1 //===- Vectorization.cpp - Implementation of linalg Vectorization ---------===//
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 the linalg dialect Vectorization transformations.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "mlir/Dialect/Linalg/Analysis/DependenceAnalysis.h"
14 #include "mlir/Dialect/Linalg/IR/LinalgOps.h"
15 #include "mlir/Dialect/Linalg/Transforms/Transforms.h"
16 #include "mlir/Dialect/Linalg/Utils/Utils.h"
17 #include "mlir/Dialect/StandardOps/EDSC/Intrinsics.h"
18 #include "mlir/Dialect/Utils/StructuredOpsUtils.h"
19 #include "mlir/Dialect/Vector/EDSC/Intrinsics.h"
20 #include "mlir/Dialect/Vector/VectorOps.h"
21 #include "mlir/IR/AffineExpr.h"
22 #include "mlir/IR/Matchers.h"
23 #include "mlir/IR/PatternMatch.h"
24 #include "mlir/Pass/Pass.h"
25 #include "mlir/Support/LLVM.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/raw_ostream.h"
28 #include <type_traits>
29 
30 using namespace mlir;
31 using namespace mlir::edsc;
32 using namespace mlir::edsc::intrinsics;
33 using namespace mlir::linalg;
34 
35 using llvm::dbgs;
36 
37 #define DEBUG_TYPE "linalg-vectorization"
38 
39 static bool hasMultiplyAddBody(Region &r) {
40   if (!llvm::hasSingleElement(r))
41     return false;
42   if (!llvm::hasNItems(r.front().begin(), r.front().end(), 3))
43     return false;
44 
45   using mlir::matchers::m_Val;
46   auto a = m_Val(r.getArgument(0));
47   auto b = m_Val(r.getArgument(1));
48   auto c = m_Val(r.getArgument(2));
49   // TODO: Update this detection once we have  matcher support for specifying
50   // that any permutation of operands matches.
51   auto pattern1 = m_Op<linalg::YieldOp>(m_Op<AddFOp>(m_Op<MulFOp>(a, b), c));
52   auto pattern2 = m_Op<linalg::YieldOp>(m_Op<AddFOp>(c, m_Op<MulFOp>(a, b)));
53   auto pattern3 = m_Op<linalg::YieldOp>(m_Op<AddFOp>(m_Op<MulFOp>(b, a), c));
54   auto pattern4 = m_Op<linalg::YieldOp>(m_Op<AddFOp>(c, m_Op<MulFOp>(b, a)));
55   auto pattern5 = m_Op<linalg::YieldOp>(m_Op<AddIOp>(m_Op<MulIOp>(a, b), c));
56   auto pattern6 = m_Op<linalg::YieldOp>(m_Op<AddIOp>(c, m_Op<MulIOp>(a, b)));
57   auto pattern7 = m_Op<linalg::YieldOp>(m_Op<AddIOp>(m_Op<MulIOp>(b, a), c));
58   auto pattern8 = m_Op<linalg::YieldOp>(m_Op<AddIOp>(c, m_Op<MulIOp>(b, a)));
59   return pattern1.match(&r.front().back()) ||
60          pattern2.match(&r.front().back()) ||
61          pattern3.match(&r.front().back()) ||
62          pattern4.match(&r.front().back()) ||
63          pattern5.match(&r.front().back()) ||
64          pattern6.match(&r.front().back()) ||
65          pattern7.match(&r.front().back()) || pattern8.match(&r.front().back());
66 }
67 
68 // TODO: Should be Tablegen'd from a single source that generates the op itself.
69 static LogicalResult isContraction(Operation *op) {
70   // TODO: interface for named ops.
