1 //===- Promotion.cpp - Implementation of linalg Promotion -----------------===// 2 // 3 // Copyright 2019 The MLIR Authors. 4 // 5 // Licensed under the Apache License, Version 2.0 (the "License"); 6 // you may not use this file except in compliance with the License. 7 // You may obtain a copy of the License at 8 // 9 // http://www.apache.org/licenses/LICENSE-2.0 10 // 11 // Unless required by applicable law or agreed to in writing, software 12 // distributed under the License is distributed on an "AS IS" BASIS, 13 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 14 // See the License for the specific language governing permissions and 15 // limitations under the License. 16 // ============================================================================= 17 // 18 // This file implements the linalg dialect Promotion pass. 19 // 20 //===----------------------------------------------------------------------===// 21 22 #include "mlir/Dialect/Linalg/IR/LinalgOps.h" 23 #include "mlir/Dialect/Linalg/IR/LinalgTypes.h" 24 #include "mlir/Dialect/Linalg/Passes.h" 25 #include "mlir/Dialect/Linalg/Utils/Intrinsics.h" 26 #include "mlir/Dialect/Linalg/Utils/Utils.h" 27 #include "mlir/Dialect/LoopOps/LoopOps.h" 28 #include "mlir/EDSC/Helpers.h" 29 #include "mlir/IR/AffineExpr.h" 30 #include "mlir/IR/AffineExprVisitor.h" 31 #include "mlir/IR/AffineMap.h" 32 #include "mlir/IR/OpImplementation.h" 33 #include "mlir/Pass/Pass.h" 34 #include "mlir/Support/LLVM.h" 35 #include "mlir/Support/STLExtras.h" 36 #include "mlir/Transforms/FoldUtils.h" 37 38 #include "llvm/ADT/SetVector.h" 39 #include "llvm/Support/CommandLine.h" 40 41 using namespace mlir; 42 using namespace mlir::edsc; 43 using namespace mlir::edsc::intrinsics; 44 using namespace mlir::linalg; 45 using namespace mlir::linalg::intrinsics; 46 using namespace mlir::loop; 47 48 using llvm::SetVector; 49 50 #define DEBUG_TYPE "linalg-promotion" 51 52 static llvm::cl::OptionCategory clOptionsCategory(DEBUG_TYPE " options"); 53 static llvm::cl::opt<bool> clPromoteDynamic( 54 "test-linalg-promote-dynamic", 55 llvm::cl::desc("Test generation of dynamic promoted buffers"), 56 llvm::cl::cat(clOptionsCategory), llvm::cl::init(false)); 57 58 static AffineMap getAffineDifferenceMap(MLIRContext *context) { 59 AffineExpr d0(getAffineDimExpr(0, context)), d1(getAffineDimExpr(1, context)); 60 return AffineMap::get(2, 0, {d0 - d1}); 61 } 62 63 static Value *allocBuffer(Type elementType, Value *size, bool dynamicBuffers) { 64 auto *ctx = size->getContext(); 65 auto width = llvm::divideCeil(elementType.getIntOrFloatBitWidth(), 8); 66 if (!dynamicBuffers) 67 if (auto cst = dyn_cast_or_null<ConstantIndexOp>(size->getDefiningOp())) 68 return alloc( 69 MemRefType::get(width * cst.getValue(), IntegerType::get(8, ctx))); 70 Value *mul = muli(constant_index(width), size); 71 return alloc(MemRefType::get(-1, IntegerType::get(8, ctx)), mul); 72 } 73 74 // Performs promotion of a `subView` into a local buffer of the size of the 75 // *ranges* of the `subView`. This produces a buffer whose size may be bigger 76 // than the actual size of the `subView` at the boundaries. 77 // This is related to the full/partial tile problem. 78 // Returns a PromotionInfo containing a `buffer`, `fullLocalView` and 79 // `partialLocalView` such that: 80 // * `buffer` is always the size of the full tile. 81 // * `fullLocalView` is a dense contiguous view into that buffer. 