| 1 | //===- GPUToSPIRV.cpp - GPU 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 GPU dialect to SPIR-V dialect. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "mlir/Conversion/GPUToSPIRV/GPUToSPIRV.h" |
| 14 | #include "mlir/Dialect/GPU/IR/GPUDialect.h" |
| 15 | #include "mlir/Dialect/SPIRV/IR/SPIRVAttributes.h" |
| 16 | #include "mlir/Dialect/SPIRV/IR/SPIRVEnums.h" |
| 17 | #include "mlir/Dialect/SPIRV/IR/SPIRVOps.h" |
| 18 | #include "mlir/Dialect/SPIRV/IR/SPIRVTypes.h" |
| 19 | #include "mlir/Dialect/SPIRV/IR/TargetAndABI.h" |
| 20 | #include "mlir/Dialect/SPIRV/Transforms/SPIRVConversion.h" |
| 21 | #include "mlir/IR/Matchers.h" |
| 22 | #include "mlir/Transforms/DialectConversion.h" |
| 23 | #include <optional> |
| 24 | |
| 25 | using namespace mlir; |
| 26 | |
| 27 | static constexpr const char kSPIRVModule[] = "__spv__" ; |
| 28 | |
| 29 | namespace { |
| 30 | /// Pattern lowering GPU block/thread size/id to loading SPIR-V invocation |
| 31 | /// builtin variables. |
| 32 | template <typename SourceOp, spirv::BuiltIn builtin> |
| 33 | class LaunchConfigConversion : public OpConversionPattern<SourceOp> { |
| 34 | public: |
| 35 | using OpConversionPattern<SourceOp>::OpConversionPattern; |
| 36 | |
| 37 | LogicalResult |
| 38 | matchAndRewrite(SourceOp op, typename SourceOp::Adaptor adaptor, |
| 39 | ConversionPatternRewriter &rewriter) const override; |
| 40 | }; |
| 41 | |
| 42 | /// Pattern lowering subgroup size/id to loading SPIR-V invocation |
| 43 | /// builtin variables. |
| 44 | template <typename SourceOp, spirv::BuiltIn builtin> |
| 45 | class SingleDimLaunchConfigConversion : public OpConversionPattern<SourceOp> { |
| 46 | public: |
| 47 | using OpConversionPattern<SourceOp>::OpConversionPattern; |
| 48 | |
| 49 | LogicalResult |
| 50 | matchAndRewrite(SourceOp op, typename SourceOp::Adaptor adaptor, |
| 51 | ConversionPatternRewriter &rewriter) const override; |
| 52 | }; |
| 53 | |
| 54 | /// This is separate because in Vulkan workgroup size is exposed to shaders via |
| 55 | /// a constant with WorkgroupSize decoration. So here we cannot generate a |
| 56 | /// builtin variable; instead the information in the `spirv.entry_point_abi` |
| 57 | /// attribute on the surrounding FuncOp is used to replace the gpu::BlockDimOp. |
| 58 | class WorkGroupSizeConversion : public OpConversionPattern<gpu::BlockDimOp> { |
| 59 | public: |
| 60 | WorkGroupSizeConversion(const TypeConverter &typeConverter, |
| 61 | MLIRContext *context) |
| 62 | : OpConversionPattern(typeConverter, context, /*benefit*/ 10) {} |
| 63 | |
| 64 | LogicalResult |
| 65 | matchAndRewrite(gpu::BlockDimOp op, OpAdaptor adaptor, |
| 66 | ConversionPatternRewriter &rewriter) const override; |
| 67 | }; |
| 68 | |
| 69 | /// Pattern to convert a kernel function in GPU dialect within a spirv.module. |
| 70 | class GPUFuncOpConversion final : public OpConversionPattern<gpu::GPUFuncOp> { |
| 71 | public: |
| 72 | using Base::Base; |
| 73 | |
| 74 | LogicalResult |
| 75 | matchAndRewrite(gpu::GPUFuncOp funcOp, OpAdaptor adaptor, |
| 76 | ConversionPatternRewriter &rewriter) const override; |
| 77 | |
| 78 | private: |
| 79 | SmallVector<int32_t, 3> workGroupSizeAsInt32; |
| 80 | }; |
| 81 | |
| 82 | /// Pattern to convert a gpu.module to a spirv.module. |
| 83 | class GPUModuleConversion final : public OpConversionPattern<gpu::GPUModuleOp> { |
| 84 | public: |
| 85 | using Base::Base; |
| 86 | |
| 87 | LogicalResult |
| 88 | matchAndRewrite(gpu::GPUModuleOp moduleOp, OpAdaptor adaptor, |
| 89 | ConversionPatternRewriter &rewriter) const override; |
| 90 | }; |
| 91 | |
| 92 | /// Pattern to convert a gpu.return into a SPIR-V return. |
| 93 | // TODO: This can go to DRR when GPU return has operands. |
| 94 | class GPUReturnOpConversion final : public OpConversionPattern<gpu::ReturnOp> { |
| 95 | public: |
| 96 | using Base::Base; |
| 97 | |
| 98 | LogicalResult |
| 99 | matchAndRewrite(gpu::ReturnOp returnOp, OpAdaptor adaptor, |
| 100 | ConversionPatternRewriter &rewriter) const override; |
| 101 | }; |
| 102 | |
| 103 | /// Pattern to convert a gpu.barrier op into a spirv.ControlBarrier or |
| 104 | /// spirv.MemoryNamedBarrier op. |
| 105 | class GPUBarrierConversion final : public OpConversionPattern<gpu::BarrierOp> { |
| 106 | public: |
| 107 | using Base::Base; |
| 108 | |
| 109 | LogicalResult |
| 110 | matchAndRewrite(gpu::BarrierOp barrierOp, OpAdaptor adaptor, |
| 111 | ConversionPatternRewriter &rewriter) const override; |
| 112 | }; |
| 113 | |
| 114 | /// Pattern to convert a gpu.initialize_named_barrier into |
| 115 | /// spirv.NamedBarrierInitialize. |
| 116 | class GPUInitializeNamedBarrierConversion final |
| 117 | : public OpConversionPattern<gpu::InitializeNamedBarrierOp> { |
| 118 | public: |
| 119 | using Base::Base; |
| 120 | |
| 121 | LogicalResult |
| 122 | matchAndRewrite(gpu::InitializeNamedBarrierOp op, OpAdaptor adaptor, |
| 123 | ConversionPatternRewriter &rewriter) const override; |
| 124 | }; |
| 125 | |
| 126 | /// Pattern to convert a gpu.shuffle op into a spirv.GroupNonUniformShuffle op. |
| 127 | class GPUShuffleConversion final : public OpConversionPattern<gpu::ShuffleOp> { |
| 128 | public: |
| 129 | using Base::Base; |
| 130 | |
| 131 | LogicalResult |
| 132 | matchAndRewrite(gpu::ShuffleOp shuffleOp, OpAdaptor adaptor, |
| 133 | ConversionPatternRewriter &rewriter) const override; |
| 134 | }; |
| 135 | |
| 136 | /// Pattern to convert a gpu.rotate op into a spirv.GroupNonUniformRotateKHROp. |
| 137 | class GPURotateConversion final : public OpConversionPattern<gpu::RotateOp> { |
| 138 | public: |
| 139 | using Base::Base; |
| 140 | |
| 141 | LogicalResult |
| 142 | matchAndRewrite(gpu::RotateOp rotateOp, OpAdaptor adaptor, |
| 143 | ConversionPatternRewriter &rewriter) const override; |
| 144 | }; |
| 145 | |
| 146 | /// Pattern to convert a gpu.subgroup_broadcast op into a |
| 147 | /// spirv.GroupNonUniformBroadcast op. |
| 148 | class GPUSubgroupBroadcastConversion final |
| 149 | : public OpConversionPattern<gpu::SubgroupBroadcastOp> { |
| 150 | public: |
| 151 | using Base::Base; |
| 152 | |
| 153 | LogicalResult |
| 154 | matchAndRewrite(gpu::SubgroupBroadcastOp op, OpAdaptor adaptor, |
| 155 | ConversionPatternRewriter &rewriter) const override; |
| 156 | }; |
| 157 | |
| 158 | class GPUBallotConversion final : public OpConversionPattern<gpu::BallotOp> { |
| 159 | public: |
| 160 | using Base::Base; |
| 161 | |
| 162 | LogicalResult |
| 163 | matchAndRewrite(gpu::BallotOp ballotOp, OpAdaptor adaptor, |
