1 //===----- CGCUDANV.cpp - Interface to NVIDIA CUDA Runtime ----------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This provides a class for CUDA code generation targeting the NVIDIA CUDA
11 // runtime library.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "CGCUDARuntime.h"
16 #include "CodeGenFunction.h"
17 #include "CodeGenModule.h"
18 #include "clang/AST/Decl.h"
19 #include "llvm/IR/BasicBlock.h"
20 #include "llvm/IR/CallSite.h"
21 #include "llvm/IR/Constants.h"
22 #include "llvm/IR/DerivedTypes.h"
23 
24 using namespace clang;
25 using namespace CodeGen;
26 
27 namespace {
28 
29 class CGNVCUDARuntime : public CGCUDARuntime {
30 
31 private:
32   llvm::Type *IntTy, *SizeTy, *VoidTy;
33   llvm::PointerType *CharPtrTy, *VoidPtrTy, *VoidPtrPtrTy;
34 
35   /// Convenience reference to LLVM Context
36   llvm::LLVMContext &Context;
37   /// Convenience reference to the current module
38   llvm::Module &TheModule;
39   /// Keeps track of kernel launch stubs emitted in this module
40   llvm::SmallVector<llvm::Function *, 16> EmittedKernels;
41   llvm::SmallVector<std::pair<llvm::GlobalVariable *, unsigned>, 16> DeviceVars;
42   /// Keeps track of variables containing handles of GPU binaries. Populated by
43   /// ModuleCtorFunction() and used to create corresponding cleanup calls in
44   /// ModuleDtorFunction()
45   llvm::SmallVector<llvm::GlobalVariable *, 16> GpuBinaryHandles;
46 
47   llvm::Constant *getSetupArgumentFn() const;
48   llvm::Constant *getLaunchFn() const;
49 
50   /// Creates a function to register all kernel stubs generated in this module.
51   llvm::Function *makeRegisterGlobalsFn();
52 
53   /// Helper function that generates a constant string and returns a pointer to
54   /// the start of the string.  The result of this function can be used anywhere
55   /// where the C code specifies const char*.
56   llvm::Constant *makeConstantString(const std::string &Str,
57                                      const std::string &Name = "",
58                                      unsigned Alignment = 0) {
59     llvm::Constant *Zeros[] = {llvm::ConstantInt::get(SizeTy, 0),
60                                llvm::ConstantInt::get(SizeTy, 0)};
61     auto ConstStr = CGM.GetAddrOfConstantCString(Str, Name.c_str());
62     return llvm::ConstantExpr::getGetElementPtr(ConstStr.getElementType(),
63                                                 ConstStr.getPointer(), Zeros);
64  }
65 
66   void emitDeviceStubBody(CodeGenFunction &CGF, FunctionArgList &Args);
67 
68 public:
69   CGNVCUDARuntime(CodeGenModule &CGM);
70 
71   void emitDeviceStub(CodeGenFunction &CGF, FunctionArgList &Args) override;
72   void registerDeviceVar(llvm::GlobalVariable &Var, unsigned Flags) override {
73     DeviceVars.push_back(std::make_pair(&Var, Flags));
74   }
75 
76   /// Creates module constructor function
77   llvm::Function *makeModuleCtorFunction() override;
78   /// Creates module destructor function
79   llvm::Function *makeModuleDtorFunction() override;
80 };
81 
82 }
83 
84 CGNVCUDARuntime::CGNVCUDARuntime(CodeGenModule &CGM)
85     : CGCUDARuntime(CGM), Context(CGM.getLLVMContext()),
86       TheModule(CGM.getModule()) {
87   CodeGen::CodeGenTypes &Types = CGM.getTypes();
88   ASTContext &Ctx = CGM.getContext();
89 
90   IntTy = Types.ConvertType(Ctx.IntTy);
91   SizeTy = Types.ConvertType(Ctx.getSizeType());
92   VoidTy = llvm::Type::getVoidTy(Context);
93 
94   CharPtrTy = llvm::PointerType::getUnqual(Types.ConvertType(Ctx.CharTy));
95   VoidPtrTy = cast<llvm::PointerType>(Types.ConvertType(Ctx.VoidPtrTy));
96   VoidPtrPtrTy = VoidPtrTy->getPointerTo();
97 }
98 
99 llvm::Constant *CGNVCUDARuntime::getSetupArgumentFn() const {
100   // cudaError_t cudaSetupArgument(void *, size_t, size_t)
101   llvm::Type *Params[] = {VoidPtrTy, SizeTy, SizeTy};
102   return CGM.CreateRuntimeFunction(llvm::FunctionType::get(IntTy,
103                                                            Params, false),
104                                    "cudaSetupArgument");
105 }
106 
107 llvm::Constant *CGNVCUDARuntime::getLaunchFn() const {
108   // cudaError_t cudaLaunch(char *)
109   return CGM.CreateRuntimeFunction(
110       llvm::FunctionType::get(IntTy, CharPtrTy, false), "cudaLaunch");
111 }
112 
113 void CGNVCUDARuntime::emitDeviceStub(CodeGenFunction &CGF,
114                                      FunctionArgList &Args) {
115   EmittedKernels.push_back(CGF.CurFn);
116   emitDeviceStubBody(CGF, Args);
117 }
118 
119 void CGNVCUDARuntime::emitDeviceStubBody(CodeGenFunction &CGF,
120                                          FunctionArgList &Args) {
121   // Emit a call to cudaSetupArgument for each arg in Args.
