| 1 | //===- VectorTransferSplitRewritePatterns.cpp - Transfer Split Rewrites ---===// |
| 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 target-independent patterns to rewrite a vector.transfer |
| 10 | // op into a fully in-bounds part and a partial part. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include <optional> |
| 15 | |
| 16 | #include "mlir/Dialect/Affine/IR/AffineOps.h" |
| 17 | #include "mlir/Dialect/Arith/IR/Arith.h" |
| 18 | #include "mlir/Dialect/MemRef/IR/MemRef.h" |
| 19 | #include "mlir/Dialect/SCF/IR/SCF.h" |
| 20 | #include "mlir/Dialect/Utils/StructuredOpsUtils.h" |
| 21 | #include "llvm/ADT/SmallVectorExtras.h" |
| 22 | |
| 23 | #include "mlir/Dialect/Vector/Transforms/VectorTransforms.h" |
| 24 | #include "mlir/IR/PatternMatch.h" |
| 25 | #include "mlir/Interfaces/VectorInterfaces.h" |
| 26 | |
| 27 | #include "llvm/ADT/STLExtras.h" |
| 28 | #include "llvm/ADT/SmallVectorExtras.h" |
| 29 | |
| 30 | #define DEBUG_TYPE "vector-transfer-split" |
| 31 | |
| 32 | using namespace mlir; |
| 33 | using namespace mlir::vector; |
| 34 | |
| 35 | /// Build the condition to ensure that a particular VectorTransferOpInterface |
| 36 | /// is in-bounds. |
| 37 | static Value createInBoundsCond(RewriterBase &b, |
| 38 | VectorTransferOpInterface xferOp) { |
| 39 | assert(xferOp.getPermutationMap().isMinorIdentity() && |
| 40 | "Expected minor identity map" ); |
| 41 | Value inBoundsCond; |
| 42 | xferOp.zipResultAndIndexing(fun: [&](int64_t resultIdx, int64_t indicesIdx) { |
| 43 | // Zip over the resulting vector shape and memref indices. |
| 44 | // If the dimension is known to be in-bounds, it does not participate in |
| 45 | // the construction of `inBoundsCond`. |
| 46 | if (xferOp.isDimInBounds(dim: resultIdx)) |
| 47 | return; |
| 48 | // Fold or create the check that `index + vector_size` <= `memref_size`. |
| 49 | Location loc = xferOp.getLoc(); |
| 50 | int64_t vectorSize = xferOp.getVectorType().getDimSize(idx: resultIdx); |
| 51 | OpFoldResult sum = affine::makeComposedFoldedAffineApply( |
| 52 | b, loc, expr: b.getAffineDimExpr(position: 0) + b.getAffineConstantExpr(constant: vectorSize), |
| 53 | operands: {xferOp.getIndices()[indicesIdx]}); |
| 54 | OpFoldResult dimSz = |
| 55 | memref::getMixedSize(builder&: b, loc, value: xferOp.getBase(), dim: indicesIdx); |
| 56 | auto maybeCstSum = getConstantIntValue(ofr: sum); |
| 57 | auto maybeCstDimSz = getConstantIntValue(ofr: dimSz); |
| 58 | if (maybeCstSum && maybeCstDimSz && *maybeCstSum <= *maybeCstDimSz) |
| 59 | return; |
| 60 | Value cond = |
| 61 | arith::CmpIOp::create(builder&: b, location: loc, predicate: arith::CmpIPredicate::sle, |
| 62 | lhs: getValueOrCreateConstantIndexOp(b, loc, ofr: sum), |
| 63 | rhs: getValueOrCreateConstantIndexOp(b, loc, ofr: dimSz)); |
| 64 | // Conjunction over all dims for which we are in-bounds. |
| 65 | if (inBoundsCond) |
| 66 | inBoundsCond = arith::AndIOp::create(builder&: b, location: loc, lhs: inBoundsCond, rhs: cond); |
| 67 | else |
| 68 | inBoundsCond = cond; |
| 69 | }); |
| 70 | return inBoundsCond; |
| 71 | } |
| 72 | |
| 73 | /// Split a vector.transfer operation into an in-bounds (i.e., no out-of-bounds |
| 74 | /// masking) fast path and a slow path. |
| 75 | /// If `ifOp` is not null and the result is `success, the `ifOp` points to the |
| 76 | /// newly created conditional upon function return. |
| 77 | /// To accommodate for the fact that the original vector.transfer indexing may |
| 78 | /// be arbitrary and the slow path indexes @[0...0] in the temporary buffer, the |
| 79 | /// scf.if op returns a view and values of type index. |
| 80 | /// At this time, only vector.transfer_read case is implemented. |
| 81 | /// |
| 82 | /// Example (a 2-D vector.transfer_read): |
| 83 | /// ``` |
| 84 | /// %1 = vector.transfer_read %0[...], %pad : memref<A...>, vector<...> |
| 85 | /// ``` |
| 86 | /// is transformed into: |
| 87 | /// ``` |
| 88 | /// %1:3 = scf.if (%inBounds) { |
| 89 | /// // fast path, direct cast |
| 90 | /// memref.cast %A: memref<A...> to compatibleMemRefType |
| 91 | /// scf.yield %view : compatibleMemRefType, index, index |
| 92 | /// } else { |
| 93 | /// // slow path, not in-bounds vector.transfer or memref.copy. |
| 94 | /// memref.cast %alloc: memref<B...> to compatibleMemRefType |
