| 1 | //===- VectorToSPIRV.cpp - Vector to SPIR-V Patterns ----------------------===// |
| 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 patterns to convert Vector dialect to SPIRV dialect. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "mlir/Conversion/VectorToSPIRV/VectorToSPIRV.h" |
| 14 | |
| 15 | #include "mlir/Dialect/Arith/IR/Arith.h" |
| 16 | #include "mlir/Dialect/SPIRV/IR/SPIRVOps.h" |
| 17 | #include "mlir/Dialect/SPIRV/IR/SPIRVTypes.h" |
| 18 | #include "mlir/Dialect/SPIRV/Transforms/SPIRVConversion.h" |
| 19 | #include "mlir/Dialect/Utils/StaticValueUtils.h" |
| 20 | #include "mlir/Dialect/Vector/IR/VectorOps.h" |
| 21 | #include "mlir/IR/Attributes.h" |
| 22 | #include "mlir/IR/BuiltinAttributes.h" |
| 23 | #include "mlir/IR/BuiltinTypes.h" |
| 24 | #include "mlir/IR/Location.h" |
| 25 | #include "mlir/IR/PatternMatch.h" |
| 26 | #include "mlir/IR/TypeUtilities.h" |
| 27 | #include "mlir/Transforms/DialectConversion.h" |
| 28 | #include "llvm/ADT/ArrayRef.h" |
| 29 | #include "llvm/ADT/STLExtras.h" |
| 30 | #include "llvm/ADT/SmallVector.h" |
| 31 | #include "llvm/ADT/SmallVectorExtras.h" |
| 32 | #include "llvm/Support/FormatVariadic.h" |
| 33 | #include <cassert> |
| 34 | #include <cstdint> |
| 35 | #include <numeric> |
| 36 | |
| 37 | using namespace mlir; |
| 38 | |
| 39 | /// Returns the integer value from the first valid input element, assuming Value |
| 40 | /// inputs are defined by a constant index ops and Attribute inputs are integer |
| 41 | /// attributes. |
| 42 | static uint64_t getFirstIntValue(ArrayAttr attr) { |
| 43 | return (*attr.getAsValueRange<IntegerAttr>().begin()).getZExtValue(); |
| 44 | } |
| 45 | |
| 46 | /// Returns the number of bits for the given scalar/vector type. |
| 47 | static int getNumBits(Type type) { |
| 48 | // TODO: This does not take into account any memory layout or widening |
| 49 | // constraints. E.g., a vector<3xi57> may report to occupy 3x57=171 bit, even |
| 50 | // though in practice it will likely be stored as in a 4xi64 vector register. |
| 51 | if (auto vectorType = dyn_cast<VectorType>(Val&: type)) |
| 52 | return vectorType.getNumElements() * vectorType.getElementTypeBitWidth(); |
| 53 | return type.getIntOrFloatBitWidth(); |
| 54 | } |
| 55 | |
| 56 | namespace { |
| 57 | struct MemoryRequirements { |
| 58 | spirv::MemoryAccessAttr memoryAccess; |
| 59 | IntegerAttr alignment; |
| 60 | }; |
| 61 | } // namespace |
| 62 | |
| 63 | /// PhysicalStorageBuffer pointers always need Aligned, derived from pointee |
| 64 | /// element size unless preferredAlignment is set. Other storage classes may |
| 65 | /// omit it. |
| 66 | static FailureOr<MemoryRequirements> |
| 67 | calculateMemoryRequirements(Value accessedPtr, uint64_t preferredAlignment) { |
| 68 | if (preferredAlignment >= std::numeric_limits<uint32_t>::max()) |
| 69 | return failure(); |
| 70 | |
| 71 | MLIRContext *ctx = accessedPtr.getContext(); |
| 72 | auto ptrType = cast<spirv::PointerType>(Val: accessedPtr.getType()); |
| 73 | bool mayOmitAlignment = |
| 74 | !preferredAlignment && |
| 75 | ptrType.getStorageClass() != spirv::StorageClass::PhysicalStorageBuffer; |
| 76 | if (mayOmitAlignment) |
| 77 | return MemoryRequirements{.memoryAccess: spirv::MemoryAccessAttr{}, .alignment: IntegerAttr{}}; |
| 78 | |
| 79 | // PhysicalStorageBuffer pointers require Aligned. The vector pointer is a |
| 80 | // bitcast of a scalar element pointer, so only element alignment is known. |
| 81 | std::optional<int64_t> sizeInBytes; |
| 82 | Type pointeeType = getElementTypeOrSelf(type: ptrType.getPointeeType()); |
| 83 | if (auto scalarType = dyn_cast<spirv::ScalarType>(Val&: pointeeType)) |
| 84 | sizeInBytes = scalarType.getSizeInBytes(); |
| 85 | |
| 86 | if (!sizeInBytes) |
| 87 | return failure(); |
| 88 | |
| 89 | auto memoryAccess = |
| 90 | spirv::MemoryAccessAttr::get(context: ctx, value: spirv::MemoryAccess::Aligned); |
| 91 | uint64_t alignmentValue = |
| 92 | preferredAlignment ? preferredAlignment : *sizeInBytes; |
| 93 | auto alignment = IntegerAttr::get(type: IntegerType::get(context: ctx, width: 32), value: alignmentValue); |
| 94 | return MemoryRequirements{.memoryAccess: memoryAccess, .alignment: alignment}; |
| 95 | } |
| 96 | |
| 97 | namespace { |
| 98 | |
| 99 | struct VectorShapeCast final : public OpConversionPattern<vector::ShapeCastOp> { |
| 100 | using Base::Base; |
| 101 | |
| 102 | LogicalResult |
| 103 | matchAndRewrite(vector::ShapeCastOp shapeCastOp, OpAdaptor adaptor, |
| 104 | ConversionPatternRewriter &rewriter) const override { |
| 105 | Type dstType = getTypeConverter()->convertType(t: shapeCastOp.getType()); |
| 106 | if (!dstType) |
| 107 | return failure(); |
| 108 | |
| 109 | // If dstType is same as the source type or the vector size is 1, it can be |
| 110 | // directly replaced by the source. |
| 111 | if (dstType == adaptor.getSource().getType() || |
| 112 | shapeCastOp.getResultVectorType().getNumElements() == 1) { |
| 113 | rewriter.replaceOp(op: shapeCastOp, newValues: adaptor.getSource()); |
| 114 | return success(); |
| 115 | } |
| 116 | |
| 117 | // Lowering for size-n vectors when n > 1 hasn't been implemented. |
| 118 | return failure(); |
| 119 | } |
| 120 | }; |
| 121 | |
| 122 | struct VectorBitcastConvert final |
| 123 | : public OpConversionPattern<vector::BitCastOp> { |
| 124 | using Base::Base; |
| 125 | |
| 126 | LogicalResult |
| 127 | matchAndRewrite(vector::BitCastOp bitcastOp, OpAdaptor adaptor, |
| 128 | ConversionPatternRewriter &rewriter) const override { |
| 129 | Type dstType = getTypeConverter()->convertType(t: bitcastOp.getType()); |
| 130 | if (!dstType) |
| 131 | return failure(); |
| 132 | |
| 133 | if (dstType == adaptor.getSource().getType()) { |
| 134 | rewriter.replaceOp(op: bitcastOp, newValues: adaptor.getSource()); |
| 135 | return success(); |
| 136 | } |
| 137 | |
| 138 | // Check that the source and destination type have the same bitwidth. |
| 139 | // Depending on the target environment, we may need to emulate certain |
| 140 | // types, which can cause issue with bitcast. |
| 141 | Type srcType = adaptor.getSource().getType(); |
| 142 | if (getNumBits(type: dstType) != getNumBits(type: srcType)) { |
| 143 | return rewriter.notifyMatchFailure( |
| 144 | arg&: bitcastOp, |
| 145 | msg: llvm::formatv(Fmt: "different source ({0}) and target ({1}) bitwidth" , |
| 146 | Vals&: srcType, Vals&: dstType)); |
| 147 | } |
| 148 | |
| 149 | rewriter.replaceOpWithNewOp<spirv::BitcastOp>(op: bitcastOp, args&: dstType, |
| 150 | args: adaptor.getSource()); |
