xref: /llvm-project/mlir/lib/Dialect/GPU/Transforms/ParallelLoopMapper.cpp (revision 7bdd3722f20c68ce14d153939bd8aa11d1c0d482)
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/ParallelLoopMapper.h"
15 
16 #include "PassDetail.h"
17 #include "mlir/Dialect/GPU/GPUDialect.h"
18 #include "mlir/Dialect/GPU/Passes.h"
19 #include "mlir/Dialect/SCF/SCF.h"
20 #include "mlir/IR/AffineMap.h"
21 
22 namespace mlir {
23 
24 using scf::ParallelOp;
25 
26 StringRef gpu::getMappingAttrName() { return "mapping"; }
27 
28 LogicalResult
29 gpu::setMappingAttr(ParallelOp ploopOp,
30                     ArrayRef<ParallelLoopDimMappingAttr> mapping) {
31   // Verify that each processor is mapped to only once.
32   llvm::DenseSet<gpu::Processor> specifiedMappings;
33   for (auto dimAttr : mapping) {
34     gpu::Processor processor = dimAttr.getProcessor();
35     if (processor != gpu::Processor::Sequential &&
36         specifiedMappings.count(processor))
37       return ploopOp.emitError(
38           "invalid mapping multiple loops to same processor");
39   }
40   ArrayRef<Attribute> mappingAsAttrs(mapping.data(), mapping.size());
41   ploopOp->setAttr(getMappingAttrName(),
42                    ArrayAttr::get(ploopOp.getContext(), mappingAsAttrs));
43   return success();
44 }
45 
46 namespace gpu {
47 namespace {
48 enum MappingLevel { MapGrid = 0, MapBlock = 1, Sequential = 2 };
49 } // namespace
50 
51 static constexpr int kNumHardwareIds = 3;
52 
53 /// Bounded increment on MappingLevel. Increments to the next
54 /// level unless Sequential was already reached.
55 static MappingLevel &operator++(MappingLevel &mappingLevel) {
56   if (mappingLevel < Sequential) {
57     mappingLevel = static_cast<MappingLevel>(mappingLevel + 1);
58   }
59   return mappingLevel;
60 }
61 
62 /// Computed the hardware id to use for a given mapping level. Will
63 /// assign x,y and z hardware ids for the first 3 dimensions and use
64 /// sequential after.
65 /// TODO: Make this use x for the inner-most loop that is
66 /// distributed to map to x, the next innermost to y and the next innermost to
67 /// z.
68 static Processor getHardwareIdForMapping(MappingLevel level, int dimension) {
69 
70   if (dimension >= kNumHardwareIds || level == Sequential)
71     return Processor::Sequential;
72   switch (level) {
73   case MapGrid:
74     switch (dimension) {
75     case 0:
76       return Processor::BlockX;
77     case 1:
78       return Processor::BlockY;
79     case 2:
80       return Processor::BlockZ;
81     default:
82       return Processor::Sequential;
83     }
84     break;
85   case MapBlock:
86     switch (dimension) {
87     case 0:
88       return Processor::ThreadX;
89     case 1:
90       return Processor::ThreadY;
91     case 2:
92       return Processor::ThreadZ;
93     default:
94       return Processor::Sequential;
95     }
96   default:;
97   }
98   return Processor::Sequential;
99 }
100 
101 /// Add mapping information to the given parallel loop. Do not add
102 /// mapping information if the loop already has it. Also, don't
103 /// start a mapping at a nested loop.
104 static void mapParallelOp(ParallelOp parallelOp,
105                           MappingLevel mappingLevel = MapGrid) {
106   // Do not try to add a mapping to already mapped loops or nested loops.
107   if (parallelOp->getAttr(getMappingAttrName()) ||
108       ((mappingLevel == MapGrid) && parallelOp->getParentOfType<ParallelOp>()))
109     return;
110 
111   MLIRContext *ctx = parallelOp.getContext();
112   Builder b(ctx);
113   SmallVector<ParallelLoopDimMappingAttr, 4> attrs;
114   attrs.reserve(parallelOp.getNumLoops());
115   for (int i = 0, e = parallelOp.getNumLoops(); i < e; ++i) {
116     attrs.push_back(b.getAttr<ParallelLoopDimMappingAttr>(
117         getHardwareIdForMapping(mappingLevel, i), b.getDimIdentityMap(),
118         b.getDimIdentityMap()));
119   }
120   (void)setMappingAttr(parallelOp, attrs);
121   ++mappingLevel;
122   // Parallel loop operations are immediately nested, so do not use
123   // walk but just iterate over the operations.
124   for (Operation &op : *parallelOp.getBody()) {
125     if (ParallelOp nested = dyn_cast<ParallelOp>(op))
126       mapParallelOp(nested, mappingLevel);
127   }
128 }
129 
130 namespace {
131 struct GpuMapParallelLoopsPass
132     : public GpuMapParallelLoopsPassBase<GpuMapParallelLoopsPass> {
133   void runOnOperation() override {
134     for (Region &region : getOperation()->getRegions()) {
135       region.walk([](ParallelOp parallelOp) { mapParallelOp(parallelOp); });
136     }
137   }
138 };
139 
140 } // namespace
141 } // namespace gpu
142 } // namespace mlir
143 
144 std::unique_ptr<mlir::OperationPass<mlir::func::FuncOp>>
145 mlir::createGpuMapParallelLoopsPass() {
146   return std::make_unique<gpu::GpuMapParallelLoopsPass>();
147 }
148