| 95 | /// scf.yield %4 : compatibleMemRefType, index, index |
| 96 | // } |
| 97 | /// %0 = vector.transfer_read %1#0[%1#1, %1#2] {in_bounds = [true ... true]} |
| 98 | /// ``` |
| 99 | /// where `alloc` is a top of the function alloca'ed buffer of one vector. |
| 100 | /// |
| 101 | /// Preconditions: |
| 102 | /// 1. `xferOp.getPermutationMap()` must be a minor identity map |
| 103 | /// 2. the rank of the `xferOp.memref()` and the rank of the |
| 104 | /// `xferOp.getVector()` must be equal. This will be relaxed in the future |
| 105 | /// but requires rank-reducing subviews. |
| 106 | static LogicalResult |
| 107 | splitFullAndPartialTransferPrecondition(VectorTransferOpInterface xferOp) { |
| 108 | // TODO: support 0-d corner case. |
| 109 | if (xferOp.getTransferRank() == 0) |
| 110 | return failure(); |
| 111 | |
| 112 | // TODO: expand support to these 2 cases. |
| 113 | if (!xferOp.getPermutationMap().isMinorIdentity()) |
| 114 | return failure(); |
| 115 | // Must have some out-of-bounds dimension to be a candidate for splitting. |
| 116 | if (!xferOp.hasOutOfBoundsDim()) |
| 117 | return failure(); |
| 118 | // Don't split transfer operations directly under IfOp, this avoids applying |
| 119 | // the pattern recursively. |
| 120 | // TODO: improve the filtering condition to make it more applicable. |
| 121 | if (isa<scf::IfOp>(Val: xferOp->getParentOp())) |
| 122 | return failure(); |
| 123 | return success(); |
| 124 | } |
| 125 | |
| 126 | /// Given two MemRefTypes `aT` and `bT`, return a MemRefType to which both can |
| 127 | /// be cast. If the MemRefTypes don't have the same rank or are not strided, |
| 128 | /// return null; otherwise: |
| 129 | /// 1. if `aT` and `bT` are cast-compatible, return `aT`. |
| 130 | /// 2. else return a new MemRefType obtained by iterating over the shape and |
| 131 | /// strides and: |
| 132 | /// a. keeping the ones that are static and equal across `aT` and `bT`. |
| 133 | /// b. using a dynamic shape and/or stride for the dimensions that don't |
| 134 | /// agree. |
| 135 | static MemRefType getCastCompatibleMemRefType(MemRefType aT, MemRefType bT) { |
| 136 | if (memref::CastOp::areCastCompatible(inputs: aT, outputs: bT)) |
| 137 | return aT; |
| 138 | if (aT.getRank() != bT.getRank()) |
| 139 | return MemRefType(); |
| 140 | int64_t aOffset, bOffset; |
| 141 | SmallVector<int64_t, 4> aStrides, bStrides; |
| 142 | if (failed(Result: aT.getStridesAndOffset(strides&: aStrides, offset&: aOffset)) || |
| 143 | failed(Result: bT.getStridesAndOffset(strides&: bStrides, offset&: bOffset)) || |
| 144 | aStrides.size() != bStrides.size()) |
| 145 | return MemRefType(); |
| 146 | |
| 147 | ArrayRef<int64_t> aShape = aT.getShape(), bShape = bT.getShape(); |
| 148 | int64_t resOffset; |
| 149 | SmallVector<int64_t, 4> resShape(aT.getRank(), 0), |
| 150 | resStrides(bT.getRank(), 0); |
| 151 | for (int64_t idx = 0, e = aT.getRank(); idx < e; ++idx) { |
| 152 | resShape[idx] = |
| 153 | (aShape[idx] == bShape[idx]) ? aShape[idx] : ShapedType::kDynamic; |
| 154 | resStrides[idx] = |
| 155 | (aStrides[idx] == bStrides[idx]) ? aStrides[idx] : ShapedType::kDynamic; |
| 156 | } |
| 157 | resOffset = (aOffset == bOffset) ? aOffset : ShapedType::kDynamic; |
| 158 | return MemRefType::get( |
| 159 | shape: resShape, elementType: aT.getElementType(), |
| 160 | layout: StridedLayoutAttr::get(context: aT.getContext(), offset: resOffset, strides: resStrides)); |
| 161 | } |
| 162 | |
| 163 | /// Casts the given memref to a compatible memref type. If the source memref has |
| 164 | /// a different address space than the target type, a `memref.memory_space_cast` |
| 165 | /// is first inserted, followed by a `memref.cast`. |
| 166 | static Value castToCompatibleMemRefType(OpBuilder &b, Value memref, |
| 167 | MemRefType compatibleMemRefType) { |
| 168 | MemRefType sourceType = cast<MemRefType>(Val: memref.getType()); |
| 169 | Value res = memref; |
| 170 | if (sourceType.getMemorySpace() != compatibleMemRefType.getMemorySpace()) { |
| 171 | sourceType = MemRefType::get( |
| 172 | shape: sourceType.getShape(), elementType: sourceType.getElementType(), |
| 173 | layout: sourceType.getLayout(), memorySpace: compatibleMemRefType.getMemorySpace()); |
| 174 | res = |