| 164 | ConversionPatternRewriter &rewriter) const override; |
| 165 | }; |
| 166 | |
| 167 | class GPUPrintfConversion final : public OpConversionPattern<gpu::PrintfOp> { |
| 168 | public: |
| 169 | using Base::Base; |
| 170 | |
| 171 | LogicalResult |
| 172 | matchAndRewrite(gpu::PrintfOp gpuPrintfOp, OpAdaptor adaptor, |
| 173 | ConversionPatternRewriter &rewriter) const override; |
| 174 | }; |
| 175 | |
| 176 | } // namespace |
| 177 | |
| 178 | //===----------------------------------------------------------------------===// |
| 179 | // Builtins. |
| 180 | //===----------------------------------------------------------------------===// |
| 181 | |
| 182 | template <typename SourceOp, spirv::BuiltIn builtin> |
| 183 | LogicalResult LaunchConfigConversion<SourceOp, builtin>::matchAndRewrite( |
| 184 | SourceOp op, typename SourceOp::Adaptor adaptor, |
| 185 | ConversionPatternRewriter &rewriter) const { |
| 186 | auto *typeConverter = this->template getTypeConverter<SPIRVTypeConverter>(); |
| 187 | Type indexType = typeConverter->getIndexType(); |
| 188 | |
| 189 | // For Vulkan, these SPIR-V builtin variables are required to be a vector of |
| 190 | // type <3xi32> by the spec: |
| 191 | // https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/NumWorkgroups.html |
| 192 | // https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/WorkgroupId.html |
| 193 | // https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/WorkgroupSize.html |
| 194 | // https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/LocalInvocationId.html |
| 195 | // https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/LocalInvocationId.html |
| 196 | // https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/GlobalInvocationId.html |
| 197 | // |
| 198 | // For OpenCL, it depends on the Physical32/Physical64 addressing model: |
| 199 | // https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_Env.html#_built_in_variables |
| 200 | bool forShader = |
| 201 | typeConverter->getTargetEnv().allows(spirv::Capability::Shader); |
| 202 | Type builtinType = forShader ? rewriter.getIntegerType(width: 32) : indexType; |
| 203 | |
| 204 | Value vector = |
| 205 | spirv::getBuiltinVariableValue(op, builtin, integerType: builtinType, builder&: rewriter); |
| 206 | Value dim = spirv::CompositeExtractOp::create( |
| 207 | rewriter, op.getLoc(), builtinType, vector, |
| 208 | rewriter.getI32ArrayAttr(values: {static_cast<int32_t>(op.getDimension())})); |
| 209 | if (forShader && builtinType != indexType) |
| 210 | dim = spirv::UConvertOp::create(rewriter, op.getLoc(), indexType, dim); |
| 211 | rewriter.replaceOp(op, dim); |
| 212 | return success(); |
| 213 | } |
| 214 | |
| 215 | template <typename SourceOp, spirv::BuiltIn builtin> |
| 216 | LogicalResult |
| 217 | SingleDimLaunchConfigConversion<SourceOp, builtin>::matchAndRewrite( |
| 218 | SourceOp op, typename SourceOp::Adaptor adaptor, |
| 219 | ConversionPatternRewriter &rewriter) const { |
| 220 | auto *typeConverter = this->template getTypeConverter<SPIRVTypeConverter>(); |
| 221 | Type indexType = typeConverter->getIndexType(); |
| 222 | Type i32Type = rewriter.getIntegerType(width: 32); |
| 223 | |
| 224 | // For Vulkan, these SPIR-V builtin variables are required to be a vector of |
| 225 | // type i32 by the spec: |
| 226 | // https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/NumSubgroups.html |
| 227 | // https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/SubgroupId.html |
| 228 | // https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/SubgroupSize.html |
| 229 | // |
| 230 | // For OpenCL, they are also required to be i32: |
| 231 | // https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_Env.html#_built_in_variables |
| 232 | Value builtinValue = |
| 233 | spirv::getBuiltinVariableValue(op, builtin, integerType: i32Type, builder&: rewriter); |
| 234 | if (i32Type != indexType) |
| 235 | builtinValue = spirv::UConvertOp::create(rewriter, op.getLoc(), indexType, |
| 236 | builtinValue); |
| 237 | rewriter.replaceOp(op, builtinValue); |
| 238 | return success(); |
| 239 | } |
| 240 | |
| 241 | LogicalResult WorkGroupSizeConversion::matchAndRewrite( |
| 242 | gpu::BlockDimOp op, OpAdaptor adaptor, |
| 243 | ConversionPatternRewriter &rewriter) const { |
| 244 | DenseI32ArrayAttr workGroupSizeAttr = spirv::lookupLocalWorkGroupSize(op); |
| 245 | if (!workGroupSizeAttr) |
| 246 | return failure(); |
| 247 | |
| 248 | int val = |
| 249 | workGroupSizeAttr.asArrayRef()[static_cast<int32_t>(op.getDimension())]; |
| 250 | auto convertedType = |
| 251 | getTypeConverter()->convertType(t: op.getResult().getType()); |
| 252 | if (!convertedType) |
| 253 | return failure(); |
| 254 | rewriter.replaceOpWithNewOp<spirv::ConstantOp>( |
| 255 | op, args&: convertedType, args: IntegerAttr::get(type: convertedType, value: val)); |
| 256 | return success(); |
| 257 | } |
| 258 | |
| 259 | //===----------------------------------------------------------------------===// |
| 260 | // GPUFuncOp |
| 261 | //===----------------------------------------------------------------------===// |
| 262 | |
| 263 | // Legalizes a GPU function as an entry SPIR-V function. |
| 264 | static spirv::FuncOp |
| 265 | lowerAsEntryFunction(gpu::GPUFuncOp funcOp, const TypeConverter &typeConverter, |
| 266 | ConversionPatternRewriter &rewriter, |
| 267 | spirv::EntryPointABIAttr entryPointInfo, |
| 268 | ArrayRef<spirv::InterfaceVarABIAttr> argABIInfo) { |
| 269 | auto fnType = funcOp.getFunctionType(); |
| 270 | if (fnType.getNumResults()) { |
| 271 | funcOp.emitError(message: "SPIR-V lowering only supports entry functions" |
| 272 | "with no return values right now" ); |
| 273 | return nullptr; |
| 274 | } |
| 275 | if (!argABIInfo.empty() && fnType.getNumInputs() != argABIInfo.size()) { |
| 276 | funcOp.emitError( |
| 277 | message: "lowering as entry functions requires ABI info for all arguments " |
| 278 | "or none of them" ); |
| 279 | return nullptr; |
| 280 | } |
| 281 | // Update the signature to valid SPIR-V types and add the ABI |
| 282 | // attributes. These will be "materialized" by using the |
| 283 | // LowerABIAttributesPass. |
| 284 | TypeConverter::SignatureConversion signatureConverter(fnType.getNumInputs()); |
| 285 | { |
| 286 | for (const auto &argType : |
| 287 | enumerate(First: funcOp.getFunctionType().getInputs())) { |
| 288 | auto convertedType = typeConverter.convertType(t: argType.value()); |
| 289 | if (!convertedType) |
