| 1 | //========- CIRGenCUDANV.cpp - Interface to NVIDIA CUDA Runtime -----=========// |
| 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 provides a class for CUDA code generation targeting the NVIDIA CUDA |
| 10 | // runtime library. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "CIRGenCUDARuntime.h" |
| 15 | #include "CIRGenCXXABI.h" |
| 16 | #include "CIRGenFunction.h" |
| 17 | #include "CIRGenModule.h" |
| 18 | #include "mlir/IR/Operation.h" |
| 19 | #include "clang/AST/ASTContext.h" |
| 20 | #include "clang/AST/Attrs.inc" |
| 21 | #include "clang/AST/Decl.h" |
| 22 | #include "clang/AST/GlobalDecl.h" |
| 23 | #include "clang/Basic/AddressSpaces.h" |
| 24 | #include "clang/Basic/Cuda.h" |
| 25 | #include "clang/Basic/DiagnosticFrontend.h" |
| 26 | #include "clang/Basic/FileManager.h" |
| 27 | #include "clang/Basic/SourceManager.h" |
| 28 | #include "clang/CIR/Dialect/IR/CIRDialect.h" |
| 29 | #include "clang/CIR/Dialect/IR/CIRTypes.h" |
| 30 | #include "llvm/Support/Casting.h" |
| 31 | #include "llvm/Support/MemoryBuffer.h" |
| 32 | #include "llvm/Support/VirtualFileSystem.h" |
| 33 | |
| 34 | using namespace clang; |
| 35 | using namespace clang::CIRGen; |
| 36 | |
| 37 | namespace { |
| 38 | |
| 39 | class CIRGenNVCUDARuntime : public CIRGenCUDARuntime { |
| 40 | protected: |
| 41 | StringRef prefix; |
| 42 | |
| 43 | // Map a device stub function to a symbol for identifying kernel in host |
| 44 | // code. For CUDA, the symbol for identifying the kernel is the same as the |
| 45 | // device stub function. For HIP, they are different. |
| 46 | llvm::StringMap<mlir::Operation *> kernelHandles; |
| 47 | |
| 48 | // Map a kernel handle to the kernel stub. |
| 49 | llvm::DenseMap<mlir::Operation *, mlir::Operation *> kernelStubs; |
| 50 | |
| 51 | struct VarInfo { |
| 52 | cir::GlobalOp var; |
| 53 | const VarDecl *d; |
| 54 | cir::CUDADeviceVarKind flags; |
| 55 | }; |
| 56 | llvm::SmallVector<VarInfo, 16> deviceVars; |
| 57 | |
| 58 | // Mangle context for device. |
| 59 | std::unique_ptr<MangleContext> deviceMC; |
| 60 | |
| 61 | private: |
| 62 | void emitDeviceStubBodyNew(CIRGenFunction &cgf, cir::FuncOp fn, |
| 63 | FunctionArgList &args); |
| 64 | void recordDeviceBinary(); |
| 65 | void transformManagedVars(); |
| 66 | mlir::Value prepareKernelArgs(CIRGenFunction &cgf, mlir::Location loc, |
| 67 | FunctionArgList &args); |
| 68 | mlir::Operation *getKernelHandle(cir::FuncOp fn, GlobalDecl gd) override; |
| 69 | |
| 70 | mlir::Operation *getKernelStub(mlir::Operation *handle) override { |
| 71 | auto it = kernelStubs.find(Val: handle); |
| 72 | assert(it != kernelStubs.end()); |
| 73 | return it->second; |
| 74 | } |
| 75 | std::string addPrefixToName(StringRef funcName) const; |
| 76 | std::string addUnderscoredPrefixToName(StringRef funcName) const; |
| 77 | |
| 78 | public: |
| 79 | CIRGenNVCUDARuntime(CIRGenModule &cgm); |
| 80 | ~CIRGenNVCUDARuntime(); |
| 81 | |
| 82 | void emitDeviceStub(CIRGenFunction &cgf, cir::FuncOp fn, |
| 83 | FunctionArgList &args) override; |
| 84 | |
| 85 | void handleVarRegistration(const VarDecl *vd, cir::GlobalOp var) override; |
| 86 | void finalizeModule() override; |
| 87 | void handleGlobalReplace(cir::GlobalOp oldGV, cir::GlobalOp newGV) override; |
| 88 | |
| 89 | void internalizeDeviceSideVar(const VarDecl *d, |
| 90 | cir::GlobalLinkageKind &linkage) override; |
| 91 | |
| 92 | std::string getDeviceSideName(const NamedDecl *nd) override; |
| 93 | |
| 94 | void registerDeviceVar(const VarDecl *vd, cir::GlobalOp &var, bool isExtern, |
| 95 | bool isConstant) { |
| 96 | // Attach the device var attribute to the GlobalOp |
| 97 | auto &builder = cgm.getBuilder(); |
| 98 | var->setAttr(name: cir::CUDAVarRegistrationInfoAttr::getMnemonic(), |
| 99 | value: cir::CUDAVarRegistrationInfoAttr::get( |
| 100 | context: builder.getContext(), |
| 101 | device_side_name: getDeviceSideName(nd: cast<NamedDecl>(Val: vd)), |
| 102 | kind: cir::CUDADeviceVarKind::Variable, isExtern, isConstant, |
| 103 | isManaged: vd->hasAttr<HIPManagedAttr>(), |
| 104 | /*surfTexType=*/0, |
| 105 | /*normalized=*/false)); |
| 106 | deviceVars.push_back(Elt: { |
| 107 | .var: var, |
| 108 | .d: vd, |
| 109 | .flags: cir::CUDADeviceVarKind::Variable, |
| 110 | }); |
| 111 | } |
| 112 | |