82 // * `partialLocalView` is a dense non-contiguous slice of `fullLocalView` 83 // that corresponds to the size of `subView` and accounting for boundary 84 // effects. 85 // The point of the full tile buffer is that constant static tile sizes are 86 // folded and result in a buffer type with statically known size and alignment 87 // properties. 88 // To account for general boundary effects, padding must be performed on the 89 // boundary tiles. For now this is done with an unconditional `fill` op followed 90 // by a partial `copy` op. 91 static PromotionInfo promoteFullTileBuffer(OpBuilder &b, Location loc, 92 mlir::linalg::SubViewOp subView, 93 bool dynamicBuffers, 94 OperationFolder *folder) { 95 auto zero = constant_index(folder, 0); 96 auto one = constant_index(folder, 1); 97 98 auto viewType = subView.getViewType(); 99 auto rank = viewType.getRank(); 100 Value *allocSize = one; 101 SmallVector<Value *, 8> fullRanges, partialRanges; 102 fullRanges.reserve(rank); 103 partialRanges.reserve(rank); 104 for (auto en : llvm::enumerate(subView.getRanges())) { 105 auto rank = en.index(); 106 auto rangeValue = en.value(); 107 Value *d = 108 isa<DimOp>(rangeValue.max->getDefiningOp()) 109 ? rangeValue.max 110 : applyMapToValues(b, loc, getAffineDifferenceMap(b.getContext()), 111 {rangeValue.max, rangeValue.min}, folder) 112 .front(); 113 allocSize = muli(folder, allocSize, d).getValue(); 114 fullRanges.push_back(d); 115 partialRanges.push_back(range(folder, zero, dim(subView, rank), one)); 116 } 117 SmallVector<int64_t, 4> dynSizes(fullRanges.size(), -1); 118 auto *buffer = 119 allocBuffer(viewType.getElementType(), allocSize, dynamicBuffers); 120 auto fullLocalView = view( 121 MemRefType::get(dynSizes, viewType.getElementType()), buffer, fullRanges); 122 auto partialLocalView = slice(fullLocalView, partialRanges); 123 return PromotionInfo{buffer, fullLocalView, partialLocalView}; 124 } 125 126 SmallVector<PromotionInfo, 8> 127 mlir::linalg::promoteSubViews(OpBuilder &b, Location loc, 128 ArrayRef<Value *> subViews, bool dynamicBuffers, 129 OperationFolder *folder) { 130 if (subViews.empty()) 131 return {}; 132 133 ScopedContext scope(b, loc); 134 SmallVector<PromotionInfo, 8> res; 135 res.reserve(subViews.size()); 136 DenseMap<Value *, PromotionInfo> promotionInfoMap; 137 for (auto *v : subViews) { 138 mlir::linalg::SubViewOp subView = 139 cast<mlir::linalg::SubViewOp>(v->getDefiningOp()); 140 auto viewType = subView.getViewType(); 141 // TODO(ntv): support more cases than just float. 142 if (!viewType.getElementType().isa<FloatType>()) 143 continue; 144 auto promotionInfo = 145 promoteFullTileBuffer(b, loc, subView, dynamicBuffers, folder); 146 promotionInfoMap.insert(std::make_pair(subView.getResult(), promotionInfo)); 147 res.push_back(promotionInfo); 148 } 149 150 for (auto *v : subViews) { 151 mlir::linalg::SubViewOp subView = 152 cast<mlir::linalg::SubViewOp>(v->getDefiningOp()); 153 auto info = promotionInfoMap.find(v); 154 if (info == promotionInfoMap.end()) 155 continue; 156 // TODO(ntv): value to fill with should be related to the operation. 157 // For now, just use APFloat(0.0f). 158 auto t = subView.getViewType().getElementType().cast<FloatType>(); 159 Value *fillVal = constant_float(folder, APFloat(0.0f), t); 160 // TODO(ntv): fill is only necessary if `promotionInfo` has a full local 161 // view that is different from the partial local view and we are on the 162 // boundary. 