xref: /llvm-project/mlir/lib/Dialect/Linalg/Transforms/Promotion.cpp (revision a7d6039f3efb02992d64164e6778b8bebf0a526b)
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/Func/IR/FuncOps.h"
17 #include "mlir/Dialect/GPU/IR/GPUDialect.h"
18 #include "mlir/Dialect/Linalg/IR/Linalg.h"
19 #include "mlir/Dialect/Linalg/Passes.h"
20 #include "mlir/Dialect/Linalg/Transforms/Transforms.h"
21 #include "mlir/Dialect/SCF/IR/SCF.h"
22 #include "mlir/IR/AffineExpr.h"
23 #include "mlir/IR/AffineExprVisitor.h"
24 #include "mlir/IR/AffineMap.h"
25 #include "mlir/IR/ImplicitLocOpBuilder.h"
26 #include "mlir/Interfaces/ValueBoundsOpInterface.h"
27 #include "mlir/Support/LLVM.h"
28 #include "mlir/Transforms/FoldUtils.h"
29 #include "llvm/ADT/MapVector.h"
30 #include "llvm/ADT/SmallBitVector.h"
31 #include "llvm/ADT/SmallSet.h"
32 #include "llvm/ADT/TypeSwitch.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/Debug.h"
35 
36 using namespace mlir;
37 using namespace mlir::linalg;
38 using namespace mlir::scf;
39 
40 using llvm::MapVector;
41 
42 #define DEBUG_TYPE "linalg-promotion"
43 
44 /// Alloc a new buffer of `size` * `width` i8; where `width` is given by the
45 /// data `layout` for `elementType`.
46 /// Use AllocOp or AllocaOp depending on `options`.
47 /// Take an optional alignment.
48 static Value allocBuffer(ImplicitLocOpBuilder &b,
49                          const LinalgPromotionOptions &options,
50                          Type elementType, Value allocSize, DataLayout &layout,
51                          std::optional<unsigned> alignment = std::nullopt) {
52   auto width = layout.getTypeSize(elementType);
53 
54   IntegerAttr alignmentAttr;
55   if (alignment.has_value())
56     alignmentAttr = b.getI64IntegerAttr(alignment.value());
57 
58   Attribute memorySpaceAttr;
59   if (options.memorySpace.has_value())
60     memorySpaceAttr = *options.memorySpace;
61 
62   // Static buffer.
63   if (std::optional<int64_t> cst = getConstantIntValue(allocSize)) {
64     auto staticBufferType =
65         MemRefType::get(width * cst.value(), b.getIntegerType(8));
66     staticBufferType =
67         MemRefType::Builder(staticBufferType).setMemorySpace(memorySpaceAttr);
68     if (options.useAlloca) {
69       return b.create<memref::AllocaOp>(staticBufferType, ValueRange{},
70                                         alignmentAttr);
71     }
72     return b.create<memref::AllocOp>(staticBufferType, ValueRange{},
73                                      alignmentAttr);
74   }
75 
76   // Fallback dynamic buffer.
77   auto dynamicBufferType =
78       MemRefType::get(ShapedType::kDynamic, b.getIntegerType(8));
79   dynamicBufferType =
80       MemRefType::Builder(dynamicBufferType).setMemorySpace(memorySpaceAttr);
81   Value mul = b.createOrFold<arith::MulIOp>(
82       b.create<arith::ConstantIndexOp>(width), allocSize);
83   if (options.useAlloca)
84     return b.create<memref::AllocaOp>(dynamicBufferType, mul, alignmentAttr);
85   return b.create<memref::AllocOp>(dynamicBufferType, mul, alignmentAttr);
86 }
87 
88 /// Default allocation callback function. This allocates a promoted buffer when
89 /// no call back to do so is provided. The default is to allocate a
90 /// memref<..xi8> and return a view to get a memref type of shape
91 /// boundingSubViewSize.
