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 "PassDetail.h" 14 #include "mlir/Dialect/Arithmetic/IR/Arithmetic.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/Linalg/Utils/Utils.h" 20 #include "mlir/Dialect/SCF/IR/SCF.h" 21 #include "mlir/IR/AffineExpr.h" 22 #include "mlir/IR/AffineExprVisitor.h" 23 #include "mlir/IR/AffineMap.h" 24 #include "mlir/IR/ImplicitLocOpBuilder.h" 25 #include "mlir/Support/LLVM.h" 26 #include "mlir/Transforms/FoldUtils.h" 27 #include "llvm/ADT/MapVector.h" 28 #include "llvm/ADT/SmallBitVector.h" 29 #include "llvm/ADT/TypeSwitch.h" 30 #include "llvm/Support/CommandLine.h" 31 #include "llvm/Support/Debug.h" 32 33 using namespace mlir; 34 using namespace mlir::linalg; 35 using namespace mlir::scf; 36 37 using llvm::MapVector; 38 39 #define DEBUG_TYPE "linalg-promotion" 40 41 /// Alloc a new buffer of `size` * `width` i8; where `width` is given by the 42 /// data `layout` for `elementType`. 43 /// Use AllocOp or AllocaOp depending on `options`. 44 /// Take an optional alignment. 45 static Value allocBuffer(ImplicitLocOpBuilder &b, 46 const LinalgPromotionOptions &options, 47 Type elementType, Value allocSize, DataLayout &layout, 48 Optional<unsigned> alignment = None) { 49 auto width = layout.getTypeSize(elementType); 50 51 IntegerAttr alignmentAttr; 52 if (alignment.hasValue()) 53 alignmentAttr = b.getI64IntegerAttr(alignment.getValue()); 54 55 // Static buffer. 56 if (auto cst = allocSize.getDefiningOp<arith::ConstantIndexOp>()) { 57 auto staticBufferType = 58 MemRefType::get(width * cst.value(), b.getIntegerType(8)); 59 if (options.useAlloca) { 60 return b.createOrFold<memref::AllocaOp>(staticBufferType, ValueRange{}, 61 alignmentAttr); 62 } 63 return b.createOrFold<memref::AllocOp>(staticBufferType, ValueRange{}, 64 alignmentAttr); 65 } 66 67 // Fallback dynamic buffer. 68 auto dynamicBufferType = MemRefType::get(-1, 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 Optional<Value> 81 defaultAllocBufferCallBack(const LinalgPromotionOptions &options, 82 OpBuilder &builder, memref::SubViewOp subView, 83 ArrayRef<Value> boundingSubViewSize, 84 Optional<unsigned> alignment, DataLayout &layout) { 85 ShapedType viewType = subView.getType(); 86 ImplicitLocOpBuilder b(subView.getLoc(), builder); 87 auto zero = b.createOrFold<arith::ConstantIndexOp>(0); 88 auto one = b.createOrFold<arith::ConstantIndexOp>(1); 89 90 Value allocSize = one; 91 for (const auto &size : llvm::enumerate(boundingSubViewSize)) 92 allocSize = b.createOrFold<arith::MulIOp>(allocSize, size.value()); 93 Value buffer = allocBuffer(b, options, viewType.getElementType(), allocSize, 94 layout, alignment); 95 SmallVector<int64_t, 4> dynSizes(boundingSubViewSize.size(), 96 ShapedType::kDynamicSize); 97 Value view = b.createOrFold<memref::ViewOp>( 98 MemRefType::get(dynSizes, viewType.getElementType()), buffer, zero, 99 boundingSubViewSize); 100 return view; 101 } 102 103 /// Default implementation of deallocation of the buffer use for promotion. It 104 /// expects to get the same value that the default allocation method returned, 105 /// i.e. result of a ViewOp. 106 static LogicalResult 107 defaultDeallocBufferCallBack(const LinalgPromotionOptions &options, 108 OpBuilder &b, Value fullLocalView) { 109 if (!options.useAlloca) { 110 auto viewOp = cast<memref::ViewOp>(fullLocalView.getDefiningOp()); 111 b.create<memref::DeallocOp>(viewOp.getSource().getLoc(), 112 viewOp.getSource()); 113 } 114 return success(); 115 } 116 117 namespace { 118 119 /// Helper struct that captures the information required to apply the 120 /// transformation on each op. This bridges the abstraction gap with the 121 /// user-facing API which exposes positional arguments to control which operands 122 /// are promoted. 123 struct LinalgOpInstancePromotionOptions { 124 LinalgOpInstancePromotionOptions(LinalgOp op, 125 const LinalgPromotionOptions &options); 126 /// SubViews to promote. 127 MapVector<int64_t, Value> subViews; 128 /// True if the full view should be used for the promoted buffer. 