| 290 | return nullptr; |
| 291 | signatureConverter.addInputs(origInputNo: argType.index(), types: convertedType); |
| 292 | } |
| 293 | } |
| 294 | auto newFuncOp = spirv::FuncOp::create( |
| 295 | builder&: rewriter, location: funcOp.getLoc(), name: funcOp.getName(), |
| 296 | type: rewriter.getFunctionType(inputs: signatureConverter.getConvertedTypes(), results: {})); |
| 297 | newFuncOp.setArgAttrsAttr(funcOp.getArgAttrsAttr()); |
| 298 | newFuncOp.setResAttrsAttr(funcOp.getResAttrsAttr()); |
| 299 | cast<SymbolOpInterface>(Val: newFuncOp.getOperation()) |
| 300 | .setVisibility( |
| 301 | cast<SymbolOpInterface>(Val: funcOp.getOperation()).getVisibility()); |
| 302 | |
| 303 | auto copyGPUProperty = [&](StringAttr name, Attribute value) { |
| 304 | if (value) |
| 305 | newFuncOp->setDiscardableAttr(name, value); |
| 306 | }; |
| 307 | copyGPUProperty(funcOp.getWorkgroupAttribAttrsAttrName(), |
| 308 | funcOp.getWorkgroupAttribAttrsAttr()); |
| 309 | copyGPUProperty(funcOp.getPrivateAttribAttrsAttrName(), |
| 310 | funcOp.getPrivateAttribAttrsAttr()); |
| 311 | copyGPUProperty(funcOp.getKnownBlockSizeAttrName(), |
| 312 | funcOp.getKnownBlockSizeAttr()); |
| 313 | copyGPUProperty(funcOp.getKnownGridSizeAttrName(), |
| 314 | funcOp.getKnownGridSizeAttr()); |
| 315 | copyGPUProperty(funcOp.getKnownClusterSizeAttrName(), |
| 316 | funcOp.getKnownClusterSizeAttr()); |
| 317 | copyGPUProperty(funcOp.getWorkgroupAttributionsAttrName(), |
| 318 | funcOp.getWorkgroupAttributionsAttr()); |
| 319 | for (const auto &discardableAttr : |
| 320 | funcOp->getDiscardableAttrDictionary().getValue()) |
| 321 | newFuncOp->setDiscardableAttr(name: discardableAttr.getName(), |
| 322 | value: discardableAttr.getValue()); |
| 323 | |
| 324 | rewriter.inlineRegionBefore(region&: funcOp.getBody(), parent&: newFuncOp.getBody(), |
| 325 | before: newFuncOp.end()); |
| 326 | if (failed(Result: rewriter.convertRegionTypes(region: &newFuncOp.getBody(), converter: typeConverter, |
| 327 | entryConversion: &signatureConverter))) |
| 328 | return nullptr; |
| 329 | rewriter.eraseOp(op: funcOp); |
| 330 | |
| 331 | // Set the attributes for argument and the function. |
| 332 | StringRef argABIAttrName = spirv::getInterfaceVarABIAttrName(); |
| 333 | for (auto argIndex : llvm::seq<unsigned>(Begin: 0, End: argABIInfo.size())) { |
| 334 | newFuncOp.setArgAttr(index: argIndex, name: argABIAttrName, value: argABIInfo[argIndex]); |
| 335 | } |
| 336 | newFuncOp->setDiscardableAttr(name: spirv::getEntryPointABIAttrName(), |
| 337 | value: entryPointInfo); |
| 338 | |
| 339 | return newFuncOp; |
| 340 | } |
| 341 | |
| 342 | /// Populates `argABI` with spirv.interface_var_abi attributes for lowering |
| 343 | /// gpu.func to spirv.func if no arguments have the attributes set |
| 344 | /// already. Returns failure if any argument has the ABI attribute set already. |
| 345 | static LogicalResult |
| 346 | getDefaultABIAttrs(const spirv::TargetEnv &targetEnv, gpu::GPUFuncOp funcOp, |
| 347 | SmallVectorImpl<spirv::InterfaceVarABIAttr> &argABI) { |
| 348 | if (!spirv::needsInterfaceVarABIAttrs(targetAttr: targetEnv)) |
| 349 | return success(); |
| 350 | |
| 351 | for (auto argIndex : llvm::seq<unsigned>(Begin: 0, End: funcOp.getNumArguments())) { |
| 352 | if (funcOp.getArgAttrOfType<spirv::InterfaceVarABIAttr>( |
| 353 | index: argIndex, name: spirv::getInterfaceVarABIAttrName())) |
| 354 | return failure(); |
| 355 | // Vulkan's interface variable requirements needs scalars to be wrapped in a |
| 356 | // struct. The struct held in storage buffer. |
| 357 | std::optional<spirv::StorageClass> sc; |
| 358 | if (funcOp.getArgument(idx: argIndex).getType().isIntOrIndexOrFloat()) |
| 359 | sc = spirv::StorageClass::StorageBuffer; |
| 360 | argABI.push_back( |
| 361 | Elt: spirv::getInterfaceVarABIAttr(descriptorSet: 0, binding: argIndex, storageClass: sc, context: funcOp.getContext())); |
| 362 | } |
| 363 | return success(); |
| 364 | } |
| 365 | |
| 366 | LogicalResult GPUFuncOpConversion::matchAndRewrite( |
| 367 | gpu::GPUFuncOp funcOp, OpAdaptor adaptor, |
| 368 | ConversionPatternRewriter &rewriter) const { |
| 369 | if (!gpu::GPUDialect::isKernel(op: funcOp)) |
| 370 | return failure(); |
| 371 | |
| 372 | auto *typeConverter = getTypeConverter<SPIRVTypeConverter>(); |
| 373 | SmallVector<spirv::InterfaceVarABIAttr, 4> argABI; |
| 374 | if (failed( |
| 375 | Result: getDefaultABIAttrs(targetEnv: typeConverter->getTargetEnv(), funcOp, argABI))) { |
| 376 | argABI.clear(); |
| 377 | for (auto argIndex : llvm::seq<unsigned>(Begin: 0, End: funcOp.getNumArguments())) { |
| 378 | // If the ABI is already specified, use it. |
| 379 | auto abiAttr = funcOp.getArgAttrOfType<spirv::InterfaceVarABIAttr>( |
| 380 | index: argIndex, name: spirv::getInterfaceVarABIAttrName()); |
| 381 | if (!abiAttr) { |
| 382 | funcOp.emitRemark( |
| 383 | message: "match failure: missing 'spirv.interface_var_abi' attribute at " |
| 384 | "argument " ) |
| 385 | << argIndex; |
| 386 | return failure(); |
| 387 | } |
| 388 | argABI.push_back(Elt: abiAttr); |
| 389 | } |
| 390 | } |
| 391 | |
| 392 | auto entryPointAttr = spirv::lookupEntryPointABI(op: funcOp); |
| 393 | if (!entryPointAttr) { |
| 394 | funcOp.emitRemark( |
| 395 | message: "match failure: missing 'spirv.entry_point_abi' attribute" ); |
| 396 | return failure(); |
| 397 | } |
| 398 | spirv::FuncOp newFuncOp = lowerAsEntryFunction( |
| 399 | funcOp, typeConverter: *getTypeConverter(), rewriter, entryPointInfo: entryPointAttr, argABIInfo: argABI); |
| 400 | if (!newFuncOp) |
| 401 | return failure(); |
| 402 | newFuncOp->removeDiscardableAttr( |
| 403 | name: rewriter.getStringAttr(bytes: gpu::GPUDialect::getKernelFuncAttrName())); |
| 404 | return success(); |
| 405 | } |
| 406 | |
| 407 | //===----------------------------------------------------------------------===// |
| 408 | // ModuleOp with gpu.module. |
| 409 | //===----------------------------------------------------------------------===// |
| 410 | |
| 411 | LogicalResult GPUModuleConversion::matchAndRewrite( |
| 412 | gpu::GPUModuleOp moduleOp, OpAdaptor adaptor, |
| 413 | ConversionPatternRewriter &rewriter) const { |
| 414 | auto *typeConverter = getTypeConverter<SPIRVTypeConverter>(); |
| 415 | const spirv::TargetEnv &targetEnv = typeConverter->getTargetEnv(); |
| 416 | spirv::AddressingModel addressingModel = spirv::getAddressingModel( |
| 417 | targetAttr: targetEnv, use64bitAddress: typeConverter->getOptions().use64bitIndex); |