| 151 | return success(); |
| 152 | } |
| 153 | }; |
| 154 | |
| 155 | struct VectorBroadcastConvert final |
| 156 | : public OpConversionPattern<vector::BroadcastOp> { |
| 157 | using Base::Base; |
| 158 | |
| 159 | LogicalResult |
| 160 | matchAndRewrite(vector::BroadcastOp castOp, OpAdaptor adaptor, |
| 161 | ConversionPatternRewriter &rewriter) const override { |
| 162 | Type resultType = |
| 163 | getTypeConverter()->convertType(t: castOp.getResultVectorType()); |
| 164 | if (!resultType) |
| 165 | return failure(); |
| 166 | |
| 167 | if (isa<spirv::ScalarType>(Val: resultType)) { |
| 168 | rewriter.replaceOp(op: castOp, newValues: adaptor.getSource()); |
| 169 | return success(); |
| 170 | } |
| 171 | |
| 172 | SmallVector<Value, 4> source(castOp.getResultVectorType().getNumElements(), |
| 173 | adaptor.getSource()); |
| 174 | rewriter.replaceOpWithNewOp<spirv::CompositeConstructOp>(op: castOp, args&: resultType, |
| 175 | args&: source); |
| 176 | return success(); |
| 177 | } |
| 178 | }; |
| 179 | |
| 180 | // SPIR-V does not have a concept of a poison index for certain instructions, |
| 181 | // which creates a UB hazard when lowering from otherwise equivalent Vector |
| 182 | // dialect instructions, because this index will be considered out-of-bounds. |
| 183 | // To avoid this, this function implements a dynamic sanitization that returns |
| 184 | // some arbitrary safe index. For power-of-two vector sizes, this uses a bitmask |
| 185 | // (presumably more efficient), and otherwise index 0 (always in-bounds). |
| 186 | static Value sanitizeDynamicIndex(ConversionPatternRewriter &rewriter, |
| 187 | Location loc, Value dynamicIndex, |
| 188 | int64_t kPoisonIndex, unsigned vectorSize) { |
| 189 | if (llvm::isPowerOf2_32(Value: vectorSize)) { |
| 190 | Value inBoundsMask = spirv::ConstantOp::create( |
| 191 | builder&: rewriter, location: loc, constant: dynamicIndex.getType(), |
| 192 | value: rewriter.getIntegerAttr(type: dynamicIndex.getType(), value: vectorSize - 1)); |
| 193 | return spirv::BitwiseAndOp::create(builder&: rewriter, location: loc, operand1: dynamicIndex, |
| 194 | operand2: inBoundsMask); |
| 195 | } |
| 196 | Value poisonIndex = spirv::ConstantOp::create( |
| 197 | builder&: rewriter, location: loc, constant: dynamicIndex.getType(), |
| 198 | value: rewriter.getIntegerAttr(type: dynamicIndex.getType(), value: kPoisonIndex)); |
| 199 | Value cmpResult = |
| 200 | spirv::IEqualOp::create(builder&: rewriter, location: loc, operand1: dynamicIndex, operand2: poisonIndex); |
| 201 | return spirv::SelectOp::create( |
| 202 | builder&: rewriter, location: loc, condition: cmpResult, |
| 203 | true_value: spirv::ConstantOp::getZero(type: dynamicIndex.getType(), loc, builder&: rewriter), |
| 204 | false_value: dynamicIndex); |
| 205 | } |
| 206 | |
| 207 | struct final |
| 208 | : public OpConversionPattern<vector::ExtractOp> { |
| 209 | using Base::Base; |
| 210 | |
| 211 | LogicalResult |
| 212 | matchAndRewrite(vector::ExtractOp , OpAdaptor adaptor, |
| 213 | ConversionPatternRewriter &rewriter) const override { |
| 214 | Type dstType = getTypeConverter()->convertType(t: extractOp.getType()); |
| 215 | if (!dstType) |
| 216 | return failure(); |
| 217 | |
| 218 | if (isa<spirv::ScalarType>(Val: adaptor.getSource().getType())) { |
| 219 | rewriter.replaceOp(op: extractOp, newValues: adaptor.getSource()); |
| 220 | return success(); |
| 221 | } |
| 222 | |
| 223 | if (std::optional<int64_t> id = |
| 224 | getConstantIntValue(ofr: extractOp.getMixedPosition()[0])) { |
| 225 | if (id == vector::ExtractOp::kPoisonIndex) |
| 226 | return rewriter.notifyMatchFailure( |
| 227 | arg&: extractOp, |
| 228 | msg: "Static use of poison index handled elsewhere (folded to poison)" ); |
| 229 | rewriter.replaceOpWithNewOp<spirv::CompositeExtractOp>( |
| 230 | op: extractOp, args&: dstType, args: adaptor.getSource(), |
| 231 | args: rewriter.getI32ArrayAttr(values: id.value())); |
| 232 | } else { |
| 233 | Value sanitizedIndex = sanitizeDynamicIndex( |
| 234 | rewriter, loc: extractOp.getLoc(), dynamicIndex: adaptor.getDynamicPosition()[0], |
| 235 | kPoisonIndex: vector::ExtractOp::kPoisonIndex, |
| 236 | vectorSize: extractOp.getSourceVectorType().getNumElements()); |
| 237 | rewriter.replaceOpWithNewOp<spirv::VectorExtractDynamicOp>( |
| 238 | op: extractOp, args&: dstType, args: adaptor.getSource(), args&: sanitizedIndex); |
| 239 | } |
| 240 | return success(); |
| 241 | } |
| 242 | }; |
| 243 | |
| 244 | struct final |
| 245 | : public OpConversionPattern<vector::ExtractStridedSliceOp> { |
| 246 | using Base::Base; |
| 247 | |
| 248 | LogicalResult |
| 249 | matchAndRewrite(vector::ExtractStridedSliceOp , OpAdaptor adaptor, |
| 250 | ConversionPatternRewriter &rewriter) const override { |
| 251 | Type dstType = getTypeConverter()->convertType(t: extractOp.getType()); |
| 252 | if (!dstType) |
| 253 | return failure(); |
| 254 | |
| 255 | uint64_t offset = getFirstIntValue(attr: extractOp.getOffsets()); |
| 256 | uint64_t size = getFirstIntValue(attr: extractOp.getSizes()); |
| 257 | uint64_t stride = getFirstIntValue(attr: extractOp.getStrides()); |
| 258 | if (stride != 1) |
| 259 | return failure(); |
| 260 | |
| 261 | Value srcVector = adaptor.getOperands().front(); |
| 262 | |
| 263 | // Extract vector<1xT> case. |
| 264 | if (isa<spirv::ScalarType>(Val: dstType)) { |
| 265 | rewriter.replaceOpWithNewOp<spirv::CompositeExtractOp>(op: extractOp, |
| 266 | args&: srcVector, args&: offset); |
| 267 | return success(); |
| 268 | } |
| 269 | |
| 270 | SmallVector<int32_t, 2> indices(size); |
| 271 | std::iota(first: indices.begin(), last: indices.end(), value: offset); |
| 272 | |
| 273 | rewriter.replaceOpWithNewOp<spirv::VectorShuffleOp>( |
| 274 | op: extractOp, args&: dstType, args&: srcVector, args&: srcVector, |
| 275 | args: rewriter.getI32ArrayAttr(values: indices)); |
| 276 | |
| 277 | return success(); |
| 278 | } |
| 279 | }; |
| 280 | |
| 281 | template <class SPIRVFMAOp> |
| 282 | struct VectorFmaOpConvert final : public OpConversionPattern<vector::FMAOp> { |
| 283 | using Base::Base; |
| 284 | |
| 285 | LogicalResult |
| 286 | matchAndRewrite(vector::FMAOp fmaOp, OpAdaptor adaptor, |
| 287 | ConversionPatternRewriter &rewriter) const override { |
| 288 | Type dstType = getTypeConverter()->convertType(fmaOp.getType()); |
| 289 | if (!dstType) |
| 290 | return failure(); |
| 291 | rewriter.replaceOpWithNewOp<SPIRVFMAOp>(fmaOp, dstType, adaptor.getLhs(), |
| 292 | adaptor.getRhs(), adaptor.getAcc()); |