71   if (isa<linalg::BatchMatmulOp, linalg::MatmulOp, linalg::MatvecOp,
72           linalg::VecmatOp, linalg::DotOp>(op))
73     return success();
74 
75   auto genericOp = dyn_cast<linalg::GenericOp>(op);
76   if (!genericOp)
77     return failure();
78 
79   auto mapRange = genericOp.indexing_maps().getAsValueRange<AffineMapAttr>();
80   return success(
81       genericOp.getNumInputs() == 2 && genericOp.getNumOutputs() == 1 &&
82       llvm::all_of(mapRange,
83                    [](AffineMap m) { return m.isProjectedPermutation(); }) &&
84       hasMultiplyAddBody(genericOp.region()));
85 }
86 
87 LogicalResult mlir::linalg::vectorizeLinalgOpPrecondition(Operation *op) {
88   auto linalgOp = cast<linalg::LinalgOp>(op);
89   // All types must be static shape to go to vector.
90   for (Value operand : linalgOp.getInputsAndOutputBuffers())
91     if (!operand.getType().cast<ShapedType>().hasStaticShape())
92       return failure();
93   for (Type outputTensorType : linalgOp.getOutputTensorTypes())
94     if (!outputTensorType.cast<ShapedType>().hasStaticShape())
95       return failure();
96 
97   if (isa<linalg::FillOp, linalg::CopyOp>(op))
98     return success();
99 
100   return isContraction(op);
101 }
102 
103 void mlir::linalg::vectorizeLinalgOp(OpBuilder &builder, Operation *op) {
104   assert(succeeded(vectorizeLinalgOpPrecondition(op)));
105 
106   StringRef dbgPref = "\n[" DEBUG_TYPE "]: ";
107   (void)dbgPref;
108   edsc::ScopedContext scope(builder, op->getLoc());
109   // In the case of 0-D memrefs, return null and special case to scalar load or
110   // store later.
111   auto extractVectorTypeFromScalarView = [](Value v) {
112     MemRefType mt = v.getType().cast<MemRefType>();
113     return mt.getShape().empty()
114                ? VectorType()
115                : VectorType::get(mt.getShape(), mt.getElementType());
116   };
117   if (auto fillOp = dyn_cast<linalg::FillOp>(op)) {
118     // Vectorize fill as a vector.broadcast.
119     LLVM_DEBUG(dbgs() << dbgPref
120                       << "Rewrite linalg.fill as vector.broadcast: " << *op);
121     Value viewOutput = fillOp.output();
122     if (VectorType outputType = extractVectorTypeFromScalarView(viewOutput)) {
123       auto vecType =
124           VectorType::get(fillOp.getOutputBufferType(0).getShape(),
125                           fillOp.getOutputBufferType(0).getElementType());
126       Value vector = vector_broadcast(vecType, fillOp.value());
127       Value zero = std_constant_index(0);
128       SmallVector<Value, 4> indicesOutput(outputType.getRank(), zero);
129       vector_transfer_write(vector, viewOutput, indicesOutput);
130     } else {
131       std_store(fillOp.value(), viewOutput);
132     }
133     return;
134   }
135   if (auto copyOp = dyn_cast<linalg::CopyOp>(op)) {
136     // Vectorize copy as a vector.transfer_read+vector.transfer_write.
137     LLVM_DEBUG(dbgs() << dbgPref
138                       << "Rewrite linalg.copy as vector.transfer_read + "
139                          "vector.transfer_write: "
140                       << *op);
141     Value zero = std_constant_index(0);
142     Value viewInput = copyOp.input();
143     Value viewOutput = copyOp.output();
144     Value vector;
145     if (VectorType inputType = extractVectorTypeFromScalarView(viewInput)) {
146       SmallVector<Value, 4> indicesInput(inputType.getRank(), zero);
147       if (copyOp.inputPermutation())
148         vector = vector_transfer_read(
149             extractVectorTypeFromScalarView(viewInput), viewInput, indicesInput,
150             copyOp.inputPermutation().getValue());
151       else
152         vector =
153             vector_transfer_read(extractVectorTypeFromScalarView(viewInput),
154                                  viewInput, indicesInput);
155     } else {
156       vector = std_load(viewInput).value;
157     }
158     if (VectorType outputType = extractVectorTypeFromScalarView(viewOutput)) {
159       SmallVector<Value, 4> indicesOutput(outputType.getRank(), zero);
160       if (copyOp.outputPermutation())
161         vector_transfer_write(vector, viewOutput, indicesOutput,
162                               copyOp.outputPermutation().getValue());
163       else
164         vector_transfer_write(vector, viewOutput, indicesOutput);
165     } else {
166       std_store(vector, viewOutput);
167     }
168     return;
169   }
170 
171   assert(succeeded(isContraction(op)) && "Expected contraction");
172 
173   // Vectorize other ops as vector contraction.