92 static std::optional<Value> defaultAllocBufferCallBack(
93     const LinalgPromotionOptions &options, OpBuilder &builder,
94     memref::SubViewOp subView, ArrayRef<Value> boundingSubViewSize,
95     std::optional<unsigned> alignment, DataLayout &layout) {
96   ShapedType viewType = subView.getType();
97   ImplicitLocOpBuilder b(subView.getLoc(), builder);
98   auto zero = b.create<arith::ConstantIndexOp>(0);
99   auto one = b.create<arith::ConstantIndexOp>(1);
100 
101   Attribute memorySpaceAttr;
102   if (options.memorySpace.has_value())
103     memorySpaceAttr = *options.memorySpace;
104 
105   Value allocSize = one;
106   for (const auto &size : llvm::enumerate(boundingSubViewSize))
107     allocSize = b.createOrFold<arith::MulIOp>(allocSize, size.value());
108   Value buffer = allocBuffer(b, options, viewType.getElementType(), allocSize,
109                              layout, alignment);
110   SmallVector<int64_t, 4> dynSizes(boundingSubViewSize.size(),
111                                    ShapedType::kDynamic);
112 
113   auto viewMemRefType = MemRefType::get(dynSizes, viewType.getElementType());
114   viewMemRefType =
115       MemRefType::Builder(viewMemRefType).setMemorySpace(memorySpaceAttr);
116   Value view = b.createOrFold<memref::ViewOp>(viewMemRefType, buffer, zero,
117                                               boundingSubViewSize);
118   return view;
119 }
120 
121 /// Default implementation of deallocation of the buffer use for promotion. It
122 /// expects to get the same value that the default allocation method returned,
123 /// i.e. result of a ViewOp.
124 static LogicalResult
125 defaultDeallocBufferCallBack(const LinalgPromotionOptions &options,
126                              OpBuilder &b, Value fullLocalView) {
127   if (!options.useAlloca) {
128     auto viewOp = cast<memref::ViewOp>(fullLocalView.getDefiningOp());
129     b.create<memref::DeallocOp>(viewOp.getSource().getLoc(),
130                                 viewOp.getSource());
131   }
132   return success();
133 }
134 
135 namespace {
136 
137 /// Helper struct that captures the information required to apply the
138 /// transformation on each op. This bridges the abstraction gap with the
139 /// user-facing API which exposes positional arguments to control which operands
140 /// are promoted.
141 struct LinalgOpInstancePromotionOptions {
142   LinalgOpInstancePromotionOptions(LinalgOp op,
143                                    const LinalgPromotionOptions &options);
144   /// SubViews to promote.
145   MapVector<int64_t, Value> subViews;
146   /// Subviews operand numbers to copy in using copyInFn.
147   llvm::SmallSet<int64_t, 4> operandsNumbersToCopyIn;
148   /// True if the full view should be used for the promoted buffer.
149   DenseMap<Value, bool> useFullTileBuffers;
150 
151   /// Callback functions for allocation and deallocation of promoted buffers, as
152   /// well as to copy the data into and out of these buffers.
153   AllocBufferCallbackFn allocationFn;
154   DeallocBufferCallbackFn deallocationFn;
155   CopyCallbackFn copyInFn;
156   CopyCallbackFn copyOutFn;
157 
158   /// Alignment of promoted buffer.
159   std::optional<unsigned> alignment;
160 };
161 } // namespace
162 
163 LinalgOpInstancePromotionOptions::LinalgOpInstancePromotionOptions(
164     LinalgOp linalgOp, const LinalgPromotionOptions &options)
165     : subViews(), alignment(options.alignment) {
166   assert(linalgOp.hasBufferSemantics() && "revisit usage of shaped operand");
167   auto vUseFullTileBuffers =
168       options.useFullTileBuffers.value_or(llvm::SmallBitVector());
169   vUseFullTileBuffers.resize(linalgOp->getNumOperands(),
170                              options.useFullTileBuffersDefault);
171 
172   for (OpOperand &opOperand : linalgOp->getOpOperands()) {
173     int64_t operandNumber = opOperand.getOperandNumber();
174     if (options.operandsToPromote &&
175         !options.operandsToPromote->count(operandNumber))
176       continue;
177     Operation *op = opOperand.get().getDefiningOp();
178     if (auto sv = dyn_cast_or_null<memref::SubViewOp>(op)) {
179       subViews[operandNumber] = sv;
180       // In case of linalg generic, copy in only if subview is used in linalg
181       // payload.
