xref: /llvm-project/mlir/lib/Dialect/Linalg/Transforms/Promotion.cpp (revision cbb0981388b6b2c4ccc574a674841ecd517115e5)
1 //===- Promotion.cpp - Implementation of linalg Promotion -----------------===//
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 Promotion pass.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "mlir/Dialect/Arith/IR/Arith.h"
14 #include "mlir/Dialect/Arith/Utils/Utils.h"
15 #include "mlir/Dialect/Complex/IR/Complex.h"
16 #include "mlir/Dialect/Linalg/IR/Linalg.h"
17 #include "mlir/Dialect/Linalg/Passes.h"
18 #include "mlir/Dialect/Linalg/Transforms/Transforms.h"
19 #include "mlir/Dialect/SCF/IR/SCF.h"
20 #include "mlir/IR/AffineExpr.h"
21 #include "mlir/IR/AffineExprVisitor.h"
22 #include "mlir/IR/AffineMap.h"
23 #include "mlir/IR/ImplicitLocOpBuilder.h"
24 #include "mlir/Support/LLVM.h"
25 #include "mlir/Transforms/FoldUtils.h"
26 #include "llvm/ADT/MapVector.h"
27 #include "llvm/ADT/SmallBitVector.h"
28 #include "llvm/ADT/TypeSwitch.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/Debug.h"
31 
32 using namespace mlir;
33 using namespace mlir::linalg;
34 using namespace mlir::scf;
35 
36 using llvm::MapVector;
37 
38 #define DEBUG_TYPE "linalg-promotion"
39 
40 /// Alloc a new buffer of `size` * `width` i8; where `width` is given by the
41 /// data `layout` for `elementType`.
42 /// Use AllocOp or AllocaOp depending on `options`.
43 /// Take an optional alignment.
44 static Value allocBuffer(ImplicitLocOpBuilder &b,
45                          const LinalgPromotionOptions &options,
46                          Type elementType, Value allocSize, DataLayout &layout,
47                          std::optional<unsigned> alignment = std::nullopt) {
48   auto width = layout.getTypeSize(elementType);
49 
50   IntegerAttr alignmentAttr;
51   if (alignment.has_value())
52     alignmentAttr = b.getI64IntegerAttr(alignment.value());
53 
54   // Static buffer.
55   if (auto cst = allocSize.getDefiningOp<arith::ConstantIndexOp>()) {
56     auto staticBufferType =
57         MemRefType::get(width * cst.value(), b.getIntegerType(8));
58     if (options.useAlloca) {
59       return b.createOrFold<memref::AllocaOp>(staticBufferType, ValueRange{},
60                                               alignmentAttr);
61     }
62     return b.createOrFold<memref::AllocOp>(staticBufferType, ValueRange{},
63                                            alignmentAttr);
64   }
65 
66   // Fallback dynamic buffer.
67   auto dynamicBufferType =
68       MemRefType::get(ShapedType::kDynamic, b.getIntegerType(8));
69   Value mul = b.createOrFold<arith::MulIOp>(
70       b.create<arith::ConstantIndexOp>(width), allocSize);
71   if (options.useAlloca)
72     return b.create<memref::AllocaOp>(dynamicBufferType, mul, alignmentAttr);
73   return b.create<memref::AllocOp>(dynamicBufferType, mul, alignmentAttr);
74 }
75 
76 /// Default allocation callback function. This allocates a promoted buffer when
77 /// no call back to do so is provided. The default is to allocate a
78 /// memref<..xi8> and return a view to get a memref type of shape
79 /// boundingSubViewSize.
80 static std::optional<Value> defaultAllocBufferCallBack(
81     const LinalgPromotionOptions &options, OpBuilder &builder,
82     memref::SubViewOp subView, ArrayRef<Value> boundingSubViewSize,
83     std::optional<unsigned> alignment, DataLayout &layout) {
84   ShapedType viewType = subView.getType();
85   ImplicitLocOpBuilder b(subView.getLoc(), builder);
86   auto zero = b.createOrFold<arith::ConstantIndexOp>(0);
87   auto one = b.createOrFold<arith::ConstantIndexOp>(1);
88 
89   Value allocSize = one;
90   for (const auto &size : llvm::enumerate(boundingSubViewSize))
91     allocSize = b.createOrFold<arith::MulIOp>(allocSize, size.value());
92   Value buffer = allocBuffer(b, options, viewType.getElementType(), allocSize,
93                              layout, alignment);
94   SmallVector<int64_t, 4> dynSizes(boundingSubViewSize.size(),
95                                    ShapedType::kDynamic);
96   Value view = b.createOrFold<memref::ViewOp>(
97       MemRefType::get(dynSizes, viewType.getElementType()), buffer, zero,
98       boundingSubViewSize);
99   return view;
100 }
101 
102 /// Default implementation of deallocation of the buffer use for promotion. It
103 /// expects to get the same value that the default allocation method returned,
104 /// i.e. result of a ViewOp.
