xref: /llvm-project/mlir/lib/Dialect/Affine/Transforms/AffineDataCopyGeneration.cpp (revision 039b969b32b64b64123dce30dd28ec4e343d893f)
1 //===- AffineDataCopyGeneration.cpp - Explicit memref copying 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 automatically promote accessed memref regions
10 // to buffers in a faster memory space that is explicitly managed, with the
11 // necessary data movement operations performed through either regular
12 // point-wise load/store's or DMAs. Such explicit copying (also referred to as
13 // array packing/unpacking in the literature), when done on arrays that exhibit
14 // reuse, results in near elimination of conflict misses, TLB misses, reduced
15 // use of hardware prefetch streams, and reduced false sharing. It is also
16 // necessary for hardware that explicitly managed levels in the memory
17 // hierarchy, and where DMAs may have to be used. This optimization is often
18 // performed on already tiled code.
19 //
20 //===----------------------------------------------------------------------===//
21 
22 #include "PassDetail.h"
23 #include "mlir/Dialect/Affine/Analysis/Utils.h"
24 #include "mlir/Dialect/Affine/IR/AffineOps.h"
25 #include "mlir/Dialect/Affine/LoopUtils.h"
26 #include "mlir/Dialect/Affine/Passes.h"
27 #include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
28 #include "mlir/Dialect/MemRef/IR/MemRef.h"
29 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
30 #include "llvm/ADT/MapVector.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include <algorithm>
34 
35 #define DEBUG_TYPE "affine-data-copy-generate"
36 
37 using namespace mlir;
38 
39 namespace {
40 
41 /// Replaces all loads and stores on memref's living in 'slowMemorySpace' by
42 /// introducing copy operations to transfer data into `fastMemorySpace` and
43 /// rewriting the original load's/store's to instead load/store from the
44 /// allocated fast memory buffers. Additional options specify the identifier
45 /// corresponding to the fast memory space and the amount of fast memory space
46 /// available. The pass traverses through the nesting structure, recursing to
47 /// inner levels if necessary to determine at what depth copies need to be
48 /// placed so that the allocated buffers fit within the memory capacity
49 /// provided.
50 // TODO: We currently can't generate copies correctly when stores
51 // are strided. Check for strided stores.
52 struct AffineDataCopyGeneration
53     : public AffineDataCopyGenerationBase<AffineDataCopyGeneration> {
54   AffineDataCopyGeneration() = default;
55   explicit AffineDataCopyGeneration(unsigned slowMemorySpace,
56                                     unsigned fastMemorySpace,
57                                     unsigned tagMemorySpace,
58                                     int minDmaTransferSize,
59                                     uint64_t fastMemCapacityBytes) {
60     this->slowMemorySpace = slowMemorySpace;
61     this->fastMemorySpace = fastMemorySpace;
62     this->tagMemorySpace = tagMemorySpace;
63     this->minDmaTransferSize = minDmaTransferSize;
64     this->fastMemoryCapacity = fastMemCapacityBytes / 1024;
65   }
66 
67   void runOnOperation() override;
68   void runOnBlock(Block *block, DenseSet<Operation *> &copyNests);
69 
70   // Constant zero index to avoid too many duplicates.
71   Value zeroIndex = nullptr;
72 };
73 
74 } // namespace
75 
76 /// Generates copies for memref's living in 'slowMemorySpace' into newly created
77 /// buffers in 'fastMemorySpace', and replaces memory operations to the former
78 /// by the latter. Only load op's handled for now.
79 /// TODO: extend this to store op's.
80 std::unique_ptr<OperationPass<func::FuncOp>>
81 mlir::createAffineDataCopyGenerationPass(unsigned slowMemorySpace,
82                                          unsigned fastMemorySpace,
83                                          unsigned tagMemorySpace,
84                                          int minDmaTransferSize,
85                                          uint64_t fastMemCapacityBytes) {
86   return std::make_unique<AffineDataCopyGeneration>(
87       slowMemorySpace, fastMemorySpace, tagMemorySpace, minDmaTransferSize,
88       fastMemCapacityBytes);
89 }
90 std::unique_ptr<OperationPass<func::FuncOp>>
91 mlir::createAffineDataCopyGenerationPass() {
92   return std::make_unique<AffineDataCopyGeneration>();
93 }
94 
95 /// Generate copies for this block. The block is partitioned into separate
96 /// ranges: each range is either a sequence of one or more operations starting
97 /// and ending with an affine load or store op, or just an affine.forop (which
98 /// could have other affine for op's nested within).
99 void AffineDataCopyGeneration::runOnBlock(Block *block,
100                                           DenseSet<Operation *> &copyNests) {
101   if (block->empty())
102     return;
103 
104   uint64_t fastMemCapacityBytes =
105       fastMemoryCapacity != std::numeric_limits<uint64_t>::max()
106           ? fastMemoryCapacity * 1024
107           : fastMemoryCapacity;
108   AffineCopyOptions copyOptions = {generateDma, slowMemorySpace,
109                                    fastMemorySpace, tagMemorySpace,
110                                    fastMemCapacityBytes};
111 
112   // Every affine.for op in the block starts and ends a block range for copying;
113   // in addition, a contiguous sequence of operations starting with a
114   // load/store op but not including any copy nests themselves is also
115   // identified as a copy block range. Straightline code (a contiguous chunk of
116   // operations excluding AffineForOp's) are always assumed to not exhaust
117   // memory. As a result, this approach is conservative in some cases at the
118   // moment; we do a check later and report an error with location info.
