1 //===- Promotion.cpp - Implementation of linalg Promotion -----------------===// 2 // 3 // Part of the MLIR 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/Linalg/IR/LinalgOps.h" 14 #include "mlir/Dialect/Linalg/IR/LinalgTypes.h" 15 #include "mlir/Dialect/Linalg/Passes.h" 16 #include "mlir/Dialect/Linalg/Utils/Intrinsics.h" 17 #include "mlir/Dialect/Linalg/Utils/Utils.h" 18 #include "mlir/Dialect/LoopOps/LoopOps.h" 19 #include "mlir/EDSC/Helpers.h" 20 #include "mlir/IR/AffineExpr.h" 21 #include "mlir/IR/AffineExprVisitor.h" 22 #include "mlir/IR/AffineMap.h" 23 #include "mlir/IR/OpImplementation.h" 24 #include "mlir/Pass/Pass.h" 25 #include "mlir/Support/LLVM.h" 26 #include "mlir/Support/STLExtras.h" 27 #include "mlir/Transforms/FoldUtils.h" 28 29 #include "llvm/ADT/SetVector.h" 30 #include "llvm/Support/CommandLine.h" 31 32 using namespace mlir; 33 using namespace mlir::edsc; 34 using namespace mlir::edsc::intrinsics; 35 using namespace mlir::linalg; 36 using namespace mlir::linalg::intrinsics; 37 using namespace mlir::loop; 38 39 using llvm::SetVector; 40 41 #define DEBUG_TYPE "linalg-promotion" 42 43 static llvm::cl::OptionCategory clOptionsCategory(DEBUG_TYPE " options"); 44 static llvm::cl::opt<bool> clPromoteDynamic( 45 "test-linalg-promote-dynamic", 46 llvm::cl::desc("Test generation of dynamic promoted buffers"), 47 llvm::cl::cat(clOptionsCategory), llvm::cl::init(false)); 48 49 static Value allocBuffer(Type elementType, Value size, bool dynamicBuffers) { 50 auto *ctx = size.getContext(); 51 auto width = llvm::divideCeil(elementType.getIntOrFloatBitWidth(), 8); 52 if (!dynamicBuffers) 53 if (auto cst = dyn_cast_or_null<ConstantIndexOp>(size.getDefiningOp())) 54 return alloc( 55 MemRefType::get(width * cst.getValue(), IntegerType::get(8, ctx))); 56 Value mul = muli(constant_index(width), size); 57 return alloc(MemRefType::get(-1, IntegerType::get(8, ctx)), mul); 58 } 59 60 // Performs promotion of a `subView` into a local buffer of the size of the 61 // *ranges* of the `subView`. This produces a buffer whose size may be bigger 62 // than the actual size of the `subView` at the boundaries. 63 // This is related to the full/partial tile problem. 64 // Returns a PromotionInfo containing a `buffer`, `fullLocalView` and 65 // `partialLocalView` such that: 66 // * `buffer` is always the size of the full tile. 67 // * `fullLocalView` is a dense contiguous view into that buffer. 68 // * `partialLocalView` is a dense non-contiguous slice of `fullLocalView` 69 // that corresponds to the size of `subView` and accounting for boundary 70 // effects. 71 // The point of the full tile buffer is that constant static tile sizes are 72 // folded and result in a buffer type with statically known size and alignment 73 // properties. 74 // To account for general boundary effects, padding must be performed on the 75 // boundary tiles. For now this is done with an unconditional `fill` op followed 76 // by a partial `copy` op. 77 static PromotionInfo promoteFullTileBuffer(OpBuilder &b, Location loc, 78 SubViewOp subView, 79 bool dynamicBuffers, 80 OperationFolder *folder) { 81 auto zero = constant_index(folder, 0); 82 auto one = constant_index(folder, 1); 83 84 auto viewType = subView.getType(); 85 auto rank = viewType.getRank(); 86 Value allocSize = one; 87 SmallVector<Value, 8> fullRanges, partialRanges; 88 fullRanges.reserve(rank); 89 partialRanges.reserve(rank); 90 for (auto en : llvm::enumerate(subView.getRanges())) { 91 auto rank = en.index(); 92 auto rangeValue = en.value(); 93 Value d = rangeValue.size; 94 allocSize = muli(folder, allocSize, d).getValue(); 95 fullRanges.push_back(d); 96 partialRanges.push_back(range(folder, zero, dim(subView, rank), one)); 97 } 98 SmallVector<int64_t, 4> dynSizes(fullRanges.size(), -1); 99 auto buffer = 100 allocBuffer(viewType.getElementType(), allocSize, dynamicBuffers); 101 auto fullLocalView = view( 102 MemRefType::get(dynSizes, viewType.getElementType()), buffer, fullRanges); 103 auto partialLocalView = slice(fullLocalView, partialRanges); 104 return PromotionInfo{buffer, fullLocalView, partialLocalView}; 105 } 106 107 SmallVector<PromotionInfo, 8> 108 mlir::linalg::promoteSubViews(OpBuilder &b, Location loc, 109 ArrayRef<Value> subViews, bool dynamicBuffers, 110 OperationFolder *folder) { 111 if (subViews.empty()) 112 return {}; 113 114 ScopedContext scope(b, loc); 115 SmallVector<PromotionInfo, 8> res; 116 res.reserve(subViews.size()); 117 DenseMap<Value, PromotionInfo> promotionInfoMap; 118 for (auto v : subViews) { 119 SubViewOp subView = cast<SubViewOp>(v.getDefiningOp()); 120 auto viewType = subView.getType(); 121 // TODO(ntv): support more cases than just float. 