| 113 | void registerDeviceSurf(const VarDecl *vd, cir::GlobalOp &var, bool isExtern, |
| 114 | int32_t surfaceType) { |
| 115 | auto &builder = cgm.getBuilder(); |
| 116 | |
| 117 | var->setAttr(name: cir::CUDAVarRegistrationInfoAttr::getMnemonic(), |
| 118 | value: cir::CUDAVarRegistrationInfoAttr::get( |
| 119 | context: builder.getContext(), |
| 120 | device_side_name: getDeviceSideName(nd: cast<NamedDecl>(Val: vd)), |
| 121 | kind: cir::CUDADeviceVarKind::Surface, isExtern, |
| 122 | /*isConstant=*/false, |
| 123 | /*isManaged=*/false, surfTexType: surfaceType, |
| 124 | /*normalized=*/false)); |
| 125 | |
| 126 | deviceVars.push_back(Elt: { |
| 127 | .var: var, |
| 128 | .d: vd, |
| 129 | .flags: cir::CUDADeviceVarKind::Surface, |
| 130 | }); |
| 131 | } |
| 132 | |
| 133 | void registerDeviceTex(const VarDecl *vd, cir::GlobalOp &var, bool isExtern, |
| 134 | int32_t textureType, bool normalized) { |
| 135 | auto &builder = cgm.getBuilder(); |
| 136 | |
| 137 | var->setAttr(name: cir::CUDAVarRegistrationInfoAttr::getMnemonic(), |
| 138 | value: cir::CUDAVarRegistrationInfoAttr::get( |
| 139 | context: builder.getContext(), |
| 140 | device_side_name: getDeviceSideName(nd: cast<NamedDecl>(Val: vd)), |
| 141 | kind: cir::CUDADeviceVarKind::Texture, isExtern, |
| 142 | /*isConstant=*/false, |
| 143 | /*isManaged=*/false, surfTexType: textureType, normalized)); |
| 144 | |
| 145 | deviceVars.push_back(Elt: { |
| 146 | .var: var, |
| 147 | .d: vd, |
| 148 | .flags: cir::CUDADeviceVarKind::Texture, |
| 149 | }); |
| 150 | } |
| 151 | }; |
| 152 | |
| 153 | } // namespace |
| 154 | |
| 155 | std::string CIRGenNVCUDARuntime::addPrefixToName(StringRef funcName) const { |
| 156 | return (prefix + funcName).str(); |
| 157 | } |
| 158 | |
| 159 | std::string |
| 160 | CIRGenNVCUDARuntime::addUnderscoredPrefixToName(StringRef funcName) const { |
| 161 | return ("__" + prefix + funcName).str(); |
| 162 | } |
| 163 | |
| 164 | CIRGenNVCUDARuntime::CIRGenNVCUDARuntime(CIRGenModule &cgm) |
| 165 | : CIRGenCUDARuntime(cgm), |
| 166 | deviceMC(cgm.getASTContext().cudaNVInitDeviceMC()) { |
| 167 | if (cgm.getLangOpts().OffloadViaLLVM) |
| 168 | cgm.errorNYI(feature: "CIRGenNVCUDARuntime: Offload via LLVM" ); |
| 169 | else if (cgm.getLangOpts().HIP) |
| 170 | prefix = "hip" ; |
| 171 | else |
| 172 | prefix = "cuda" ; |
| 173 | } |
| 174 | |
| 175 | mlir::Value CIRGenNVCUDARuntime::prepareKernelArgs(CIRGenFunction &cgf, |
| 176 | mlir::Location loc, |
| 177 | FunctionArgList &args) { |
| 178 | CIRGenBuilderTy &builder = cgm.getBuilder(); |
| 179 | |
| 180 | // Build void *args[] and populate with the addresses of kernel arguments. |
| 181 | auto voidPtrArrayTy = cir::ArrayType::get(elementType: cgm.voidPtrTy, size: args.size()); |
| 182 | mlir::Value kernelArgs = |
| 183 | builder.createAlloca(loc, addrType: cir::PointerType::get(pointee: voidPtrArrayTy), |
| 184 | name: "kernel_args" , alignment: CharUnits::fromQuantity(Quantity: 16)); |
| 185 | |
| 186 | mlir::Value kernelArgsDecayed = |
| 187 | builder.createCast(kind: cir::CastKind::array_to_ptrdecay, src: kernelArgs, |
| 188 | newTy: cir::PointerType::get(pointee: cgm.voidPtrTy)); |
| 189 | |
| 190 | for (const auto &[i, arg] : llvm::enumerate(First&: args)) { |
| 191 | mlir::Value index = |
| 192 | builder.getConstInt(loc, intVal: llvm::APInt(/*numBits=*/32, i)); |
| 193 | mlir::Value storePos = |
| 194 | builder.createPtrStride(loc, base: kernelArgsDecayed, stride: index); |
| 195 | mlir::Value argAddr = cgf.getAddrOfLocalVar(vd: arg).getPointer(); |
| 196 | mlir::Value argAsVoid = builder.createBitcast(src: argAddr, newTy: cgm.voidPtrTy); |
| 197 | |
| 198 | builder.CIRBaseBuilderTy::createStore(loc, val: argAsVoid, dst: storePos); |
| 199 | } |
| 200 | |
| 201 | return kernelArgsDecayed; |
| 202 | } |
| 203 | |
| 204 | // CUDA 9.0+ uses new way to launch kernels. Parameters are packed in a local |
| 205 | // array and kernels are launched using cudaLaunchKernel(). |
| 206 | void CIRGenNVCUDARuntime::emitDeviceStubBodyNew(CIRGenFunction &cgf, |
| 207 | cir::FuncOp fn, |
| 208 | FunctionArgList &args) { |
| 209 | |
| 210 | // This requires arguments to be sent to kernels in a different way. |
| 211 | if (cgm.getLangOpts().OffloadViaLLVM) |
| 212 | cgm.errorNYI(feature: "CIRGenNVCUDARuntime: Offload via LLVM" ); |
| 213 | |