| 175 | memref::MemorySpaceCastOp::create(builder&: b, location: memref.getLoc(), dest: sourceType, source: res); |
| 176 | } |
| 177 | if (sourceType == compatibleMemRefType) |
| 178 | return res; |
| 179 | return memref::CastOp::create(builder&: b, location: memref.getLoc(), dest: compatibleMemRefType, source: res); |
| 180 | } |
| 181 | |
| 182 | /// Operates under a scoped context to build the intersection between the |
| 183 | /// view `xferOp.getbase()` @ `xferOp.getIndices()` and the view `alloc`. |
| 184 | // TODO: view intersection/union/differences should be a proper std op. |
| 185 | static std::pair<Value, Value> |
| 186 | createSubViewIntersection(RewriterBase &b, VectorTransferOpInterface xferOp, |
| 187 | Value alloc) { |
| 188 | Location loc = xferOp.getLoc(); |
| 189 | int64_t memrefRank = xferOp.getShapedType().getRank(); |
| 190 | // TODO: relax this precondition, will require rank-reducing subviews. |
| 191 | assert(memrefRank == cast<MemRefType>(alloc.getType()).getRank() && |
| 192 | "Expected memref rank to match the alloc rank" ); |
| 193 | ValueRange leadingIndices = |
| 194 | xferOp.getIndices().take_front(n: xferOp.getLeadingShapedRank()); |
| 195 | SmallVector<OpFoldResult, 4> sizes; |
| 196 | sizes.append(in_start: leadingIndices.begin(), in_end: leadingIndices.end()); |
| 197 | auto isaWrite = isa<vector::TransferWriteOp>(Val: xferOp); |
| 198 | xferOp.zipResultAndIndexing(fun: [&](int64_t resultIdx, int64_t indicesIdx) { |
| 199 | using MapList = ArrayRef<ArrayRef<AffineExpr>>; |
| 200 | Value dimMemRef = |
| 201 | memref::DimOp::create(builder&: b, location: xferOp.getLoc(), source: xferOp.getBase(), index: indicesIdx); |
| 202 | Value dimAlloc = memref::DimOp::create(builder&: b, location: loc, source: alloc, index: resultIdx); |
| 203 | Value index = xferOp.getIndices()[indicesIdx]; |
| 204 | AffineExpr i, j, k; |
| 205 | bindDims(ctx: xferOp.getContext(), exprs&: i, exprs&: j, exprs&: k); |
| 206 | SmallVector<AffineMap, 4> maps = |
| 207 | AffineMap::inferFromExprList(exprsList: MapList{{i - j, k}}, context: b.getContext()); |
| 208 | // affine_min(%dimMemRef - %index, %dimAlloc) |
| 209 | Value affineMin = |
| 210 | affine::AffineMinOp::create(builder&: b, location: loc, resultType0: index.getType(), map: maps[0], |
| 211 | operands: ValueRange{dimMemRef, index, dimAlloc}); |
| 212 | sizes.push_back(Elt: affineMin); |
| 213 | }); |
| 214 | |
| 215 | SmallVector<OpFoldResult> srcIndices = llvm::map_to_vector<4>( |
| 216 | C: xferOp.getIndices(), F: [](Value idx) -> OpFoldResult { return idx; }); |
| 217 | SmallVector<OpFoldResult> destIndices(memrefRank, b.getIndexAttr(value: 0)); |
| 218 | SmallVector<OpFoldResult> strides(memrefRank, b.getIndexAttr(value: 1)); |
| 219 | auto copySrc = memref::SubViewOp::create( |
| 220 | builder&: b, location: loc, source: isaWrite ? alloc : xferOp.getBase(), offsets: srcIndices, sizes, strides); |
| 221 | auto copyDest = memref::SubViewOp::create( |
| 222 | builder&: b, location: loc, source: isaWrite ? xferOp.getBase() : alloc, offsets: destIndices, sizes, strides); |
| 223 | return std::make_pair(x&: copySrc, y&: copyDest); |
| 224 | } |
| 225 | |
| 226 | /// Given an `xferOp` for which: |
| 227 | /// 1. `inBoundsCond` and a `compatibleMemRefType` have been computed. |
| 228 | /// 2. a memref of single vector `alloc` has been allocated. |
| 229 | /// Produce IR resembling: |
| 230 | /// ``` |
| 231 | /// %1:3 = scf.if (%inBounds) { |
| 232 | /// (memref.memory_space_cast %A: memref<A..., addr_space> to memref<A...>) |
| 233 | /// %view = memref.cast %A: memref<A...> to compatibleMemRefType |
| 234 | /// scf.yield %view, ... : compatibleMemRefType, index, index |
| 235 | /// } else { |
| 236 | /// %2 = vector.broadcast %pad : elt to vector<...> |
| 237 | /// %3 = vector.type_cast %alloc : memref<B...> to memref<vector<...>> |
| 238 | /// memref.store %2, %3[] : memref<vector<...>> |
| 239 | /// %4 = subview %view [...][...][...] |
| 240 | /// %5 = subview %alloc [0, 0] [...] [...] |
| 241 | /// memref.copy %4, %5 |
| 242 | /// %6 = memref.cast %alloc: memref<B...> to compatibleMemRefType |
| 243 | /// scf.yield %6, ... : compatibleMemRefType, index, index |
| 244 | /// } |
| 245 | /// ``` |
| 246 | /// Return the produced scf::IfOp. |
| 247 | static scf::IfOp |