| 418 | FailureOr<spirv::MemoryModel> memoryModel = spirv::getMemoryModel(targetAttr: targetEnv); |
| 419 | if (failed(Result: memoryModel)) |
| 420 | return moduleOp.emitRemark( |
| 421 | message: "cannot deduce memory model from 'spirv.target_env'" ); |
| 422 | |
| 423 | // Add a keyword to the module name to avoid symbolic conflict. |
| 424 | std::string spvModuleName = (kSPIRVModule + moduleOp.getName()).str(); |
| 425 | auto spvModule = spirv::ModuleOp::create( |
| 426 | builder&: rewriter, location: moduleOp.getLoc(), addressing_model: addressingModel, memory_model: *memoryModel, vce_triple: std::nullopt, |
| 427 | name: StringRef(spvModuleName)); |
| 428 | |
| 429 | // Move the region from the module op into the SPIR-V module. |
| 430 | Region &spvModuleRegion = spvModule.getRegion(); |
| 431 | rewriter.inlineRegionBefore(region&: moduleOp.getBodyRegion(), parent&: spvModuleRegion, |
| 432 | before: spvModuleRegion.begin()); |
| 433 | // The spirv.module build method adds a block. Remove that. |
| 434 | rewriter.eraseBlock(block: &spvModuleRegion.back()); |
| 435 | |
| 436 | // Some of the patterns call `lookupTargetEnv` during conversion and they |
| 437 | // will fail if called after GPUModuleConversion and we don't preserve |
| 438 | // `TargetEnv` attribute. |
| 439 | // Copy TargetEnvAttr only if it is attached directly to the GPUModuleOp. |
| 440 | if (auto attr = moduleOp->getDiscardableAttrOfType<spirv::TargetEnvAttr>( |
| 441 | name: spirv::getTargetEnvAttrName())) |
| 442 | spvModule->setDiscardableAttr(name: spirv::getTargetEnvAttrName(), value: attr); |
| 443 | if (ArrayAttr targets = moduleOp.getTargetsAttr()) { |
| 444 | for (Attribute targetAttr : targets) |
| 445 | if (auto spirvTargetEnvAttr = |
| 446 | dyn_cast<spirv::TargetEnvAttr>(Val&: targetAttr)) { |
| 447 | spvModule->setDiscardableAttr(name: spirv::getTargetEnvAttrName(), |
| 448 | value: spirvTargetEnvAttr); |
| 449 | break; |
| 450 | } |
| 451 | } |
| 452 | |
| 453 | rewriter.eraseOp(op: moduleOp); |
| 454 | return success(); |
| 455 | } |
| 456 | |
| 457 | //===----------------------------------------------------------------------===// |
| 458 | // GPU return inside kernel functions to SPIR-V return. |
| 459 | //===----------------------------------------------------------------------===// |
| 460 | |
| 461 | LogicalResult GPUReturnOpConversion::matchAndRewrite( |
| 462 | gpu::ReturnOp returnOp, OpAdaptor adaptor, |
| 463 | ConversionPatternRewriter &rewriter) const { |
| 464 | if (!adaptor.getOperands().empty()) |
| 465 | return failure(); |
| 466 | |
| 467 | rewriter.replaceOpWithNewOp<spirv::ReturnOp>(op: returnOp); |
| 468 | return success(); |
| 469 | } |
| 470 | |
| 471 | //===----------------------------------------------------------------------===// |
| 472 | // Barrier. |
| 473 | //===----------------------------------------------------------------------===// |
| 474 | |
| 475 | /// Map gpu::BarrierScope to spirv::Scope. |
| 476 | static FailureOr<spirv::Scope> |
| 477 | mapGPUBarrierScopeToSPIRV(gpu::BarrierScope gpuScope) { |
| 478 | switch (gpuScope) { |
| 479 | case gpu::BarrierScope::Subgroup: |
| 480 | return spirv::Scope::Subgroup; |
| 481 | case gpu::BarrierScope::Workgroup: |
| 482 | return spirv::Scope::Workgroup; |
| 483 | case gpu::BarrierScope::Cluster: |
| 484 | return failure(); |
| 485 | } |
| 486 | return failure(); |
| 487 | } |
| 488 | |
| 489 | LogicalResult GPUBarrierConversion::matchAndRewrite( |
| 490 | gpu::BarrierOp barrierOp, OpAdaptor adaptor, |
| 491 | ConversionPatternRewriter &rewriter) const { |
| 492 | MLIRContext *context = getContext(); |
| 493 | |
| 494 | // Map GPU scope to SPIR-V scope. |
| 495 | auto spirvScope = mapGPUBarrierScopeToSPIRV(gpuScope: barrierOp.getScope()); |
| 496 | if (failed(Result: spirvScope)) |
| 497 | return rewriter.notifyMatchFailure( |
| 498 | arg&: barrierOp, msg: "cluster scope is not supported in SPIR-V" ); |
| 499 | |
| 500 | auto scopeAttr = spirv::ScopeAttr::get(context, value: *spirvScope); |
| 501 | auto memoryScopeAttr = |
| 502 | spirv::ScopeAttr::get(context, value: spirv::Scope::Workgroup); |
| 503 | |
| 504 | // Require acquire and release memory semantics for workgroup memory. |
| 505 | auto memorySemantics = spirv::MemorySemanticsAttr::get( |
| 506 | context, value: spirv::MemorySemantics::WorkgroupMemory | |
| 507 | spirv::MemorySemantics::AcquireRelease); |
| 508 | |
| 509 | if (adaptor.getNamedBarrier()) { |
| 510 | spirv::MemoryNamedBarrierOp::create(builder&: rewriter, location: barrierOp.getLoc(), |
| 511 | named_barrier: adaptor.getNamedBarrier(), |
| 512 | memory_scope: memoryScopeAttr, memory_semantics: memorySemantics); |
| 513 | rewriter.eraseOp(op: barrierOp); |
| 514 | } else { |
| 515 | rewriter.replaceOpWithNewOp<spirv::ControlBarrierOp>( |
| 516 | op: barrierOp, args&: scopeAttr, args&: memoryScopeAttr, args&: memorySemantics); |
| 517 | } |
| 518 | return success(); |
| 519 | } |
| 520 | |
| 521 | LogicalResult GPUInitializeNamedBarrierConversion::matchAndRewrite( |
| 522 | gpu::InitializeNamedBarrierOp op, OpAdaptor adaptor, |
| 523 | ConversionPatternRewriter &rewriter) const { |
| 524 | auto nbType = spirv::NamedBarrierType::get(context: getContext()); |
| 525 | rewriter.replaceOpWithNewOp<spirv::NamedBarrierInitializeOp>( |
| 526 | op, args&: nbType, args: adaptor.getMemberCount()); |
| 527 | return success(); |
| 528 | } |
| 529 | |
| 530 | //===----------------------------------------------------------------------===// |
| 531 | // Shuffle |
| 532 | //===----------------------------------------------------------------------===// |
| 533 | |
| 534 | LogicalResult GPUShuffleConversion::matchAndRewrite( |
| 535 | gpu::ShuffleOp shuffleOp, OpAdaptor adaptor, |
| 536 | ConversionPatternRewriter &rewriter) const { |
| 537 | // Require the shuffle width to be the same as the target's subgroup size, |
| 538 | // given that for SPIR-V non-uniform subgroup ops, we cannot select |
| 539 | // participating invocations. |
| 540 | const spirv::TargetEnv &targetEnv = |
| 541 | getTypeConverter<SPIRVTypeConverter>()->getTargetEnv(); |
| 542 | unsigned subgroupSize = |
| 543 | targetEnv.getAttr().getResourceLimits().getSubgroupSize(); |
| 544 | IntegerAttr widthAttr; |
| 545 | if (!matchPattern(value: shuffleOp.getWidth(), pattern: m_Constant(bind_value: &widthAttr)) || |
| 546 | widthAttr.getValue().getZExtValue() != subgroupSize) |
| 547 | return rewriter.notifyMatchFailure( |
| 548 | arg&: shuffleOp, msg: "shuffle width and target subgroup size mismatch" ); |
| 549 | |
