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/AffineOps/EDSC/Intrinsics.h" 14 #include "mlir/Dialect/Linalg/EDSC/Intrinsics.h" 15 #include "mlir/Dialect/Linalg/IR/LinalgOps.h" 16 #include "mlir/Dialect/Linalg/IR/LinalgTypes.h" 17 #include "mlir/Dialect/Linalg/Passes.h" 18 #include "mlir/Dialect/Linalg/Utils/Utils.h" 19 #include "mlir/Dialect/LoopOps/LoopOps.h" 20 #include "mlir/Dialect/StandardOps/EDSC/Intrinsics.h" 21 #include "mlir/IR/AffineExpr.h" 22 #include "mlir/IR/AffineExprVisitor.h" 23 #include "mlir/IR/AffineMap.h" 24 #include "mlir/IR/OpImplementation.h" 25 #include "mlir/Pass/Pass.h" 26 #include "mlir/Support/LLVM.h" 27 #include "mlir/Support/STLExtras.h" 28 #include "mlir/Transforms/FoldUtils.h" 29 30 #include "llvm/ADT/SetVector.h" 31 #include "llvm/Support/CommandLine.h" 32 33 using namespace mlir; 34 using namespace mlir::edsc; 35 using namespace mlir::edsc::intrinsics; 36 using namespace mlir::linalg; 37 using namespace mlir::loop; 38 39 using llvm::SetVector; 40 41 using folded_affine_min = folded::ValueBuilder<AffineMinOp>; 42 using folded_linalg_range = folded::ValueBuilder<linalg::RangeOp>; 43 44 #define DEBUG_TYPE "linalg-promotion" 45 46 static llvm::cl::OptionCategory clOptionsCategory(DEBUG_TYPE " options"); 47 static llvm::cl::opt<bool> clPromoteDynamic( 48 "test-linalg-promote-dynamic", 49 llvm::cl::desc("Test generation of dynamic promoted buffers"), 50 llvm::cl::cat(clOptionsCategory), llvm::cl::init(false)); 51 52 static Value allocBuffer(Type elementType, Value size, bool dynamicBuffers) { 53 auto *ctx = size.getContext(); 54 auto width = llvm::divideCeil(elementType.getIntOrFloatBitWidth(), 8); 55 if (!dynamicBuffers) 56 if (auto cst = dyn_cast_or_null<ConstantIndexOp>(size.getDefiningOp())) 57 return std_alloc( 58 MemRefType::get(width * cst.getValue(), IntegerType::get(8, ctx))); 59 Value mul = std_muli(std_constant_index(width), size); 60 return std_alloc(MemRefType::get(-1, IntegerType::get(8, ctx)), mul); 61 } 62 63 // Performs promotion of a `subView` into a local buffer of the size of the 64 // *ranges* of the `subView`. This produces a buffer whose size may be bigger 65 // than the actual size of the `subView` at the boundaries. 66 // This is related to the full/partial tile problem. 67 // Returns a PromotionInfo containing a `buffer`, `fullLocalView` and 68 // `partialLocalView` such that: 69 // * `buffer` is always the size of the full tile. 70 // * `fullLocalView` is a dense contiguous view into that buffer. 71 // * `partialLocalView` is a dense non-contiguous slice of `fullLocalView` 72 // that corresponds to the size of `subView` and accounting for boundary 73 // effects. 74 // The point of the full tile buffer is that constant static tile sizes are 75 // folded and result in a buffer type with statically known size and alignment 76 // properties. 77 // To account for general boundary effects, padding must be performed on the 78 // boundary tiles. For now this is done with an unconditional `fill` op followed 79 // by a partial `copy` op. 80 static PromotionInfo promoteFullTileBuffer(OpBuilder &b, Location loc, 81 SubViewOp subView, 82 bool dynamicBuffers, 83 OperationFolder *folder) { 84 auto zero = folded_std_constant_index(folder, 0); 85 auto one = folded_std_constant_index(folder, 1); 86 87 auto viewType = subView.getType(); 88 auto rank = viewType.getRank(); 89 Value allocSize = one; 90 SmallVector<Value, 8> fullRanges, partialRanges; 91 fullRanges.reserve(rank); 92 partialRanges.reserve(rank); 93 for (auto en : llvm::enumerate(subView.getRanges())) { 94 auto rank = en.index(); 95 auto rangeValue = en.value(); 96 Value d = rangeValue.size; 97 allocSize = folded_std_muli(folder, allocSize, d).getValue(); 98 fullRanges.push_back(d); 99 partialRanges.push_back( 100 folded_linalg_range(folder, zero, std_dim(subView, rank), one)); 101 } 102 SmallVector<int64_t, 4> dynSizes(fullRanges.size(), -1); 103 auto buffer = 104 allocBuffer(viewType.getElementType(), allocSize, dynamicBuffers); 105 auto fullLocalView = std_view( 106 MemRefType::get(dynSizes, viewType.getElementType()), buffer, fullRanges); 107 auto partialLocalView = linalg_slice(fullLocalView, partialRanges); 108 return PromotionInfo{buffer, fullLocalView, partialLocalView}; 109 } 110 111 SmallVector<PromotionInfo, 8> 112 mlir::linalg::promoteSubViews(OpBuilder &b, Location loc, 113 ArrayRef<Value> subViews, bool dynamicBuffers, 114 OperationFolder *folder) { 115 if (subViews.empty()) 116 return {}; 117 118 ScopedContext scope(b, loc); 119 SmallVector<PromotionInfo, 8> res; 120 res.reserve(subViews.size()); 121 DenseMap<Value, PromotionInfo> promotionInfoMap; 122 for (auto v : subViews) { 123 SubViewOp subView = cast<SubViewOp>(v.getDefiningOp()); 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 Value fillVal; 136 if (auto t = subView.getType().getElementType().dyn_cast<FloatType>()) 137 fillVal = folded_std_constant(folder, FloatAttr::get(t, 0.0)); 138 else if (auto t = 139 subView.getType().getElementType().dyn_cast<IntegerType>()) 140 fillVal = folded_std_constant_int(folder, 0, t); 141 // TODO(ntv): fill is only necessary if `promotionInfo` has a full local 142 // view that is different from the partial local view and we are on the 143 // boundary. 