| 248 | createFullPartialMemRefCopy(RewriterBase &b, vector::TransferReadOp xferOp, |
| 249 | TypeRange returnTypes, Value inBoundsCond, |
| 250 | MemRefType compatibleMemRefType, Value alloc) { |
| 251 | Location loc = xferOp.getLoc(); |
| 252 | Value zero = arith::ConstantIndexOp::create(builder&: b, location: loc, value: 0); |
| 253 | Value memref = xferOp.getBase(); |
| 254 | return scf::IfOp::create( |
| 255 | builder&: b, location: loc, cond: inBoundsCond, |
| 256 | thenBuilder: [&](OpBuilder &b, Location loc) { |
| 257 | Value res = castToCompatibleMemRefType(b, memref, compatibleMemRefType); |
| 258 | scf::ValueVector viewAndIndices{res}; |
| 259 | llvm::append_range(C&: viewAndIndices, R: xferOp.getIndices()); |
| 260 | scf::YieldOp::create(builder&: b, location: loc, results: viewAndIndices); |
| 261 | }, |
| 262 | elseBuilder: [&](OpBuilder &b, Location loc) { |
| 263 | // Fill the whole buffer with the padding value. The buffer holds |
| 264 | // exactly one vector, so a single vector store covers it. |
| 265 | Value padding = vector::BroadcastOp::create( |
| 266 | builder&: b, location: loc, result: xferOp.getVectorType(), source: xferOp.getPadding()); |
| 267 | memref::StoreOp::create( |
| 268 | builder&: b, location: loc, valueToStore: padding, |
| 269 | memref: vector::TypeCastOp::create( |
| 270 | builder&: b, location: loc, result: MemRefType::get(shape: {}, elementType: padding.getType()), memref: alloc)); |
| 271 | // Take partial subview of memref which guarantees no dimension |
| 272 | // overflows. |
| 273 | IRRewriter rewriter(b); |
| 274 | std::pair<Value, Value> copyArgs = createSubViewIntersection( |
| 275 | b&: rewriter, xferOp: cast<VectorTransferOpInterface>(Val: xferOp.getOperation()), |
| 276 | alloc); |
| 277 | memref::CopyOp::create(builder&: b, location: loc, source: copyArgs.first, target: copyArgs.second); |
| 278 | Value casted = |
| 279 | castToCompatibleMemRefType(b, memref: alloc, compatibleMemRefType); |
| 280 | scf::ValueVector viewAndIndices{casted}; |
| 281 | viewAndIndices.insert(I: viewAndIndices.end(), NumToInsert: xferOp.getTransferRank(), |
| 282 | Elt: zero); |
| 283 | scf::YieldOp::create(builder&: b, location: loc, results: viewAndIndices); |
| 284 | }); |
| 285 | } |
| 286 | |
| 287 | /// Given an `xferOp` for which: |
| 288 | /// 1. `inBoundsCond` and a `compatibleMemRefType` have been computed. |
| 289 | /// 2. a memref of single vector `alloc` has been allocated. |
| 290 | /// Produce IR resembling: |
| 291 | /// ``` |
| 292 | /// %1:3 = scf.if (%inBounds) { |
| 293 | /// (memref.memory_space_cast %A: memref<A..., addr_space> to memref<A...>) |
| 294 | /// memref.cast %A: memref<A...> to compatibleMemRefType |
| 295 | /// scf.yield %view, ... : compatibleMemRefType, index, index |
| 296 | /// } else { |
| 297 | /// %2 = vector.transfer_read %view[...], %pad : memref<A...>, vector<...> |
| 298 | /// %3 = vector.type_cast %extra_alloc : |
| 299 | /// memref<...> to memref<vector<...>> |
| 300 | /// store %2, %3[] : memref<vector<...>> |
| 301 | /// %4 = memref.cast %alloc: memref<B...> to compatibleMemRefType |
| 302 | /// scf.yield %4, ... : compatibleMemRefType, index, index |
| 303 | /// } |
| 304 | /// ``` |
| 305 | /// Return the produced scf::IfOp. |
| 306 | static scf::IfOp createFullPartialVectorTransferRead( |
| 307 | RewriterBase &b, vector::TransferReadOp xferOp, TypeRange returnTypes, |
| 308 | Value inBoundsCond, MemRefType compatibleMemRefType, Value alloc) { |
| 309 | Location loc = xferOp.getLoc(); |
| 310 | scf::IfOp fullPartialIfOp; |
| 311 | Value zero = arith::ConstantIndexOp::create(builder&: b, location: loc, value: 0); |
| 312 | Value memref = xferOp.getBase(); |
| 313 | return scf::IfOp::create( |
| 314 | builder&: b, location: loc, cond: inBoundsCond, |
| 315 | thenBuilder: [&](OpBuilder &b, Location loc) { |
| 316 | Value res = castToCompatibleMemRefType(b, memref, compatibleMemRefType); |
| 317 | scf::ValueVector viewAndIndices{res}; |
| 318 | llvm::append_range(C&: viewAndIndices, R: xferOp.getIndices()); |
| 319 | scf::YieldOp::create(builder&: b, location: loc, results: viewAndIndices); |
| 320 | }, |
| 321 | elseBuilder: [&](OpBuilder &b, Location loc) { |
| 322 | Operation *newXfer = b.clone(op&: *xferOp.getOperation()); |
| 323 | Value vector = cast<VectorTransferOpInterface>(Val: newXfer).getVector(); |