| 293 | return success(); |
| 294 | } |
| 295 | }; |
| 296 | |
| 297 | struct VectorFromElementsOpConvert final |
| 298 | : public OpConversionPattern<vector::FromElementsOp> { |
| 299 | using Base::Base; |
| 300 | |
| 301 | LogicalResult |
| 302 | matchAndRewrite(vector::FromElementsOp op, OpAdaptor adaptor, |
| 303 | ConversionPatternRewriter &rewriter) const override { |
| 304 | Type resultType = getTypeConverter()->convertType(t: op.getType()); |
| 305 | if (!resultType) |
| 306 | return failure(); |
| 307 | ValueRange elements = adaptor.getElements(); |
| 308 | if (isa<spirv::ScalarType>(Val: resultType)) { |
| 309 | // In the case with a single scalar operand / single-element result, |
| 310 | // pass through the scalar. |
| 311 | rewriter.replaceOp(op, newValues: elements[0]); |
| 312 | return success(); |
| 313 | } |
| 314 | // SPIRVTypeConverter rejects vectors with rank > 1, so multi-dimensional |
| 315 | // vector.from_elements cases should not need to be handled, only 1d. |
| 316 | assert(cast<VectorType>(resultType).getRank() == 1); |
| 317 | rewriter.replaceOpWithNewOp<spirv::CompositeConstructOp>(op, args&: resultType, |
| 318 | args&: elements); |
| 319 | return success(); |
| 320 | } |
| 321 | }; |
| 322 | |
| 323 | struct VectorInsertOpConvert final |
| 324 | : public OpConversionPattern<vector::InsertOp> { |
| 325 | using Base::Base; |
| 326 | |
| 327 | LogicalResult |
| 328 | matchAndRewrite(vector::InsertOp insertOp, OpAdaptor adaptor, |
| 329 | ConversionPatternRewriter &rewriter) const override { |
| 330 | if (isa<VectorType>(Val: insertOp.getValueToStoreType())) |
| 331 | return rewriter.notifyMatchFailure(arg&: insertOp, msg: "unsupported vector source" ); |
| 332 | if (!getTypeConverter()->convertType(t: insertOp.getDestVectorType())) |
| 333 | return rewriter.notifyMatchFailure(arg&: insertOp, |
| 334 | msg: "unsupported dest vector type" ); |
| 335 | |
| 336 | // Special case for inserting scalar values into size-1 vectors. |
| 337 | if (insertOp.getValueToStoreType().isIntOrFloat() && |
| 338 | insertOp.getDestVectorType().getNumElements() == 1) { |
| 339 | rewriter.replaceOp(op: insertOp, newValues: adaptor.getValueToStore()); |
| 340 | return success(); |
| 341 | } |
| 342 | |
| 343 | if (std::optional<int64_t> id = |
| 344 | getConstantIntValue(ofr: insertOp.getMixedPosition()[0])) { |
| 345 | if (id == vector::InsertOp::kPoisonIndex) |
| 346 | return rewriter.notifyMatchFailure( |
| 347 | arg&: insertOp, |
| 348 | msg: "Static use of poison index handled elsewhere (folded to poison)" ); |
| 349 | rewriter.replaceOpWithNewOp<spirv::CompositeInsertOp>( |
| 350 | op: insertOp, args: adaptor.getValueToStore(), args: adaptor.getDest(), args&: id.value()); |
| 351 | } else { |
| 352 | Value sanitizedIndex = sanitizeDynamicIndex( |
| 353 | rewriter, loc: insertOp.getLoc(), dynamicIndex: adaptor.getDynamicPosition()[0], |
| 354 | kPoisonIndex: vector::InsertOp::kPoisonIndex, |
| 355 | vectorSize: insertOp.getDestVectorType().getNumElements()); |
| 356 | rewriter.replaceOpWithNewOp<spirv::VectorInsertDynamicOp>( |
| 357 | op: insertOp, args: insertOp.getDest(), args: adaptor.getValueToStore(), |
| 358 | args&: sanitizedIndex); |
| 359 | } |
| 360 | return success(); |
| 361 | } |
| 362 | }; |
| 363 | |
| 364 | struct VectorInsertStridedSliceOpConvert final |
| 365 | : public OpConversionPattern<vector::InsertStridedSliceOp> { |
| 366 | using Base::Base; |
| 367 | |
| 368 | LogicalResult |
| 369 | matchAndRewrite(vector::InsertStridedSliceOp insertOp, OpAdaptor adaptor, |
| 370 | ConversionPatternRewriter &rewriter) const override { |
| 371 | Value srcVector = adaptor.getOperands().front(); |
| 372 | Value dstVector = adaptor.getOperands().back(); |
| 373 | |
| 374 | uint64_t stride = getFirstIntValue(attr: insertOp.getStrides()); |
| 375 | if (stride != 1) |
| 376 | return failure(); |
| 377 | uint64_t offset = getFirstIntValue(attr: insertOp.getOffsets()); |
| 378 | |
| 379 | if (isa<spirv::ScalarType>(Val: srcVector.getType())) { |
| 380 | assert(!isa<spirv::ScalarType>(dstVector.getType())); |
| 381 | rewriter.replaceOpWithNewOp<spirv::CompositeInsertOp>( |
| 382 | op: insertOp, args: dstVector.getType(), args&: srcVector, args&: dstVector, |
| 383 | args: rewriter.getI32ArrayAttr(values: offset)); |
| 384 | return success(); |
| 385 | } |
| 386 | |
| 387 | uint64_t totalSize = cast<VectorType>(Val: dstVector.getType()).getNumElements(); |
| 388 | uint64_t insertSize = |
| 389 | cast<VectorType>(Val: srcVector.getType()).getNumElements(); |
| 390 | |
| 391 | SmallVector<int32_t, 2> indices(totalSize); |
| 392 | std::iota(first: indices.begin(), last: indices.end(), value: 0); |
| 393 | std::iota(first: indices.begin() + offset, last: indices.begin() + offset + insertSize, |
| 394 | value: totalSize); |
| 395 | |
| 396 | rewriter.replaceOpWithNewOp<spirv::VectorShuffleOp>( |
| 397 | op: insertOp, args: dstVector.getType(), args&: dstVector, args&: srcVector, |
| 398 | args: rewriter.getI32ArrayAttr(values: indices)); |
| 399 | |
| 400 | return success(); |
| 401 | } |
| 402 | }; |
| 403 | |
| 404 | static SmallVector<Value> ( |
| 405 | vector::ReductionOp reduceOp, vector::ReductionOp::Adaptor adaptor, |
| 406 | VectorType srcVectorType, ConversionPatternRewriter &rewriter) { |
| 407 | int numElements = static_cast<int>(srcVectorType.getDimSize(idx: 0)); |
| 408 | SmallVector<Value> values; |
| 409 | values.reserve(N: numElements + (adaptor.getAcc() ? 1 : 0)); |
| 410 | Location loc = reduceOp.getLoc(); |
| 411 | |
| 412 | for (int i = 0; i < numElements; ++i) { |
| 413 | values.push_back(Elt: spirv::CompositeExtractOp::create( |
| 414 | builder&: rewriter, location: loc, component: srcVectorType.getElementType(), composite: adaptor.getVector(), |
| 415 | indices: rewriter.getI32ArrayAttr(values: {i}))); |
| 416 | } |
| 417 | if (Value acc = adaptor.getAcc()) |
| 418 | values.push_back(Elt: acc); |
| 419 | |
| 420 | return values; |
| 421 | } |
| 422 | |
| 423 | struct ReductionRewriteInfo { |
| 424 | Type resultType; |
| 425 | SmallVector<Value> ; |
| 426 | }; |
| 427 | |
| 428 | FailureOr<ReductionRewriteInfo> static getReductionInfo( |
| 429 | vector::ReductionOp op, vector::ReductionOp::Adaptor adaptor, |
| 430 | ConversionPatternRewriter &rewriter, const TypeConverter &typeConverter) { |
| 431 | Type resultType = typeConverter.convertType(t: op.getType()); |
| 432 | if (!resultType) |
| 433 | return failure(); |
| 434 | |
| 435 | auto srcVectorType = dyn_cast<VectorType>(Val: adaptor.getVector().getType()); |