| 214 | CIRGenBuilderTy &builder = cgm.getBuilder(); |
| 215 | mlir::Location loc = fn.getLoc(); |
| 216 | |
| 217 | // For [cuda|hip]LaunchKernel, we must add another layer of indirection |
| 218 | // to arguments. For example, for function `add(int a, float b)`, |
| 219 | // we need to pass it as `void *args[2] = { &a, &b }`. |
| 220 | mlir::Value kernelArgs = prepareKernelArgs(cgf, loc, args); |
| 221 | |
| 222 | // Lookup cudaLaunchKernel/hipLaunchKernel function. |
| 223 | // HIP kernel launching API name depends on -fgpu-default-stream option. For |
| 224 | // the default value 'legacy', it is hipLaunchKernel. For 'per-thread', |
| 225 | // it is hipLaunchKernel_spt. |
| 226 | // cudaError_t cudaLaunchKernel(const void *func, dim3 gridDim, dim3 blockDim, |
| 227 | // void **args, size_t sharedMem, |
| 228 | // cudaStream_t stream); |
| 229 | // hipError_t hipLaunchKernel[_spt](const void *func, dim3 gridDim, |
| 230 | // dim3 blockDim, void **args, |
| 231 | // size_t sharedMem, hipStream_t stream); |
| 232 | TranslationUnitDecl *tuDecl = cgm.getASTContext().getTranslationUnitDecl(); |
| 233 | DeclContext *dc = TranslationUnitDecl::castToDeclContext(D: tuDecl); |
| 234 | |
| 235 | // The default stream is usually stream 0 (the legacy default stream). |
| 236 | // For per-thread default stream, we need a different LaunchKernel function. |
| 237 | std::string kernelLaunchAPI = "LaunchKernel" ; |
| 238 | if (cgm.getLangOpts().GPUDefaultStream == |
| 239 | LangOptions::GPUDefaultStreamKind::PerThread) { |
| 240 | if (cgm.getLangOpts().HIP) |
| 241 | kernelLaunchAPI += "_spt" ; |
| 242 | else if (cgm.getLangOpts().CUDA) |
| 243 | kernelLaunchAPI += "_ptsz" ; |
| 244 | } |
| 245 | |
| 246 | std::string launchKernelName = addPrefixToName(funcName: kernelLaunchAPI); |
| 247 | const IdentifierInfo &launchII = |
| 248 | cgm.getASTContext().Idents.get(Name: launchKernelName); |
| 249 | FunctionDecl *cudaLaunchKernelFD = nullptr; |
| 250 | for (NamedDecl *result : dc->lookup(Name: &launchII)) { |
| 251 | if (FunctionDecl *fd = dyn_cast<FunctionDecl>(Val: result)) |
| 252 | cudaLaunchKernelFD = fd; |
| 253 | } |
| 254 | |
| 255 | if (cudaLaunchKernelFD == nullptr) { |
| 256 | cgm.error(loc: cgf.curFuncDecl->getLocation(), |
| 257 | error: "Can't find declaration for " + launchKernelName); |
| 258 | return; |
| 259 | } |
| 260 | |
| 261 | // Use this function to retrieve arguments for cudaLaunchKernel: |
| 262 | // int __[cuda|hip]PopCallConfiguration(dim3 *gridDim, dim3 *blockDim, size_t |
| 263 | // *sharedMem, cudaStream_t *stream) |
| 264 | // |
| 265 | // Here [cuda|hip]Stream_t, while also being the 6th argument of |
| 266 | // [cuda|hip]LaunchKernel, is a pointer to some opaque struct. |
| 267 | |
| 268 | mlir::Type dim3Ty = cgf.getTypes().convertType( |
| 269 | type: cudaLaunchKernelFD->getParamDecl(i: 1)->getType()); |
| 270 | mlir::Type streamTy = cgf.getTypes().convertType( |
| 271 | type: cudaLaunchKernelFD->getParamDecl(i: 5)->getType()); |
| 272 | |
| 273 | mlir::Value gridDim = |
| 274 | builder.createAlloca(loc, addrType: cir::PointerType::get(pointee: dim3Ty), name: "grid_dim" , |
| 275 | alignment: CharUnits::fromQuantity(Quantity: 8)); |
| 276 | mlir::Value blockDim = |
| 277 | builder.createAlloca(loc, addrType: cir::PointerType::get(pointee: dim3Ty), name: "block_dim" , |
| 278 | alignment: CharUnits::fromQuantity(Quantity: 8)); |
| 279 | mlir::Value sharedMem = builder.createAlloca( |
| 280 | loc, addrType: cir::PointerType::get(pointee: cgm.sizeTy), name: "shared_mem" , alignment: cgm.getSizeAlign()); |
| 281 | mlir::Value stream = builder.createAlloca( |
| 282 | loc, addrType: cir::PointerType::get(pointee: streamTy), name: "stream" , alignment: cgm.getPointerAlign()); |
| 283 | |
| 284 | cir::FuncOp popConfig = cgm.createRuntimeFunction( |
| 285 | ty: cir::FuncType::get(inputs: {gridDim.getType(), blockDim.getType(), |
| 286 | sharedMem.getType(), stream.getType()}, |
| 287 | returnType: cgm.sInt32Ty), |
| 288 | name: addUnderscoredPrefixToName(funcName: "PopCallConfiguration" )); |
| 289 | cgf.emitRuntimeCall(loc, callee: popConfig, args: {gridDim, blockDim, sharedMem, stream}); |
| 290 | |
| 291 | // Now emit the call to cudaLaunchKernel |
| 292 | // [cuda|hip]Error_t [cuda|hip]LaunchKernel(const void *func, dim3 gridDim, |
| 293 | // dim3 blockDim, |