182       if (!isa<linalg::GenericOp>(linalgOp) ||
183           linalgOp.payloadUsesValueFromOperand(&opOperand))
184         operandsNumbersToCopyIn.insert(operandNumber);
185       useFullTileBuffers[sv] = vUseFullTileBuffers[operandNumber];
186     }
187   }
188 
189   if (options.allocationFn) {
190     allocationFn = *options.allocationFn;
191   } else {
192     allocationFn = [&](OpBuilder &b, memref::SubViewOp subViewOp,
193                        ArrayRef<Value> boundingSubViewSize,
194                        DataLayout &layout) -> std::optional<Value> {
195       return defaultAllocBufferCallBack(options, b, subViewOp,
196                                         boundingSubViewSize, alignment, layout);
197     };
198   }
199 
200   if (options.deallocationFn) {
201     deallocationFn = *options.deallocationFn;
202   } else {
203     deallocationFn = [&](OpBuilder &b, Value buffer) {
204       return defaultDeallocBufferCallBack(options, b, buffer);
205     };
206   }
207 
208   // Save the loc because `linalgOp` goes out of scope.
209   Location loc = linalgOp.getLoc();
210   auto defaultCopyCallBack = [loc](OpBuilder &b, Value src,
211                                    Value dst) -> LogicalResult {
212     b.create<linalg::CopyOp>(loc, src, dst);
213     return success();
214   };
215   copyInFn = (options.copyInFn ? *(options.copyInFn) : defaultCopyCallBack);
216   copyOutFn = (options.copyOutFn ? *(options.copyOutFn) : defaultCopyCallBack);
217 }
218 
219 // Performs promotion of a `subView` into a local buffer of the size of the
220 // *ranges* of the `subView`. This produces a buffer whose size may be bigger
221 // than the actual size of the `subView` at the boundaries.
222 // This is related to the full/partial tile problem.
223 // Returns a PromotionInfo containing a `buffer`, `fullLocalView` and
224 // `partialLocalView` such that:
225 //   * `buffer` is always the size of the full tile.
226 //   * `fullLocalView` is a dense contiguous view into that buffer.
227 //   * `partialLocalView` is a dense non-contiguous slice of `fullLocalView`
228 //     that corresponds to the size of `subView` and accounting for boundary
229 //     effects.
230 // The point of the full tile buffer is that constant static tile sizes are
231 // folded and result in a buffer type with statically known size and alignment
232 // properties.
233 // To account for general boundary effects, padding must be performed on the
234 // boundary tiles. For now this is done with an unconditional `fill` op followed
235 // by a partial `copy` op.
236 FailureOr<PromotionInfo> mlir::linalg::promoteSubviewAsNewBuffer(
237     OpBuilder &b, Location loc, memref::SubViewOp subView,
238     const AllocBufferCallbackFn &allocationFn, DataLayout &layout) {
239   auto viewType = subView.getType();
240   auto rank = viewType.getRank();
241   SmallVector<Value, 4> fullSizes;
242   SmallVector<OpFoldResult> partialSizes;
243   fullSizes.reserve(rank);
244   partialSizes.reserve(rank);
245   llvm::SmallBitVector droppedDims = subView.getDroppedDims();
246   int64_t resultDimIdx = 0;
247   for (const auto &en : llvm::enumerate(subView.getOrCreateRanges(b, loc))) {
248     if (droppedDims[en.index()])
249       continue;
250     auto rangeValue = en.value();
251     // Try to extract a tight constant. If the size is known statically, no need
252     // to look for the bound.
