xref: /llvm-project/mlir/lib/Dialect/Affine/Transforms/LoopTiling.cpp (revision 4c48f016effde67d500fc95290096aec9f3bdb70)
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/Builders.h"
25 #include "mlir/IR/IRMapping.h"
26 #include "llvm/Support/CommandLine.h"
27 #include "llvm/Support/Debug.h"
28 #include <optional>
29 
30 namespace mlir {
31 namespace affine {
32 #define GEN_PASS_DEF_AFFINELOOPTILING
33 #include "mlir/Dialect/Affine/Passes.h.inc"
34 } // namespace affine
35 } // namespace mlir
36 
37 using namespace mlir;
38 using namespace mlir::affine;
39 
40 #define DEBUG_TYPE "affine-loop-tile"
41 
42 namespace {
43 
44 /// A pass to perform loop tiling on all suitable loop nests of a Function.
45 struct LoopTiling : public affine::impl::AffineLoopTilingBase<LoopTiling> {
46   LoopTiling() = default;
47   explicit LoopTiling(uint64_t cacheSizeBytes, bool avoidMaxMinBounds = true)
48       : avoidMaxMinBounds(avoidMaxMinBounds) {
49     this->cacheSizeInKiB = cacheSizeBytes / 1024;
50   }
51 
52   void runOnOperation() override;
53   void getTileSizes(ArrayRef<AffineForOp> band,
54                     SmallVectorImpl<unsigned> *tileSizes);
55 
56   // Default tile size if nothing is provided.
57   constexpr static unsigned kDefaultTileSize = 4;
58 
59   // If true, tile sizes are set to avoid max/min in bounds if possible.
60   bool avoidMaxMinBounds = true;
61 };
62 
63 } // namespace
64 
65 /// Creates a pass to perform loop tiling on all suitable loop nests of a
66 /// Function.
67 std::unique_ptr<OperationPass<func::FuncOp>>
68 mlir::affine::createLoopTilingPass(uint64_t cacheSizeBytes) {
69   return std::make_unique<LoopTiling>(cacheSizeBytes);
70 }
71 std::unique_ptr<OperationPass<func::FuncOp>>
72 mlir::affine::createLoopTilingPass() {
73   return std::make_unique<LoopTiling>();
74 }
75 
76 /// Reduces each tile size to the largest divisor of the corresponding trip
77 /// count (if the trip count is known).
78 static void adjustToDivisorsOfTripCounts(ArrayRef<AffineForOp> band,
79                                          SmallVectorImpl<unsigned> *tileSizes) {
80   assert(band.size() == tileSizes->size() && "invalid tile size count");
81   for (unsigned i = 0, e = band.size(); i < e; i++) {
82     unsigned &tSizeAdjusted = (*tileSizes)[i];
83     std::optional<uint64_t> mayConst = getConstantTripCount(band[i]);
84     if (!mayConst)
85       continue;
86     // Adjust the tile size to largest factor of the trip count less than
87     // tSize.
88     uint64_t constTripCount = *mayConst;
89     if (constTripCount > 1 && tSizeAdjusted > constTripCount / 2)
90       tSizeAdjusted = constTripCount / 2;
91     while (constTripCount % tSizeAdjusted != 0)
92       tSizeAdjusted--;
93   }
94 }
95 
96 /// Checks whether hyper-rectangular loop tiling of the nest represented by
97 /// `origLoops` is valid. The validity condition is from Irigoin and Triolet,
98 /// which states that two tiles cannot depend on each other. We simplify such
99 /// condition to just checking whether there is any negative dependence
100 /// direction, since we have the prior knowledge that the tiling results will be
101 /// hyper-rectangles, which are scheduled in the lexicographically increasing
102 /// order on the vector of loop indices. This function will return failure when
103 /// any dependence component is negative along any of `origLoops`.
104 static bool checkTilingLegality(MutableArrayRef<AffineForOp> origLoops) {
105   assert(!origLoops.empty() && "no original loops provided");
106 
107   // We first find out all dependences we intend to check.