122   llvm::Constant *cudaSetupArgFn = getSetupArgumentFn();
123   llvm::BasicBlock *EndBlock = CGF.createBasicBlock("setup.end");
124   CharUnits Offset = CharUnits::Zero();
125   for (const VarDecl *A : Args) {
126     CharUnits TyWidth, TyAlign;
127     std::tie(TyWidth, TyAlign) =
128         CGM.getContext().getTypeInfoInChars(A->getType());
129     Offset = Offset.alignTo(TyAlign);
130     llvm::Value *Args[] = {
131         CGF.Builder.CreatePointerCast(CGF.GetAddrOfLocalVar(A).getPointer(),
132                                       VoidPtrTy),
133         llvm::ConstantInt::get(SizeTy, TyWidth.getQuantity()),
134         llvm::ConstantInt::get(SizeTy, Offset.getQuantity()),
135     };
136     llvm::CallSite CS = CGF.EmitRuntimeCallOrInvoke(cudaSetupArgFn, Args);
137     llvm::Constant *Zero = llvm::ConstantInt::get(IntTy, 0);
138     llvm::Value *CSZero = CGF.Builder.CreateICmpEQ(CS.getInstruction(), Zero);
139     llvm::BasicBlock *NextBlock = CGF.createBasicBlock("setup.next");
140     CGF.Builder.CreateCondBr(CSZero, NextBlock, EndBlock);
141     CGF.EmitBlock(NextBlock);
142     Offset += TyWidth;
143   }
144 
145   // Emit the call to cudaLaunch
146   llvm::Constant *cudaLaunchFn = getLaunchFn();
147   llvm::Value *Arg = CGF.Builder.CreatePointerCast(CGF.CurFn, CharPtrTy);
148   CGF.EmitRuntimeCallOrInvoke(cudaLaunchFn, Arg);
149   CGF.EmitBranch(EndBlock);
150 
151   CGF.EmitBlock(EndBlock);
152 }
153 
154 /// Creates a function that sets up state on the host side for CUDA objects that
155 /// have a presence on both the host and device sides. Specifically, registers
156 /// the host side of kernel functions and device global variables with the CUDA
157 /// runtime.
158 /// \code
159 /// void __cuda_register_globals(void** GpuBinaryHandle) {
160 ///    __cudaRegisterFunction(GpuBinaryHandle,Kernel0,...);
161 ///    ...
162 ///    __cudaRegisterFunction(GpuBinaryHandle,KernelM,...);
163 ///    __cudaRegisterVar(GpuBinaryHandle, GlobalVar0, ...);
164 ///    ...
165 ///    __cudaRegisterVar(GpuBinaryHandle, GlobalVarN, ...);
166 /// }
167 /// \endcode
168 llvm::Function *CGNVCUDARuntime::makeRegisterGlobalsFn() {
169   // No need to register anything
170   if (EmittedKernels.empty() && DeviceVars.empty())
171     return nullptr;
172 
173   llvm::Function *RegisterKernelsFunc = llvm::Function::Create(
174       llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false),
175       llvm::GlobalValue::InternalLinkage, "__cuda_register_globals", &TheModule);
176   llvm::BasicBlock *EntryBB =
177       llvm::BasicBlock::Create(Context, "entry", RegisterKernelsFunc);
178   CGBuilderTy Builder(CGM, Context);
179   Builder.SetInsertPoint(EntryBB);
180 
181   // void __cudaRegisterFunction(void **, const char *, char *, const char *,
182   //                             int, uint3*, uint3*, dim3*, dim3*, int*)
183   llvm::Type *RegisterFuncParams[] = {
184       VoidPtrPtrTy, CharPtrTy, CharPtrTy, CharPtrTy, IntTy,
185       VoidPtrTy,    VoidPtrTy, VoidPtrTy, VoidPtrTy, IntTy->getPointerTo()};
186   llvm::Constant *RegisterFunc = CGM.CreateRuntimeFunction(
187       llvm::FunctionType::get(IntTy, RegisterFuncParams, false),
188       "__cudaRegisterFunction");
189 
190   // Extract GpuBinaryHandle passed as the first argument passed to
191   // __cuda_register_globals() and generate __cudaRegisterFunction() call for
192   // each emitted kernel.