129 DenseMap<Value, bool> useFullTileBuffers; 130 131 /// Callback functions for allocation and deallocation of promoted buffers, as 132 /// well as to copy the data into and out of these buffers. 133 AllocBufferCallbackFn allocationFn; 134 DeallocBufferCallbackFn deallocationFn; 135 CopyCallbackFn copyInFn; 136 CopyCallbackFn copyOutFn; 137 138 /// Allow the use of dynamically-sized buffers. 139 bool dynamicBuffers; 140 141 /// Alignment of promoted buffer. 142 Optional<unsigned> alignment; 143 }; 144 } // namespace 145 146 LinalgOpInstancePromotionOptions::LinalgOpInstancePromotionOptions( 147 LinalgOp linalgOp, const LinalgPromotionOptions &options) 148 : subViews(), dynamicBuffers(options.dynamicBuffers), 149 alignment(options.alignment) { 150 assert(linalgOp.hasBufferSemantics() && "revisit usage of shaped operand"); 151 auto vUseFullTileBuffers = 152 options.useFullTileBuffers.value_or(llvm::SmallBitVector()); 153 vUseFullTileBuffers.resize(linalgOp.getNumInputsAndOutputs(), 154 options.useFullTileBuffersDefault); 155 156 for (OpOperand *opOperand : linalgOp.getInputAndOutputOperands()) { 157 int64_t operandNumber = opOperand->getOperandNumber(); 158 if (options.operandsToPromote && 159 !options.operandsToPromote->count(operandNumber)) 160 continue; 161 Operation *op = opOperand->get().getDefiningOp(); 162 if (auto sv = dyn_cast_or_null<memref::SubViewOp>(op)) { 163 subViews[operandNumber] = sv; 164 useFullTileBuffers[sv] = vUseFullTileBuffers[operandNumber]; 165 } 166 } 167 168 if (options.allocationFn) { 169 allocationFn = *options.allocationFn; 170 } else { 171 allocationFn = [&](OpBuilder &b, memref::SubViewOp subViewOp, 172 ArrayRef<Value> boundingSubViewSize, 173 DataLayout &layout) -> Optional<Value> { 174 return defaultAllocBufferCallBack(options, b, subViewOp, 175 boundingSubViewSize, alignment, layout); 176 }; 177 } 178 179 if (options.deallocationFn) { 180 deallocationFn = *options.deallocationFn; 181 } else { 182 deallocationFn = [&](OpBuilder &b, Value buffer) { 183 return defaultDeallocBufferCallBack(options, b, buffer); 184 }; 185 } 186 187 // Save the loc because `linalgOp` goes out of scope. 188 Location loc = linalgOp.getLoc(); 189 auto defaultCopyCallBack = [loc](OpBuilder &b, Value src, 190 Value dst) -> LogicalResult { 191 b.create<memref::CopyOp>(loc, src, dst); 192 return success(); 193 }; 194 copyInFn = (options.copyInFn ? *(options.copyInFn) : defaultCopyCallBack); 195 copyOutFn = (options.copyOutFn ? *(options.copyOutFn) : defaultCopyCallBack); 196 } 197 198 // Performs promotion of a `subView` into a local buffer of the size of the 199 // *ranges* of the `subView`. This produces a buffer whose size may be bigger 200 // than the actual size of the `subView` at the boundaries. 201 // This is related to the full/partial tile problem. 202 // Returns a PromotionInfo containing a `buffer`, `fullLocalView` and 203 // `partialLocalView` such that: 204 // * `buffer` is always the size of the full tile. 205 // * `fullLocalView` is a dense contiguous view into that buffer. 206 // * `partialLocalView` is a dense non-contiguous slice of `fullLocalView` 207 // that corresponds to the size of `subView` and accounting for boundary 208 // effects. 209 // The point of the full tile buffer is that constant static tile sizes are 210 // folded and result in a buffer type with statically known size and alignment 211 // properties. 212 // To account for general boundary effects, padding must be performed on the 213 // boundary tiles. For now this is done with an unconditional `fill` op followed 214 // by a partial `copy` op. 215 FailureOr<PromotionInfo> mlir::linalg::promoteSubviewAsNewBuffer( 216 OpBuilder &b, Location loc, memref::SubViewOp subView, 217 const AllocBufferCallbackFn &allocationFn, DataLayout &layout) { 218 auto viewType = subView.getType(); 219 auto rank = viewType.getRank(); 220 SmallVector<Value, 4> fullSizes; 221 SmallVector<OpFoldResult> partialSizes; 222 fullSizes.reserve(rank); 223 partialSizes.reserve(rank); 224 llvm::SmallBitVector droppedDims = subView.getDroppedDims(); 225 int64_t resultDimIdx = 0; 226 for (const auto &en : llvm::enumerate(subView.getOrCreateRanges(b, loc))) { 227 if (droppedDims[en.index()]) 228 continue; 229 auto rangeValue = en.value(); 230 // Try to extract a tight constant. 