253     LLVM_DEBUG(llvm::dbgs() << "Extract tightest: " << rangeValue.size << "\n");
254     Value size;
255     if (auto attr = llvm::dyn_cast_if_present<Attribute>(rangeValue.size)) {
256       size = getValueOrCreateConstantIndexOp(b, loc, rangeValue.size);
257     } else {
258       Value materializedSize =
259           getValueOrCreateConstantIndexOp(b, loc, rangeValue.size);
260       FailureOr<int64_t> upperBound =
261           ValueBoundsConstraintSet::computeConstantBound(
262               presburger::BoundType::UB, materializedSize, /*dim=*/std::nullopt,
263               /*stopCondition=*/nullptr, /*closedUB=*/true);
264       size = failed(upperBound)
265                  ? materializedSize
266                  : b.create<arith::ConstantIndexOp>(loc, *upperBound);
267     }
268     LLVM_DEBUG(llvm::dbgs() << "Extracted tightest: " << size << "\n");
269     fullSizes.push_back(size);
270     partialSizes.push_back(
271         b.createOrFold<memref::DimOp>(loc, subView, resultDimIdx++));
272   }
273   SmallVector<int64_t, 4> dynSizes(fullSizes.size(), ShapedType::kDynamic);
274   // If a callback is not specified, then use the default implementation for
275   // allocating the promoted buffer.
276   std::optional<Value> fullLocalView =
277       allocationFn(b, subView, fullSizes, layout);
278   if (!fullLocalView)
279     return failure();
280   SmallVector<OpFoldResult, 4> zeros(fullSizes.size(), b.getIndexAttr(0));
281   SmallVector<OpFoldResult, 4> ones(fullSizes.size(), b.getIndexAttr(1));
282   auto partialLocalView = b.createOrFold<memref::SubViewOp>(
283       loc, *fullLocalView, zeros, partialSizes, ones);
284   return PromotionInfo{*fullLocalView, partialLocalView};
285 }
286 
287 static FailureOr<MapVector<int64_t, PromotionInfo>>
288 promoteSubViews(ImplicitLocOpBuilder &b,
289                 LinalgOpInstancePromotionOptions options, DataLayout &layout) {
290   if (options.subViews.empty())
291     return failure();
292 
293   MapVector<int64_t, PromotionInfo> promotionInfoMap;
294 
295   for (auto v : options.subViews) {
296     memref::SubViewOp subView =
297         cast<memref::SubViewOp>(v.second.getDefiningOp());
298     auto promotionInfo = promoteSubviewAsNewBuffer(
299         b, b.getLoc(), subView, options.allocationFn, layout);
300     if (failed(promotionInfo))
301       return failure();
302     promotionInfoMap[v.first] = *promotionInfo;
303 
304     // Only fill the buffer if the full local view is used
305     if (!options.useFullTileBuffers[v.second])
306       continue;
307     Type subviewEltType = subView.getType().getElementType();
308     Value fillVal =
309         llvm::TypeSwitch<Type, Value>(subviewEltType)
310             .Case([&](FloatType t) {
311               return b.create<arith::ConstantOp>(FloatAttr::get(t, 0.0));
312             })
313             .Case([&](IntegerType t) {
314               return b.create<arith::ConstantOp>(IntegerAttr::get(t, 0));
315             })
316             .Case([&](ComplexType t) {
317               Value tmp;
318               if (auto et = dyn_cast<FloatType>(t.getElementType()))
319                 tmp = b.create<arith::ConstantOp>(FloatAttr::get(et, 0.0));
320               else if (auto et = cast<IntegerType>(t.getElementType()))
321                 tmp = b.create<arith::ConstantOp>(IntegerAttr::get(et, 0));
322               return b.create<complex::CreateOp>(t, tmp, tmp);
323             })
324             .Default([](auto) { return Value(); });
325     if (!fillVal)
326       return failure();
327     b.create<linalg::FillOp>(fillVal, promotionInfo->fullLocalView);
328   }
329 
330   // Copy data into the promoted buffers. Use callback if provided.