193   llvm::Argument &GpuBinaryHandlePtr = *RegisterKernelsFunc->arg_begin();
194   for (llvm::Function *Kernel : EmittedKernels) {
195     llvm::Constant *KernelName = makeConstantString(Kernel->getName());
196     llvm::Constant *NullPtr = llvm::ConstantPointerNull::get(VoidPtrTy);
197     llvm::Value *Args[] = {
198         &GpuBinaryHandlePtr, Builder.CreateBitCast(Kernel, VoidPtrTy),
199         KernelName, KernelName, llvm::ConstantInt::get(IntTy, -1), NullPtr,
200         NullPtr, NullPtr, NullPtr,
201         llvm::ConstantPointerNull::get(IntTy->getPointerTo())};
202     Builder.CreateCall(RegisterFunc, Args);
203   }
204 
205   // void __cudaRegisterVar(void **, char *, char *, const char *,
206   //                        int, int, int, int)
207   llvm::Type *RegisterVarParams[] = {VoidPtrPtrTy, CharPtrTy, CharPtrTy,
208                                      CharPtrTy,    IntTy,     IntTy,
209                                      IntTy,        IntTy};
210   llvm::Constant *RegisterVar = CGM.CreateRuntimeFunction(
211       llvm::FunctionType::get(IntTy, RegisterVarParams, false),
212       "__cudaRegisterVar");
213   for (auto &Pair : DeviceVars) {
214     llvm::GlobalVariable *Var = Pair.first;
215     unsigned Flags = Pair.second;
216     llvm::Constant *VarName = makeConstantString(Var->getName());
217     uint64_t VarSize =
218         CGM.getDataLayout().getTypeAllocSize(Var->getValueType());
219     llvm::Value *Args[] = {
220         &GpuBinaryHandlePtr,
221         Builder.CreateBitCast(Var, VoidPtrTy),
222         VarName,
223         VarName,
224         llvm::ConstantInt::get(IntTy, (Flags & ExternDeviceVar) ? 1 : 0),
225         llvm::ConstantInt::get(IntTy, VarSize),
226         llvm::ConstantInt::get(IntTy, (Flags & ConstantDeviceVar) ? 1 : 0),
227         llvm::ConstantInt::get(IntTy, 0)};
228     Builder.CreateCall(RegisterVar, Args);
229   }
230 
231   Builder.CreateRetVoid();
232   return RegisterKernelsFunc;
233 }
234 
235 /// Creates a global constructor function for the module:
236 /// \code
237 /// void __cuda_module_ctor(void*) {
238 ///     Handle0 = __cudaRegisterFatBinary(GpuBinaryBlob0);
239 ///     __cuda_register_globals(Handle0);
240 ///     ...
241 ///     HandleN = __cudaRegisterFatBinary(GpuBinaryBlobN);
242 ///     __cuda_register_globals(HandleN);
243 /// }
244 /// \endcode
245 llvm::Function *CGNVCUDARuntime::makeModuleCtorFunction() {
246   // No need to generate ctors/dtors if there are no GPU binaries.
247   if (CGM.getCodeGenOpts().CudaGpuBinaryFileNames.empty())
248     return nullptr;
249 
250   // void __cuda_register_globals(void* handle);
251   llvm::Function *RegisterGlobalsFunc = makeRegisterGlobalsFn();
252   // void ** __cudaRegisterFatBinary(void *);
253   llvm::Constant *RegisterFatbinFunc = CGM.CreateRuntimeFunction(
254       llvm::FunctionType::get(VoidPtrPtrTy, VoidPtrTy, false),
255       "__cudaRegisterFatBinary");
256   // struct { int magic, int version, void * gpu_binary, void * dont_care };
257   llvm::StructType *FatbinWrapperTy =
258       llvm::StructType::get(IntTy, IntTy, VoidPtrTy, VoidPtrTy, nullptr);
259 
260   llvm::Function *ModuleCtorFunc = llvm::Function::Create(
261       llvm::FunctionType::get(VoidTy, VoidPtrTy, false),
262       llvm::GlobalValue::InternalLinkage, "__cuda_module_ctor", &TheModule);
263   llvm::BasicBlock *CtorEntryBB =
264       llvm::BasicBlock::Create(Context, "entry", ModuleCtorFunc);
265   CGBuilderTy CtorBuilder(CGM, Context);
266 
267   CtorBuilder.SetInsertPoint(CtorEntryBB);
268 
269   // For each GPU binary, register it with the CUDA runtime and store returned
270   // handle in a global variable and save the handle in GpuBinaryHandles vector
271   // to be cleaned up in destructor on exit. Then associate all known kernels
272   // with the GPU binary handle so CUDA runtime can figure out what to call on
273   // the GPU side.