| 550 | assert(!adaptor.getOffset().getType().isSignedInteger() && |
| 551 | "shuffle offset must be a signless/unsigned integer" ); |
| 552 | |
| 553 | Location loc = shuffleOp.getLoc(); |
| 554 | auto scope = rewriter.getAttr<spirv::ScopeAttr>(args: spirv::Scope::Subgroup); |
| 555 | Value result; |
| 556 | Value validVal; |
| 557 | |
| 558 | switch (shuffleOp.getMode()) { |
| 559 | case gpu::ShuffleMode::XOR: { |
| 560 | result = spirv::GroupNonUniformShuffleXorOp::create( |
| 561 | builder&: rewriter, location: loc, execution_scope: scope, value: adaptor.getValue(), mask: adaptor.getOffset()); |
| 562 | validVal = spirv::ConstantOp::getOne(type: rewriter.getI1Type(), |
| 563 | loc: shuffleOp.getLoc(), builder&: rewriter); |
| 564 | break; |
| 565 | } |
| 566 | case gpu::ShuffleMode::IDX: { |
| 567 | result = spirv::GroupNonUniformShuffleOp::create( |
| 568 | builder&: rewriter, location: loc, execution_scope: scope, value: adaptor.getValue(), id: adaptor.getOffset()); |
| 569 | validVal = spirv::ConstantOp::getOne(type: rewriter.getI1Type(), |
| 570 | loc: shuffleOp.getLoc(), builder&: rewriter); |
| 571 | break; |
| 572 | } |
| 573 | case gpu::ShuffleMode::DOWN: { |
| 574 | result = spirv::GroupNonUniformShuffleDownOp::create( |
| 575 | builder&: rewriter, location: loc, execution_scope: scope, value: adaptor.getValue(), delta: adaptor.getOffset()); |
| 576 | |
| 577 | Value laneId = gpu::LaneIdOp::create(builder&: rewriter, location: loc, upper_bound: widthAttr); |
| 578 | Value resultLaneId = |
| 579 | arith::AddIOp::create(builder&: rewriter, location: loc, lhs: laneId, rhs: adaptor.getOffset()); |
| 580 | validVal = arith::CmpIOp::create(builder&: rewriter, location: loc, predicate: arith::CmpIPredicate::ult, |
| 581 | lhs: resultLaneId, rhs: adaptor.getWidth()); |
| 582 | break; |
| 583 | } |
| 584 | case gpu::ShuffleMode::UP: { |
| 585 | result = spirv::GroupNonUniformShuffleUpOp::create( |
| 586 | builder&: rewriter, location: loc, execution_scope: scope, value: adaptor.getValue(), delta: adaptor.getOffset()); |
| 587 | |
| 588 | Value laneId = gpu::LaneIdOp::create(builder&: rewriter, location: loc, upper_bound: widthAttr); |
| 589 | Value resultLaneId = |
| 590 | arith::SubIOp::create(builder&: rewriter, location: loc, lhs: laneId, rhs: adaptor.getOffset()); |
| 591 | auto i32Type = rewriter.getIntegerType(width: 32); |
| 592 | validVal = arith::CmpIOp::create( |
| 593 | builder&: rewriter, location: loc, predicate: arith::CmpIPredicate::sge, lhs: resultLaneId, |
| 594 | rhs: arith::ConstantOp::create(builder&: rewriter, location: loc, result: i32Type, |
| 595 | value: rewriter.getIntegerAttr(type: i32Type, value: 0))); |
| 596 | break; |
| 597 | } |
| 598 | } |
| 599 | |
| 600 | rewriter.replaceOp(op: shuffleOp, newValues: {result, validVal}); |
| 601 | return success(); |
| 602 | } |
| 603 | |
| 604 | //===----------------------------------------------------------------------===// |
| 605 | // Rotate |
| 606 | //===----------------------------------------------------------------------===// |
| 607 | |
| 608 | LogicalResult GPURotateConversion::matchAndRewrite( |
| 609 | gpu::RotateOp rotateOp, OpAdaptor adaptor, |
| 610 | ConversionPatternRewriter &rewriter) const { |
| 611 | const spirv::TargetEnv &targetEnv = |
| 612 | getTypeConverter<SPIRVTypeConverter>()->getTargetEnv(); |
| 613 | unsigned subgroupSize = |
| 614 | targetEnv.getAttr().getResourceLimits().getSubgroupSize(); |
| 615 | unsigned width = rotateOp.getWidth(); |
| 616 | if (width > subgroupSize) |
| 617 | return rewriter.notifyMatchFailure( |
| 618 | arg&: rotateOp, msg: "rotate width is larger than target subgroup size" ); |
| 619 | |
| 620 | Location loc = rotateOp.getLoc(); |
| 621 | auto scope = rewriter.getAttr<spirv::ScopeAttr>(args: spirv::Scope::Subgroup); |
| 622 | Value offsetVal = |
| 623 | arith::ConstantOp::create(builder&: rewriter, location: loc, value: adaptor.getOffsetAttr()); |
| 624 | Value widthVal = |
| 625 | arith::ConstantOp::create(builder&: rewriter, location: loc, value: adaptor.getWidthAttr()); |
| 626 | Value rotateResult = spirv::GroupNonUniformRotateKHROp::create( |
| 627 | builder&: rewriter, location: loc, execution_scope: scope, value: adaptor.getValue(), delta: offsetVal, cluster_size: widthVal); |
| 628 | Value validVal; |
| 629 | if (width == subgroupSize) { |
| 630 | validVal = spirv::ConstantOp::getOne(type: rewriter.getI1Type(), loc, builder&: rewriter); |
| 631 | } else { |
| 632 | IntegerAttr widthAttr = adaptor.getWidthAttr(); |
| 633 | Value laneId = gpu::LaneIdOp::create(builder&: rewriter, location: loc, upper_bound: widthAttr); |
| 634 | validVal = arith::CmpIOp::create(builder&: rewriter, location: loc, predicate: arith::CmpIPredicate::ult, |
| 635 | lhs: laneId, rhs: widthVal); |
| 636 | } |
| 637 | |
| 638 | rewriter.replaceOp(op: rotateOp, newValues: {rotateResult, validVal}); |
| 639 | return success(); |
| 640 | } |
| 641 | |
| 642 | //===----------------------------------------------------------------------===// |
| 643 | // Subgroup broadcast |
| 644 | //===----------------------------------------------------------------------===// |
| 645 | |
| 646 | LogicalResult GPUSubgroupBroadcastConversion::matchAndRewrite( |
| 647 | gpu::SubgroupBroadcastOp op, OpAdaptor adaptor, |
| 648 | ConversionPatternRewriter &rewriter) const { |
| 649 | Location loc = op.getLoc(); |
| 650 | auto scope = rewriter.getAttr<spirv::ScopeAttr>(args: spirv::Scope::Subgroup); |
| 651 | Value result; |
| 652 | |
| 653 | switch (op.getBroadcastType()) { |
| 654 | case gpu::BroadcastType::specific_lane: |
| 655 | result = spirv::GroupNonUniformBroadcastOp::create( |
| 656 | builder&: rewriter, location: loc, execution_scope: scope, value: adaptor.getSrc(), id: adaptor.getLane()); |
| 657 | break; |
| 658 | case gpu::BroadcastType::first_active_lane: |
| 659 | result = spirv::GroupNonUniformBroadcastFirstOp::create( |
| 660 | builder&: rewriter, location: loc, execution_scope: scope, value: adaptor.getSrc()); |
| 661 | break; |
| 662 | } |
| 663 | |
| 664 | rewriter.replaceOp(op, newValues: result); |
| 665 | return success(); |
| 666 | } |
| 667 | |
| 668 | LogicalResult GPUBallotConversion::matchAndRewrite( |
| 669 | gpu::BallotOp ballotOp, OpAdaptor adaptor, |
| 670 | ConversionPatternRewriter &rewriter) const { |
| 671 | Location loc = ballotOp.getLoc(); |
| 672 | auto scope = rewriter.getAttr<spirv::ScopeAttr>(args: spirv::Scope::Subgroup); |
| 673 | auto int32Type = rewriter.getI32Type(); |