| 324 | memref::StoreOp::create( |
| 325 | builder&: b, location: loc, valueToStore: vector, |
| 326 | memref: vector::TypeCastOp::create( |
| 327 | builder&: b, location: loc, result: MemRefType::get(shape: {}, elementType: vector.getType()), memref: alloc)); |
| 328 | |
| 329 | Value casted = |
| 330 | castToCompatibleMemRefType(b, memref: alloc, compatibleMemRefType); |
| 331 | scf::ValueVector viewAndIndices{casted}; |
| 332 | viewAndIndices.insert(I: viewAndIndices.end(), NumToInsert: xferOp.getTransferRank(), |
| 333 | Elt: zero); |
| 334 | scf::YieldOp::create(builder&: b, location: loc, results: viewAndIndices); |
| 335 | }); |
| 336 | } |
| 337 | |
| 338 | /// Given an `xferOp` for which: |
| 339 | /// 1. `inBoundsCond` and a `compatibleMemRefType` have been computed. |
| 340 | /// 2. a memref of single vector `alloc` has been allocated. |
| 341 | /// Produce IR resembling: |
| 342 | /// ``` |
| 343 | /// %1:3 = scf.if (%inBounds) { |
| 344 | /// memref.cast %A: memref<A...> to compatibleMemRefType |
| 345 | /// scf.yield %view, ... : compatibleMemRefType, index, index |
| 346 | /// } else { |
| 347 | /// %3 = vector.type_cast %extra_alloc : |
| 348 | /// memref<...> to memref<vector<...>> |
| 349 | /// %4 = memref.cast %alloc: memref<B...> to compatibleMemRefType |
| 350 | /// scf.yield %4, ... : compatibleMemRefType, index, index |
| 351 | /// } |
| 352 | /// ``` |
| 353 | static ValueRange |
| 354 | getLocationToWriteFullVec(RewriterBase &b, vector::TransferWriteOp xferOp, |
| 355 | TypeRange returnTypes, Value inBoundsCond, |
| 356 | MemRefType compatibleMemRefType, Value alloc) { |
| 357 | Location loc = xferOp.getLoc(); |
| 358 | Value zero = arith::ConstantIndexOp::create(builder&: b, location: loc, value: 0); |
| 359 | Value memref = xferOp.getBase(); |
| 360 | return scf::IfOp::create( |
| 361 | builder&: b, location: loc, cond: inBoundsCond, |
| 362 | thenBuilder: [&](OpBuilder &b, Location loc) { |
| 363 | Value res = |
| 364 | castToCompatibleMemRefType(b, memref, compatibleMemRefType); |
| 365 | scf::ValueVector viewAndIndices{res}; |
| 366 | llvm::append_range(C&: viewAndIndices, R: xferOp.getIndices()); |
| 367 | scf::YieldOp::create(builder&: b, location: loc, results: viewAndIndices); |
| 368 | }, |
| 369 | elseBuilder: [&](OpBuilder &b, Location loc) { |
| 370 | Value casted = |
| 371 | castToCompatibleMemRefType(b, memref: alloc, compatibleMemRefType); |
| 372 | scf::ValueVector viewAndIndices{casted}; |
| 373 | viewAndIndices.insert(I: viewAndIndices.end(), |
| 374 | NumToInsert: xferOp.getTransferRank(), Elt: zero); |
| 375 | scf::YieldOp::create(builder&: b, location: loc, results: viewAndIndices); |
| 376 | }) |
| 377 | ->getResults(); |
| 378 | } |
| 379 | |
| 380 | /// Given an `xferOp` for which: |
| 381 | /// 1. `inBoundsCond` has been computed. |
| 382 | /// 2. a memref of single vector `alloc` has been allocated. |
| 383 | /// 3. it originally wrote to %view |
| 384 | /// Produce IR resembling: |
| 385 | /// ``` |
| 386 | /// %notInBounds = arith.xori %inBounds, %true |
| 387 | /// scf.if (%notInBounds) { |
| 388 | /// %3 = subview %alloc [...][...][...] |
| 389 | /// %4 = subview %view [0, 0][...][...] |
| 390 | /// memref.copy %3, %4 |
| 391 | /// } |
| 392 | /// ``` |
| 393 | static void createFullPartialMemRefCopy(RewriterBase &b, |
| 394 | vector::TransferWriteOp xferOp, |
| 395 | Value inBoundsCond, Value alloc) { |
| 396 | Location loc = xferOp.getLoc(); |
| 397 | auto notInBounds = arith::XOrIOp::create( |
| 398 | builder&: b, location: loc, lhs: inBoundsCond, rhs: arith::ConstantIntOp::create(builder&: b, location: loc, value: true, width: 1)); |
| 399 | scf::IfOp::create(builder&: b, location: loc, cond: notInBounds, thenBuilder: [&](OpBuilder &b, Location loc) { |
| 400 | IRRewriter rewriter(b); |
| 401 | std::pair<Value, Value> copyArgs = createSubViewIntersection( |
| 402 | b&: rewriter, xferOp: cast<VectorTransferOpInterface>(Val: xferOp.getOperation()), |
| 403 | alloc); |
| 404 | memref::CopyOp::create(builder&: b, location: loc, source: copyArgs.first, target: copyArgs.second); |
| 405 | scf::YieldOp::create(builder&: b, location: loc, results: ValueRange{}); |
| 406 | }); |
| 407 | } |
| 408 | |
| 409 | /// Given an `xferOp` for which: |
