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/CodeGen/ConstantInitBuilder.h"
19 #include "clang/AST/Decl.h"
20 #include "llvm/IR/BasicBlock.h"
21 #include "llvm/IR/CallSite.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DerivedTypes.h"
24 
25 using namespace clang;
26 using namespace CodeGen;
27 
28 namespace {
29 
30 class CGNVCUDARuntime : public CGCUDARuntime {
31 
32 private:
33   llvm::IntegerType *IntTy, *SizeTy;
34   llvm::Type *VoidTy;
35   llvm::PointerType *CharPtrTy, *VoidPtrTy, *VoidPtrPtrTy;
36 
37   /// Convenience reference to LLVM Context
38   llvm::LLVMContext &Context;
39   /// Convenience reference to the current module
40   llvm::Module &TheModule;
41   /// Keeps track of kernel launch stubs emitted in this module
42   llvm::SmallVector<llvm::Function *, 16> EmittedKernels;
43   llvm::SmallVector<std::pair<llvm::GlobalVariable *, unsigned>, 16> DeviceVars;
44   /// Keeps track of variables containing handles of GPU binaries. Populated by
45   /// ModuleCtorFunction() and used to create corresponding cleanup calls in
46   /// ModuleDtorFunction()
47   llvm::SmallVector<llvm::GlobalVariable *, 16> GpuBinaryHandles;
48 
49   llvm::Constant *getSetupArgumentFn() const;
50   llvm::Constant *getLaunchFn() const;
51 
52   /// Creates a function to register all kernel stubs generated in this module.
53   llvm::Function *makeRegisterGlobalsFn();
54 
55   /// Helper function that generates a constant string and returns a pointer to
56   /// the start of the string.  The result of this function can be used anywhere
57   /// where the C code specifies const char*.
58   llvm::Constant *makeConstantString(const std::string &Str,
59                                      const std::string &Name = "",
60                                      const std::string &SectionName = "",
61                                      unsigned Alignment = 0) {
62     llvm::Constant *Zeros[] = {llvm::ConstantInt::get(SizeTy, 0),
63                                llvm::ConstantInt::get(SizeTy, 0)};
64     auto ConstStr = CGM.GetAddrOfConstantCString(Str, Name.c_str());
65     llvm::GlobalVariable *GV =
66         cast<llvm::GlobalVariable>(ConstStr.getPointer());
67     if (!SectionName.empty())
68       GV->setSection(SectionName);
69     if (Alignment)
70       GV->setAlignment(Alignment);
71 
72     return llvm::ConstantExpr::getGetElementPtr(ConstStr.getElementType(),
73                                                 ConstStr.getPointer(), Zeros);
74  }
75 
76   void emitDeviceStubBody(CodeGenFunction &CGF, FunctionArgList &Args);
77 
78 public:
79   CGNVCUDARuntime(CodeGenModule &CGM);
80 
81   void emitDeviceStub(CodeGenFunction &CGF, FunctionArgList &Args) override;
82   void registerDeviceVar(llvm::GlobalVariable &Var, unsigned Flags) override {
83     DeviceVars.push_back(std::make_pair(&Var, Flags));
84   }
85 
86   /// Creates module constructor function
87   llvm::Function *makeModuleCtorFunction() override;
88   /// Creates module destructor function
89   llvm::Function *makeModuleDtorFunction() override;
90 };
91 
92 }
93 
94 CGNVCUDARuntime::CGNVCUDARuntime(CodeGenModule &CGM)
95     : CGCUDARuntime(CGM), Context(CGM.getLLVMContext()),
96       TheModule(CGM.getModule()) {
97   CodeGen::CodeGenTypes &Types = CGM.getTypes();
98   ASTContext &Ctx = CGM.getContext();
99 
100   IntTy = CGM.IntTy;
101   SizeTy = CGM.SizeTy;
102   VoidTy = CGM.VoidTy;
103 
104   CharPtrTy = llvm::PointerType::getUnqual(Types.ConvertType(Ctx.CharTy));
105   VoidPtrTy = cast<llvm::PointerType>(Types.ConvertType(Ctx.VoidPtrTy));
106   VoidPtrPtrTy = VoidPtrTy->getPointerTo();
107 }
108 
109 llvm::Constant *CGNVCUDARuntime::getSetupArgumentFn() const {
110   // cudaError_t cudaSetupArgument(void *, size_t, size_t)
111   llvm::Type *Params[] = {VoidPtrTy, SizeTy, SizeTy};
112   return CGM.CreateRuntimeFunction(llvm::FunctionType::get(IntTy,
113                                                            Params, false),
114                                    "cudaSetupArgument");
115 }
116 
117 llvm::Constant *CGNVCUDARuntime::getLaunchFn() const {
118   // cudaError_t cudaLaunch(char *)
119   return CGM.CreateRuntimeFunction(
120       llvm::FunctionType::get(IntTy, CharPtrTy, false), "cudaLaunch");
121 }
122 
123 void CGNVCUDARuntime::emitDeviceStub(CodeGenFunction &CGF,
124                                      FunctionArgList &Args) {
125   EmittedKernels.push_back(CGF.CurFn);
126   emitDeviceStubBody(CGF, Args);
127 }
128 
129 void CGNVCUDARuntime::emitDeviceStubBody(CodeGenFunction &CGF,
130                                          FunctionArgList &Args) {
131   // Emit a call to cudaSetupArgument for each arg in Args.