| 294 | // void **args, size_t sharedMem, |
| 295 | // [cuda|hip]Stream_t stream); |
| 296 | |
| 297 | // We now either pick the function or the stub global for cuda, hip |
| 298 | // respectively. |
| 299 | mlir::Value kernel = [&]() -> mlir::Value { |
| 300 | if (cir::GlobalOp globalOp = llvm::dyn_cast_or_null<cir::GlobalOp>( |
| 301 | Val: kernelHandles[fn.getSymName()])) { |
| 302 | cir::PointerType kernelTy = cir::PointerType::get(pointee: globalOp.getSymType()); |
| 303 | mlir::Value kernelVal = cir::GetGlobalOp::create(builder, location: loc, addr: kernelTy, |
| 304 | name: globalOp.getSymName()); |
| 305 | mlir::Value func = builder.createBitcast(src: kernelVal, newTy: cgm.voidPtrTy); |
| 306 | return func; |
| 307 | } |
| 308 | if (cir::FuncOp funcOp = llvm::dyn_cast_or_null<cir::FuncOp>( |
| 309 | Val: kernelHandles[fn.getSymName()])) { |
| 310 | cir::PointerType kernelTy = |
| 311 | cir::PointerType::get(pointee: funcOp.getFunctionType()); |
| 312 | mlir::Value kernelVal = |
| 313 | cir::GetGlobalOp::create(builder, location: loc, addr: kernelTy, name: funcOp.getSymName()); |
| 314 | mlir::Value func = builder.createBitcast(src: kernelVal, newTy: cgm.voidPtrTy); |
| 315 | return func; |
| 316 | } |
| 317 | llvm_unreachable("Expected stub handle to be cir::GlobalOp or FuncOp" ); |
| 318 | }(); |
| 319 | |
| 320 | CallArgList launchArgs; |
| 321 | launchArgs.add(rvalue: RValue::get(v: kernel), |
| 322 | type: cudaLaunchKernelFD->getParamDecl(i: 0)->getType()); |
| 323 | launchArgs.add( |
| 324 | rvalue: RValue::getAggregate(addr: Address(gridDim, CharUnits::fromQuantity(Quantity: 8))), |
| 325 | type: cudaLaunchKernelFD->getParamDecl(i: 1)->getType()); |
| 326 | launchArgs.add( |
| 327 | rvalue: RValue::getAggregate(addr: Address(blockDim, CharUnits::fromQuantity(Quantity: 8))), |
| 328 | type: cudaLaunchKernelFD->getParamDecl(i: 2)->getType()); |
| 329 | launchArgs.add(rvalue: RValue::get(v: kernelArgs), |
| 330 | type: cudaLaunchKernelFD->getParamDecl(i: 3)->getType()); |
| 331 | launchArgs.add( |
| 332 | rvalue: RValue::get(v: builder.CIRBaseBuilderTy::createLoad(loc, ptr: sharedMem)), |
| 333 | type: cudaLaunchKernelFD->getParamDecl(i: 4)->getType()); |
| 334 | launchArgs.add(rvalue: RValue::get(v: builder.CIRBaseBuilderTy::createLoad(loc, ptr: stream)), |
| 335 | type: cudaLaunchKernelFD->getParamDecl(i: 5)->getType()); |
| 336 | |
| 337 | mlir::Type launchTy = |
| 338 | cgm.getTypes().convertType(type: cudaLaunchKernelFD->getType()); |
| 339 | mlir::Operation *cudaKernelLauncherFn = cgm.createRuntimeFunction( |
| 340 | ty: cast<cir::FuncType>(Val&: launchTy), name: launchKernelName); |
| 341 | const CIRGenFunctionInfo &callInfo = |
| 342 | cgm.getTypes().arrangeFunctionDeclaration(fd: cudaLaunchKernelFD); |
| 343 | cgf.emitCall(funcInfo: callInfo, callee: CIRGenCallee::forDirect(funcPtr: cudaKernelLauncherFn), |
| 344 | returnValue: ReturnValueSlot(), args: launchArgs, /*isMustTail=*/false); |
| 345 | |
| 346 | if (cgm.getASTContext().getTargetInfo().getCXXABI().isMicrosoft() && |
| 347 | !cgf.getLangOpts().HIP) |
| 348 | cgm.errorNYI(feature: "MSVC CUDA stub handling" ); |
| 349 | } |
| 350 | |
| 351 | void CIRGenNVCUDARuntime::emitDeviceStub(CIRGenFunction &cgf, cir::FuncOp fn, |
| 352 | FunctionArgList &args) { |
| 353 | CIRGenBuilderTy &builder = cgm.getBuilder(); |
| 354 | StringRef kernelName = cgm.getMangledName( |
| 355 | gd: cgf.curGD.getWithKernelReferenceKind(Kind: KernelReferenceKind::Kernel)); |
| 356 | auto kernelNameAttr = cir::CUDAKernelNameAttr::get( |
| 357 | context: builder.getContext(), kernel_name: builder.getStringAttr(bytes: kernelName)); |
| 358 | fn->setDiscardableAttr(name: cir::CUDAKernelNameAttr::getMnemonic(), |
| 359 | value: kernelNameAttr); |
| 360 | |
| 361 | if (auto globalOp = |
| 362 | llvm::dyn_cast<cir::GlobalOp>(Val: kernelHandles[fn.getSymName()])) { |
| 363 | mlir::Type fnPtrTy = globalOp.getSymType(); |
| 364 | auto sym = mlir::FlatSymbolRefAttr::get(value: fn.getSymNameAttr()); |
| 365 | auto gv = cir::GlobalViewAttr::get(type: fnPtrTy, symbol: sym); |
| 366 | |
| 367 | globalOp->setAttr(name: "initial_value" , value: gv); |
| 368 | globalOp->setAttr(name: "alignment" , value: builder.getI64IntegerAttr( |
| 369 | value: cgm.getPointerAlign().getQuantity())); |
| 370 | |