108   SmallVector<Operation *, 8> loadAndStoreOps;
109   origLoops[0]->walk([&](Operation *op) {
110     if (isa<AffineReadOpInterface, AffineWriteOpInterface>(op))
111       loadAndStoreOps.push_back(op);
112   });
113 
114   unsigned numOps = loadAndStoreOps.size();
115   unsigned numLoops = origLoops.size();
116   for (unsigned d = 1; d <= numLoops + 1; ++d) {
117     for (unsigned i = 0; i < numOps; ++i) {
118       Operation *srcOp = loadAndStoreOps[i];
119       MemRefAccess srcAccess(srcOp);
120       for (unsigned j = 0; j < numOps; ++j) {
121         Operation *dstOp = loadAndStoreOps[j];
122         MemRefAccess dstAccess(dstOp);
123 
124         SmallVector<DependenceComponent, 2> depComps;
125         DependenceResult result = checkMemrefAccessDependence(
126             srcAccess, dstAccess, d, /*dependenceConstraints=*/nullptr,
127             &depComps);
128 
129         // Skip if there is no dependence in this case.
130         if (!hasDependence(result))
131           continue;
132 
133         // Check whether there is any negative direction vector in the
134         // dependence components found above, which means that dependence is
135         // violated by the default hyper-rect tiling method.
136         LLVM_DEBUG(llvm::dbgs() << "Checking whether tiling legality violated "
137                                    "for dependence at depth: "
138                                 << Twine(d) << " between:\n";);
139         LLVM_DEBUG(srcAccess.opInst->dump(););
140         LLVM_DEBUG(dstAccess.opInst->dump(););
141         for (unsigned k = 0, e = depComps.size(); k < e; k++) {
142           DependenceComponent depComp = depComps[k];
143           if (depComp.lb.has_value() && depComp.ub.has_value() &&
144               *depComp.lb < *depComp.ub && *depComp.ub < 0) {
145             LLVM_DEBUG(llvm::dbgs()
146                        << "Dependence component lb = " << Twine(*depComp.lb)
147                        << " ub = " << Twine(*depComp.ub)
148                        << " is negative  at depth: " << Twine(d)
149                        << " and thus violates the legality rule.\n");
150             return false;
151           }
152         }
153       }
154     }
155   }
156 
157   return true;
158 }
159 
160 // Returns tile sizes to use. Checks CL options; if none are specified, sets it
161 // based on a simple model that looks at the memory footprint and determines
162 // tile sizes assuming identity accesses / 1:1 tile size proportional footprint
163 // along each of the dimensions being tiled.
164 // TODO: evolve this model. Tile size determination is a large area
165 // to play with in general.
166 void LoopTiling::getTileSizes(ArrayRef<AffineForOp> band,
167                               SmallVectorImpl<unsigned> *tileSizes) {
168   if (band.empty())
169     return;
170 
171   // Use command-line tileSize for all loops if specified.
172   if (tileSize) {
173     tileSizes->assign(band.size(), tileSize);
174     return;
175   }
176 
177   // Use tileSizes and fill them with default tile size if it's short.
178   if (!this->tileSizes.empty()) {
179     tileSizes->assign(this->tileSizes.begin(), this->tileSizes.end());
180     tileSizes->resize(band.size(), kDefaultTileSize);
181     return;
182   }
183   tileSizes->resize(band.size());
184 
185   // The first loop in the band.
186   AffineForOp rootForOp = band[0];
187   (void)rootForOp;
188 
189   // Obtain memory footprint and set tile sizes so that a tile fits in
190   // the cache size. This is an approximation with the assumption that the
191   // footprint increases with the tile size linearly in that dimension (i.e.,
192   // assumes one-to-one access function).
193   std::optional<int64_t> fp = getMemoryFootprintBytes(band[0], 0);
194   if (!fp) {
195     // Fill with default tile sizes if footprint is unknown.