231 LLVM_DEBUG(llvm::dbgs() << "Extract tightest: " << rangeValue.size << "\n"); 232 FailureOr<int64_t> upperBound = 233 getConstantUpperBoundForIndex(rangeValue.size); 234 Value size = 235 failed(upperBound) 236 ? rangeValue.size 237 : b.create<arith::ConstantIndexOp>(loc, upperBound.getValue()); 238 LLVM_DEBUG(llvm::dbgs() << "Extracted tightest: " << size << "\n"); 239 fullSizes.push_back(size); 240 partialSizes.push_back( 241 b.createOrFold<memref::DimOp>(loc, subView, resultDimIdx++)); 242 } 243 SmallVector<int64_t, 4> dynSizes(fullSizes.size(), -1); 244 // If a callback is not specified, then use the default implementation for 245 // allocating the promoted buffer. 246 Optional<Value> fullLocalView = allocationFn(b, subView, fullSizes, layout); 247 if (!fullLocalView) 248 return failure(); 249 SmallVector<OpFoldResult, 4> zeros(fullSizes.size(), b.getIndexAttr(0)); 250 SmallVector<OpFoldResult, 4> ones(fullSizes.size(), b.getIndexAttr(1)); 251 auto partialLocalView = b.createOrFold<memref::SubViewOp>( 252 loc, *fullLocalView, zeros, partialSizes, ones); 253 return PromotionInfo{*fullLocalView, partialLocalView}; 254 } 255 256 static FailureOr<MapVector<int64_t, PromotionInfo>> 257 promoteSubViews(ImplicitLocOpBuilder &b, 258 LinalgOpInstancePromotionOptions options, DataLayout &layout) { 259 if (options.subViews.empty()) 260 return failure(); 261 262 MapVector<int64_t, PromotionInfo> promotionInfoMap; 263 264 for (auto v : options.subViews) { 265 memref::SubViewOp subView = 266 cast<memref::SubViewOp>(v.second.getDefiningOp()); 267 auto promotionInfo = promoteSubviewAsNewBuffer( 268 b, b.getLoc(), subView, options.allocationFn, layout); 269 if (failed(promotionInfo)) 270 return failure(); 271 promotionInfoMap[v.first] = *promotionInfo; 272 273 // Only fill the buffer if the full local view is used 274 if (!options.useFullTileBuffers[v.second]) 275 continue; 276 Type subviewEltType = subView.getType().getElementType(); 277 Value fillVal = 278 llvm::TypeSwitch<Type, Value>(subviewEltType) 279 .Case([&](FloatType t) { 280 return b.create<arith::ConstantOp>(FloatAttr::get(t, 0.0)); 281 }) 282 .Case([&](IntegerType t) { 283 return b.create<arith::ConstantOp>(IntegerAttr::get(t, 0)); 284 }) 285 .Case([&](ComplexType t) { 286 Value tmp; 287 if (auto et = t.getElementType().dyn_cast<FloatType>()) 288 tmp = b.create<arith::ConstantOp>(FloatAttr::get(et, 0.0)); 289 else if (auto et = t.getElementType().cast<IntegerType>()) 290 tmp = b.create<arith::ConstantOp>(IntegerAttr::get(et, 0)); 291 return b.create<complex::CreateOp>(t, tmp, tmp); 292 }) 293 .Default([](auto) { return Value(); }); 294 if (!fillVal) 295 return failure(); 296 b.create<linalg::FillOp>(fillVal, promotionInfo->fullLocalView); 297 } 298 299 // Copy data into the promoted buffers. Use callback if provided. 300 for (auto v : options.subViews) { 301 auto info = promotionInfoMap.find(v.first); 302 if (info == promotionInfoMap.end()) 303 continue; 304 if (failed(options.copyInFn( 305 b, cast<memref::SubViewOp>(v.second.getDefiningOp()), 306 info->second.partialLocalView))) 307 return failure(); 308 } 309 return promotionInfoMap; 310 } 311 312 static FailureOr<LinalgOp> 313 promoteSubViews(ImplicitLocOpBuilder &b, LinalgOp op, 314 LinalgOpInstancePromotionOptions options, DataLayout &layout) { 315 assert(op.hasBufferSemantics() && "expected linalg op with buffer semantics"); 316 317 // 1. Promote the specified views and use them in the new op. 318 auto promotedBuffersAndViews = promoteSubViews(b, options, layout); 319 if (failed(promotedBuffersAndViews) || 320 promotedBuffersAndViews->size() != options.subViews.size()) 321 return failure(); 322 323 // 2. Append all other operands as they appear, this enforces that such 324 // operands are not views. This is to support cases