132   llvm::Constant *cudaSetupArgFn = getSetupArgumentFn();
133   llvm::BasicBlock *EndBlock = CGF.createBasicBlock("setup.end");
134   CharUnits Offset = CharUnits::Zero();
135   for (const VarDecl *A : Args) {
136     CharUnits TyWidth, TyAlign;
137     std::tie(TyWidth, TyAlign) =
138         CGM.getContext().getTypeInfoInChars(A->getType());
139     Offset = Offset.alignTo(TyAlign);
140     llvm::Value *Args[] = {
141         CGF.Builder.CreatePointerCast(CGF.GetAddrOfLocalVar(A).getPointer(),
142                                       VoidPtrTy),
143         llvm::ConstantInt::get(SizeTy, TyWidth.getQuantity()),
144         llvm::ConstantInt::get(SizeTy, Offset.getQuantity()),
145     };
146     llvm::CallSite CS = CGF.EmitRuntimeCallOrInvoke(cudaSetupArgFn, Args);
147     llvm::Constant *Zero = llvm::ConstantInt::get(IntTy, 0);
148     llvm::Value *CSZero = CGF.Builder.CreateICmpEQ(CS.getInstruction(), Zero);
149     llvm::BasicBlock *NextBlock = CGF.createBasicBlock("setup.next");
150     CGF.Builder.CreateCondBr(CSZero, NextBlock, EndBlock);
151     CGF.EmitBlock(NextBlock);
152     Offset += TyWidth;
153   }
154 
155   // Emit the call to cudaLaunch
156   llvm::Constant *cudaLaunchFn = getLaunchFn();
157   llvm::Value *Arg = CGF.Builder.CreatePointerCast(CGF.CurFn, CharPtrTy);
158   CGF.EmitRuntimeCallOrInvoke(cudaLaunchFn, Arg);
159   CGF.EmitBranch(EndBlock);
160 
161   CGF.EmitBlock(EndBlock);
162 }
163 
164 /// Creates a function that sets up state on the host side for CUDA objects that
165 /// have a presence on both the host and device sides. Specifically, registers
166 /// the host side of kernel functions and device global variables with the CUDA
167 /// runtime.
168 /// \code
169 /// void __cuda_register_globals(void** GpuBinaryHandle) {
170 ///    __cudaRegisterFunction(GpuBinaryHandle,Kernel0,...);
171 ///    ...
172 ///    __cudaRegisterFunction(GpuBinaryHandle,KernelM,...);
173 ///    __cudaRegisterVar(GpuBinaryHandle, GlobalVar0, ...);
174 ///    ...
175 ///    __cudaRegisterVar(GpuBinaryHandle, GlobalVarN, ...);
176 /// }
177 /// \endcode
178 llvm::Function *CGNVCUDARuntime::makeRegisterGlobalsFn() {
179   // No need to register anything
180   if (EmittedKernels.empty() && DeviceVars.empty())
181     return nullptr;
182 
183   llvm::Function *RegisterKernelsFunc = llvm::Function::Create(
184       llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false),
185       llvm::GlobalValue::InternalLinkage, "__cuda_register_globals", &TheModule);
186   llvm::BasicBlock *EntryBB =
187       llvm::BasicBlock::Create(Context, "entry", RegisterKernelsFunc);
188   CGBuilderTy Builder(CGM, Context);
189   Builder.SetInsertPoint(EntryBB);
190 
191   // void __cudaRegisterFunction(void **, const char *, char *, const char *,
192   //                             int, uint3*, uint3*, dim3*, dim3*, int*)
193   llvm::Type *RegisterFuncParams[] = {
194       VoidPtrPtrTy, CharPtrTy, CharPtrTy, CharPtrTy, IntTy,
195       VoidPtrTy,    VoidPtrTy, VoidPtrTy, VoidPtrTy, IntTy->getPointerTo()};
196   llvm::Constant *RegisterFunc = CGM.CreateRuntimeFunction(
197       llvm::FunctionType::get(IntTy, RegisterFuncParams, false),
198       "__cudaRegisterFunction");
199 
200   // Extract GpuBinaryHandle passed as the first argument passed to
201   // __cuda_register_globals() and generate __cudaRegisterFunction() call for
202   // each emitted kernel.