174   // TODO: interface.
175   LLVM_DEBUG(dbgs() << dbgPref
176                     << "Rewrite linalg op as vector.contract: " << *op);
177   auto linalgOp = cast<linalg::LinalgOp>(op);
178   Value viewA = linalgOp.getInput(0);
179   Value viewB = linalgOp.getInput(1);
180   Value viewC = linalgOp.getOutputBuffer(0);
181   VectorType vtA = extractVectorTypeFromScalarView(viewA);
182   VectorType vtB = extractVectorTypeFromScalarView(viewB);
183   VectorType vtC = extractVectorTypeFromScalarView(viewC);
184   Value zero = std_constant_index(0);
185   SmallVector<Value, 4> indicesA, indicesB, indicesC;
186   if (vtA)
187     indicesA = SmallVector<Value, 4>(vtA.getRank(), zero);
188   if (vtB)
189     indicesB = SmallVector<Value, 4>(vtB.getRank(), zero);
190   if (vtC)
191     indicesC = SmallVector<Value, 4>(vtC.getRank(), zero);
192   Value a = vtA ? vector_transfer_read(vtA, viewA, indicesA).value
193                 : std_load(viewA, indicesA).value;
194   Value b = vtB ? vector_transfer_read(vtB, viewB, indicesB).value
195                 : std_load(viewB, indicesB).value;
196   Value c = vtC ? vector_transfer_read(vtC, viewC, indicesC).value
197                 : std_load(viewC, indicesC).value;
198   Value res = vector_contract(a, b, c, linalgOp.indexing_maps(),
199                               linalgOp.iterator_types());
200   if (vtC)
201     vector_transfer_write(res, viewC, indicesC);
202   else
203     std_store(res, viewC, indicesC);
204 }
205 
206 /// Check whether there is any interleaved use of any `values` between `firstOp`
207 /// and `secondOp`. Conservatively return `true` if any op or value is in a
208 /// different block.
209 static bool mayExistInterleavedUses(Operation *firstOp, Operation *secondOp,
210                                     ValueRange values) {
211   StringRef dbgPref = "\n[" DEBUG_TYPE "]: ";
212   (void)dbgPref;
213   if (firstOp->getBlock() != secondOp->getBlock() ||
214       !firstOp->isBeforeInBlock(secondOp)) {
215     LLVM_DEBUG(llvm::dbgs()
216                << dbgPref << "interleavedUses precondition failed, firstOp: "
217                << *firstOp << ", second op: " << *secondOp);
218     return true;
219   }
220   for (auto v : values) {
221     for (auto &u : v.getUses()) {
222       Operation *owner = u.getOwner();
223       if (owner == firstOp || owner == secondOp)
224         continue;
225       // TODO: this is too conservative, use dominance info in the future.
226       if (owner->getBlock() == firstOp->getBlock() &&
227           (owner->isBeforeInBlock(firstOp) || secondOp->isBeforeInBlock(owner)))
228         continue;
229       LLVM_DEBUG(llvm::dbgs()
230                  << dbgPref << " found interleaved op " << *owner
231                  << ", firstOp: " << *firstOp << ", second op: " << *secondOp);
232       return true;
233     }
234   }
235   return false;
236 }
237 
238 /// Return the unique subview use of `v` if it is indeed unique, null otherwise.
239 static SubViewOp getSubViewUseIfUnique(Value v) {
240   SubViewOp subViewOp;
241   for (auto &u : v.getUses()) {
242     if (auto newSubViewOp = dyn_cast<SubViewOp>(u.getOwner())) {
243       if (subViewOp)
244         return SubViewOp();
245       subViewOp = newSubViewOp;
246     }
247   }
248   return subViewOp;
249 }
250 
251 /// TODO: use interfaces, side-effects and aliasing analysis as appropriate,
252 /// when available.