119   // TODO: An 'affine.if' operation is being treated similar to an
120   // operation. 'affine.if''s could have 'affine.for's in them;
121   // treat them separately.
122 
123   // Get to the first load, store, or for op (that is not a copy nest itself).
124   auto curBegin =
125       std::find_if(block->begin(), block->end(), [&](Operation &op) {
126         return isa<AffineLoadOp, AffineStoreOp, AffineForOp>(op) &&
127                copyNests.count(&op) == 0;
128       });
129 
130   // Create [begin, end) ranges.
131   auto it = curBegin;
132   while (it != block->end()) {
133     AffineForOp forOp;
134     // If you hit a non-copy for loop, we will split there.
135     if ((forOp = dyn_cast<AffineForOp>(&*it)) && copyNests.count(forOp) == 0) {
136       // Perform the copying up unti this 'for' op first.
137       (void)affineDataCopyGenerate(/*begin=*/curBegin, /*end=*/it, copyOptions,
138                                    /*filterMemRef=*/llvm::None, copyNests);
139 
140       // Returns true if the footprint is known to exceed capacity.
141       auto exceedsCapacity = [&](AffineForOp forOp) {
142         Optional<int64_t> footprint =
143             getMemoryFootprintBytes(forOp,
144                                     /*memorySpace=*/0);
145         return (footprint.has_value() &&
146                 static_cast<uint64_t>(footprint.value()) >
147                     fastMemCapacityBytes);
148       };
149 
150       // If the memory footprint of the 'affine.for' loop is higher than fast
151       // memory capacity (when provided), we recurse to copy at an inner level
152       // until we find a depth at which footprint fits in fast mem capacity. If
153       // the footprint can't be calculated, we assume for now it fits. Recurse
154       // inside if footprint for 'forOp' exceeds capacity, or when
155       // skipNonUnitStrideLoops is set and the step size is not one.
156       bool recurseInner = skipNonUnitStrideLoops ? forOp.getStep() != 1
157                                                  : exceedsCapacity(forOp);
158       if (recurseInner) {
159         // We'll recurse and do the copies at an inner level for 'forInst'.
160         // Recurse onto the body of this loop.
161         runOnBlock(forOp.getBody(), copyNests);
162       } else {
163         // We have enough capacity, i.e., copies will be computed for the
164         // portion of the block until 'it', and for 'it', which is 'forOp'. Note
165         // that for the latter, the copies are placed just before this loop (for
166         // incoming copies) and right after (for outgoing ones).
167 
168         // Inner loop copies have their own scope - we don't thus update
169         // consumed capacity. The footprint check above guarantees this inner
170         // loop's footprint fits.
171         (void)affineDataCopyGenerate(/*begin=*/it, /*end=*/std::next(it),
172                                      copyOptions,
173                                      /*filterMemRef=*/llvm::None, copyNests);
174       }
175       // Get to the next load or store op after 'forOp'.
176       curBegin = std::find_if(std::next(it), block->end(), [&](Operation &op) {
177         return isa<AffineLoadOp, AffineStoreOp, AffineForOp>(op) &&
178                copyNests.count(&op) == 0;
179       });
180       it = curBegin;
181     } else {
182       assert(copyNests.count(&*it) == 0 &&
183              "all copy nests generated should have been skipped above");
184       // We simply include this op in the current range and continue for more.
185       ++it;
186     }
187   }
188 
189   // Generate the copy for the final block range.
190   if (curBegin != block->end()) {
191     // Can't be a terminator because it would have been skipped above.
192     assert(!curBegin->hasTrait<OpTrait::IsTerminator>() &&
193            "can't be a terminator");
194     // Exclude the affine.yield - hence, the std::prev.
195     (void)affineDataCopyGenerate(/*begin=*/curBegin,
196                                  /*end=*/std::prev(block->end()), copyOptions,
197                                  /*filterMemRef=*/llvm::None, copyNests);
198   }
199 }
200 
201 void AffineDataCopyGeneration::runOnOperation() {
202   func::FuncOp f = getOperation();
203   OpBuilder topBuilder(f.getBody());
204   zeroIndex = topBuilder.create<arith::ConstantIndexOp>(f.getLoc(), 0);
205 
206   // Nests that are copy-in's or copy-out's; the root AffineForOps of those
207   // nests are stored herein.
208   DenseSet<Operation *> copyNests;
209 
210   // Clear recorded copy nests.
211   copyNests.clear();
212 
213   for (auto &block : f)
214     runOnBlock(&block, copyNests);
215 
216   // Promote any single iteration loops in the copy nests and collect
217   // load/stores to simplify.
218   SmallVector<Operation *, 4> copyOps;
219   for (Operation *nest : copyNests)
220     // With a post order walk, the erasure of loops does not affect
221     // continuation of the walk or the collection of load/store ops.
222     nest->walk([&](Operation *op) {
223       if (auto forOp = dyn_cast<AffineForOp>(op))
224         (void)promoteIfSingleIteration(forOp);
225       else if (isa<AffineLoadOp, AffineStoreOp>(op))
226         copyOps.push_back(op);
227     });
228 
229   // Promoting single iteration loops could lead to simplification of
230   // contained load's/store's, and the latter could anyway also be
231   // canonicalized.
232   RewritePatternSet patterns(&getContext());
233   AffineLoadOp::getCanonicalizationPatterns(patterns, &getContext());
234   AffineStoreOp::getCanonicalizationPatterns(patterns, &getContext());
235   FrozenRewritePatternSet frozenPatterns(std::move(patterns));
236   (void)applyOpPatternsAndFold(copyOps, frozenPatterns, /*strict=*/true);
237 }
238