105 static LogicalResult
106 defaultDeallocBufferCallBack(const LinalgPromotionOptions &options,
107                              OpBuilder &b, Value fullLocalView) {
108   if (!options.useAlloca) {
109     auto viewOp = cast<memref::ViewOp>(fullLocalView.getDefiningOp());
110     b.create<memref::DeallocOp>(viewOp.getSource().getLoc(),
111                                 viewOp.getSource());
112   }
113   return success();
114 }
115 
116 namespace {
117 
118 /// Helper struct that captures the information required to apply the
119 /// transformation on each op. This bridges the abstraction gap with the
120 /// user-facing API which exposes positional arguments to control which operands
121 /// are promoted.
122 struct LinalgOpInstancePromotionOptions {
123   LinalgOpInstancePromotionOptions(LinalgOp op,
124                                    const LinalgPromotionOptions &options);
125   /// SubViews to promote.
126   MapVector<int64_t, Value> subViews;
127   /// True if the full view should be used for the promoted buffer.
128   DenseMap<Value, bool> useFullTileBuffers;
129 
130   /// Callback functions for allocation and deallocation of promoted buffers, as
131   /// well as to copy the data into and out of these buffers.
132   AllocBufferCallbackFn allocationFn;
133   DeallocBufferCallbackFn deallocationFn;
134   CopyCallbackFn copyInFn;
135   CopyCallbackFn copyOutFn;
136 
137   /// Alignment of promoted buffer.
138   std::optional<unsigned> alignment;
139 };
140 } // namespace
141 
142 LinalgOpInstancePromotionOptions::LinalgOpInstancePromotionOptions(
143     LinalgOp linalgOp, const LinalgPromotionOptions &options)
144     : subViews(), alignment(options.alignment) {
145   assert(linalgOp.hasBufferSemantics() && "revisit usage of shaped operand");
146   auto vUseFullTileBuffers =
147       options.useFullTileBuffers.value_or(llvm::SmallBitVector());
148   vUseFullTileBuffers.resize(linalgOp->getNumOperands(),
149                              options.useFullTileBuffersDefault);
150 
151   for (OpOperand &opOperand : linalgOp->getOpOperands()) {
152     int64_t operandNumber = opOperand.getOperandNumber();
153     if (options.operandsToPromote &&
154         !options.operandsToPromote->count(operandNumber))
155       continue;
156     Operation *op = opOperand.get().getDefiningOp();
157     if (auto sv = dyn_cast_or_null<memref::SubViewOp>(op)) {
158       subViews[operandNumber] = sv;
159       useFullTileBuffers[sv] = vUseFullTileBuffers[operandNumber];
160     }
161   }
162 
163   if (options.allocationFn) {
164     allocationFn = *options.allocationFn;
165   } else {
166     allocationFn = [&](OpBuilder &b, memref::SubViewOp subViewOp,
167                        ArrayRef<Value> boundingSubViewSize,
168                        DataLayout &layout) -> std::optional<Value> {
169       return defaultAllocBufferCallBack(options, b, subViewOp,
170                                         boundingSubViewSize, alignment, layout);
171     };
172   }
173 
174   if (options.deallocationFn) {
175     deallocationFn = *options.deallocationFn;
176   } else {
177     deallocationFn = [&](OpBuilder &b, Value buffer) {
178       return defaultDeallocBufferCallBack(options, b, buffer);
179     };
180   }
181 
182   // Save the loc because `linalgOp` goes out of scope.
183   Location loc = linalgOp.getLoc();
184   auto defaultCopyCallBack = [loc](OpBuilder &b, Value src,
185                                    Value dst) -> LogicalResult {
186     b.create<memref::CopyOp>(loc, src, dst);
187     return success();
188   };
189   copyInFn = (options.copyInFn ? *(options.copyInFn) : defaultCopyCallBack);
190   copyOutFn = (options.copyOutFn ? *(options.copyOutFn) : defaultCopyCallBack);
191 }
192 
193 // Performs promotion of a `subView` into a local buffer of the size of the
194 // *ranges* of the `subView`. This produces a buffer whose size may be bigger
195 // than the actual size of the `subView` at the boundaries.
196 // This is related to the full/partial tile problem.
197 // Returns a PromotionInfo containing a `buffer`, `fullLocalView` and
198 // `partialLocalView` such that:
199 //   * `buffer` is always the size of the full tile.
200 //   * `fullLocalView` is a dense contiguous view into that buffer.
201 //   * `partialLocalView` is a dense non-contiguous slice of `fullLocalView`
202 //     that corresponds to the size of `subView` and accounting for boundary
203 //     effects.