122 if (!viewType.getElementType().isa<FloatType>()) 123 continue; 124 auto promotionInfo = 125 promoteFullTileBuffer(b, loc, subView, dynamicBuffers, folder); 126 promotionInfoMap.insert(std::make_pair(subView.getResult(), promotionInfo)); 127 res.push_back(promotionInfo); 128 } 129 130 for (auto v : subViews) { 131 SubViewOp subView = cast<SubViewOp>(v.getDefiningOp()); 132 auto info = promotionInfoMap.find(v); 133 if (info == promotionInfoMap.end()) 134 continue; 135 // TODO(ntv): value to fill with should be related to the operation. 136 // For now, just use APFloat(0.0f). 137 auto t = subView.getType().getElementType().cast<FloatType>(); 138 Value fillVal = constant_float(folder, APFloat(0.0f), t); 139 // TODO(ntv): fill is only necessary if `promotionInfo` has a full local 140 // view that is different from the partial local view and we are on the 141 // boundary. 142 fill(info->second.fullLocalView, fillVal); 143 } 144 145 for (auto v : subViews) { 146 auto info = promotionInfoMap.find(v); 147 if (info == promotionInfoMap.end()) 148 continue; 149 copy(cast<SubViewOp>(v.getDefiningOp()), info->second.partialLocalView); 150 } 151 return res; 152 } 153 154 LinalgOp mlir::linalg::promoteSubViewOperands(OpBuilder &b, LinalgOp op, 155 SetVector<Value> subViews, 156 bool dynamicBuffers, 157 OperationFolder *folder) { 158 assert(op.hasBufferSemantics() && "expected linalg op with buffer semantics"); 159 160 // 1. Promote the specified views and use them in the new op. 161 ScopedContext scope(b, op.getLoc()); 162 auto promotedBufferAndViews = promoteSubViews( 163 b, op.getLoc(), subViews.getArrayRef(), dynamicBuffers, folder); 164 SmallVector<Value, 8> opViews; 165 opViews.reserve(op.getNumInputsAndOutputs()); 166 SmallVector<std::pair<Value, Value>, 8> writebackViews; 167 writebackViews.reserve(subViews.size()); 168 unsigned promotedIdx = 0; 169 for (auto view : op.getInputsAndOutputBuffers()) { 170 if (subViews.count(view) != 0) { 171 opViews.push_back(promotedBufferAndViews[promotedIdx].fullLocalView); 172 writebackViews.emplace_back(std::make_pair( 173 view, promotedBufferAndViews[promotedIdx].partialLocalView)); 174 promotedIdx++; 175 } else { 176 opViews.push_back(view); 177 } 178 } 179 180 // 2. Append all other operands as they appear, this enforces that such 181 // operands are not views. This is to support cases such as FillOp taking 182 // extra scalars etc. 183 auto operands = getAssumedNonViewOperands(op); 184 opViews.append(operands.begin(), operands.end()); 185 LinalgOp res = op.clone(b, op.getLoc(), opViews); 186 187 // 3. Emit write-back for the promoted output views: copy the partial view. 188 for (auto viewAndPartialLocalView : writebackViews) { 189 // WARNING: MUST use the old op to determine whether the operand view is an 190 // output. 191 bool isOutput = 192 op.getIndexOfOutputBuffer(viewAndPartialLocalView.first).hasValue(); 193 if (isOutput) 194 copy(viewAndPartialLocalView.second, viewAndPartialLocalView.first); 195 } 196 197 // 4. Dealloc local buffers. 198 for (const auto &pi : promotedBufferAndViews) 199 dealloc(pi.buffer); 200 201 return res; 202 } 203 204 static void promoteSubViews(FuncOp f, bool dynamicBuffers) { 205 SmallVector<LinalgOp, 8> toErase; 206 OperationFolder folder(f.getContext()); 207 f.walk([dynamicBuffers, &folder, &toErase](LinalgOp op) { 208 if (!op.hasBufferSemantics()) 209 return; 210 211 // TODO(ntv) some heuristic here to decide what to promote. Atm it is all or 212 // nothing. 213 SetVector<Value> subViews; 214 OpBuilder b(op); 215 for (auto it : op.getInputsAndOutputBuffers()) 216 if (auto sv = dyn_cast_or_null<SubViewOp>(it.getDefiningOp())) 217 subViews.insert(sv); 218 if (!subViews.empty()) { 219 promoteSubViewOperands(b, op, subViews, dynamicBuffers, &folder); 220 toErase.push_back(op); 221 } 222 }); 223 for (auto op : toErase) 224 op.erase(); 225 } 226 227 namespace { 228 struct LinalgPromotionPass : public FunctionPass<LinalgPromotionPass> { 229 LinalgPromotionPass() = default; 230 LinalgPromotionPass(bool dynamicBuffers) : dynamicBuffers(dynamicBuffers) {} 231 232 void runOnFunction() override { 233 promoteSubViews(getFunction(), dynamicBuffers); 234 } 235 236 bool dynamicBuffers; 237 }; 238 } // namespace 239 240 std::unique_ptr<OpPassBase<FuncOp>> 241 mlir::linalg::createLinalgPromotionPass(bool dynamicBuffers) { 242 return std::make_unique<LinalgPromotionPass>(dynamicBuffers); 243 } 244 245 static PassRegistration<LinalgPromotionPass> 246 pass("linalg-promote-subviews", "promote subview ops to local buffers", [] { 247 return std::make_unique<LinalgPromotionPass>(clPromoteDynamic); 248 }); 249