| 436 | if (!srcVectorType || srcVectorType.getRank() != 1) |
| 437 | return rewriter.notifyMatchFailure(arg&: op, msg: "not a 1-D vector source" ); |
| 438 | |
| 439 | SmallVector<Value> = |
| 440 | extractAllElements(reduceOp: op, adaptor, srcVectorType, rewriter); |
| 441 | |
| 442 | return ReductionRewriteInfo{.resultType: resultType, .extractedElements: std::move(extractedElements)}; |
| 443 | } |
| 444 | |
| 445 | template <typename SPIRVUMaxOp, typename SPIRVUMinOp, typename SPIRVSMaxOp, |
| 446 | typename SPIRVSMinOp> |
| 447 | struct VectorReductionPattern final : OpConversionPattern<vector::ReductionOp> { |
| 448 | using Base::Base; |
| 449 | |
| 450 | LogicalResult |
| 451 | matchAndRewrite(vector::ReductionOp reduceOp, OpAdaptor adaptor, |
| 452 | ConversionPatternRewriter &rewriter) const override { |
| 453 | auto reductionInfo = |
| 454 | getReductionInfo(reduceOp, adaptor, rewriter, *getTypeConverter()); |
| 455 | if (failed(reductionInfo)) |
| 456 | return failure(); |
| 457 | |
| 458 | auto [resultType, extractedElements] = *reductionInfo; |
| 459 | Location loc = reduceOp->getLoc(); |
| 460 | |
| 461 | // Handle boolean reductions with spirv.Any / spirv.All. |
| 462 | if (resultType.isInteger(1)) { |
| 463 | vector::CombiningKind kind = reduceOp.getKind(); |
| 464 | |
| 465 | if (kind == vector::CombiningKind::OR) { |
| 466 | Value result = spirv::AnyOp::create(rewriter, loc, resultType, |
| 467 | adaptor.getVector()); |
| 468 | if (Value acc = adaptor.getAcc()) |
| 469 | result = spirv::LogicalOrOp::create(rewriter, loc, resultType, result, |
| 470 | acc); |
| 471 | rewriter.replaceOp(op: reduceOp, newValues: result); |
| 472 | return success(); |
| 473 | } |
| 474 | |
| 475 | if (kind == vector::CombiningKind::AND) { |
| 476 | Value result = spirv::AllOp::create(rewriter, loc, resultType, |
| 477 | adaptor.getVector()); |
| 478 | if (Value acc = adaptor.getAcc()) |
| 479 | result = spirv::LogicalAndOp::create(rewriter, loc, resultType, |
| 480 | result, acc); |
| 481 | rewriter.replaceOp(op: reduceOp, newValues: result); |
| 482 | return success(); |
| 483 | } |
| 484 | } |
| 485 | |
| 486 | Value result = extractedElements.front(); |
| 487 | for (Value next : llvm::drop_begin(extractedElements)) { |
| 488 | switch (reduceOp.getKind()) { |
| 489 | |
| 490 | #define INT_AND_FLOAT_CASE(kind, iop, fop) \ |
| 491 | case vector::CombiningKind::kind: \ |
| 492 | if (isa<IntegerType>(resultType)) { \ |
| 493 | result = spirv::iop::create(rewriter, loc, resultType, result, next); \ |
| 494 | } else { \ |
| 495 | assert(isa<FloatType>(resultType)); \ |
| 496 | result = spirv::fop::create(rewriter, loc, resultType, result, next); \ |
| 497 | } \ |
| 498 | break |
| 499 | |
| 500 | #define INT_OR_FLOAT_CASE(kind, fop) \ |
| 501 | case vector::CombiningKind::kind: \ |
| 502 | result = fop::create(rewriter, loc, resultType, result, next); \ |
| 503 | break |
| 504 | |
| 505 | #define INT_CASE(kind, iop) \ |
| 506 | case vector::CombiningKind::kind: \ |
| 507 | assert(isa<IntegerType>(resultType)); \ |
| 508 | result = spirv::iop::create(rewriter, loc, resultType, result, next); \ |
| 509 | break |
| 510 | |
| 511 | INT_AND_FLOAT_CASE(ADD, IAddOp, FAddOp); |
| 512 | INT_AND_FLOAT_CASE(MUL, IMulOp, FMulOp); |
| 513 | INT_OR_FLOAT_CASE(MINUI, SPIRVUMinOp); |
| 514 | INT_OR_FLOAT_CASE(MINSI, SPIRVSMinOp); |
| 515 | INT_OR_FLOAT_CASE(MAXUI, SPIRVUMaxOp); |
| 516 | INT_OR_FLOAT_CASE(MAXSI, SPIRVSMaxOp); |
| 517 | INT_CASE(AND, BitwiseAndOp); |
| 518 | INT_CASE(OR, BitwiseOrOp); |
| 519 | INT_CASE(XOR, BitwiseXorOp); |
| 520 | |
| 521 | default: |
| 522 | return rewriter.notifyMatchFailure(arg&: reduceOp, msg: "not handled here" ); |
| 523 | } |
| 524 | #undef INT_AND_FLOAT_CASE |
| 525 | #undef INT_OR_FLOAT_CASE |
| 526 | #undef INT_CASE |
| 527 | } |
| 528 | |
| 529 | rewriter.replaceOp(op: reduceOp, newValues: result); |
| 530 | return success(); |
| 531 | } |
| 532 | }; |
| 533 | |
| 534 | template <typename SPIRVFMaxOp, typename SPIRVFMinOp> |
| 535 | struct VectorReductionFloatMinMax final |
| 536 | : OpConversionPattern<vector::ReductionOp> { |
| 537 | using Base::Base; |
| 538 | |
| 539 | LogicalResult |
| 540 | matchAndRewrite(vector::ReductionOp reduceOp, OpAdaptor adaptor, |
| 541 | ConversionPatternRewriter &rewriter) const override { |
| 542 | auto reductionInfo = |
| 543 | getReductionInfo(reduceOp, adaptor, rewriter, *getTypeConverter()); |
| 544 | if (failed(reductionInfo)) |
| 545 | return failure(); |
| 546 | |
| 547 | auto [resultType, extractedElements] = *reductionInfo; |
| 548 | Location loc = reduceOp->getLoc(); |
| 549 | Value result = extractedElements.front(); |
| 550 | for (Value next : llvm::drop_begin(extractedElements)) { |
| 551 | switch (reduceOp.getKind()) { |
| 552 | |
| 553 | #define INT_OR_FLOAT_CASE(kind, fop) \ |
| 554 | case vector::CombiningKind::kind: \ |
| 555 | result = fop::create(rewriter, loc, resultType, result, next); \ |
| 556 | break |
| 557 | |
| 558 | INT_OR_FLOAT_CASE(MAXIMUMF, SPIRVFMaxOp); |
| 559 | INT_OR_FLOAT_CASE(MINIMUMF, SPIRVFMinOp); |
| 560 | INT_OR_FLOAT_CASE(MAXNUMF, SPIRVFMaxOp); |
| 561 | INT_OR_FLOAT_CASE(MINNUMF, SPIRVFMinOp); |
| 562 | |
| 563 | default: |
| 564 | return rewriter.notifyMatchFailure(arg&: reduceOp, msg: "not handled here" ); |
| 565 | } |
| 566 | #undef INT_OR_FLOAT_CASE |
| 567 | } |
| 568 | |
| 569 | rewriter.replaceOp(op: reduceOp, newValues: result); |
| 570 | return success(); |
| 571 | } |
| 572 | }; |
| 573 | |
| 574 | class VectorScalarBroadcastPattern final |
| 575 | : public OpConversionPattern<vector::BroadcastOp> { |
| 576 | public: |
| 577 | using Base::Base; |
| 578 | |
| 579 | LogicalResult |
| 580 | matchAndRewrite(vector::BroadcastOp op, OpAdaptor adaptor, |
| 581 | ConversionPatternRewriter &rewriter) const override { |
| 582 | if (isa<VectorType>(Val: op.getSourceType())) { |
| 583 | return rewriter.notifyMatchFailure( |
| 584 | arg&: op, msg: "only conversion of 'broadcast from scalar' is supported" ); |
| 585 | } |
| 586 | Type dstType = getTypeConverter()->convertType(t: op.getType()); |
| 587 | if (!dstType) |
| 588 | return failure(); |
| 589 | if (isa<spirv::ScalarType>(Val: dstType)) { |
| 590 | rewriter.replaceOp(op, newValues: adaptor.getSource()); |
| 591 | } else { |
| 592 | auto dstVecType = cast<VectorType>(Val&: dstType); |
| 593 | SmallVector<Value, 4> source(dstVecType.getNumElements(), |
| 594 | adaptor.getSource()); |
| 595 | rewriter.replaceOpWithNewOp<spirv::CompositeConstructOp>(op, args&: dstType, |
| 596 | args&: source); |
| 597 | } |