204 // The point of the full tile buffer is that constant static tile sizes are
205 // folded and result in a buffer type with statically known size and alignment
206 // properties.
207 // To account for general boundary effects, padding must be performed on the
208 // boundary tiles. For now this is done with an unconditional `fill` op followed
209 // by a partial `copy` op.
210 FailureOr<PromotionInfo> mlir::linalg::promoteSubviewAsNewBuffer(
211     OpBuilder &b, Location loc, memref::SubViewOp subView,
212     const AllocBufferCallbackFn &allocationFn, DataLayout &layout) {
213   auto viewType = subView.getType();
214   auto rank = viewType.getRank();
215   SmallVector<Value, 4> fullSizes;
216   SmallVector<OpFoldResult> partialSizes;
217   fullSizes.reserve(rank);
218   partialSizes.reserve(rank);
219   llvm::SmallBitVector droppedDims = subView.getDroppedDims();
220   int64_t resultDimIdx = 0;
221   for (const auto &en : llvm::enumerate(subView.getOrCreateRanges(b, loc))) {
222     if (droppedDims[en.index()])
223       continue;
224     auto rangeValue = en.value();
225     // Try to extract a tight constant. If the size is known statically, no need
226     // to look for the bound.
227     LLVM_DEBUG(llvm::dbgs() << "Extract tightest: " << rangeValue.size << "\n");
228     Value size;
229     if (auto attr = rangeValue.size.dyn_cast<Attribute>()) {
230       size = getValueOrCreateConstantIndexOp(b, loc, rangeValue.size);
231     } else {
232       Value materializedSize =
233           getValueOrCreateConstantIndexOp(b, loc, rangeValue.size);
234       FailureOr<int64_t> upperBound =
235           getConstantUpperBoundForIndex(materializedSize);
236       size = failed(upperBound)
237                  ? materializedSize
238                  : b.create<arith::ConstantIndexOp>(loc, *upperBound);
239     }
240     LLVM_DEBUG(llvm::dbgs() << "Extracted tightest: " << size << "\n");
241     fullSizes.push_back(size);
242     partialSizes.push_back(
243         b.createOrFold<memref::DimOp>(loc, subView, resultDimIdx++));
244   }
245   SmallVector<int64_t, 4> dynSizes(fullSizes.size(), ShapedType::kDynamic);
246   // If a callback is not specified, then use the default implementation for
247   // allocating the promoted buffer.
248   std::optional<Value> fullLocalView =
249       allocationFn(b, subView, fullSizes, layout);
250   if (!fullLocalView)
251     return failure();
252   SmallVector<OpFoldResult, 4> zeros(fullSizes.size(), b.getIndexAttr(0));
253   SmallVector<OpFoldResult, 4> ones(fullSizes.size(), b.getIndexAttr(1));
254   auto partialLocalView = b.createOrFold<memref::SubViewOp>(
255       loc, *fullLocalView, zeros, partialSizes, ones);
256   return PromotionInfo{*fullLocalView, partialLocalView};
257 }
258 
259 static FailureOr<MapVector<int64_t, PromotionInfo>>
260 promoteSubViews(ImplicitLocOpBuilder &b,
261                 LinalgOpInstancePromotionOptions options, DataLayout &layout) {
262   if (options.subViews.empty())
263     return failure();
264 
265   MapVector<int64_t, PromotionInfo> promotionInfoMap;
266 
267   for (auto v : options.subViews) {
268     memref::SubViewOp subView =
269         cast<memref::SubViewOp>(v.second.getDefiningOp());
270     auto promotionInfo = promoteSubviewAsNewBuffer(
271         b, b.getLoc(), subView, options.allocationFn, layout);
272     if (failed(promotionInfo))
273       return failure();
274     promotionInfoMap[v.first] = *promotionInfo;
275 
276     // Only fill the buffer if the full local view is used
277     if (!options.useFullTileBuffers[v.second])
278       continue;
279     Type subviewEltType = subView.getType().getElementType();
280     Value fillVal =
281         llvm::TypeSwitch<Type, Value>(subviewEltType)
282             .Case([&](FloatType t) {
283               return b.create<arith::ConstantOp>(FloatAttr::get(t, 0.0));
284             })
285             .Case([&](IntegerType t) {
286               return b.create<arith::ConstantOp>(IntegerAttr::get(t, 0));
287             })
288             .Case([&](ComplexType t) {
289               Value tmp;
290               if (auto et = t.getElementType().dyn_cast<FloatType>())
291                 tmp = b.create<arith::ConstantOp>(FloatAttr::get(et, 0.0));
292               else if (auto et = t.getElementType().cast<IntegerType>())
293                 tmp = b.create<arith::ConstantOp>(IntegerAttr::get(et, 0));
294               return b.create<complex::CreateOp>(t, tmp, tmp);
295             })
296             .Default([](auto) { return Value(); });
297     if (!fillVal)
298       return failure();
299     b.create<linalg::FillOp>(fillVal, promotionInfo->fullLocalView);
300   }
301 
302   // Copy data into the promoted buffers. Use callback if provided.