163 fill(info->second.fullLocalView, fillVal); 164 } 165 166 for (auto *v : subViews) { 167 auto info = promotionInfoMap.find(v); 168 if (info == promotionInfoMap.end()) 169 continue; 170 copy(cast<mlir::linalg::SubViewOp>(v->getDefiningOp()), 171 info->second.partialLocalView); 172 } 173 return res; 174 } 175 176 static void promoteSubViewOperands(LinalgOp op, SetVector<Value *> subViews, 177 bool dynamicBuffers, 178 OperationFolder *folder) { 179 // 1. Promote the specified views and use them in the new op. 180 OpBuilder b(op); 181 ScopedContext scope(b, op.getLoc()); 182 auto promotedBufferAndViews = promoteSubViews( 183 b, op.getLoc(), subViews.getArrayRef(), dynamicBuffers, folder); 184 SmallVector<Value *, 8> opViews; 185 opViews.reserve(op.getNumInputsAndOutputs()); 186 SmallVector<std::pair<Value *, Value *>, 8> writebackViews; 187 writebackViews.reserve(subViews.size()); 188 unsigned promotedIdx = 0; 189 for (auto *view : op.getInputsAndOutputs()) { 190 if (subViews.count(view) != 0) { 191 opViews.push_back(promotedBufferAndViews[promotedIdx].fullLocalView); 192 writebackViews.emplace_back(std::make_pair( 193 view, promotedBufferAndViews[promotedIdx].partialLocalView)); 194 promotedIdx++; 195 } else { 196 opViews.push_back(view); 197 } 198 } 199 200 // 2. Append all other operands as they appear, this enforces that such 201 // operands are not views. This is to support cases such as FillOp taking 202 // extra scalars etc. 203 auto operands = getAssumedNonViewOperands(op); 204 opViews.append(operands.begin(), operands.end()); 205 op.clone(b, op.getLoc(), opViews); 206 207 // 3. Emit write-back for the promoted output views: copy the partial view. 208 for (auto viewAndPartialLocalView : writebackViews) { 209 // Note: use the old op to determine whether the operand view is an output. 210 bool isOutput = 211 op.getIndexOfOutput(viewAndPartialLocalView.first).hasValue(); 212 if (isOutput) 213 copy(viewAndPartialLocalView.second, viewAndPartialLocalView.first); 214 } 215 216 // 4. Dealloc local buffers. 217 for (const auto &pi : promotedBufferAndViews) 218 dealloc(pi.buffer); 219 } 220 221 static void promoteSubViews(FuncOp f, bool dynamicBuffers) { 222 SmallVector<LinalgOp, 8> toErase; 223 OperationFolder folder(f.getContext()); 224 f.walk([dynamicBuffers, &folder, &toErase](LinalgOp op) { 225 // TODO(ntv) some heuristic here to decide what to promote. Atm it is all or 226 // nothing. 227 SetVector<Value *> subViews; 228 for (auto it : op.getInputsAndOutputs()) 229 if (auto sv = 230 dyn_cast_or_null<mlir::linalg::SubViewOp>(it->getDefiningOp())) 231 subViews.insert(sv); 232 if (!subViews.empty()) { 233 promoteSubViewOperands(op, subViews, dynamicBuffers, &folder); 234 toErase.push_back(op); 235 } 236 }); 237 for (auto op : toErase) 238 op.erase(); 239 } 240 241 namespace { 242 struct LinalgPromotionPass : public FunctionPass<LinalgPromotionPass> { 243 LinalgPromotionPass() = default; 244 LinalgPromotionPass(bool dynamicBuffers) : dynamicBuffers(dynamicBuffers) {} 245 246 void runOnFunction() override { 247 promoteSubViews(getFunction(), dynamicBuffers); 248 } 249 250 bool dynamicBuffers; 251 }; 252 } // namespace 253 254 std::unique_ptr<OpPassBase<FuncOp>> 255 mlir::linalg::createLinalgPromotionPass(bool dynamicBuffers) { 256 return std::make_unique<LinalgPromotionPass>(dynamicBuffers); 257 } 258 259 static PassRegistration<LinalgPromotionPass> 260 pass("linalg-promote-subviews", "promote subview ops to local buffers", [] { 261 return std::make_unique<LinalgPromotionPass>(clPromoteDynamic); 262 }); 263