331   for (auto v : options.subViews) {
332     auto info = promotionInfoMap.find(v.first);
333     if (info == promotionInfoMap.end())
334       continue;
335     if (options.operandsNumbersToCopyIn.count(v.first) == 0)
336       continue;
337     if (failed(options.copyInFn(
338             b, cast<memref::SubViewOp>(v.second.getDefiningOp()),
339             info->second.partialLocalView)))
340       return failure();
341   }
342   return promotionInfoMap;
343 }
344 
345 static FailureOr<LinalgOp>
346 promoteSubViews(ImplicitLocOpBuilder &b, LinalgOp op,
347                 LinalgOpInstancePromotionOptions options, DataLayout &layout) {
348   assert(op.hasBufferSemantics() && "expected linalg op with buffer semantics");
349 
350   // 1. Promote the specified views and use them in the new op.
351   auto promotedBuffersAndViews = promoteSubViews(b, options, layout);
352   if (failed(promotedBuffersAndViews) ||
353       promotedBuffersAndViews->size() != options.subViews.size())
354     return failure();
355 
356   // 2. Append all other operands as they appear, this enforces that such
357   // operands are not views. This is to support cases such as FillOp taking
358   // extra scalars etc.  Keep a reference to output buffers;
359   SmallVector<Value, 8> opViews;
360   opViews.reserve(op->getNumOperands());
361   SmallVector<std::pair<Value, Value>, 8> writebackViews;
362   writebackViews.reserve(promotedBuffersAndViews->size());
363   for (OpOperand &opOperand : op->getOpOperands()) {
364     int64_t operandNumber = opOperand.getOperandNumber();
365     if (options.subViews.count(operandNumber) != 0) {
366       if (options.useFullTileBuffers[opOperand.get()])
367         opViews.push_back(
368             (*promotedBuffersAndViews)[operandNumber].fullLocalView);
369       else
370         opViews.push_back(
371             (*promotedBuffersAndViews)[operandNumber].partialLocalView);
372       if (operandNumber >= op.getNumDpsInputs())
373         writebackViews.emplace_back(std::make_pair(
374             opOperand.get(),
375             (*promotedBuffersAndViews)[operandNumber].partialLocalView));
376     } else {
377       opViews.push_back(opOperand.get());
378     }
379   }
380   op->setOperands(0, opViews.size(), opViews);
381 
382   OpBuilder::InsertionGuard guard(b);
383   b.setInsertionPointAfter(op);
384   // 3. Emit write-back for the promoted output views: copy the partial view.
385   for (auto viewAndPartialLocalView : writebackViews) {
386     if (failed(options.copyOutFn(b, viewAndPartialLocalView.second,
387                                  viewAndPartialLocalView.first)))
388       return failure();
389   }
390 
391   // 4. Dealloc all local buffers.
392   for (const auto &pi : *promotedBuffersAndViews)
393     (void)options.deallocationFn(b, pi.second.fullLocalView);
394   return op;
395 }
396 
397 LogicalResult
398 mlir::linalg::promoteSubviewsPrecondition(Operation *op,
399                                           LinalgPromotionOptions options) {
400   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
401   // Transformation applies to buffers only.
402   if (!linalgOp || !linalgOp.hasBufferSemantics())
403     return failure();
404   // Check that at least one of the requested operands is indeed a subview.
405   for (OpOperand &opOperand : linalgOp->getOpOperands()) {
406     auto sv =
407         isa_and_nonnull<memref::SubViewOp>(opOperand.get().getDefiningOp());
408     if (sv) {
409       if (!options.operandsToPromote ||
410           options.operandsToPromote->count(opOperand.getOperandNumber()))
411         return success();
412     }
413   }
414   // TODO: Check all subviews requested are bound by a static constant.
415   // TODO: Check that the total footprint fits within a given size.
416   return failure();
417 }
418 
419 FailureOr<LinalgOp>
420 mlir::linalg::promoteSubViews(OpBuilder &builder, LinalgOp linalgOp,
421                               const LinalgPromotionOptions &options) {
422   LinalgOpInstancePromotionOptions linalgOptions(linalgOp, options);
423   auto layout = DataLayout::closest(linalgOp);
424   ImplicitLocOpBuilder b(linalgOp.getLoc(), builder);
425   auto res = ::promoteSubViews(b, linalgOp, linalgOptions, layout);
426   if (failed(res))
427     return failure();
428   return res;
429 }
430 
431 /// Allocate the given subview to a memory address space in GPU by creating a
432 /// allocation operation and setting the memref type address space to desired
433 /// address space.