| 598 | return success(); |
| 599 | } |
| 600 | }; |
| 601 | |
| 602 | struct VectorShuffleOpConvert final |
| 603 | : public OpConversionPattern<vector::ShuffleOp> { |
| 604 | using Base::Base; |
| 605 | |
| 606 | LogicalResult |
| 607 | matchAndRewrite(vector::ShuffleOp shuffleOp, OpAdaptor adaptor, |
| 608 | ConversionPatternRewriter &rewriter) const override { |
| 609 | VectorType oldResultType = shuffleOp.getResultVectorType(); |
| 610 | Type newResultType = getTypeConverter()->convertType(t: oldResultType); |
| 611 | if (!newResultType) |
| 612 | return rewriter.notifyMatchFailure(arg&: shuffleOp, |
| 613 | msg: "unsupported result vector type" ); |
| 614 | |
| 615 | auto mask = llvm::to_vector_of<int32_t>(Range: shuffleOp.getMask()); |
| 616 | |
| 617 | VectorType oldV1Type = shuffleOp.getV1VectorType(); |
| 618 | VectorType oldV2Type = shuffleOp.getV2VectorType(); |
| 619 | |
| 620 | // When both operands and the result are SPIR-V vectors, emit a SPIR-V |
| 621 | // shuffle. |
| 622 | if (oldV1Type.getNumElements() > 1 && oldV2Type.getNumElements() > 1 && |
| 623 | oldResultType.getNumElements() > 1) { |
| 624 | rewriter.replaceOpWithNewOp<spirv::VectorShuffleOp>( |
| 625 | op: shuffleOp, args&: newResultType, args: adaptor.getV1(), args: adaptor.getV2(), |
| 626 | args: rewriter.getI32ArrayAttr(values: mask)); |
| 627 | return success(); |
| 628 | } |
| 629 | |
| 630 | // When at least one of the operands or the result becomes a scalar after |
| 631 | // type conversion for SPIR-V, extract all the required elements and |
| 632 | // construct the result vector. |
| 633 | auto getElementAtIdx = [&rewriter, loc = shuffleOp.getLoc(), |
| 634 | scalarType = oldV1Type.getElementType()]( |
| 635 | Value scalarOrVec, int32_t idx) -> Value { |
| 636 | if (idx == vector::ShuffleOp::kPoisonIndex) |
| 637 | return spirv::UndefOp::create(builder&: rewriter, location: loc, result: scalarType); |
| 638 | |
| 639 | if (auto vecTy = dyn_cast<VectorType>(Val: scalarOrVec.getType())) |
| 640 | return spirv::CompositeExtractOp::create(builder&: rewriter, location: loc, composite: scalarOrVec, |
| 641 | indices: idx); |
| 642 | |
| 643 | assert(idx == 0 && "Invalid scalar element index" ); |
| 644 | return scalarOrVec; |
| 645 | }; |
| 646 | |
| 647 | int32_t numV1Elems = oldV1Type.getNumElements(); |
| 648 | SmallVector<Value> newOperands(mask.size()); |
| 649 | for (auto [shuffleIdx, newOperand] : llvm::zip_equal(t&: mask, u&: newOperands)) { |
| 650 | Value vec = adaptor.getV1(); |
| 651 | int32_t elementIdx = shuffleIdx; |
| 652 | if (elementIdx >= numV1Elems) { |
| 653 | vec = adaptor.getV2(); |
| 654 | elementIdx -= numV1Elems; |
| 655 | } |
| 656 | |
| 657 | newOperand = getElementAtIdx(vec, elementIdx); |
| 658 | } |
| 659 | |
| 660 | // Handle the scalar result corner case. |
| 661 | if (newOperands.size() == 1) { |
| 662 | rewriter.replaceOp(op: shuffleOp, newValues: newOperands.front()); |
| 663 | return success(); |
| 664 | } |
| 665 | |
| 666 | rewriter.replaceOpWithNewOp<spirv::CompositeConstructOp>( |
| 667 | op: shuffleOp, args&: newResultType, args&: newOperands); |
| 668 | return success(); |
| 669 | } |
| 670 | }; |
| 671 | |
| 672 | struct VectorInterleaveOpConvert final |
| 673 | : public OpConversionPattern<vector::InterleaveOp> { |
| 674 | using Base::Base; |
| 675 | |
| 676 | LogicalResult |
| 677 | matchAndRewrite(vector::InterleaveOp interleaveOp, OpAdaptor adaptor, |
| 678 | ConversionPatternRewriter &rewriter) const override { |
| 679 | // Check the result vector type. |
| 680 | VectorType oldResultType = interleaveOp.getResultVectorType(); |
| 681 | Type newResultType = getTypeConverter()->convertType(t: oldResultType); |
| 682 | if (!newResultType) |
| 683 | return rewriter.notifyMatchFailure(arg&: interleaveOp, |
| 684 | msg: "unsupported result vector type" ); |
| 685 | |
| 686 | // Interleave the indices. |
| 687 | VectorType sourceType = interleaveOp.getSourceVectorType(); |
| 688 | int n = sourceType.getNumElements(); |
| 689 | |
| 690 | // Input vectors of size 1 are converted to scalars by the type converter. |
| 691 | // We cannot use `spirv::VectorShuffleOp` directly in this case, and need to |
| 692 | // use `spirv::CompositeConstructOp`. |
| 693 | if (n == 1) { |
| 694 | Value newOperands[] = {adaptor.getLhs(), adaptor.getRhs()}; |
| 695 | rewriter.replaceOpWithNewOp<spirv::CompositeConstructOp>( |
| 696 | op: interleaveOp, args&: newResultType, args&: newOperands); |
| 697 | return success(); |
| 698 | } |
| 699 | |
| 700 | auto seq = llvm::seq<int64_t>(Size: 2 * n); |
| 701 | auto indices = llvm::map_to_vector( |
| 702 | C&: seq, F: [n](int i) { return (i % 2 ? n : 0) + i / 2; }); |
| 703 | |
| 704 | // Emit a SPIR-V shuffle. |
| 705 | rewriter.replaceOpWithNewOp<spirv::VectorShuffleOp>( |
| 706 | op: interleaveOp, args&: newResultType, args: adaptor.getLhs(), args: adaptor.getRhs(), |
| 707 | args: rewriter.getI32ArrayAttr(values: indices)); |
| 708 | |
| 709 | return success(); |
| 710 | } |
| 711 | }; |
| 712 | |
| 713 | struct VectorDeinterleaveOpConvert final |
| 714 | : public OpConversionPattern<vector::DeinterleaveOp> { |
| 715 | using Base::Base; |
| 716 | |
| 717 | LogicalResult |
| 718 | matchAndRewrite(vector::DeinterleaveOp deinterleaveOp, OpAdaptor adaptor, |
| 719 | ConversionPatternRewriter &rewriter) const override { |
| 720 | |
| 721 | // Check the result vector type. |
| 722 | VectorType oldResultType = deinterleaveOp.getResultVectorType(); |
| 723 | Type newResultType = getTypeConverter()->convertType(t: oldResultType); |
| 724 | if (!newResultType) |
| 725 | return rewriter.notifyMatchFailure(arg&: deinterleaveOp, |
| 726 | msg: "unsupported result vector type" ); |
| 727 | |
| 728 | Location loc = deinterleaveOp->getLoc(); |
| 729 | |
| 730 | // Deinterleave the indices. |
| 731 | Value sourceVector = adaptor.getSource(); |
| 732 | VectorType sourceType = deinterleaveOp.getSourceVectorType(); |
| 733 | int n = sourceType.getNumElements(); |
| 734 | |
| 735 | // Output vectors of size 1 are converted to scalars by the type converter. |
| 736 | // We cannot use `spirv::VectorShuffleOp` directly in this case, and need to |
| 737 | // use `spirv::CompositeExtractOp`. |
| 738 | if (n == 2) { |
| 739 | auto elem0 = spirv::CompositeExtractOp::create( |
| 740 | builder&: rewriter, location: loc, component: newResultType, composite: sourceVector, |
| 741 | indices: rewriter.getI32ArrayAttr(values: {0})); |
| 742 | |
| 743 | auto elem1 = spirv::CompositeExtractOp::create( |
