xref: /llvm-project/mlir/lib/Dialect/Affine/Transforms/LoopTiling.cpp (revision 2be8af8f0e0780901213b6fd3013a5268ddc3359)
1 //===- LoopTiling.cpp --- Loop tiling pass ------------------------------*-===//
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 a pass to tile loop nests.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "mlir/Dialect/Affine/Passes.h"
14 
15 #include "mlir/Dialect/Affine/Analysis/AffineAnalysis.h"
16 #include "mlir/Dialect/Affine/Analysis/AffineStructures.h"
17 #include "mlir/Dialect/Affine/Analysis/LoopAnalysis.h"
18 #include "mlir/Dialect/Affine/Analysis/Utils.h"
19 #include "mlir/Dialect/Affine/IR/AffineOps.h"
20 #include "mlir/Dialect/Affine/IR/AffineValueMap.h"
21 #include "mlir/Dialect/Affine/LoopUtils.h"
22 #include "mlir/Dialect/Affine/Utils.h"
23 #include "mlir/Dialect/Func/IR/FuncOps.h"
24 #include "mlir/IR/BlockAndValueMapping.h"
25 #include "mlir/IR/Builders.h"
26 #include "llvm/Support/CommandLine.h"
27 #include "llvm/Support/Debug.h"
28 
29 namespace mlir {
30 #define GEN_PASS_DEF_AFFINELOOPTILINGPASS
31 #include "mlir/Dialect/Affine/Passes.h.inc"
32 } // namespace mlir
33 
34 #define DEBUG_TYPE "affine-loop-tile"
35 
36 using namespace mlir;
37 
38 namespace {
39 
40 /// A pass to perform loop tiling on all suitable loop nests of a Function.
41 struct LoopTiling : public impl::AffineLoopTilingPassBase<LoopTiling> {
42   LoopTiling() = default;
43   explicit LoopTiling(uint64_t cacheSizeBytes, bool avoidMaxMinBounds = true)
44       : avoidMaxMinBounds(avoidMaxMinBounds) {
45     this->cacheSizeInKiB = cacheSizeBytes / 1024;
46   }
47 
48   void runOnOperation() override;
49   void getTileSizes(ArrayRef<AffineForOp> band,
50                     SmallVectorImpl<unsigned> *tileSizes);
51 
52   // Default tile size if nothing is provided.
53   constexpr static unsigned kDefaultTileSize = 4;
54 
55   // If true, tile sizes are set to avoid max/min in bounds if possible.
56   bool avoidMaxMinBounds = true;
57 };
58 
59 } // namespace
60 
61 /// Creates a pass to perform loop tiling on all suitable loop nests of a
62 /// Function.
63 std::unique_ptr<OperationPass<func::FuncOp>>
64 mlir::createLoopTilingPass(uint64_t cacheSizeBytes) {
65   return std::make_unique<LoopTiling>(cacheSizeBytes);
66 }
67 std::unique_ptr<OperationPass<func::FuncOp>> mlir::createLoopTilingPass() {
68   return std::make_unique<LoopTiling>();
69 }
70 
71 /// Reduces each tile size to the largest divisor of the corresponding trip
72 /// count (if the trip count is known).
73 static void adjustToDivisorsOfTripCounts(ArrayRef<AffineForOp> band,
74                                          SmallVectorImpl<unsigned> *tileSizes) {
75   assert(band.size() == tileSizes->size() && "invalid tile size count");
76   for (unsigned i = 0, e = band.size(); i < e; i++) {
77     unsigned &tSizeAdjusted = (*tileSizes)[i];
78     Optional<uint64_t> mayConst = getConstantTripCount(band[i]);
79     if (!mayConst)
80       continue;
81     // Adjust the tile size to largest factor of the trip count less than
82     // tSize.
83     uint64_t constTripCount = *mayConst;
84     if (constTripCount > 1 && tSizeAdjusted > constTripCount / 2)
85       tSizeAdjusted = constTripCount / 2;
86     while (constTripCount % tSizeAdjusted != 0)
87       tSizeAdjusted--;
88   }
89 }
90 
91 // Returns tile sizes to use. Checks CL options; if none are specified, sets it
92 // based on a simple model that looks at the memory footprint and determines
93 // tile sizes assuming identity accesses / 1:1 tile size proportional footprint
94 // along each of the dimensions being tiled.
95 // TODO: evolve this model. Tile size determination is a large area
96 // to play with in general.
97 void LoopTiling::getTileSizes(ArrayRef<AffineForOp> band,
98                               SmallVectorImpl<unsigned> *tileSizes) {
99   if (band.empty())
100     return;
101 
102   // Use command-line tileSize for all loops if specified.