| 371 | // The stub's linkage is only final once it is defined, so the handle |
| 372 | // takes it over here. |
| 373 | globalOp.setLinkage(fn.getLinkage()); |
| 374 | mlir::SymbolTable::setSymbolVisibility( |
| 375 | symbol: globalOp, vis: cgm.getMLIRVisibilityFromCIRLinkage(GLK: fn.getLinkage())); |
| 376 | cgm.setDSOLocal(static_cast<mlir::Operation *>(globalOp)); |
| 377 | } |
| 378 | |
| 379 | // CUDA 9.0 changed the way to launch kernels. |
| 380 | if (CudaFeatureEnabled(cgm.getTarget().getSDKVersion(), |
| 381 | CudaFeature::CUDA_USES_NEW_LAUNCH) || |
| 382 | (cgm.getLangOpts().HIP && cgm.getLangOpts().HIPUseNewLaunchAPI) || |
| 383 | cgm.getLangOpts().OffloadViaLLVM) |
| 384 | emitDeviceStubBodyNew(cgf, fn, args); |
| 385 | else |
| 386 | cgm.errorNYI(feature: "Emit Stub Body Legacy" ); |
| 387 | } |
| 388 | |
| 389 | CIRGenCUDARuntime *clang::CIRGen::createNVCUDARuntime(CIRGenModule &cgm) { |
| 390 | return new CIRGenNVCUDARuntime(cgm); |
| 391 | } |
| 392 | |
| 393 | CIRGenNVCUDARuntime::~CIRGenNVCUDARuntime() {} |
| 394 | |
| 395 | mlir::Operation *CIRGenNVCUDARuntime::getKernelHandle(cir::FuncOp fn, |
| 396 | GlobalDecl gd) { |
| 397 | |
| 398 | // Check if we already have a kernel handle for this function |
| 399 | auto it = kernelHandles.find(Key: fn.getSymName()); |
| 400 | if (it != kernelHandles.end()) { |
| 401 | mlir::Operation *oldHandle = it->second; |
| 402 | // Here we know that the fn did not change. Return it |
| 403 | if (kernelStubs[oldHandle] == fn) |
| 404 | return oldHandle; |
| 405 | |
| 406 | // We've found the function name, but F itself has changed, so we need to |
| 407 | // update the references. |
| 408 | if (cgm.getLangOpts().HIP) { |
| 409 | // For HIP compilation the handle itself does not change, so we only need |
| 410 | // to update the Stub value. |
| 411 | kernelStubs[oldHandle] = fn; |
| 412 | return oldHandle; |
| 413 | } |
| 414 | // For non-HIP compilation, erase the old Stub and fall-through to creating |
| 415 | // new entries. |
| 416 | kernelStubs.erase(Val: oldHandle); |
| 417 | } |
| 418 | |
| 419 | // If not targeting HIP, store the function itself |
| 420 | if (!cgm.getLangOpts().HIP) { |
| 421 | kernelHandles[fn.getSymName()] = fn; |
| 422 | kernelStubs[fn] = fn; |
| 423 | return fn; |
| 424 | } |
| 425 | |
| 426 | // Create a new CIR global variable to represent the kernel handle |
| 427 | CIRGenBuilderTy &builder = cgm.getBuilder(); |
| 428 | StringRef globalName = cgm.getMangledName( |
| 429 | gd: gd.getWithKernelReferenceKind(Kind: KernelReferenceKind::Kernel)); |
| 430 | cir::PointerType fnPtrTy = builder.getPointerTo(ty: fn.getFunctionType()); |
| 431 | cir::GlobalOp globalOp = |
| 432 | cgm.createGlobalOp(loc: fn.getLoc(), name: globalName, t: fnPtrTy, /*isConstant=*/true); |
| 433 | |
| 434 | globalOp->setAttr(name: "alignment" , value: builder.getI64IntegerAttr( |
| 435 | value: cgm.getPointerAlign().getQuantity())); |
| 436 | |
| 437 | // Inherit visibility and comdat from the kernel's declaration. |
| 438 | auto *fd = cast<FunctionDecl>(Val: gd.getDecl()); |
| 439 | cgm.setGVProperties(op: globalOp, d: fd); |
| 440 | FunctionTemplateDecl *ft = fd->getPrimaryTemplate(); |
| 441 | if (!ft || ft->isThisDeclarationADefinition()) |
| 442 | cgm.maybeSetTrivialComdat(d: *fd, op: globalOp); |
| 443 | |
| 444 | // Store references |
| 445 | kernelHandles[fn.getSymName()] = globalOp; |
| 446 | kernelStubs[globalOp] = fn; |
| 447 | |
| 448 | return globalOp; |
| 449 | } |
| 450 | |
| 451 | void CIRGenNVCUDARuntime::internalizeDeviceSideVar( |
| 452 | const VarDecl *d, cir::GlobalLinkageKind &linkage) { |
| 453 | if (cgm.getLangOpts().GPURelocatableDeviceCode) |
| 454 | cgm.errorNYI(d->getSourceRange(), |
| 455 | "internalizeDeviceSideVar: GPU Relocatable Device Code (RDC)" ); |
| 456 | |
| 457 | // __shared__ variables are odd. Shadows do get created, but |
| 458 | // they are not registered with the CUDA runtime, so they |
| 459 | // can't really be used to access their device-side |
| 460 | // counterparts. It's not clear yet whether it's nvcc's bug or |
| 461 | // a feature, but we've got to do the same for compatibility. |
| 462 | if (d->hasAttr<CUDADeviceAttr>() || d->hasAttr<CUDAConstantAttr>() || |
| 463 | d->hasAttr<CUDASharedAttr>() || |
| 464 | d->getType()->isCUDADeviceBuiltinSurfaceType() || |