| 744 | builder&: rewriter, location: loc, component: newResultType, composite: sourceVector, |
| 745 | indices: rewriter.getI32ArrayAttr(values: {1})); |
| 746 | |
| 747 | rewriter.replaceOp(op: deinterleaveOp, newValues: {elem0, elem1}); |
| 748 | return success(); |
| 749 | } |
| 750 | |
| 751 | // Indices for `shuffleEven` (result 0). |
| 752 | auto seqEven = llvm::seq<int64_t>(Size: n / 2); |
| 753 | auto indicesEven = |
| 754 | llvm::map_to_vector(C&: seqEven, F: [](int i) { return i * 2; }); |
| 755 | |
| 756 | // Indices for `shuffleOdd` (result 1). |
| 757 | auto seqOdd = llvm::seq<int64_t>(Size: n / 2); |
| 758 | auto indicesOdd = |
| 759 | llvm::map_to_vector(C&: seqOdd, F: [](int i) { return i * 2 + 1; }); |
| 760 | |
| 761 | // Create two SPIR-V shuffles. |
| 762 | auto shuffleEven = spirv::VectorShuffleOp::create( |
| 763 | builder&: rewriter, location: loc, result: newResultType, vector1: sourceVector, vector2: sourceVector, |
| 764 | components: rewriter.getI32ArrayAttr(values: indicesEven)); |
| 765 | |
| 766 | auto shuffleOdd = spirv::VectorShuffleOp::create( |
| 767 | builder&: rewriter, location: loc, result: newResultType, vector1: sourceVector, vector2: sourceVector, |
| 768 | components: rewriter.getI32ArrayAttr(values: indicesOdd)); |
| 769 | |
| 770 | rewriter.replaceOp(op: deinterleaveOp, newValues: {shuffleEven, shuffleOdd}); |
| 771 | return success(); |
| 772 | } |
| 773 | }; |
| 774 | |
| 775 | struct VectorLoadOpConverter final |
| 776 | : public OpConversionPattern<vector::LoadOp> { |
| 777 | using Base::Base; |
| 778 | |
| 779 | LogicalResult |
| 780 | matchAndRewrite(vector::LoadOp loadOp, OpAdaptor adaptor, |
| 781 | ConversionPatternRewriter &rewriter) const override { |
| 782 | auto memrefType = loadOp.getMemRefType(); |
| 783 | auto attr = |
| 784 | dyn_cast_or_null<spirv::StorageClassAttr>(Val: memrefType.getMemorySpace()); |
| 785 | if (!attr) |
| 786 | return rewriter.notifyMatchFailure( |
| 787 | arg&: loadOp, msg: "expected spirv.storage_class memory space" ); |
| 788 | |
| 789 | const auto &typeConverter = *getTypeConverter<SPIRVTypeConverter>(); |
| 790 | auto loc = loadOp.getLoc(); |
| 791 | Value accessChain = spirv::getElementPtr( |
| 792 | typeConverter, baseType: memrefType, basePtr: adaptor.getBase(), indices: adaptor.getIndices(), loc, |
| 793 | builder&: rewriter, accessElementCount: loadOp.getVectorType().getNumElements()); |
| 794 | if (!accessChain) |
| 795 | return rewriter.notifyMatchFailure( |
| 796 | arg&: loadOp, msg: "failed to get memref element pointer" ); |
| 797 | |
| 798 | spirv::StorageClass storageClass = attr.getValue(); |
| 799 | auto vectorType = loadOp.getVectorType(); |
| 800 | // Use the converted vector type instead of original (single element vector |
| 801 | // would get converted to scalar). |
| 802 | auto spirvVectorType = typeConverter.convertType(t: vectorType); |
| 803 | if (!spirvVectorType) |
| 804 | return rewriter.notifyMatchFailure(arg&: loadOp, msg: "unsupported vector type" ); |
| 805 | |
| 806 | auto vectorPtrType = spirv::PointerType::get(pointeeType: spirvVectorType, storageClass); |
| 807 | |
| 808 | // For single element vectors, we don't need to bitcast the access chain to |
| 809 | // the original vector type. Both is going to be the same, a pointer |
| 810 | // to a scalar. |
| 811 | Value castedAccessChain = |
| 812 | (vectorType.getNumElements() == 1) |
| 813 | ? accessChain |
| 814 | : spirv::BitcastOp::create(builder&: rewriter, location: loc, result: vectorPtrType, |
| 815 | operand: accessChain); |
| 816 | |
| 817 | auto memoryRequirements = calculateMemoryRequirements( |
| 818 | accessedPtr: castedAccessChain, preferredAlignment: loadOp.getAlignment().value_or(u: 0)); |
| 819 | if (failed(Result: memoryRequirements)) |
| 820 | return rewriter.notifyMatchFailure( |
| 821 | arg&: loadOp, msg: "failed to determine memory requirements" ); |
| 822 | |
| 823 | auto [memoryAccess, alignment] = *memoryRequirements; |
| 824 | rewriter.replaceOpWithNewOp<spirv::LoadOp>( |
| 825 | op: loadOp, args&: spirvVectorType, args&: castedAccessChain, args&: memoryAccess, args&: alignment); |
| 826 | |
| 827 | return success(); |
| 828 | } |
| 829 | }; |
| 830 | |
| 831 | struct VectorStoreOpConverter final |
| 832 | : public OpConversionPattern<vector::StoreOp> { |
| 833 | using Base::Base; |
| 834 | |
| 835 | LogicalResult |
| 836 | matchAndRewrite(vector::StoreOp storeOp, OpAdaptor adaptor, |
| 837 | ConversionPatternRewriter &rewriter) const override { |
| 838 | auto memrefType = storeOp.getMemRefType(); |
| 839 | auto attr = |
| 840 | dyn_cast_or_null<spirv::StorageClassAttr>(Val: memrefType.getMemorySpace()); |
| 841 | if (!attr) |
| 842 | return rewriter.notifyMatchFailure( |
| 843 | arg&: storeOp, msg: "expected spirv.storage_class memory space" ); |
| 844 | |
| 845 | const auto &typeConverter = *getTypeConverter<SPIRVTypeConverter>(); |
| 846 | auto loc = storeOp.getLoc(); |
| 847 | Value accessChain = spirv::getElementPtr( |
| 848 | typeConverter, baseType: memrefType, basePtr: adaptor.getBase(), indices: adaptor.getIndices(), loc, |
| 849 | builder&: rewriter, accessElementCount: storeOp.getVectorType().getNumElements()); |
| 850 | if (!accessChain) |
| 851 | return rewriter.notifyMatchFailure( |
| 852 | arg&: storeOp, msg: "failed to get memref element pointer" ); |
| 853 | |
| 854 | spirv::StorageClass storageClass = attr.getValue(); |
| 855 | auto vectorType = storeOp.getVectorType(); |
| 856 | // Use the converted vector type instead of original (single element vector |
| 857 | // would get converted to scalar). |
| 858 | auto spirvVectorType = typeConverter.convertType(t: vectorType); |
| 859 | if (!spirvVectorType) |
| 860 | return rewriter.notifyMatchFailure(arg&: storeOp, msg: "unsupported vector type" ); |
| 861 | |
| 862 | auto vectorPtrType = spirv::PointerType::get(pointeeType: spirvVectorType, storageClass); |
| 863 | |
| 864 | // For single element vectors, we don't need to bitcast the access chain to |
| 865 | // the original vector type. Both is going to be the same, a pointer |
| 866 | // to a scalar. |
| 867 | Value castedAccessChain = |
| 868 | (vectorType.getNumElements() == 1) |
| 869 | ? accessChain |
| 870 | : spirv::BitcastOp::create(builder&: rewriter, location: loc, result: vectorPtrType, |
| 871 | operand: accessChain); |
| 872 | |
| 873 | auto memoryRequirements = calculateMemoryRequirements( |
| 874 | accessedPtr: castedAccessChain, preferredAlignment: storeOp.getAlignment().value_or(u: 0)); |
| 875 | if (failed(Result: memoryRequirements)) |