103   if (tileSize) {
104     tileSizes->assign(band.size(), tileSize);
105     return;
106   }
107 
108   // Use tileSizes and fill them with default tile size if it's short.
109   if (!this->tileSizes.empty()) {
110     tileSizes->assign(this->tileSizes.begin(), this->tileSizes.end());
111     tileSizes->resize(band.size(), kDefaultTileSize);
112     return;
113   }
114   tileSizes->resize(band.size());
115 
116   // The first loop in the band.
117   AffineForOp rootForOp = band[0];
118   (void)rootForOp;
119 
120   // Obtain memory footprint and set tile sizes so that a tile fits in
121   // the cache size. This is an approximation with the assumption that the
122   // footprint increases with the tile size linearly in that dimension (i.e.,
123   // assumes one-to-one access function).
124   Optional<int64_t> fp = getMemoryFootprintBytes(band[0], 0);
125   if (!fp) {
126     // Fill with default tile sizes if footprint is unknown.
127     std::fill(tileSizes->begin(), tileSizes->end(),
128               LoopTiling::kDefaultTileSize);
129     if (avoidMaxMinBounds)
130       adjustToDivisorsOfTripCounts(band, tileSizes);
131     LLVM_DEBUG(
132         rootForOp.emitWarning("memory footprint unknown: using default tile "
133                               "sizes adjusted to trip count divisors"));
134     return;
135   }
136 
137   // Check how many times larger the cache size is when compared to footprint.
138   uint64_t cacheSizeBytes = cacheSizeInKiB * 1024;
139   uint64_t excessFactor = llvm::divideCeil(*fp, cacheSizeBytes);
140   if (excessFactor <= 1) {
141     // No need of any tiling - set tile size to 1.
142     std::fill(tileSizes->begin(), tileSizes->end(), 1);
143     return;
144   }
145 
146   // Divide all loops equally in an attempt to reduce footprint.
147   // TODO: this is approximate. Ideally, obtain reuse factor /
148   // profitability along each dimension and weight tile sizes based on that as
149   // one possible approach. Or compute a polynomial in tile sizes and solve for
150   // it.
151 
152   // For an n-d tileable band, compute the n^th root of the excess.
153   unsigned tSize =
154       static_cast<unsigned>(floorl(std::pow(excessFactor, 1.0 / band.size())));
155   // We'll keep a running product to determine the last tile size better.
156   unsigned cumulProductOfTileSizes = 1;
157   for (unsigned i = 0, e = band.size(); i < e; i++) {
158     if (i < e - 1)
159       (*tileSizes)[i] = tSize;
160     else
161       // Set last tile size to cover the balance.
162       (*tileSizes)[i] = std::max(
163           1U, static_cast<unsigned>(excessFactor / cumulProductOfTileSizes));
164     cumulProductOfTileSizes *= (*tileSizes)[i];
165   }
166   if (avoidMaxMinBounds)
167     adjustToDivisorsOfTripCounts(band, tileSizes);
168 }
169 
170 void LoopTiling::runOnOperation() {
171   // Bands of loops to tile.
172   std::vector<SmallVector<AffineForOp, 6>> bands;
173   getTileableBands(getOperation(), &bands);
174 
175   // Tile each band.
176   for (auto &band : bands) {
177     // Set up tile sizes; fill missing tile sizes at the end with default tile
178     // size or tileSize if one was provided.
179     SmallVector<unsigned, 6> tileSizes;
180     getTileSizes(band, &tileSizes);
181     if (llvm::DebugFlag) {
182       auto diag = band[0].emitRemark("using tile sizes [");
183       for (unsigned tSize : tileSizes)
184         diag << tSize << ' ';
185       diag << "]\n";
186     }
187     SmallVector<AffineForOp, 6> tiledNest;
188     if (failed(tilePerfectlyNested(band, tileSizes, &tiledNest))) {
189       // An empty band always succeeds.
190       assert(!band.empty() && "guaranteed to succeed on empty bands");
191       LLVM_DEBUG(band.front()->emitRemark("loop tiling failed!\n"));
192       continue;
193     }
194 
195     // Separate full and partial tiles.
196     if (separate) {
197       auto intraTileLoops =
198           MutableArrayRef<AffineForOp>(tiledNest).drop_front(band.size());
199       if (failed(separateFullTiles(intraTileLoops))) {
200         assert(!intraTileLoops.empty() &&
201                "guaranteed to succeed on empty bands");
202         LLVM_DEBUG(intraTileLoops.front()->emitRemark(
203             "separation post tiling failed!\n"));
204       }
205     }
206   }
207 }
208 
209 constexpr unsigned LoopTiling::kDefaultTileSize;
210