253 LogicalResult LinalgCopyVTRForwardingPattern::matchAndRewrite(
254     vector::TransferReadOp xferOp, PatternRewriter &rewriter) const {
255 
256   // Transfer into `view`.
257   Value viewOrAlloc = xferOp.memref();
258   if (!viewOrAlloc.getDefiningOp<ViewOp>() &&
259       !viewOrAlloc.getDefiningOp<AllocOp>())
260     return failure();
261 
262   StringRef dbgPref = "\n[" DEBUG_TYPE "]: VTRForwarding: ";
263   (void)dbgPref;
264   LLVM_DEBUG(llvm::dbgs() << dbgPref << viewOrAlloc);
265 
266   // Ensure there is exactly one subview of `viewOrAlloc` defining `subView`.
267   SubViewOp subViewOp = getSubViewUseIfUnique(viewOrAlloc);
268   if (!subViewOp)
269     return failure();
270   Value subView = subViewOp.getResult();
271   LLVM_DEBUG(llvm::dbgs() << dbgPref << "with subView " << subView);
272 
273   // Find the copy into `subView` without interleaved uses.
274   CopyOp copyOp;
275   for (auto &u : subView.getUses()) {
276     if (auto newCopyOp = dyn_cast<CopyOp>(u.getOwner())) {
277       if (newCopyOp.getOutputBuffer(0) != subView)
278         continue;
279       LLVM_DEBUG(llvm::dbgs() << dbgPref << "copy candidate " << *newCopyOp);
280       if (mayExistInterleavedUses(newCopyOp, xferOp, {viewOrAlloc, subView}))
281         continue;
282       copyOp = newCopyOp;
283       break;
284     }
285   }
286   if (!copyOp)
287     return failure();
288   LLVM_DEBUG(llvm::dbgs() << dbgPref << "with copy " << *copyOp);
289 
290   // Find the fill into `viewOrAlloc` without interleaved uses before the copy.
291   FillOp maybeFillOp;
292   for (auto &u : viewOrAlloc.getUses()) {
293     if (auto newFillOp = dyn_cast<FillOp>(u.getOwner())) {
294       if (newFillOp.getOutputBuffer(0) != viewOrAlloc)
295         continue;
296       LLVM_DEBUG(llvm::dbgs() << dbgPref << "fill candidate " << *newFillOp);
297       if (mayExistInterleavedUses(newFillOp, copyOp, {viewOrAlloc, subView}))
298         continue;
299       maybeFillOp = newFillOp;
300       break;
301     }
302   }
303   // Ensure padding matches.
304   if (maybeFillOp && xferOp.padding() != maybeFillOp.value())
305     return failure();
306   if (maybeFillOp)
307     LLVM_DEBUG(llvm::dbgs() << dbgPref << "with maybeFillOp " << *maybeFillOp);
308 
309   // `in` is the subview that linalg.copy reads. Replace it.
310   Value in = copyOp.getInput(0);
311 
312   // linalg.copy + linalg.fill can be used to create a padded local buffer.
313   // The `masked` attribute is only valid on this padded buffer.
314   // When forwarding to vector.transfer_read, the attribute must be reset
315   // conservatively.
316   Value res = rewriter.create<vector::TransferReadOp>(
317       xferOp.getLoc(), xferOp.getVectorType(), in, xferOp.indices(),
318       xferOp.permutation_map(), xferOp.padding(), ArrayAttr());
319 
320   if (maybeFillOp)
321     rewriter.eraseOp(maybeFillOp);
322   rewriter.eraseOp(copyOp);
323   rewriter.replaceOp(xferOp, res);
324 
325   return success();
326 }
327 
328 /// TODO: use interfaces, side-effects and aliasing analysis as appropriate,
329 /// when available.
330 LogicalResult LinalgCopyVTWForwardingPattern::matchAndRewrite(
331     vector::TransferWriteOp xferOp, PatternRewriter &rewriter) const {
332   // Transfer into `viewOrAlloc`.