| 876 | return rewriter.notifyMatchFailure( |
| 877 | arg&: storeOp, msg: "failed to determine memory requirements" ); |
| 878 | |
| 879 | auto [memoryAccess, alignment] = *memoryRequirements; |
| 880 | rewriter.replaceOpWithNewOp<spirv::StoreOp>(op: storeOp, args&: castedAccessChain, |
| 881 | args: adaptor.getValueToStore(), |
| 882 | args&: memoryAccess, args&: alignment); |
| 883 | |
| 884 | return success(); |
| 885 | } |
| 886 | }; |
| 887 | |
| 888 | struct VectorReductionToIntDotProd final |
| 889 | : OpRewritePattern<vector::ReductionOp> { |
| 890 | using Base::Base; |
| 891 | |
| 892 | LogicalResult matchAndRewrite(vector::ReductionOp op, |
| 893 | PatternRewriter &rewriter) const override { |
| 894 | if (op.getKind() != vector::CombiningKind::ADD) |
| 895 | return rewriter.notifyMatchFailure(arg&: op, msg: "combining kind is not 'add'" ); |
| 896 | |
| 897 | auto resultType = dyn_cast<IntegerType>(Val: op.getType()); |
| 898 | if (!resultType) |
| 899 | return rewriter.notifyMatchFailure(arg&: op, msg: "result is not an integer" ); |
| 900 | |
| 901 | int64_t resultBitwidth = resultType.getIntOrFloatBitWidth(); |
| 902 | if (!llvm::is_contained(Set: {32, 64}, Element: resultBitwidth)) |
| 903 | return rewriter.notifyMatchFailure(arg&: op, msg: "unsupported integer bitwidth" ); |
| 904 | |
| 905 | VectorType inVecTy = op.getSourceVectorType(); |
| 906 | if (!llvm::is_contained(Set: {4, 3}, Element: inVecTy.getNumElements()) || |
| 907 | inVecTy.getShape().size() != 1 || inVecTy.isScalable()) |
| 908 | return rewriter.notifyMatchFailure(arg&: op, msg: "unsupported vector shape" ); |
| 909 | |
| 910 | auto mul = op.getVector().getDefiningOp<arith::MulIOp>(); |
| 911 | if (!mul) |
| 912 | return rewriter.notifyMatchFailure( |
| 913 | arg&: op, msg: "reduction operand is not 'arith.muli'" ); |
| 914 | |
| 915 | if (succeeded(Result: handleCase<arith::ExtSIOp, arith::ExtSIOp, spirv::SDotOp, |
| 916 | spirv::SDotAccSatOp, false>(op, mul, rewriter))) |
| 917 | return success(); |
| 918 | |
| 919 | if (succeeded(Result: handleCase<arith::ExtUIOp, arith::ExtUIOp, spirv::UDotOp, |
| 920 | spirv::UDotAccSatOp, false>(op, mul, rewriter))) |
| 921 | return success(); |
| 922 | |
| 923 | if (succeeded(Result: handleCase<arith::ExtSIOp, arith::ExtUIOp, spirv::SUDotOp, |
| 924 | spirv::SUDotAccSatOp, false>(op, mul, rewriter))) |
| 925 | return success(); |
| 926 | |
| 927 | if (succeeded(Result: handleCase<arith::ExtUIOp, arith::ExtSIOp, spirv::SUDotOp, |
| 928 | spirv::SUDotAccSatOp, true>(op, mul, rewriter))) |
| 929 | return success(); |
| 930 | |
| 931 | return failure(); |
| 932 | } |
| 933 | |
| 934 | private: |
| 935 | template <typename LhsExtensionOp, typename RhsExtensionOp, typename DotOp, |
| 936 | typename DotAccOp, bool SwapOperands> |
| 937 | static LogicalResult handleCase(vector::ReductionOp op, arith::MulIOp mul, |
| 938 | PatternRewriter &rewriter) { |
| 939 | auto lhs = mul.getLhs().getDefiningOp<LhsExtensionOp>(); |
| 940 | if (!lhs) |
| 941 | return failure(); |
| 942 | Value lhsIn = lhs.getIn(); |
| 943 | auto lhsInType = cast<VectorType>(Val: lhsIn.getType()); |
| 944 | if (!lhsInType.getElementType().isInteger(width: 8)) |
| 945 | return failure(); |
| 946 | |
| 947 | auto rhs = mul.getRhs().getDefiningOp<RhsExtensionOp>(); |
| 948 | if (!rhs) |
| 949 | return failure(); |
| 950 | Value rhsIn = rhs.getIn(); |
| 951 | auto rhsInType = cast<VectorType>(Val: rhsIn.getType()); |
| 952 | if (!rhsInType.getElementType().isInteger(width: 8)) |
| 953 | return failure(); |
| 954 | |
| 955 | if (op.getSourceVectorType().getNumElements() == 3) { |
| 956 | IntegerType i8Type = rewriter.getI8Type(); |
| 957 | auto v4i8Type = VectorType::get(shape: {4}, elementType: i8Type); |
| 958 | Location loc = op.getLoc(); |
| 959 | Value zero = spirv::ConstantOp::getZero(type: i8Type, loc, builder&: rewriter); |
| 960 | lhsIn = spirv::CompositeConstructOp::create(builder&: rewriter, location: loc, result: v4i8Type, |
| 961 | constituents: ValueRange{lhsIn, zero}); |
| 962 | rhsIn = spirv::CompositeConstructOp::create(builder&: rewriter, location: loc, result: v4i8Type, |
| 963 | constituents: ValueRange{rhsIn, zero}); |
| 964 | } |
| 965 | |
| 966 | // There's no variant of dot prod ops for unsigned LHS and signed RHS, so |
| 967 | // we have to swap operands instead in that case. |
| 968 | if (SwapOperands) |
| 969 | std::swap(a&: lhsIn, b&: rhsIn); |
| 970 | |
| 971 | if (Value acc = op.getAcc()) { |
| 972 | rewriter.replaceOpWithNewOp<DotAccOp>(op, op.getType(), lhsIn, rhsIn, acc, |
| 973 | nullptr); |
| 974 | } else { |
| 975 | rewriter.replaceOpWithNewOp<DotOp>(op, op.getType(), lhsIn, rhsIn, |
| 976 | nullptr); |
| 977 | } |
| 978 | |
| 979 | return success(); |
| 980 | } |
| 981 | }; |
| 982 | |
| 983 | struct VectorReductionToFPDotProd final |
| 984 | : OpConversionPattern<vector::ReductionOp> { |
| 985 | using Base::Base; |
| 986 | |
| 987 | LogicalResult |
| 988 | matchAndRewrite(vector::ReductionOp op, OpAdaptor adaptor, |
| 989 | ConversionPatternRewriter &rewriter) const override { |
| 990 | if (op.getKind() != vector::CombiningKind::ADD) |
| 991 | return rewriter.notifyMatchFailure(arg&: op, msg: "combining kind is not 'add'" ); |
| 992 | |
| 993 | auto resultType = getTypeConverter()->convertType<FloatType>(t: op.getType()); |
| 994 | if (!resultType) |
| 995 | return rewriter.notifyMatchFailure(arg&: op, msg: "result is not a float" ); |
| 996 | |
| 997 | Value vec = adaptor.getVector(); |
| 998 | Value acc = adaptor.getAcc(); |
| 999 | |
| 1000 | auto vectorType = dyn_cast<VectorType>(Val: vec.getType()); |
| 1001 | if (!vectorType) { |
| 1002 | assert(isa<FloatType>(vec.getType()) && |
| 1003 | "Expected the vector to be scalarized" ); |
| 1004 | if (acc) { |
| 1005 | rewriter.replaceOpWithNewOp<spirv::FAddOp>(op, args&: acc, args&: vec); |
| 1006 | return success(); |
| 1007 | } |
| 1008 | |
| 1009 | rewriter.replaceOp(op, newValues: vec); |
| 1010 | return success(); |
| 1011 | } |
| 1012 | |
| 1013 | Location loc = op.getLoc(); |
| 1014 | Value lhs; |
| 1015 | Value rhs; |
| 1016 | if (auto mul = vec.getDefiningOp<arith::MulFOp>()) { |
| 1017 | lhs = mul.getLhs(); |
| 1018 | rhs = mul.getRhs(); |
| 1019 | } else { |
| 1020 | // If the operand is not a mul, use a vector of ones for the dot operand |
| 1021 | // to just sum up all values. |
| 1022 | lhs = vec; |
| 1023 | Attribute oneAttr = |
| 1024 | rewriter.getFloatAttr(type: vectorType.getElementType(), value: 1.0); |
| 1025 | oneAttr = SplatElementsAttr::get(type: vectorType, values: oneAttr); |