| 674 | auto vec4i32Type = VectorType::get(shape: {4}, elementType: int32Type); |
| 675 | |
| 676 | // SPIR-V ballot returns vector<4xi32> to support subgroups up to 128 lanes. |
| 677 | Value ballot = spirv::GroupNonUniformBallotOp::create( |
| 678 | builder&: rewriter, location: loc, result: vec4i32Type, execution_scope: scope, predicate: adaptor.getPredicate()); |
| 679 | |
| 680 | auto intType = cast<IntegerType>(Val: ballotOp.getType()); |
| 681 | unsigned width = intType.getWidth(); |
| 682 | |
| 683 | if (width == 32) { |
| 684 | Value result = |
| 685 | spirv::CompositeExtractOp::create(builder&: rewriter, location: loc, composite: ballot, indices: {0}); |
| 686 | rewriter.replaceOp(op: ballotOp, newValues: result); |
| 687 | } else if (width == 64) { |
| 688 | // Combine first two vector elements: low 32 bits + (high 32 bits << 32). |
| 689 | Value low = spirv::CompositeExtractOp::create(builder&: rewriter, location: loc, composite: ballot, indices: {0}); |
| 690 | Value high = spirv::CompositeExtractOp::create(builder&: rewriter, location: loc, composite: ballot, indices: {1}); |
| 691 | |
| 692 | auto int64Type = rewriter.getI64Type(); |
| 693 | Value lowExt = spirv::UConvertOp::create(builder&: rewriter, location: loc, result: int64Type, operand: low); |
| 694 | Value highExt = spirv::UConvertOp::create(builder&: rewriter, location: loc, result: int64Type, operand: high); |
| 695 | |
| 696 | Value shift32 = spirv::ConstantOp::create( |
| 697 | builder&: rewriter, location: loc, constant: int64Type, value: rewriter.getIntegerAttr(type: int64Type, value: 32)); |
| 698 | Value highShifted = |
| 699 | spirv::ShiftLeftLogicalOp::create(builder&: rewriter, location: loc, operand1: highExt, operand2: shift32); |
| 700 | |
| 701 | Value result = |
| 702 | spirv::BitwiseOrOp::create(builder&: rewriter, location: loc, operand1: lowExt, operand2: highShifted); |
| 703 | rewriter.replaceOp(op: ballotOp, newValues: result); |
| 704 | } else { |
| 705 | return rewriter.notifyMatchFailure( |
| 706 | arg&: ballotOp, msg: "only i32 and i64 result types are supported for SPIR-V" ); |
| 707 | } |
| 708 | |
| 709 | return success(); |
| 710 | } |
| 711 | |
| 712 | //===----------------------------------------------------------------------===// |
| 713 | // Group ops |
| 714 | //===----------------------------------------------------------------------===// |
| 715 | |
| 716 | template <typename UniformOp, typename NonUniformOp> |
| 717 | static Value createGroupReduceOpImpl(OpBuilder &builder, Location loc, |
| 718 | Value arg, bool isGroup, bool isUniform, |
| 719 | std::optional<uint32_t> clusterSize) { |
| 720 | spirv::Scope scope = |
| 721 | isGroup ? spirv::Scope::Workgroup : spirv::Scope::Subgroup; |
| 722 | // GroupNonUniform* ops only support Subgroup scope. |
| 723 | if (!isUniform && scope != spirv::Scope::Subgroup) |
| 724 | return Value(); |
| 725 | |
| 726 | Type type = arg.getType(); |
| 727 | auto scopeAttr = mlir::spirv::ScopeAttr::get(context: builder.getContext(), value: scope); |
| 728 | auto groupOp = spirv::GroupOperationAttr::get( |
| 729 | context: builder.getContext(), value: clusterSize.has_value() |
| 730 | ? spirv::GroupOperation::ClusteredReduce |
| 731 | : spirv::GroupOperation::Reduce); |
| 732 | if (isUniform) { |
| 733 | return UniformOp::create(builder, loc, type, scopeAttr, groupOp, arg) |
| 734 | .getResult(); |
| 735 | } |
| 736 | |
| 737 | Value clusterSizeValue; |
| 738 | if (clusterSize.has_value()) |
| 739 | clusterSizeValue = spirv::ConstantOp::create( |
| 740 | builder, location: loc, constant: builder.getI32Type(), |
| 741 | value: builder.getIntegerAttr(type: builder.getI32Type(), value: *clusterSize)); |
| 742 | |
| 743 | return NonUniformOp::create(builder, loc, type, scopeAttr, groupOp, arg, |
| 744 | clusterSizeValue) |
| 745 | .getResult(); |
| 746 | } |
| 747 | |
| 748 | template <typename NonUniformOp> |
| 749 | static Value createGroupNonUniformBitwiseReduceOpImpl( |
| 750 | OpBuilder &builder, Location loc, Value arg, bool isGroup, bool isUniform, |
| 751 | std::optional<uint32_t> clusterSize) { |
| 752 | spirv::Scope scope = |
| 753 | isGroup ? spirv::Scope::Workgroup : spirv::Scope::Subgroup; |
| 754 | if (isUniform || scope != spirv::Scope::Subgroup) |
| 755 | return Value(); |
| 756 | |
| 757 | Type type = arg.getType(); |
| 758 | auto scopeAttr = mlir::spirv::ScopeAttr::get(context: builder.getContext(), value: scope); |
| 759 | auto groupOp = spirv::GroupOperationAttr::get( |
| 760 | context: builder.getContext(), value: clusterSize.has_value() |
| 761 | ? spirv::GroupOperation::ClusteredReduce |
| 762 | : spirv::GroupOperation::Reduce); |
| 763 | |
| 764 | Value clusterSizeValue; |
| 765 | if (clusterSize.has_value()) |
| 766 | clusterSizeValue = spirv::ConstantOp::create( |
| 767 | builder, location: loc, constant: builder.getI32Type(), |
| 768 | value: builder.getIntegerAttr(type: builder.getI32Type(), value: *clusterSize)); |
| 769 | |
| 770 | return NonUniformOp::create(builder, loc, type, scopeAttr, groupOp, arg, |
| 771 | clusterSizeValue) |
| 772 | .getResult(); |
| 773 | } |
| 774 | |
| 775 | static std::optional<Value> |
| 776 | createGroupReduceOp(OpBuilder &builder, Location loc, Value arg, |
| 777 | gpu::AllReduceOperation opType, bool isGroup, |
| 778 | bool isUniform, std::optional<uint32_t> clusterSize) { |
| 779 | enum class ElemType { Float, Boolean, Integer }; |
| 780 | using FuncT = Value (*)(OpBuilder &, Location, Value, bool, bool, |
| 781 | std::optional<uint32_t>); |
| 782 | struct OpHandler { |
| 783 | gpu::AllReduceOperation kind; |
| 784 | ElemType elemType; |
| 785 | FuncT func; |
| 786 | }; |
| 787 | |
| 788 | Type type = arg.getType(); |
| 789 | ElemType elementType; |
| 790 | if (isa<FloatType>(Val: type)) { |
| 791 | elementType = ElemType::Float; |
| 792 | } else if (auto intTy = dyn_cast<IntegerType>(Val&: type)) { |
| 793 | elementType = (intTy.getIntOrFloatBitWidth() == 1) ? ElemType::Boolean |
| 794 | : ElemType::Integer; |
| 795 | } else { |
| 796 | return std::nullopt; |
| 797 | } |
| 798 | |
| 799 | // TODO(https://github.com/llvm/llvm-project/issues/73459): The SPIR-V spec |
| 800 | // does not specify how -0.0 / +0.0 and NaN values are handled in *FMin/*FMax |
| 801 | // reduction ops. We should account possible precision requirements in this |
| 802 | // conversion. |
| 803 | |
| 804 | using ReduceType = gpu::AllReduceOperation; |
| 805 | const OpHandler handlers[] = { |
| 806 | {.kind: ReduceType::ADD, .elemType: ElemType::Integer, |