333   Value viewOrAlloc = xferOp.memref();
334   if (!viewOrAlloc.getDefiningOp<ViewOp>() &&
335       !viewOrAlloc.getDefiningOp<AllocOp>())
336     return failure();
337 
338   // Ensure there is exactly one subview of `viewOrAlloc` defining `subView`.
339   SubViewOp subViewOp = getSubViewUseIfUnique(viewOrAlloc);
340   if (!subViewOp)
341     return failure();
342   Value subView = subViewOp.getResult();
343 
344   // Find the copy from `subView` without interleaved uses.
345   CopyOp copyOp;
346   for (auto &u : subViewOp.getResult().getUses()) {
347     if (auto newCopyOp = dyn_cast<CopyOp>(u.getOwner())) {
348       if (newCopyOp.getInput(0) != subView)
349         continue;
350       if (mayExistInterleavedUses(xferOp, newCopyOp, {viewOrAlloc, subView}))
351         continue;
352       copyOp = newCopyOp;
353       break;
354     }
355   }
356   if (!copyOp)
357     return failure();
358 
359   // `out` is the subview copied into that we replace.
360   Value out = copyOp.getOutputBuffer(0);
361 
362   // Forward vector.transfer into copy.
363   // linalg.copy + linalg.fill can be used to create a padded local buffer.
364   // The `masked` attribute is only valid on this padded buffer.
365   // When forwarding to vector.transfer_write, the attribute must be reset
366   // conservatively.
367   rewriter.create<vector::TransferWriteOp>(
368       xferOp.getLoc(), xferOp.vector(), out, xferOp.indices(),
369       xferOp.permutation_map(), ArrayAttr());
370 
371   rewriter.eraseOp(copyOp);
372   rewriter.eraseOp(xferOp);
373 
374   return success();
375 }
376 
377 template <class ConvOp, int N>
378 LogicalResult ConvOpVectorization<ConvOp, N>::matchAndRewrite(
379     ConvOp op, PatternRewriter &rewriter) const {
380   Location loc = op.getLoc();
381   MLIRContext *context = op.getContext();
382   edsc::ScopedContext scope(rewriter, loc);
383 
384   ShapedType inShapeType = op.getInputShapedType(0);
385   ShapedType kShapeType = op.getInputShapedType(1);
386 
387   ArrayRef<int64_t> inShape = inShapeType.getShape();
388   ArrayRef<int64_t> kShape = kShapeType.getShape();
389 
390   if (!inShapeType.hasStaticShape() || !kShapeType.hasStaticShape())
391     return failure();
392 
393   SmallVector<AffineExpr, 4> mapping;
394   SmallVector<int64_t, 4> vectorDims;
395   // Fail to apply when the size of not vectorized dimension is not 1.
396   for (unsigned i = 0; i < N; i++) {
397     if (!mask[i] && (inShape[i] != 1 || kShape[i] != 1))
398       return failure();
399 
400     if (mask[i] && inShape[i] != kShape[i])
401       return failure();
402 
403     if (mask[i]) {
404       mapping.push_back(getAffineDimExpr(i, context));
405       vectorDims.push_back(inShape[i]);
406     }
407   }
408 
409   Value input = op.getInput(0);
410   Value kernel = op.getInput(1);
411   Value output = op.getOutputBuffer(0);
412 
413   unsigned rank = inShapeType.getRank();
414   unsigned numDims = mapping.size();
415   Type elemType = inShapeType.getElementType();
416 
417   auto map = AffineMap::get(rank, 0, mapping, context);
418   SmallVector<Value, 4> zeros(rank, std_constant_index(0));
419   auto vecType = VectorType::get(vectorDims, elemType);
420 
421   auto inputVec = vector_transfer_read(vecType, input, zeros, map);
422   auto kernelVec = vector_transfer_read(vecType, kernel, zeros, map);
423 
424   auto acc = std_constant(elemType, rewriter.getZeroAttr(elemType));
425 
426   std::array<AffineMap, 3> indexingMaps{
427       AffineMap::getMultiDimIdentityMap(numDims, context),
428       AffineMap::getMultiDimIdentityMap(numDims, context),
429       AffineMap::get(numDims, 0, {}, context)};
430 
431   std::vector<StringRef> iteratorTypes(numDims, "reduction");
432 
433   auto result = rewriter.create<vector::ContractionOp>(
434       loc, inputVec, kernelVec, acc,
435       rewriter.getAffineMapArrayAttr(indexingMaps),
436       rewriter.getStrArrayAttr(iteratorTypes));
437 
438   rewriter.create<StoreOp>(loc, result, output, ValueRange(zeros));
439   rewriter.eraseOp(op);
440   return success();
441 }
442 
443 using ConvOpConst = ConvOpVectorization<ConvWOp, 1>;
444 
445 /// Inserts tiling, promotion and vectorization pattern for ConvOp
446 /// conversion into corresponding pattern lists.