| 410 | /// 1. `inBoundsCond` has been computed. |
| 411 | /// 2. a memref of single vector `alloc` has been allocated. |
| 412 | /// 3. it originally wrote to %view |
| 413 | /// Produce IR resembling: |
| 414 | /// ``` |
| 415 | /// %notInBounds = arith.xori %inBounds, %true |
| 416 | /// scf.if (%notInBounds) { |
| 417 | /// %2 = load %alloc : memref<vector<...>> |
| 418 | /// vector.transfer_write %2, %view[...] : memref<A...>, vector<...> |
| 419 | /// } |
| 420 | /// ``` |
| 421 | static void createFullPartialVectorTransferWrite(RewriterBase &b, |
| 422 | vector::TransferWriteOp xferOp, |
| 423 | Value inBoundsCond, |
| 424 | Value alloc) { |
| 425 | Location loc = xferOp.getLoc(); |
| 426 | auto notInBounds = arith::XOrIOp::create( |
| 427 | builder&: b, location: loc, lhs: inBoundsCond, rhs: arith::ConstantIntOp::create(builder&: b, location: loc, value: true, width: 1)); |
| 428 | scf::IfOp::create(builder&: b, location: loc, cond: notInBounds, thenBuilder: [&](OpBuilder &b, Location loc) { |
| 429 | IRMapping mapping; |
| 430 | Value load = memref::LoadOp::create( |
| 431 | builder&: b, location: loc, |
| 432 | memref: vector::TypeCastOp::create( |
| 433 | builder&: b, location: loc, result: MemRefType::get(shape: {}, elementType: xferOp.getVector().getType()), memref: alloc), |
| 434 | indices: ValueRange()); |
| 435 | mapping.map(from: xferOp.getVector(), to: load); |
| 436 | b.clone(op&: *xferOp.getOperation(), mapper&: mapping); |
| 437 | scf::YieldOp::create(builder&: b, location: loc, results: ValueRange{}); |
| 438 | }); |
| 439 | } |
| 440 | |
| 441 | // TODO: Parallelism and threadlocal considerations with a ParallelScope trait. |
| 442 | static Operation *getAutomaticAllocationScope(Operation *op) { |
| 443 | // Find the closest surrounding allocation scope that is not a known looping |
| 444 | // construct (putting alloca's in loops doesn't always lower to deallocation |
| 445 | // until the end of the loop). |
| 446 | Operation *scope = nullptr; |
| 447 | for (Operation *parent = op->getParentOp(); parent != nullptr; |
| 448 | parent = parent->getParentOp()) { |
| 449 | if (parent->hasTrait<OpTrait::AutomaticAllocationScope>()) |
| 450 | scope = parent; |
| 451 | if (!isa<scf::ForOp, affine::AffineForOp>(Val: parent)) |
| 452 | break; |
| 453 | } |
| 454 | assert(scope && "Expected op to be inside automatic allocation scope" ); |
| 455 | return scope; |
| 456 | } |
| 457 | |
| 458 | /// Split a vector.transfer operation into an in-bounds (i.e., no out-of-bounds |
| 459 | /// masking) fastpath and a slowpath. |
| 460 | /// |
| 461 | /// For vector.transfer_read: |
| 462 | /// If `ifOp` is not null and the result is `success, the `ifOp` points to the |
| 463 | /// newly created conditional upon function return. |
| 464 | /// To accomodate for the fact that the original vector.transfer indexing may be |
| 465 | /// arbitrary and the slow path indexes @[0...0] in the temporary buffer, the |
| 466 | /// scf.if op returns a view and values of type index. |
| 467 | /// |
| 468 | /// Example (a 2-D vector.transfer_read): |
| 469 | /// ``` |
| 470 | /// %1 = vector.transfer_read %0[...], %pad : memref<A...>, vector<...> |
| 471 | /// ``` |
| 472 | /// is transformed into: |
| 473 | /// ``` |
| 474 | /// %1:3 = scf.if (%inBounds) { |
| 475 | /// // fastpath, direct cast |
| 476 | /// memref.cast %A: memref<A...> to compatibleMemRefType |
| 477 | /// scf.yield %view : compatibleMemRefType, index, index |
| 478 | /// } else { |
| 479 | /// // slowpath, not in-bounds vector.transfer or memref.copy. |
| 480 | /// memref.cast %alloc: memref<B...> to compatibleMemRefType |
| 481 | /// scf.yield %4 : compatibleMemRefType, index, index |
| 482 | // } |
| 483 | /// %0 = vector.transfer_read %1#0[%1#1, %1#2] {in_bounds = [true ... true]} |
| 484 | /// ``` |
| 485 | /// where `alloc` is a top of the function alloca'ed buffer of one vector. |
| 486 | /// |
| 487 | /// For vector.transfer_write: |
| 488 | /// There are 2 conditional blocks. First a block to decide which memref and |
| 489 | /// indices to use for an unmasked, inbounds write. Then a conditional block to |
| 490 | /// further copy a partial buffer into the final result in the slow path case. |
| 491 | /// |
| 492 | /// Example (a 2-D vector.transfer_write): |
| 493 | /// ``` |