| 807 | .func: &createGroupReduceOpImpl<spirv::GroupIAddOp, |
| 808 | spirv::GroupNonUniformIAddOp>}, |
| 809 | {.kind: ReduceType::ADD, .elemType: ElemType::Float, |
| 810 | .func: &createGroupReduceOpImpl<spirv::GroupFAddOp, |
| 811 | spirv::GroupNonUniformFAddOp>}, |
| 812 | {.kind: ReduceType::MUL, .elemType: ElemType::Integer, |
| 813 | .func: &createGroupReduceOpImpl<spirv::GroupIMulKHROp, |
| 814 | spirv::GroupNonUniformIMulOp>}, |
| 815 | {.kind: ReduceType::MUL, .elemType: ElemType::Float, |
| 816 | .func: &createGroupReduceOpImpl<spirv::GroupFMulKHROp, |
| 817 | spirv::GroupNonUniformFMulOp>}, |
| 818 | {.kind: ReduceType::MINUI, .elemType: ElemType::Integer, |
| 819 | .func: &createGroupReduceOpImpl<spirv::GroupUMinOp, |
| 820 | spirv::GroupNonUniformUMinOp>}, |
| 821 | {.kind: ReduceType::MINSI, .elemType: ElemType::Integer, |
| 822 | .func: &createGroupReduceOpImpl<spirv::GroupSMinOp, |
| 823 | spirv::GroupNonUniformSMinOp>}, |
| 824 | {.kind: ReduceType::MINNUMF, .elemType: ElemType::Float, |
| 825 | .func: &createGroupReduceOpImpl<spirv::GroupFMinOp, |
| 826 | spirv::GroupNonUniformFMinOp>}, |
| 827 | {.kind: ReduceType::MAXUI, .elemType: ElemType::Integer, |
| 828 | .func: &createGroupReduceOpImpl<spirv::GroupUMaxOp, |
| 829 | spirv::GroupNonUniformUMaxOp>}, |
| 830 | {.kind: ReduceType::MAXSI, .elemType: ElemType::Integer, |
| 831 | .func: &createGroupReduceOpImpl<spirv::GroupSMaxOp, |
| 832 | spirv::GroupNonUniformSMaxOp>}, |
| 833 | {.kind: ReduceType::MAXNUMF, .elemType: ElemType::Float, |
| 834 | .func: &createGroupReduceOpImpl<spirv::GroupFMaxOp, |
| 835 | spirv::GroupNonUniformFMaxOp>}, |
| 836 | {.kind: ReduceType::MINIMUMF, .elemType: ElemType::Float, |
| 837 | .func: &createGroupReduceOpImpl<spirv::GroupFMinOp, |
| 838 | spirv::GroupNonUniformFMinOp>}, |
| 839 | {.kind: ReduceType::MAXIMUMF, .elemType: ElemType::Float, |
| 840 | .func: &createGroupReduceOpImpl<spirv::GroupFMaxOp, |
| 841 | spirv::GroupNonUniformFMaxOp>}, |
| 842 | {.kind: ReduceType::AND, .elemType: ElemType::Integer, |
| 843 | .func: &createGroupNonUniformBitwiseReduceOpImpl< |
| 844 | spirv::GroupNonUniformBitwiseAndOp>}, |
| 845 | {.kind: ReduceType::OR, .elemType: ElemType::Integer, |
| 846 | .func: &createGroupNonUniformBitwiseReduceOpImpl< |
| 847 | spirv::GroupNonUniformBitwiseOrOp>}, |
| 848 | {.kind: ReduceType::XOR, .elemType: ElemType::Integer, |
| 849 | .func: &createGroupNonUniformBitwiseReduceOpImpl< |
| 850 | spirv::GroupNonUniformBitwiseXorOp>}}; |
| 851 | |
| 852 | for (const OpHandler &handler : handlers) |
| 853 | if (handler.kind == opType && elementType == handler.elemType) |
| 854 | if (Value result = |
| 855 | handler.func(builder, loc, arg, isGroup, isUniform, clusterSize)) |
| 856 | return result; |
| 857 | |
| 858 | return std::nullopt; |
| 859 | } |
| 860 | |
| 861 | /// Pattern to convert a gpu.all_reduce op into a SPIR-V group op. |
| 862 | class GPUAllReduceConversion final |
| 863 | : public OpConversionPattern<gpu::AllReduceOp> { |
| 864 | public: |
| 865 | using Base::Base; |
| 866 | |
| 867 | LogicalResult |
| 868 | matchAndRewrite(gpu::AllReduceOp op, OpAdaptor adaptor, |
| 869 | ConversionPatternRewriter &rewriter) const override { |
| 870 | auto opType = op.getOp(); |
| 871 | |
| 872 | // gpu.all_reduce can have either reduction op attribute or reduction |
| 873 | // region. Only attribute version is supported. |
| 874 | if (!opType) |
| 875 | return failure(); |
| 876 | |
| 877 | auto result = |
| 878 | createGroupReduceOp(builder&: rewriter, loc: op.getLoc(), arg: adaptor.getValue(), opType: *opType, |
| 879 | /*isGroup*/ true, isUniform: op.getUniform(), clusterSize: std::nullopt); |
| 880 | if (!result) |
| 881 | return failure(); |
| 882 | |
| 883 | rewriter.replaceOp(op, newValues: *result); |
| 884 | return success(); |
| 885 | } |
| 886 | }; |
| 887 | |
| 888 | /// Pattern to convert a gpu.subgroup_reduce op into a SPIR-V group op. |
| 889 | class GPUSubgroupReduceConversion final |
| 890 | : public OpConversionPattern<gpu::SubgroupReduceOp> { |
| 891 | public: |
| 892 | using Base::Base; |
| 893 | |
| 894 | LogicalResult |
| 895 | matchAndRewrite(gpu::SubgroupReduceOp op, OpAdaptor adaptor, |
| 896 | ConversionPatternRewriter &rewriter) const override { |
| 897 | if (op.getClusterStride() > 1) { |
| 898 | return rewriter.notifyMatchFailure( |
| 899 | arg&: op, msg: "lowering for cluster stride > 1 is not implemented" ); |
| 900 | } |
| 901 | |
| 902 | if (!isa<spirv::ScalarType>(Val: adaptor.getValue().getType())) |
| 903 | return rewriter.notifyMatchFailure(arg&: op, msg: "reduction type is not a scalar" ); |
| 904 | |
| 905 | auto result = createGroupReduceOp( |
| 906 | builder&: rewriter, loc: op.getLoc(), arg: adaptor.getValue(), opType: adaptor.getOp(), |
| 907 | /*isGroup=*/false, isUniform: adaptor.getUniform(), clusterSize: op.getClusterSize()); |
| 908 | if (!result) |
| 909 | return failure(); |
| 910 | |
| 911 | rewriter.replaceOp(op, newValues: *result); |
| 912 | return success(); |
| 913 | } |
| 914 | }; |
| 915 | |
| 916 | // Formulate a unique variable/constant name after |
| 917 | // searching in the module for existing variable/constant names. |
| 918 | // This is to avoid name collision with existing variables. |
| 919 | // Example: printfMsg0, printfMsg1, printfMsg2, ... |
| 920 | static std::string makeVarName(spirv::ModuleOp moduleOp, llvm::Twine prefix) { |
| 921 | std::string name; |
| 922 | unsigned number = 0; |
| 923 | |
| 924 | do { |
| 925 | name.clear(); |
| 926 | name = (prefix + llvm::Twine(number++)).str(); |
| 927 | } while (moduleOp.lookupSymbol(name)); |
| 928 | |
| 929 | return name; |
| 930 | } |
| 931 | |
| 932 | /// Pattern to convert a gpu.printf op into a SPIR-V CLPrintf op. |
| 933 | |
| 934 | LogicalResult GPUPrintfConversion::matchAndRewrite( |
| 935 | gpu::PrintfOp gpuPrintfOp, OpAdaptor adaptor, |
| 936 | ConversionPatternRewriter &rewriter) const { |
| 937 | |
| 938 | Location loc = gpuPrintfOp.getLoc(); |
| 939 | |
| 940 | auto moduleOp = gpuPrintfOp->getParentOfType<spirv::ModuleOp>(); |
| 941 | if (!moduleOp) |
| 942 | return failure(); |
| 943 | |
| 944 | // SPIR-V global variable is used to initialize printf |
| 945 | // format string value, if there are multiple printf messages, |
| 946 | // each global var needs to be created with a unique name. |
| 947 | std::string globalVarName = makeVarName(moduleOp, prefix: llvm::Twine("printfMsg" )); |
| 948 | spirv::GlobalVariableOp globalVar; |
| 949 | |
| 950 | IntegerType i8Type = rewriter.getI8Type(); |