| 465 | d->getType()->isCUDADeviceBuiltinTextureType()) { |
| 466 | linkage = cir::GlobalLinkageKind::InternalLinkage; |
| 467 | } |
| 468 | } |
| 469 | |
| 470 | std::string CIRGenNVCUDARuntime::getDeviceSideName(const NamedDecl *nd) { |
| 471 | GlobalDecl gd; |
| 472 | // nd could be either a kernel or a variable. |
| 473 | if (auto *fd = dyn_cast<FunctionDecl>(Val: nd)) |
| 474 | gd = GlobalDecl(fd, KernelReferenceKind::Kernel); |
| 475 | else |
| 476 | gd = GlobalDecl(nd); |
| 477 | std::string deviceSideName; |
| 478 | MangleContext *mc; |
| 479 | if (cgm.getLangOpts().CUDAIsDevice) |
| 480 | mc = &cgm.getCXXABI().getMangleContext(); |
| 481 | else |
| 482 | mc = deviceMC.get(); |
| 483 | if (mc->shouldMangleDeclName(D: nd)) { |
| 484 | SmallString<256> buffer; |
| 485 | llvm::raw_svector_ostream out(buffer); |
| 486 | mc->mangleName(GD: gd, out); |
| 487 | deviceSideName = std::string(out.str()); |
| 488 | } else |
| 489 | deviceSideName = std::string(nd->getIdentifier()->getName()); |
| 490 | |
| 491 | // Make unique name for device side static file-scope variable for HIP. |
| 492 | if (cgm.getASTContext().shouldExternalize(D: nd) && |
| 493 | cgm.getLangOpts().GPURelocatableDeviceCode) { |
| 494 | SmallString<256> buffer; |
| 495 | llvm::raw_svector_ostream out(buffer); |
| 496 | out << deviceSideName; |
| 497 | cgm.printPostfixForExternalizedDecl(os&: out, d: nd); |
| 498 | deviceSideName = std::string(out.str()); |
| 499 | } |
| 500 | return deviceSideName; |
| 501 | } |
| 502 | |
| 503 | void CIRGenNVCUDARuntime::handleVarRegistration(const VarDecl *vd, |
| 504 | cir::GlobalOp var) { |
| 505 | if (vd->hasAttr<CUDADeviceAttr>() || vd->hasAttr<CUDAConstantAttr>()) { |
| 506 | // Shadow variables and their properties must be registered with CUDA |
| 507 | // runtime. Skip Extern global variables, which will be registered in |
| 508 | // the TU where they are defined. |
| 509 | // |
| 510 | // Don't register a C++17 inline variable. The local symbol can be |
| 511 | // discarded and referencing a discarded local symbol from outside the |
| 512 | // comdat (__cuda_register_globals) is disallowed by the ELF spec. |
| 513 | // |
| 514 | // HIP managed variables need to be always recorded in device and host |
| 515 | // compilations for transformation. |
| 516 | // |
| 517 | // HIP managed variables and variables in CUDADeviceVarODRUsedByHost are |
| 518 | // added to llvm.compiler-used, therefore they are safe to be registered. |
| 519 | if ((!vd->hasExternalStorage() && !vd->isInline()) || |
| 520 | cgm.getASTContext().CUDADeviceVarODRUsedByHost.contains(key: vd) || |
| 521 | vd->hasAttr<HIPManagedAttr>()) { |
| 522 | registerDeviceVar(vd, var, isExtern: !vd->hasDefinition(), |
| 523 | isConstant: vd->hasAttr<CUDAConstantAttr>()); |
| 524 | } |
| 525 | } else if (vd->getType()->isCUDADeviceBuiltinSurfaceType()) { |
| 526 | // Builtin surfaces and their template arguments are also registered |
| 527 | // with CUDA runtime. |
| 528 | const auto *td = cast<ClassTemplateSpecializationDecl>( |
| 529 | Val: vd->getType()->castAsCXXRecordDecl()); |
| 530 | const TemplateArgumentList &args = td->getTemplateArgs(); |
| 531 | |
| 532 | assert(args.size() == 2 && |
| 533 | "Unexpected number of template arguments of CUDA device " |
| 534 | "builtin surface type." ); |
| 535 | |
| 536 | llvm::APSInt surfaceType = args[1].getAsIntegral(); |
| 537 | |
| 538 | if (!vd->hasExternalStorage()) |
| 539 | registerDeviceSurf(vd, var, isExtern: !vd->hasDefinition(), |
| 540 | surfaceType: surfaceType.getSExtValue()); |
| 541 | |
| 542 | } else if (vd->getType()->isCUDADeviceBuiltinTextureType()) { |
| 543 | // Builtin textures and their template arguments are also registered |
| 544 | // with CUDA runtime. |
| 545 | const auto *td = cast<ClassTemplateSpecializationDecl>( |
| 546 | Val: vd->getType()->castAsCXXRecordDecl()); |
| 547 | const TemplateArgumentList &args = td->getTemplateArgs(); |
| 548 | |
| 549 | assert(args.size() == 3 && |
| 550 | "Unexpected number of template arguments of CUDA device " |
| 551 | "builtin texture type." ); |
| 552 | |
| 553 | llvm::APSInt textureType = args[1].getAsIntegral(); |
| 554 | llvm::APSInt normalized = args[2].getAsIntegral(); |
| 555 | |
| 556 | if (!vd->hasExternalStorage()) |
| 557 | registerDeviceTex(vd, var, isExtern: !vd->hasDefinition(), |
