt: b.getAttr<ParallelLoopDimMappingAttr>( |
| 154 | args: getHardwareIdForMapping(level: mappingLevel, dimension: hwMapping), args: b.getDimIdentityMap(), |
| 155 | args: b.getDimIdentityMap())); |
| 156 | } |
| 157 | (void)setMappingAttr(ploopOp: parallelOp, mapping: attrs); |
| 158 | ++mappingLevel; |
| 159 | // Parallel loop operations are immediately nested, so do not use |
| 160 | // walk but just iterate over the operations. |
| 161 | for (Operation &op : *parallelOp.getBody()) { |
| 162 | if (ParallelOp nested = dyn_cast<ParallelOp>(Val&: op)) |
| 163 | mapParallelOp(parallelOp: nested, mappingLevel, mappingPolicy); |
| 164 | } |
| 165 | } |
| 166 | |
| 167 | namespace { |
| 168 | struct GpuMapParallelLoopsPass |
| 169 | : public impl::GpuMapParallelLoopsPassBase<GpuMapParallelLoopsPass> { |
| 170 | using Base::Base; |
| 171 | |
| 172 | void runOnOperation() override { |
| 173 | // Parse the mapping policy. |
| 174 | FailureOr<MappingPolicy> policyOrFailure = |
| 175 | getMappingPolicyFromStr(policy: mappingPolicyStr); |
| 176 | if (failed(Result: policyOrFailure)) { |
| 177 | getOperation()->emitError() << "Invalid mapping policy specified." ; |
| 178 | return signalPassFailure(); |
| 179 | } |
| 180 | |
| 181 | MappingPolicy policy = *policyOrFailure; |
| 182 | MappingLevel topLevel = MappingLevel::MapGrid; |
| 183 | |
| 184 | for (Region ®ion : getOperation()->getRegions()) { |
| 185 | region.walk(callback: [&](ParallelOp parallelOp) { |
| 186 | mapParallelOp(parallelOp, mappingLevel: topLevel, mappingPolicy: policy); |
| 187 | }); |
| 188 | } |
| 189 | } |
| 190 | }; |
| 191 | |
| 192 | } // namespace |
| 193 | } // namespace gpu |
| 194 | } // namespace mlir |
| 195 | |