274   for (const std::string &GpuBinaryFileName :
275        CGM.getCodeGenOpts().CudaGpuBinaryFileNames) {
276     llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> GpuBinaryOrErr =
277         llvm::MemoryBuffer::getFileOrSTDIN(GpuBinaryFileName);
278     if (std::error_code EC = GpuBinaryOrErr.getError()) {
279       CGM.getDiags().Report(diag::err_cannot_open_file) << GpuBinaryFileName
280                                                         << EC.message();
281       continue;
282     }
283 
284     // Create initialized wrapper structure that points to the loaded GPU binary
285     llvm::Constant *Values[] = {
286         llvm::ConstantInt::get(IntTy, 0x466243b1), // Fatbin wrapper magic.
287         llvm::ConstantInt::get(IntTy, 1),          // Fatbin version.
288         makeConstantString(GpuBinaryOrErr.get()->getBuffer(), "", 16), // Data.
289         llvm::ConstantPointerNull::get(VoidPtrTy)}; // Unused in fatbin v1.
290     llvm::GlobalVariable *FatbinWrapper = new llvm::GlobalVariable(
291         TheModule, FatbinWrapperTy, true, llvm::GlobalValue::InternalLinkage,
292         llvm::ConstantStruct::get(FatbinWrapperTy, Values),
293         "__cuda_fatbin_wrapper");
294     // NVIDIA's cuobjdump looks for fatbins in this section.
295     FatbinWrapper->setSection(".nvFatBinSegment");
296 
297     // GpuBinaryHandle = __cudaRegisterFatBinary(&FatbinWrapper);
298     llvm::CallInst *RegisterFatbinCall = CtorBuilder.CreateCall(
299         RegisterFatbinFunc,
300         CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy));
301     llvm::GlobalVariable *GpuBinaryHandle = new llvm::GlobalVariable(
302         TheModule, VoidPtrPtrTy, false, llvm::GlobalValue::InternalLinkage,
303         llvm::ConstantPointerNull::get(VoidPtrPtrTy), "__cuda_gpubin_handle");
304     CtorBuilder.CreateAlignedStore(RegisterFatbinCall, GpuBinaryHandle,
305                                    CGM.getPointerAlign());
306 
307     // Call __cuda_register_globals(GpuBinaryHandle);
308     if (RegisterGlobalsFunc)
309       CtorBuilder.CreateCall(RegisterGlobalsFunc, RegisterFatbinCall);
310 
311     // Save GpuBinaryHandle so we can unregister it in destructor.
312     GpuBinaryHandles.push_back(GpuBinaryHandle);
313   }
314 
315   CtorBuilder.CreateRetVoid();
316   return ModuleCtorFunc;
317 }
318 
319 /// Creates a global destructor function that unregisters all GPU code blobs
320 /// registered by constructor.
321 /// \code
322 /// void __cuda_module_dtor(void*) {
323 ///     __cudaUnregisterFatBinary(Handle0);
324 ///     ...
325 ///     __cudaUnregisterFatBinary(HandleN);
326 /// }
327 /// \endcode
328 llvm::Function *CGNVCUDARuntime::makeModuleDtorFunction() {
329   // No need for destructor if we don't have handles to unregister.
330   if (GpuBinaryHandles.empty())
331     return nullptr;
332 
333   // void __cudaUnregisterFatBinary(void ** handle);
334   llvm::Constant *UnregisterFatbinFunc = CGM.CreateRuntimeFunction(
335       llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false),
336       "__cudaUnregisterFatBinary");
337 
338   llvm::Function *ModuleDtorFunc = llvm::Function::Create(
339       llvm::FunctionType::get(VoidTy, VoidPtrTy, false),
340       llvm::GlobalValue::InternalLinkage, "__cuda_module_dtor", &TheModule);
341   llvm::BasicBlock *DtorEntryBB =
342       llvm::BasicBlock::Create(Context, "entry", ModuleDtorFunc);
343   CGBuilderTy DtorBuilder(CGM, Context);
344   DtorBuilder.SetInsertPoint(DtorEntryBB);
345 
346   for (llvm::GlobalVariable *GpuBinaryHandle : GpuBinaryHandles) {
347     auto HandleValue =
348       DtorBuilder.CreateAlignedLoad(GpuBinaryHandle, CGM.getPointerAlign());
349     DtorBuilder.CreateCall(UnregisterFatbinFunc, HandleValue);
350   }
351 
352   DtorBuilder.CreateRetVoid();
353   return ModuleDtorFunc;
354 }
355 
356 CGCUDARuntime *CodeGen::CreateNVCUDARuntime(CodeGenModule &CGM) {
357   return new CGNVCUDARuntime(CGM);
358 }
359