196     std::fill(tileSizes->begin(), tileSizes->end(),
197               LoopTiling::kDefaultTileSize);
198     if (avoidMaxMinBounds)
199       adjustToDivisorsOfTripCounts(band, tileSizes);
200     LLVM_DEBUG(
201         rootForOp.emitWarning("memory footprint unknown: using default tile "
202                               "sizes adjusted to trip count divisors"));
203     return;
204   }
205 
206   // Check how many times larger the cache size is when compared to footprint.
207   uint64_t cacheSizeBytes = cacheSizeInKiB * 1024;
208   uint64_t excessFactor = llvm::divideCeil(*fp, cacheSizeBytes);
209   if (excessFactor <= 1) {
210     // No need of any tiling - set tile size to 1.
211     std::fill(tileSizes->begin(), tileSizes->end(), 1);
212     return;
213   }
214 
215   // Divide all loops equally in an attempt to reduce footprint.
216   // TODO: this is approximate. Ideally, obtain reuse factor /
217   // profitability along each dimension and weight tile sizes based on that as
218   // one possible approach. Or compute a polynomial in tile sizes and solve for
219   // it.
220 
221   // For an n-d tileable band, compute the n^th root of the excess.
222   unsigned tSize =
223       static_cast<unsigned>(floorl(std::pow(excessFactor, 1.0 / band.size())));
224   // We'll keep a running product to determine the last tile size better.
225   unsigned cumulProductOfTileSizes = 1;
226   for (unsigned i = 0, e = band.size(); i < e; i++) {
227     if (i < e - 1)
228       (*tileSizes)[i] = tSize;
229     else
230       // Set last tile size to cover the balance.
231       (*tileSizes)[i] = std::max(
232           1U, static_cast<unsigned>(excessFactor / cumulProductOfTileSizes));
233     cumulProductOfTileSizes *= (*tileSizes)[i];
234   }
235   if (avoidMaxMinBounds)
236     adjustToDivisorsOfTripCounts(band, tileSizes);
237 }
238 
239 void LoopTiling::runOnOperation() {
240   // Bands of loops to tile.
241   std::vector<SmallVector<AffineForOp, 6>> bands;
242   getTileableBands(getOperation(), &bands);
243 
244   // Tile each band.
245   for (auto &band : bands) {
246     if (!checkTilingLegality(band)) {
247       band.front().emitRemark("tiling code is illegal due to dependences");
248       continue;
249     }
250 
251     // Set up tile sizes; fill missing tile sizes at the end with default tile
252     // size or tileSize if one was provided.
253     SmallVector<unsigned, 6> tileSizes;
254     getTileSizes(band, &tileSizes);
255     if (llvm::DebugFlag) {
256       auto diag = band[0].emitRemark("using tile sizes [");
257       for (unsigned tSize : tileSizes)
258         diag << tSize << ' ';
259       diag << "]\n";
260     }
261     SmallVector<AffineForOp, 6> tiledNest;
262     if (failed(tilePerfectlyNested(band, tileSizes, &tiledNest))) {
263       // An empty band always succeeds.
264       assert(!band.empty() && "guaranteed to succeed on empty bands");
265       LLVM_DEBUG(band.front()->emitRemark("loop tiling failed!\n"));
266       continue;
267     }
268 
269     // Separate full and partial tiles.
270     if (separate) {
271       auto intraTileLoops =
272           MutableArrayRef<AffineForOp>(tiledNest).drop_front(band.size());
273       if (failed(separateFullTiles(intraTileLoops))) {
274         assert(!intraTileLoops.empty() &&
275                "guaranteed to succeed on empty bands");
276         LLVM_DEBUG(intraTileLoops.front()->emitRemark(
277             "separation post tiling failed!\n"));
278       }
279     }
280   }
281 }
282 
283 constexpr unsigned LoopTiling::kDefaultTileSize;
284