| 558 | textureType: textureType.getSExtValue(), normalized: normalized.getZExtValue()); |
| 559 | } |
| 560 | } |
| 561 | |
| 562 | void CIRGenNVCUDARuntime::handleGlobalReplace(cir::GlobalOp oldGV, |
| 563 | cir::GlobalOp newGV) { |
| 564 | for (auto &info : deviceVars) { |
| 565 | if (info.var == oldGV) |
| 566 | info.var = newGV; |
| 567 | } |
| 568 | } |
| 569 | |
| 570 | /// Whether this translation unit has anything for the CUDA runtime to register. |
| 571 | /// Like LoweringPrepare, count only kernel stub definitions and device |
| 572 | /// variables carrying registration metadata, not kernel declarations or |
| 573 | /// launch-site calls. |
| 574 | static bool hasEntitiesToRegister(mlir::ModuleOp module) { |
| 575 | // A walk, not a scan of the module body, so this stays in agreement with the |
| 576 | // recursive walk LoweringPrepare collects them with. |
| 577 | return module |
| 578 | ->walk(callback: [](mlir::Operation *op) { |
| 579 | if (auto fn = mlir::dyn_cast<cir::FuncOp>(Val: op); |
| 580 | fn && !fn.isDeclaration() && |
| 581 | fn->hasAttr(name: cir::CUDAKernelNameAttr::getMnemonic())) |
| 582 | return mlir::WalkResult::interrupt(); |
| 583 | if (op->hasAttr(name: cir::CUDAVarRegistrationInfoAttr::getMnemonic())) |
| 584 | return mlir::WalkResult::interrupt(); |
| 585 | return mlir::WalkResult::advance(); |
| 586 | }) |
| 587 | .wasInterrupted(); |
| 588 | } |
| 589 | |
| 590 | /// Read the device-side fat binary and record its contents on the module as |
| 591 | /// `cir.cu.device_binary`, for LoweringPrepare to build the fatbin global from. |
| 592 | /// |
| 593 | /// This mirrors the read in CGNVCUDARuntime::makeModuleCtorFunction, guards |
| 594 | /// included: nothing is read in a compilation that would build no module |
| 595 | /// constructor. |
| 596 | void CIRGenNVCUDARuntime::recordDeviceBinary() { |
| 597 | StringRef binaryName = cgm.getCodeGenOpts().OffloadBinaryToEmbedFile; |
| 598 | if (binaryName.empty()) |
| 599 | return; |
| 600 | |
| 601 | const LangOptions &langOpts = cgm.getLangOpts(); |
| 602 | if ((langOpts.HIP || !langOpts.GPURelocatableDeviceCode) && |
| 603 | !hasEntitiesToRegister(module: cgm.getModule())) |
| 604 | return; |
| 605 | |
| 606 | llvm::vfs::FileSystem &fs = cgm.getASTContext() |
| 607 | .getSourceManager() |
| 608 | .getFileManager() |
| 609 | .getVirtualFileSystem(); |
| 610 | llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> binaryOrErr = |
| 611 | fs.getBufferForFile(Name: binaryName, /*FileSize=*/-1, |
| 612 | /*RequiresNullTerminator=*/false); |
| 613 | if (std::error_code ec = binaryOrErr.getError()) { |
| 614 | cgm.getDiags().Report(DiagID: diag::err_cannot_open_file) |
| 615 | << binaryName << ec.message(); |
| 616 | return; |
| 617 | } |
| 618 | |
| 619 | // Typed as the fatbin global's array type so LoweringPrepare can use this |
| 620 | // attribute as the initializer as-is: attributes are uniqued on {value, type} |
| 621 | // and never freed, so building a second, typed copy there would keep the fat |
| 622 | // binary in memory twice. |
| 623 | StringRef bytes = binaryOrErr.get()->getBuffer(); |
| 624 | mlir::MLIRContext &ctx = cgm.getMLIRContext(); |
| 625 | auto charTy = cir::IntType::get(context: &ctx, width: cgm.getTarget().getCharWidth(), |
| 626 | /*isSigned=*/is_signed: false); |
| 627 | auto fatbinTy = cir::ArrayType::get(elementType: charTy, size: bytes.size()); |
| 628 | cgm.getModule()->setAttr(name: cir::CIRDialect::getCUDADeviceBinaryAttrName(), |
| 629 | value: mlir::StringAttr::get(bytes, type: fatbinTy)); |
| 630 | } |
| 631 | |
| 632 | // Uses of a managed variable load its address from a pointer the runtime fills. |
| 633 | void CIRGenNVCUDARuntime::transformManagedVars() { |
| 634 | CIRGenBuilderTy &builder = cgm.getBuilder(); |
| 635 | mlir::ModuleOp module = cgm.getModule(); |
| 636 | for (auto &&info : deviceVars) { |
| 637 | cir::GlobalOp var = info.var; |
| 638 | if (info.flags != cir::CUDADeviceVarKind::Variable || |
| 639 | !info.d->hasAttr<HIPManagedAttr>()) |
| 640 | continue; |
| 641 | |
| 642 | cir::PointerType varPtrTy = |
| 643 | builder.getPointerTo(ty: var.getSymType(), as: var.getAddrSpaceAttr()); |
| 644 | mlir::ptr::MemorySpaceAttrInterface managedAS = cir::toCIRAddressSpaceAttr( |
| 645 | ctx&: cgm.getMLIRContext(), |