303   for (auto v : options.subViews) {
304     auto info = promotionInfoMap.find(v.first);
305     if (info == promotionInfoMap.end())
306       continue;
307     if (failed(options.copyInFn(
308             b, cast<memref::SubViewOp>(v.second.getDefiningOp()),
309             info->second.partialLocalView)))
310       return failure();
311   }
312   return promotionInfoMap;
313 }
314 
315 static FailureOr<LinalgOp>
316 promoteSubViews(ImplicitLocOpBuilder &b, LinalgOp op,
317                 LinalgOpInstancePromotionOptions options, DataLayout &layout) {
318   assert(op.hasBufferSemantics() && "expected linalg op with buffer semantics");
319 
320   // 1. Promote the specified views and use them in the new op.
321   auto promotedBuffersAndViews = promoteSubViews(b, options, layout);
322   if (failed(promotedBuffersAndViews) ||
323       promotedBuffersAndViews->size() != options.subViews.size())
324     return failure();
325 
326   // 2. Append all other operands as they appear, this enforces that such
327   // operands are not views. This is to support cases such as FillOp taking
328   // extra scalars etc.  Keep a reference to output buffers;
329   SmallVector<Value, 8> opViews;
330   opViews.reserve(op->getNumOperands());
331   SmallVector<std::pair<Value, Value>, 8> writebackViews;
332   writebackViews.reserve(promotedBuffersAndViews->size());
333   for (OpOperand &opOperand : op->getOpOperands()) {
334     int64_t operandNumber = opOperand.getOperandNumber();
335     if (options.subViews.count(operandNumber) != 0) {
336       if (options.useFullTileBuffers[opOperand.get()])
337         opViews.push_back(
338             (*promotedBuffersAndViews)[operandNumber].fullLocalView);
339       else
340         opViews.push_back(
341             (*promotedBuffersAndViews)[operandNumber].partialLocalView);
342       if (operandNumber >= op.getNumDpsInputs())
343         writebackViews.emplace_back(std::make_pair(
344             opOperand.get(),
345             (*promotedBuffersAndViews)[operandNumber].partialLocalView));
346     } else {
347       opViews.push_back(opOperand.get());
348     }
349   }
350   op->setOperands(0, opViews.size(), opViews);
351 
352   OpBuilder::InsertionGuard guard(b);
353   b.setInsertionPointAfter(op);
354   // 3. Emit write-back for the promoted output views: copy the partial view.
355   for (auto viewAndPartialLocalView : writebackViews) {
356     if (failed(options.copyOutFn(b, viewAndPartialLocalView.second,
357                                  viewAndPartialLocalView.first)))
358       return failure();
359   }
360 
361   // 4. Dealloc all local buffers.
362   for (const auto &pi : *promotedBuffersAndViews)
363     (void)options.deallocationFn(b, pi.second.fullLocalView);
364   return op;
365 }
366 
367 LogicalResult
368 mlir::linalg::promoteSubviewsPrecondition(Operation *op,
369                                           LinalgPromotionOptions options) {
370   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
371   // Transformation applies to buffers only.
372   if (!linalgOp || !linalgOp.hasBufferSemantics())
373     return failure();
374   // Check that at least one of the requested operands is indeed a subview.
375   for (OpOperand &opOperand : linalgOp->getOpOperands()) {
376     auto sv =
377         isa_and_nonnull<memref::SubViewOp>(opOperand.get().getDefiningOp());
378     if (sv) {
379       if (!options.operandsToPromote ||
380           options.operandsToPromote->count(opOperand.getOperandNumber()))
381         return success();
382     }
383   }
384   // TODO: Check all subviews requested are bound by a static constant.
385   // TODO: Check that the total footprint fits within a given size.
386   return failure();
387 }
388 
389 FailureOr<LinalgOp>
390 mlir::linalg::promoteSubViews(OpBuilder &builder, LinalgOp linalgOp,
391                               const LinalgPromotionOptions &options) {
392   LinalgOpInstancePromotionOptions linalgOptions(linalgOp, options);
393   auto layout = DataLayout::closest(linalgOp);
394   ImplicitLocOpBuilder b(linalgOp.getLoc(), builder);
395   auto res = ::promoteSubViews(b, linalgOp, linalgOptions, layout);
396   if (failed(res))
397     return failure();
398   return res;
399 }
400