| 951 | IntegerType i32Type = rewriter.getI32Type(); |
| 952 | |
| 953 | // Each character of printf format string is |
| 954 | // stored as a spec constant. We need to create |
| 955 | // unique name for this spec constant like |
| 956 | // @printfMsg0_sc0, @printfMsg0_sc1, ... by searching in the module |
| 957 | // for existing spec constant names. |
| 958 | auto createSpecConstant = [&](unsigned value) { |
| 959 | auto attr = rewriter.getI8IntegerAttr(value); |
| 960 | std::string specCstName = |
| 961 | makeVarName(moduleOp, prefix: llvm::Twine(globalVarName) + "_sc" ); |
| 962 | |
| 963 | return spirv::SpecConstantOp::create( |
| 964 | builder&: rewriter, location: loc, sym_name: rewriter.getStringAttr(bytes: specCstName), default_value: attr, |
| 965 | /*sym_visibility=*/nullptr); |
| 966 | }; |
| 967 | { |
| 968 | Operation *parent = |
| 969 | SymbolTable::getNearestSymbolTable(from: gpuPrintfOp->getParentOp()); |
| 970 | |
| 971 | ConversionPatternRewriter::InsertionGuard guard(rewriter); |
| 972 | |
| 973 | Block &entryBlock = *parent->getRegion(index: 0).begin(); |
| 974 | rewriter.setInsertionPointToStart( |
| 975 | &entryBlock); // insertion point at module level |
| 976 | |
| 977 | // Create Constituents with SpecConstant by scanning format string |
| 978 | // Each character of format string is stored as a spec constant |
| 979 | // and then these spec constants are used to create a |
| 980 | // SpecConstantCompositeOp. |
| 981 | llvm::SmallString<20> formatString(adaptor.getFormat()); |
| 982 | formatString.push_back(Elt: '\0'); // Null terminate for C. |
| 983 | SmallVector<Attribute, 4> constituents; |
| 984 | for (char c : formatString) { |
| 985 | spirv::SpecConstantOp cSpecConstantOp = createSpecConstant(c); |
| 986 | constituents.push_back(Elt: SymbolRefAttr::get(symbol: cSpecConstantOp)); |
| 987 | } |
| 988 | |
| 989 | // Create SpecConstantCompositeOp to initialize the global variable |
| 990 | size_t contentSize = constituents.size(); |
| 991 | auto globalType = spirv::ArrayType::get(elementType: i8Type, elementCount: contentSize); |
| 992 | spirv::SpecConstantCompositeOp specCstComposite; |
| 993 | // There will be one SpecConstantCompositeOp per printf message/global var, |
| 994 | // so no need do lookup for existing ones. |
| 995 | std::string specCstCompositeName = |
| 996 | (llvm::Twine(globalVarName) + "_scc" ).str(); |
| 997 | |
| 998 | specCstComposite = spirv::SpecConstantCompositeOp::create( |
| 999 | builder&: rewriter, location: loc, type: TypeAttr::get(type: globalType), |
| 1000 | sym_name: rewriter.getStringAttr(bytes: specCstCompositeName), |
| 1001 | constituents: rewriter.getArrayAttr(value: constituents), /*sym_visibility=*/nullptr); |
| 1002 | |
| 1003 | auto ptrType = spirv::PointerType::get( |
| 1004 | pointeeType: globalType, storageClass: spirv::StorageClass::UniformConstant); |
| 1005 | |
| 1006 | // Define a GlobalVarOp initialized using specialized constants |
| 1007 | // that is used to specify the printf format string |
| 1008 | // to be passed to the SPIRV CLPrintfOp. |
| 1009 | globalVar = spirv::GlobalVariableOp::create( |
| 1010 | builder&: rewriter, location: loc, type: ptrType, sym_name: globalVarName, |
| 1011 | initializer: FlatSymbolRefAttr::get(symbol: specCstComposite)); |
| 1012 | |
| 1013 | globalVar->setDiscardableAttr(name: "Constant" , value: rewriter.getUnitAttr()); |
| 1014 | } |
| 1015 | // Get SSA value of Global variable and create pointer to i8 to point to |
| 1016 | // the format string. |
| 1017 | Value globalPtr = spirv::AddressOfOp::create(builder&: rewriter, location: loc, var: globalVar); |
| 1018 | Value fmtStr = spirv::BitcastOp::create( |
| 1019 | builder&: rewriter, location: loc, |
| 1020 | result: spirv::PointerType::get(pointeeType: i8Type, storageClass: spirv::StorageClass::UniformConstant), |
| 1021 | operand: globalPtr); |
| 1022 | |
| 1023 | // Get printf arguments. |
| 1024 | auto printfArgs = llvm::to_vector_of<Value, 4>(Range: adaptor.getArgs()); |
| 1025 | |
| 1026 | spirv::CLPrintfOp::create(builder&: rewriter, location: loc, result: i32Type, format: fmtStr, arguments: printfArgs); |
| 1027 | |
| 1028 | // Need to erase the gpu.printf op as gpu.printf does not use result vs |
| 1029 | // spirv::CLPrintfOp has i32 resultType so cannot replace with new SPIR-V |
| 1030 | // printf op. |
| 1031 | rewriter.eraseOp(op: gpuPrintfOp); |
| 1032 | |
| 1033 | return success(); |
| 1034 | } |
| 1035 | |
| 1036 | //===----------------------------------------------------------------------===// |
| 1037 | // GPU To SPIRV Patterns. |
| 1038 | //===----------------------------------------------------------------------===// |
| 1039 | |
| 1040 | void mlir::populateGPUToSPIRVPatterns(const SPIRVTypeConverter &typeConverter, |
| 1041 | RewritePatternSet &patterns) { |
| 1042 | patterns.add< |
| 1043 | GPUBarrierConversion, GPUInitializeNamedBarrierConversion, |
| 1044 | GPUBallotConversion, GPUFuncOpConversion, GPUModuleConversion, |
| 1045 | GPUReturnOpConversion, GPUShuffleConversion, GPURotateConversion, |
| 1046 | GPUSubgroupBroadcastConversion, |
| 1047 | LaunchConfigConversion<gpu::BlockIdOp, spirv::BuiltIn::WorkgroupId>, |
| 1048 | LaunchConfigConversion<gpu::GridDimOp, spirv::BuiltIn::NumWorkgroups>, |
| 1049 | LaunchConfigConversion<gpu::BlockDimOp, spirv::BuiltIn::WorkgroupSize>, |
| 1050 | LaunchConfigConversion<gpu::ThreadIdOp, |
| 1051 | spirv::BuiltIn::LocalInvocationId>, |
| 1052 | LaunchConfigConversion<gpu::GlobalIdOp, |
| 1053 | spirv::BuiltIn::GlobalInvocationId>, |
| 1054 | SingleDimLaunchConfigConversion<gpu::SubgroupIdOp, |
| 1055 | spirv::BuiltIn::SubgroupId>, |
| 1056 | SingleDimLaunchConfigConversion<gpu::NumSubgroupsOp, |
| 1057 | spirv::BuiltIn::NumSubgroups>, |
| 1058 | SingleDimLaunchConfigConversion<gpu::SubgroupSizeOp, |
| 1059 | spirv::BuiltIn::SubgroupSize>, |
| 1060 | SingleDimLaunchConfigConversion< |
| 1061 | gpu::LaneIdOp, spirv::BuiltIn::SubgroupLocalInvocationId>, |
| 1062 | WorkGroupSizeConversion, GPUAllReduceConversion, |
| 1063 | GPUSubgroupReduceConversion, GPUPrintfConversion>(arg: typeConverter, |
| 1064 | args: patterns.getContext()); |
| 1065 | } |
| 1066 | |
| 1067 | void mlir::populateGPUNamedBarrierToSPIRVTypeConversion( |
| 1068 | SPIRVTypeConverter &typeConverter) { |
| 1069 | typeConverter.addConversion(callback: [](gpu::NamedBarrierType type) { |
| 1070 | return spirv::NamedBarrierType::get(context: type.getContext()); |
| 1071 | }); |
| 1072 | } |
| 1073 | |