| 1026 | rhs = spirv::ConstantOp::create(builder&: rewriter, location: loc, constant: vectorType, value: oneAttr); |
| 1027 | } |
| 1028 | assert(lhs); |
| 1029 | assert(rhs); |
| 1030 | |
| 1031 | Value res = spirv::DotOp::create(builder&: rewriter, location: loc, result: resultType, vector1: lhs, vector2: rhs); |
| 1032 | if (acc) |
| 1033 | res = spirv::FAddOp::create(builder&: rewriter, location: loc, operand1: acc, operand2: res); |
| 1034 | |
| 1035 | rewriter.replaceOp(op, newValues: res); |
| 1036 | return success(); |
| 1037 | } |
| 1038 | }; |
| 1039 | |
| 1040 | struct VectorStepOpConvert final : OpConversionPattern<vector::StepOp> { |
| 1041 | using Base::Base; |
| 1042 | |
| 1043 | LogicalResult |
| 1044 | matchAndRewrite(vector::StepOp stepOp, OpAdaptor adaptor, |
| 1045 | ConversionPatternRewriter &rewriter) const override { |
| 1046 | const auto &typeConverter = *getTypeConverter<SPIRVTypeConverter>(); |
| 1047 | Type dstType = typeConverter.convertType(t: stepOp.getType()); |
| 1048 | if (!dstType) |
| 1049 | return failure(); |
| 1050 | |
| 1051 | Location loc = stepOp.getLoc(); |
| 1052 | int64_t numElements = stepOp.getType().getNumElements(); |
| 1053 | // Use the element type; the type converter collapses size-1 vectors to |
| 1054 | // scalars, so `dstType` may already be the scalar element type. |
| 1055 | Type intType = isa<VectorType>(Val: dstType) |
| 1056 | ? cast<VectorType>(Val&: dstType).getElementType() |
| 1057 | : dstType; |
| 1058 | |
| 1059 | // Input vectors of size 1 are converted to scalars by the type converter. |
| 1060 | // We just create a constant in this case. |
| 1061 | if (numElements == 1) { |
| 1062 | Value zero = spirv::ConstantOp::getZero(type: intType, loc, builder&: rewriter); |
| 1063 | rewriter.replaceOp(op: stepOp, newValues: zero); |
| 1064 | return success(); |
| 1065 | } |
| 1066 | |
| 1067 | SmallVector<Value> source; |
| 1068 | source.reserve(N: numElements); |
| 1069 | for (int64_t i = 0; i < numElements; ++i) { |
| 1070 | Attribute intAttr = rewriter.getIntegerAttr(type: intType, value: i); |
| 1071 | Value constOp = |
| 1072 | spirv::ConstantOp::create(builder&: rewriter, location: loc, constant: intType, value: intAttr); |
| 1073 | source.push_back(Elt: constOp); |
| 1074 | } |
| 1075 | rewriter.replaceOpWithNewOp<spirv::CompositeConstructOp>(op: stepOp, args&: dstType, |
| 1076 | args&: source); |
| 1077 | return success(); |
| 1078 | } |
| 1079 | }; |
| 1080 | |
| 1081 | struct VectorToElementOpConvert final |
| 1082 | : OpConversionPattern<vector::ToElementsOp> { |
| 1083 | using Base::Base; |
| 1084 | |
| 1085 | LogicalResult |
| 1086 | matchAndRewrite(vector::ToElementsOp toElementsOp, OpAdaptor adaptor, |
| 1087 | ConversionPatternRewriter &rewriter) const override { |
| 1088 | |
| 1089 | SmallVector<Value> results(toElementsOp->getNumResults()); |
| 1090 | Location loc = toElementsOp.getLoc(); |
| 1091 | |
| 1092 | // Input vectors of size 1 are converted to scalars by the type converter. |
| 1093 | // We cannot use `spirv::CompositeExtractOp` directly in this case. |
| 1094 | // For a scalar source, the result is just the scalar itself. |
| 1095 | if (isa<spirv::ScalarType>(Val: adaptor.getSource().getType())) { |
| 1096 | results[0] = adaptor.getSource(); |
| 1097 | rewriter.replaceOp(op: toElementsOp, newValues: results); |
| 1098 | return success(); |
| 1099 | } |
| 1100 | |
| 1101 | Type srcElementType = toElementsOp.getElements().getType().front(); |
| 1102 | Type elementType = getTypeConverter()->convertType(t: srcElementType); |
| 1103 | if (!elementType) |
| 1104 | return rewriter.notifyMatchFailure( |
| 1105 | arg&: toElementsOp, |
| 1106 | msg: llvm::formatv(Fmt: "failed to convert element type '{0}' to SPIR-V" , |
| 1107 | Vals&: srcElementType)); |
| 1108 | |
| 1109 | for (auto [idx, element] : llvm::enumerate(First: toElementsOp.getElements())) { |
| 1110 | // Create an CompositeExtract operation only for results that are not |
| 1111 | // dead. |
| 1112 | if (element.use_empty()) |
| 1113 | continue; |
| 1114 | |
| 1115 | Value result = spirv::CompositeExtractOp::create( |
| 1116 | builder&: rewriter, location: loc, component: elementType, composite: adaptor.getSource(), |
| 1117 | indices: rewriter.getI32ArrayAttr(values: {static_cast<int32_t>(idx)})); |
| 1118 | results[idx] = result; |
| 1119 | } |
| 1120 | |
| 1121 | rewriter.replaceOp(op: toElementsOp, newValues: results); |
| 1122 | return success(); |
| 1123 | } |
| 1124 | }; |
| 1125 | |
| 1126 | } // namespace |
| 1127 | #define CL_INT_MAX_MIN_OPS \ |
| 1128 | spirv::CLUMaxOp, spirv::CLUMinOp, spirv::CLSMaxOp, spirv::CLSMinOp |
| 1129 | |
| 1130 | #define GL_INT_MAX_MIN_OPS \ |
| 1131 | spirv::GLUMaxOp, spirv::GLUMinOp, spirv::GLSMaxOp, spirv::GLSMinOp |
| 1132 | |
| 1133 | #define CL_FLOAT_MAX_MIN_OPS spirv::CLFMaxOp, spirv::CLFMinOp |
| 1134 | #define GL_FLOAT_MAX_MIN_OPS spirv::GLFMaxOp, spirv::GLFMinOp |
| 1135 | |
| 1136 | void mlir::populateVectorToSPIRVPatterns( |
| 1137 | const SPIRVTypeConverter &typeConverter, RewritePatternSet &patterns) { |
| 1138 | patterns.add< |
| 1139 | VectorBitcastConvert, VectorBroadcastConvert, VectorExtractOpConvert, |
| 1140 | VectorExtractStridedSliceOpConvert, VectorFmaOpConvert<spirv::GLFmaOp>, |
| 1141 | VectorFmaOpConvert<spirv::CLFmaOp>, VectorFromElementsOpConvert, |
| 1142 | VectorToElementOpConvert, VectorInsertOpConvert, |
| 1143 | VectorReductionPattern<GL_INT_MAX_MIN_OPS>, |
| 1144 | VectorReductionPattern<CL_INT_MAX_MIN_OPS>, |
| 1145 | VectorReductionFloatMinMax<CL_FLOAT_MAX_MIN_OPS>, |
| 1146 | VectorReductionFloatMinMax<GL_FLOAT_MAX_MIN_OPS>, VectorShapeCast, |
| 1147 | VectorInsertStridedSliceOpConvert, VectorShuffleOpConvert, |
| 1148 | VectorInterleaveOpConvert, VectorDeinterleaveOpConvert, |
| 1149 | VectorScalarBroadcastPattern, VectorLoadOpConverter, |
| 1150 | VectorStoreOpConverter, VectorStepOpConvert>( |
| 1151 | arg: typeConverter, args: patterns.getContext(), args: PatternBenefit(1)); |
| 1152 | |
| 1153 | // Make sure that the more specialized dot product pattern has higher benefit |
| 1154 | // than the generic one that extracts all elements. |
| 1155 | patterns.add<VectorReductionToFPDotProd>(arg: typeConverter, args: patterns.getContext(), |
| 1156 | args: PatternBenefit(2)); |
| 1157 | } |
| 1158 | |
| 1159 | void mlir::populateVectorReductionToSPIRVDotProductPatterns( |
| 1160 | RewritePatternSet &patterns) { |
| 1161 | patterns.add<VectorReductionToIntDotProd>(arg: patterns.getContext()); |
| 1162 | } |
| 1163 | |