447 template <typename ConvOp, unsigned N>
448 static void
449 populateVectorizationPatterns(OwningRewritePatternList &tilingPatterns,
450                               OwningRewritePatternList &promotionPatterns,
451                               OwningRewritePatternList &vectorizationPatterns,
452                               ArrayRef<int64_t> tileSizes,
453                               MLIRContext *context) {
454   if (tileSizes.size() < N)
455     return;
456 
457   constexpr static StringRef kTiledMarker = "TILED";
458   constexpr static StringRef kPromotedMarker = "PROMOTED";
459   tilingPatterns.insert<LinalgTilingPattern<ConvOp>>(
460       context, LinalgTilingOptions().setTileSizes(tileSizes),
461       LinalgMarker({}, Identifier::get(kTiledMarker, context)));
462 
463   promotionPatterns.insert<LinalgPromotionPattern<ConvOp>>(
464       context, LinalgPromotionOptions().setUseFullTileBuffersByDefault(true),
465       LinalgMarker(Identifier::get(kTiledMarker, context),
466                    Identifier::get(kPromotedMarker, context)));
467 
468   SmallVector<bool, 4> mask(N);
469   int offset = tileSizes.size() - N;
470   std::transform(tileSizes.begin() + offset, tileSizes.end(), mask.begin(),
471                  [](int64_t i) -> bool { return i > 1; });
472 
473   vectorizationPatterns.insert<ConvOpVectorization<ConvOp, N>>(context, mask);
474 }
475 
476 void mlir::linalg::populateConvVectorizationPatterns(
477     MLIRContext *context, SmallVectorImpl<OwningRewritePatternList> &patterns,
478     ArrayRef<int64_t> tileSizes) {
479   OwningRewritePatternList tiling, promotion, vectorization;
480   populateVectorizationPatterns<ConvWOp, 1>(tiling, promotion, vectorization,
481                                             tileSizes, context);
482 
483   populateVectorizationPatterns<ConvNWCOp, 3>(tiling, promotion, vectorization,
484                                               tileSizes, context);
485 
486   populateVectorizationPatterns<ConvNCWOp, 3>(tiling, promotion, vectorization,
487                                               tileSizes, context);
488 
489   populateVectorizationPatterns<ConvHWOp, 2>(tiling, promotion, vectorization,
490                                              tileSizes, context);
491 
492   populateVectorizationPatterns<ConvNHWCOp, 4>(tiling, promotion, vectorization,
493                                                tileSizes, context);
494 
495   populateVectorizationPatterns<ConvNCHWOp, 4>(tiling, promotion, vectorization,
496                                                tileSizes, context);
497 
498   populateVectorizationPatterns<ConvDHWOp, 3>(tiling, promotion, vectorization,
499                                               tileSizes, context);
500 
501   populateVectorizationPatterns<ConvNDHWCOp, 5>(
502       tiling, promotion, vectorization, tileSizes, context);
503 
504   populateVectorizationPatterns<ConvNCDHWOp, 5>(
505       tiling, promotion, vectorization, tileSizes, context);
506 
507   patterns.push_back(std::move(tiling));
508   patterns.push_back(std::move(promotion));
509   patterns.push_back(std::move(vectorization));
510 }
511