| 494 | /// vector.transfer_write %arg, %0[...], %pad : memref<A...>, vector<...> |
| 495 | /// ``` |
| 496 | /// is transformed into: |
| 497 | /// ``` |
| 498 | /// %1:3 = scf.if (%inBounds) { |
| 499 | /// memref.cast %A: memref<A...> to compatibleMemRefType |
| 500 | /// scf.yield %view : compatibleMemRefType, index, index |
| 501 | /// } else { |
| 502 | /// memref.cast %alloc: memref<B...> to compatibleMemRefType |
| 503 | /// scf.yield %4 : compatibleMemRefType, index, index |
| 504 | /// } |
| 505 | /// %0 = vector.transfer_write %arg, %1#0[%1#1, %1#2] {in_bounds = [true ... |
| 506 | /// true]} |
| 507 | /// scf.if (%notInBounds) { |
| 508 | /// // slowpath: not in-bounds vector.transfer or memref.copy. |
| 509 | /// } |
| 510 | /// ``` |
| 511 | /// where `alloc` is a top of the function alloca'ed buffer of one vector. |
| 512 | /// |
| 513 | /// Preconditions: |
| 514 | /// 1. `xferOp.getPermutationMap()` must be a minor identity map |
| 515 | /// 2. the rank of the `xferOp.getBase()` and the rank of the |
| 516 | /// `xferOp.getVector()` must be equal. This will be relaxed in the future |
| 517 | /// but requires rank-reducing subviews. |
| 518 | LogicalResult mlir::vector::splitFullAndPartialTransfer( |
| 519 | RewriterBase &b, VectorTransferOpInterface xferOp, |
| 520 | VectorTransformsOptions options, scf::IfOp *ifOp) { |
| 521 | if (options.vectorTransferSplit == VectorTransferSplit::None) |
| 522 | return failure(); |
| 523 | |
| 524 | SmallVector<bool, 4> bools(xferOp.getTransferRank(), true); |
| 525 | auto inBoundsAttr = b.getBoolArrayAttr(values: bools); |
| 526 | if (options.vectorTransferSplit == VectorTransferSplit::ForceInBounds) { |
| 527 | b.modifyOpInPlace(root: xferOp, callable: [&]() { |
| 528 | xferOp->setInherentAttr(name: b.getStringAttr(bytes: xferOp.getInBoundsAttrName()), |
| 529 | value: inBoundsAttr); |
| 530 | }); |
| 531 | return success(); |
| 532 | } |
| 533 | |
| 534 | // Assert preconditions. Additionally, keep the variables in an inner scope to |
| 535 | // ensure they aren't used in the wrong scopes further down. |
| 536 | { |
| 537 | assert(succeeded(splitFullAndPartialTransferPrecondition(xferOp)) && |
| 538 | "Expected splitFullAndPartialTransferPrecondition to hold" ); |
| 539 | |
| 540 | auto xferReadOp = dyn_cast<vector::TransferReadOp>(Val: xferOp.getOperation()); |
| 541 | auto xferWriteOp = dyn_cast<vector::TransferWriteOp>(Val: xferOp.getOperation()); |
| 542 | |
| 543 | if (!(xferReadOp || xferWriteOp)) |
| 544 | return failure(); |
| 545 | if (xferWriteOp && xferWriteOp.getMask()) |
| 546 | return failure(); |
| 547 | if (xferReadOp && xferReadOp.getMask()) |
| 548 | return failure(); |
| 549 | } |
| 550 | |
| 551 | RewriterBase::InsertionGuard guard(b); |
| 552 | b.setInsertionPoint(xferOp); |
| 553 | Value inBoundsCond = createInBoundsCond( |
| 554 | b, xferOp: cast<VectorTransferOpInterface>(Val: xferOp.getOperation())); |
| 555 | if (!inBoundsCond) |
| 556 | return failure(); |
| 557 | |
| 558 | // Top of the function `alloc` for transient storage. |
| 559 | Value alloc; |
| 560 | { |
| 561 | RewriterBase::InsertionGuard guard(b); |
| 562 | Operation *scope = getAutomaticAllocationScope(op: xferOp); |
| 563 | assert(scope->getNumRegions() == 1 && |
| 564 | "AutomaticAllocationScope with >1 regions" ); |
| 565 | b.setInsertionPointToStart(&scope->getRegion(index: 0).front()); |
| 566 | auto shape = xferOp.getVectorType().getShape(); |
| 567 | Type elementType = xferOp.getVectorType().getElementType(); |
| 568 | alloc = memref::AllocaOp::create(builder&: b, location: scope->getLoc(), |
| 569 | memrefType: MemRefType::get(shape, elementType), |
| 570 | dynamicSizes: ValueRange{}, alignment: b.getI64IntegerAttr(value: 32)); |
| 571 | } |
| 572 | |
| 573 | MemRefType compatibleMemRefType = |
| 574 | getCastCompatibleMemRefType(aT: cast<MemRefType>(Val: xferOp.getShapedType()), |
| 575 | bT: cast<MemRefType>(Val: alloc.getType())); |
| 576 | if (!compatibleMemRefType) |
| 577 | return failure(); |
| 578 | |
| 579 | SmallVector<Type, 4> returnTypes(1 + xferOp.getTransferRank(), |
| 580 | b.getIndexType()); |
| 581 | returnTypes[0] = compatibleMemRefType; |
| 582 | |
| 583 | if (auto xferReadOp = |
| 584 | dyn_cast<vector::TransferReadOp>(Val: xferOp.getOperation())) { |