such as FillOp taking 325 // extra scalars etc. Keep a reference to output buffers; 326 SmallVector<Value, 8> opViews; 327 opViews.reserve(op.getNumInputsAndOutputs()); 328 SmallVector<std::pair<Value, Value>, 8> writebackViews; 329 writebackViews.reserve(promotedBuffersAndViews->size()); 330 for (OpOperand *opOperand : op.getInputAndOutputOperands()) { 331 int64_t operandNumber = opOperand->getOperandNumber(); 332 if (options.subViews.count(operandNumber) != 0) { 333 if (options.useFullTileBuffers[opOperand->get()]) 334 opViews.push_back( 335 (*promotedBuffersAndViews)[operandNumber].fullLocalView); 336 else 337 opViews.push_back( 338 (*promotedBuffersAndViews)[operandNumber].partialLocalView); 339 if (operandNumber >= op.getNumInputs()) 340 writebackViews.emplace_back(std::make_pair( 341 opOperand->get(), 342 (*promotedBuffersAndViews)[operandNumber].partialLocalView)); 343 } else { 344 opViews.push_back(opOperand->get()); 345 } 346 } 347 op->setOperands(0, opViews.size(), opViews); 348 349 OpBuilder::InsertionGuard guard(b); 350 b.setInsertionPointAfter(op); 351 // 3. Emit write-back for the promoted output views: copy the partial view. 352 for (auto viewAndPartialLocalView : writebackViews) { 353 if (failed(options.copyOutFn(b, viewAndPartialLocalView.second, 354 viewAndPartialLocalView.first))) 355 return failure(); 356 } 357 358 // 4. Dealloc all local buffers. 359 for (const auto &pi : *promotedBuffersAndViews) 360 (void)options.deallocationFn(b, pi.second.fullLocalView); 361 return op; 362 } 363 364 LogicalResult 365 mlir::linalg::promoteSubviewsPrecondition(Operation *op, 366 LinalgPromotionOptions options) { 367 LinalgOp linalgOp = dyn_cast<LinalgOp>(op); 368 // Transformation applies to buffers only. 369 if (!linalgOp || !linalgOp.hasBufferSemantics()) 370 return failure(); 371 // Check that at least one of the requested operands is indeed a subview. 372 for (OpOperand *opOperand : linalgOp.getInputAndOutputOperands()) { 373 auto sv = 374 isa_and_nonnull<memref::SubViewOp>(opOperand->get().getDefiningOp()); 375 if (sv) { 376 if (!options.operandsToPromote || 377 options.operandsToPromote->count(opOperand->getOperandNumber())) 378 return success(); 379 } 380 } 381 // TODO: Check all subviews requested are bound by a static constant. 382 // TODO: Check that the total footprint fits within a given size. 383 return failure(); 384 } 385 386 FailureOr<LinalgOp> 387 mlir::linalg::promoteSubViews(OpBuilder &builder, LinalgOp linalgOp, 388 const LinalgPromotionOptions &options) { 389 LinalgOpInstancePromotionOptions linalgOptions(linalgOp, options); 390 auto layout = DataLayout::closest(linalgOp); 391 ImplicitLocOpBuilder b(linalgOp.getLoc(), builder); 392 auto res = ::promoteSubViews(b, linalgOp, linalgOptions, layout); 393 if (failed(res)) 394 return failure(); 395 return res; 396 } 397 398 namespace { 399 struct LinalgPromotionPass : public LinalgPromotionBase<LinalgPromotionPass> { 400 LinalgPromotionPass() = default; 401 LinalgPromotionPass(bool dynamicBuffers, bool useAlloca) { 402 this->dynamicBuffers = dynamicBuffers; 403 this->useAlloca = useAlloca; 404 } 405 406 void runOnOperation() override { 407 getOperation().walk([&](LinalgOp op) { 408 auto options = LinalgPromotionOptions() 409 .setDynamicBuffers(dynamicBuffers) 410 .setUseAlloca(useAlloca); 411 if (failed(promoteSubviewsPrecondition(op, options))) 412 return; 413 LLVM_DEBUG(llvm::dbgs() << "Promote: " << *(op.getOperation()) << "\n"); 414 ImplicitLocOpBuilder b(op.getLoc(), op); 415 // TODO: signalPassFailure() ? 416 (void)promoteSubViews(b, op, options); 417 }); 418 } 419 }; 420 } // namespace 421 422 // TODO: support more transformation options in the pass. 423 std::unique_ptr<OperationPass<func::FuncOp>> 424 mlir::createLinalgPromotionPass(bool dynamicBuffers, bool useAlloca) { 425 return std::make_unique<LinalgPromotionPass>(dynamicBuffers, useAlloca); 426 } 427 std::unique_ptr<OperationPass<func::FuncOp>> mlir::createLinalgPromotionPass() { 428 return std::make_unique<LinalgPromotionPass>(); 429 } 430