144 linalg_fill(info->second.fullLocalView, fillVal); 145 } 146 147 for (auto v : subViews) { 148 auto info = promotionInfoMap.find(v); 149 if (info == promotionInfoMap.end()) 150 continue; 151 linalg_copy(cast<SubViewOp>(v.getDefiningOp()), 152 info->second.partialLocalView); 153 } 154 return res; 155 } 156 157 LinalgOp mlir::linalg::promoteSubViewOperands(OpBuilder &b, LinalgOp op, 158 SetVector<Value> subViews, 159 bool dynamicBuffers, 160 OperationFolder *folder) { 161 assert(op.hasBufferSemantics() && "expected linalg op with buffer semantics"); 162 163 if (auto convOp = dyn_cast<linalg::ConvOp>(op.getOperation())) { 164 // TODO(ntv): add a level of indirection to linalg.generic. 165 if (convOp.padding()) 166 llvm_unreachable("Unexpected conv with padding"); 167 } 168 169 // 1. Promote the specified views and use them in the new op. 170 ScopedContext scope(b, op.getLoc()); 171 auto promotedBufferAndViews = promoteSubViews( 172 b, op.getLoc(), subViews.getArrayRef(), dynamicBuffers, folder); 173 SmallVector<Value, 8> opViews; 174 opViews.reserve(op.getNumInputsAndOutputs()); 175 SmallVector<std::pair<Value, Value>, 8> writebackViews; 176 writebackViews.reserve(subViews.size()); 177 unsigned promotedIdx = 0; 178 for (auto view : op.getInputsAndOutputBuffers()) { 179 if (subViews.count(view) != 0) { 180 opViews.push_back(promotedBufferAndViews[promotedIdx].fullLocalView); 181 writebackViews.emplace_back(std::make_pair( 182 view, promotedBufferAndViews[promotedIdx].partialLocalView)); 183 promotedIdx++; 184 } else { 185 opViews.push_back(view); 186 } 187 } 188 189 // 2. Append all other operands as they appear, this enforces that such 190 // operands are not views. This is to support cases such as FillOp taking 191 // extra scalars etc. 192 auto operands = getAssumedNonViewOperands(op); 193 opViews.append(operands.begin(), operands.end()); 194 LinalgOp res = op.clone(b, op.getLoc(), opViews); 195 196 // 3. Emit write-back for the promoted output views: copy the partial view. 197 for (auto viewAndPartialLocalView : writebackViews) { 198 // WARNING: MUST use the old op to determine whether the operand view is an 199 // output. 200 bool isOutput = 201 op.getIndexOfOutputBuffer(viewAndPartialLocalView.first).hasValue(); 202 if (isOutput) 203 linalg_copy(viewAndPartialLocalView.second, 204 viewAndPartialLocalView.first); 205 } 206 207 // 4. Dealloc local buffers. 208 for (const auto &pi : promotedBufferAndViews) 209 std_dealloc(pi.buffer); 210 211 return res; 212 } 213 214 static void promoteSubViews(FuncOp f, bool dynamicBuffers) { 215 SmallVector<LinalgOp, 8> toErase; 216 OperationFolder folder(f.getContext()); 217 f.walk([dynamicBuffers, &folder, &toErase](LinalgOp op) { 218 if (!op.hasBufferSemantics()) 219 return; 220 221 // TODO(ntv) some heuristic here to decide what to promote. Atm only float 222 // and integer buffers can be promoted. 223 SetVector<Value> subViews; 224 OpBuilder b(op); 225 for (auto it : op.getInputsAndOutputBuffers()) 226 if (auto sv = dyn_cast_or_null<SubViewOp>(it.getDefiningOp())) 227 if (sv.getType().getElementType().isSignlessIntOrFloat()) 228 subViews.insert(sv); 229 if (!subViews.empty()) { 230 promoteSubViewOperands(b, op, subViews, dynamicBuffers, &folder); 231 toErase.push_back(op); 232 } 233 }); 234 for (auto op : toErase) 235 op.erase(); 236 } 237 238 namespace { 239 struct LinalgPromotionPass : public FunctionPass<LinalgPromotionPass> { 240 LinalgPromotionPass() = default; 241 LinalgPromotionPass(bool dynamicBuffers) : dynamicBuffers(dynamicBuffers) {} 242 243 void runOnFunction() override { 244 promoteSubViews(getFunction(), dynamicBuffers); 245 } 246 247 bool dynamicBuffers; 248 }; 249 } // namespace 250 251 std::unique_ptr<OpPassBase<FuncOp>> 252 mlir::createLinalgPromotionPass(bool dynamicBuffers) { 253 return std::make_unique<LinalgPromotionPass>(dynamicBuffers); 254 } 255 256 static PassRegistration<LinalgPromotionPass> 257 pass("linalg-promote-subviews", "promote subview ops to local buffers", [] { 258 return std::make_unique<LinalgPromotionPass>(clPromoteDynamic); 259 }); 260