434 static std::optional<Value> allocateSubviewGPUMemoryInAddressSpace(
435     OpBuilder &builder, memref::SubViewOp subview, ArrayRef<Value> sizeBounds,
436     gpu::AddressSpace addressSpace) {
437   OpBuilder::InsertionGuard guard(builder);
438 
439   func::FuncOp funcOp = subview->getParentOfType<func::FuncOp>();
440   if (!funcOp)
441     return std::nullopt;
442 
443   // The subview size bounds are expected to be constant; they specify the shape
444   // of the allocation.
445   SmallVector<int64_t> shape;
446   for (Value bound : sizeBounds) {
447     APInt value;
448     if (!matchPattern(bound, m_ConstantInt(&value)))
449       return std::nullopt;
450     shape.push_back(value.getSExtValue());
451   }
452 
453   builder.setInsertionPoint(&funcOp.front(), funcOp.front().begin());
454   auto type = MemRefType::get(
455       shape, subview.getType().getElementType(), MemRefLayoutAttrInterface{},
456       gpu::AddressSpaceAttr::get(builder.getContext(), addressSpace));
457   Value buffer;
458   if (addressSpace == gpu::GPUDialect::getWorkgroupAddressSpace()) {
459     buffer = builder.create<memref::AllocOp>(funcOp.getLoc(), type);
460   } else if (addressSpace == gpu::GPUDialect::getPrivateAddressSpace()) {
461     buffer = builder.create<memref::AllocaOp>(funcOp.getLoc(), type);
462   } else {
463     return std::nullopt;
464   }
465   return buffer;
466 }
467 
468 /// Allocate the subview in the GPU workgroup memory.
469 std::optional<Value> mlir::linalg::allocateWorkgroupMemory(
470     OpBuilder &builder, memref::SubViewOp subview, ArrayRef<Value> sizeBounds,
471     DataLayout &) {
472   return allocateSubviewGPUMemoryInAddressSpace(
473       builder, subview, sizeBounds,
474       gpu::GPUDialect::getWorkgroupAddressSpace());
475 }
476 
477 /// In case of GPU group memory there is no need to deallocate.
478 LogicalResult mlir::linalg::deallocateWorkgroupMemory(OpBuilder &,
479                                                       Value /*buffer*/) {
480   return success();
481 }
482 
483 /// Create Memref copy operations and add gpu barrier guards before and after
484 /// the copy operation to ensure data integrity.
485 LogicalResult mlir::linalg::copyToWorkgroupMemory(OpBuilder &b, Value src,
486                                                   Value dst) {
487   b.create<gpu::BarrierOp>(src.getLoc());
488   Operation *copyOp = b.create<memref::CopyOp>(src.getLoc(), src, dst);
489   b.create<gpu::BarrierOp>(copyOp->getLoc());
490   return success();
491 }
492 
493 /// Allocate the subview in the GPU private memory.
494 std::optional<Value> mlir::linalg::allocateGPUPrivateMemory(
495     OpBuilder &builder, memref::SubViewOp subview, ArrayRef<Value> sizeBounds,
496     DataLayout &) {
497   return allocateSubviewGPUMemoryInAddressSpace(
498       builder, subview, sizeBounds, gpu::GPUDialect::getPrivateAddressSpace());
499 }
500 
501 /// Normal copy to between src and dst.
502 LogicalResult mlir::linalg::copyToGPUPrivateMemory(OpBuilder &b, Value src,
503                                                    Value dst) {
504   b.create<memref::CopyOp>(src.getLoc(), src, dst);
505   return success();
506 }
507 
508 /// In case of GPU private memory there is no need to deallocate since the
509 /// memory is freed when going outside of the scope.
510 LogicalResult mlir::linalg::deallocateGPUPrivateMemory(OpBuilder &,
511                                                        Value /*buffer*/) {
512   return success();
513 }
514