| 585 | // Read case: full fill + partial copy -> in-bounds vector.xfer_read. |
| 586 | scf::IfOp fullPartialIfOp = |
| 587 | options.vectorTransferSplit == VectorTransferSplit::VectorTransfer |
| 588 | ? createFullPartialVectorTransferRead(b, xferOp: xferReadOp, returnTypes, |
| 589 | inBoundsCond, |
| 590 | compatibleMemRefType, alloc) |
| 591 | : createFullPartialMemRefCopy(b, xferOp: xferReadOp, returnTypes, |
| 592 | inBoundsCond, compatibleMemRefType, |
| 593 | alloc); |
| 594 | if (ifOp) |
| 595 | *ifOp = fullPartialIfOp; |
| 596 | |
| 597 | // Set existing read op to in-bounds, it always reads from a full buffer. |
| 598 | for (unsigned i = 0, e = returnTypes.size(); i != e; ++i) |
| 599 | xferReadOp.setOperand(i, value: fullPartialIfOp.getResult(i)); |
| 600 | |
| 601 | b.modifyOpInPlace(root: xferOp, callable: [&]() { |
| 602 | xferOp->setInherentAttr(name: b.getStringAttr(bytes: xferOp.getInBoundsAttrName()), |
| 603 | value: inBoundsAttr); |
| 604 | }); |
| 605 | |
| 606 | return success(); |
| 607 | } |
| 608 | |
| 609 | auto xferWriteOp = cast<vector::TransferWriteOp>(Val: xferOp.getOperation()); |
| 610 | |
| 611 | // Decide which location to write the entire vector to. |
| 612 | auto memrefAndIndices = getLocationToWriteFullVec( |
| 613 | b, xferOp: xferWriteOp, returnTypes, inBoundsCond, compatibleMemRefType, alloc); |
| 614 | |
| 615 | // Do an in bounds write to either the output or the extra allocated buffer. |
| 616 | // The operation is cloned to prevent deleting information needed for the |
| 617 | // later IR creation. |
| 618 | IRMapping mapping; |
| 619 | mapping.map(from: xferWriteOp.getBase(), to: memrefAndIndices.front()); |
| 620 | mapping.map(from: xferWriteOp.getIndices(), to: memrefAndIndices.drop_front()); |
| 621 | auto *clone = b.clone(op&: *xferWriteOp, mapper&: mapping); |
| 622 | clone->setInherentAttr(name: xferWriteOp.getInBoundsAttrName(), value: inBoundsAttr); |
| 623 | |
| 624 | // Create a potential copy from the allocated buffer to the final output in |
| 625 | // the slow path case. |
| 626 | if (options.vectorTransferSplit == VectorTransferSplit::VectorTransfer) |
| 627 | createFullPartialVectorTransferWrite(b, xferOp: xferWriteOp, inBoundsCond, alloc); |
| 628 | else |
| 629 | createFullPartialMemRefCopy(b, xferOp: xferWriteOp, inBoundsCond, alloc); |
| 630 | |
| 631 | b.eraseOp(op: xferOp); |
| 632 | |
| 633 | return success(); |
| 634 | } |
| 635 | |
| 636 | namespace { |
| 637 | /// Apply `splitFullAndPartialTransfer` selectively via a pattern. This pattern |
| 638 | /// may take an extra filter to perform selection at a finer granularity. |
| 639 | struct VectorTransferFullPartialRewriter : public RewritePattern { |
| 640 | using FilterConstraintType = |
| 641 | std::function<LogicalResult(VectorTransferOpInterface op)>; |
| 642 | |
| 643 | explicit VectorTransferFullPartialRewriter( |
| 644 | MLIRContext *context, |
| 645 | VectorTransformsOptions options = VectorTransformsOptions(), |
| 646 | FilterConstraintType filter = |
| 647 | [](VectorTransferOpInterface op) { return success(); }, |
| 648 | PatternBenefit benefit = 1) |
| 649 | : RewritePattern(MatchAnyOpTypeTag(), benefit, context), options(options), |
| 650 | filter(std::move(filter)) {} |
| 651 | |
| 652 | /// Performs the rewrite. |
| 653 | LogicalResult matchAndRewrite(Operation *op, |
| 654 | PatternRewriter &rewriter) const override; |
| 655 | |
| 656 | private: |
| 657 | VectorTransformsOptions options; |
| 658 | FilterConstraintType filter; |
| 659 | }; |
| 660 | |
| 661 | } // namespace |
| 662 | |
| 663 | LogicalResult VectorTransferFullPartialRewriter::matchAndRewrite( |
| 664 | Operation *op, PatternRewriter &rewriter) const { |
| 665 | auto xferOp = dyn_cast<VectorTransferOpInterface>(Val: op); |
| 666 | if (!xferOp || failed(Result: splitFullAndPartialTransferPrecondition(xferOp)) || |
| 667 | failed(Result: filter(xferOp))) |
| 668 | return failure(); |
| 669 | return splitFullAndPartialTransfer(b&: rewriter, xferOp, options); |
| 670 | } |
| 671 | |
| 672 | void mlir::vector::populateVectorTransferFullPartialPatterns( |
| 673 | RewritePatternSet &patterns, const VectorTransformsOptions &options) { |
| 674 | patterns.add<VectorTransferFullPartialRewriter>(arg: patterns.getContext(), |
| 675 | args: options); |
| 676 | } |
| 677 | |