| 646 | langAS: cgm.getLangOpts().CUDAIsDevice ? LangAS::cuda_device : LangAS::Default); |
| 647 | std::optional<mlir::SymbolTable::UseRange> uses = var.getSymbolUses(from: module); |
| 648 | |
| 649 | std::string name = var.getSymName().str(); |
| 650 | // An empty comdat would follow the rename. |
| 651 | std::optional<llvm::StringRef> comdat = var.getComdat(); |
| 652 | if (comdat && comdat->empty()) |
| 653 | var.setComdat(llvm::StringRef(name)); |
| 654 | cgm.eraseGlobalSymbol(op: var); |
| 655 | var.setSymName(name + ".managed" ); |
| 656 | cgm.insertGlobalSymbol(op: var); |
| 657 | |
| 658 | cir::GlobalOp managedVar = |
| 659 | cgm.createGlobalOp(loc: var.getLoc(), name, t: varPtrTy, |
| 660 | /*isConstant=*/false, addrSpace: managedAS); |
| 661 | managedVar.setLinkage(var.getLinkage()); |
| 662 | if (!var.isDeclaration()) |
| 663 | managedVar.setInitialValueAttr(builder.getConstNullPtrAttr(t: varPtrTy)); |
| 664 | managedVar.setDsoLocal(var.getDsoLocal()); |
| 665 | managedVar.setGlobalVisibility(var.getGlobalVisibility()); |
| 666 | mlir::SymbolTable::setSymbolVisibility( |
| 667 | symbol: managedVar, vis: mlir::SymbolTable::getSymbolVisibility(symbol: var)); |
| 668 | managedVar->setAttr( |
| 669 | name: cir::CUDAExternallyInitializedAttr::getMnemonic(), |
| 670 | value: cir::CUDAExternallyInitializedAttr::get(ctx: &cgm.getMLIRContext())); |
| 671 | |
| 672 | // The uses now name the pointer global. |
| 673 | cir::PointerType managedPtrTy = |
| 674 | builder.getPointerTo(ty: varPtrTy, as: managedVar.getAddrSpaceAttr()); |
| 675 | for (mlir::SymbolTable::SymbolUse use : *uses) { |
| 676 | auto getGlobalOp = dyn_cast<cir::GetGlobalOp>(Val: use.getUser()); |
| 677 | if (!getGlobalOp) { |
| 678 | cgm.errorNYI(info.d->getSourceRange(), |
| 679 | "transformManagedVars: non-instruction use" ); |
| 680 | continue; |
| 681 | } |
| 682 | mlir::Value addr = getGlobalOp.getAddr(); |
| 683 | addr.setType(managedPtrTy); |
| 684 | mlir::OpBuilder::InsertionGuard guard(builder); |
| 685 | builder.setInsertionPointAfter(getGlobalOp); |
| 686 | mlir::Value ld = builder.createAlignedLoad( |
| 687 | loc: getGlobalOp.getLoc(), ty: varPtrTy, ptr: addr, |
| 688 | align: CharUnits::fromQuantity(Quantity: var.getAlignment().value_or(u: 1))); |
| 689 | addr.replaceAllUsesExcept(newValue: ld, exceptedUser: ld.getDefiningOp()); |
| 690 | } |
| 691 | |
| 692 | // The runtime allocates them even if device code never uses them. |
| 693 | if (cgm.getLangOpts().CUDAIsDevice && !var.isDeclaration()) { |
| 694 | cgm.addCompilerUsedGlobal(gv: var); |
| 695 | cgm.addCompilerUsedGlobal(gv: managedVar); |
| 696 | } |
| 697 | } |
| 698 | } |
| 699 | |
| 700 | void CIRGenNVCUDARuntime::finalizeModule() { |
| 701 | transformManagedVars(); |
| 702 | if (!cgm.getLangOpts().CUDAIsDevice) { |
| 703 | recordDeviceBinary(); |
| 704 | return; |
| 705 | } |
| 706 | |
| 707 | // Mark ODR-used device variables as compiler used to prevent them from being |
| 708 | // eliminated by optimization. This is necessary for device variables |
| 709 | // ODR-used by host functions. Sema correctly marks them as ODR-used no |
| 710 | // matter whether they are ODR-used by device or host functions. |
| 711 | // |
| 712 | // We do not need to do this if the variable has used attribute since it |
| 713 | // has already been added. |
| 714 | // |
| 715 | // Static device variables have been externalized at this point, therefore |
| 716 | // variables with private or internal linkage need not be added. |
| 717 | for (auto &&info : deviceVars) { |
| 718 | auto kind = info.flags; |
| 719 | bool isDecl = info.var.isDeclaration(); |
| 720 | bool isLocalLinkage = cir::isLocalLinkage(linkage: info.var.getLinkage()); |
| 721 | bool isVarOrSurfaceOrTexture = (kind == cir::CUDADeviceVarKind::Variable || |
| 722 | kind == cir::CUDADeviceVarKind::Surface || |
| 723 | kind == cir::CUDADeviceVarKind::Texture); |
| 724 | bool isUsed = info.d->isUsed(); |
| 725 | bool hasUsedAttr = info.d->hasAttr<UsedAttr>(); |
| 726 | if (!isDecl && !isLocalLinkage && isVarOrSurfaceOrTexture && isUsed && |
| 727 | !hasUsedAttr) { |
| 728 | if (auto globalValue = mlir::dyn_cast<cir::CIRGlobalValueInterface>( |
| 729 | Val: info.var.getOperation())) { |
| 730 | cgm.addCompilerUsedGlobal(gv: globalValue); |
| 731 | } |
| 732 | } |
| 733 | } |
| 734 | } |
| 735 | |