203   llvm::Argument &GpuBinaryHandlePtr = *RegisterKernelsFunc->arg_begin();
204   for (llvm::Function *Kernel : EmittedKernels) {
205     llvm::Constant *KernelName = makeConstantString(Kernel->getName());
206     llvm::Constant *NullPtr = llvm::ConstantPointerNull::get(VoidPtrTy);
207     llvm::Value *Args[] = {
208         &GpuBinaryHandlePtr, Builder.CreateBitCast(Kernel, VoidPtrTy),
209         KernelName, KernelName, llvm::ConstantInt::get(IntTy, -1), NullPtr,
210         NullPtr, NullPtr, NullPtr,
211         llvm::ConstantPointerNull::get(IntTy->getPointerTo())};
212     Builder.CreateCall(RegisterFunc, Args);
213   }
214 
215   // void __cudaRegisterVar(void **, char *, char *, const char *,
216   //                        int, int, int, int)
217   llvm::Type *RegisterVarParams[] = {VoidPtrPtrTy, CharPtrTy, CharPtrTy,
218                                      CharPtrTy,    IntTy,     IntTy,
219                                      IntTy,        IntTy};
220   llvm::Constant *RegisterVar = CGM.CreateRuntimeFunction(
221       llvm::FunctionType::get(IntTy, RegisterVarParams, false),
222       "__cudaRegisterVar");
223   for (auto &Pair : DeviceVars) {
224     llvm::GlobalVariable *Var = Pair.first;
225     unsigned Flags = Pair.second;
226     llvm::Constant *VarName = makeConstantString(Var->getName());
227     uint64_t VarSize =
228         CGM.getDataLayout().getTypeAllocSize(Var->getValueType());
229     llvm::Value *Args[] = {
230         &GpuBinaryHandlePtr,
231         Builder.CreateBitCast(Var, VoidPtrTy),
232         VarName,
233         VarName,
234         llvm::ConstantInt::get(IntTy, (Flags & ExternDeviceVar) ? 1 : 0),
235         llvm::ConstantInt::get(IntTy, VarSize),
236         llvm::ConstantInt::get(IntTy, (Flags & ConstantDeviceVar) ? 1 : 0),
237         llvm::ConstantInt::get(IntTy, 0)};
238     Builder.CreateCall(RegisterVar, Args);
239   }
240 
241   Builder.CreateRetVoid();
242   return RegisterKernelsFunc;
243 }
244 
245 /// Creates a global constructor function for the module:
246 /// \code
247 /// void __cuda_module_ctor(void*) {
248 ///     Handle0 = __cudaRegisterFatBinary(GpuBinaryBlob0);
249 ///     __cuda_register_globals(Handle0);
250 ///     ...
251 ///     HandleN = __cudaRegisterFatBinary(GpuBinaryBlobN);
252 ///     __cuda_register_globals(HandleN);
253 /// }
254 /// \endcode
255 llvm::Function *CGNVCUDARuntime::makeModuleCtorFunction() {
256   // No need to generate ctors/dtors if there are no GPU binaries.
257   if (CGM.getCodeGenOpts().CudaGpuBinaryFileNames.empty())
258     return nullptr;
259 
260   // void __cuda_register_globals(void* handle);
261   llvm::Function *RegisterGlobalsFunc = makeRegisterGlobalsFn();
262   // void ** __cudaRegisterFatBinary(void *);
263   llvm::Constant *RegisterFatbinFunc = CGM.CreateRuntimeFunction(
264       llvm::FunctionType::get(VoidPtrPtrTy, VoidPtrTy, false),
265       "__cudaRegisterFatBinary");
266   // struct { int magic, int version, void * gpu_binary, void * dont_care };
267   llvm::StructType *FatbinWrapperTy =
268       llvm::StructType::get(IntTy, IntTy, VoidPtrTy, VoidPtrTy, nullptr);
269 
270   llvm::Function *ModuleCtorFunc = llvm::Function::Create(
271       llvm::FunctionType::get(VoidTy, VoidPtrTy, false),
272       llvm::GlobalValue::InternalLinkage, "__cuda_module_ctor", &TheModule);
273   llvm::BasicBlock *CtorEntryBB =
274       llvm::BasicBlock::Create(Context, "entry", ModuleCtorFunc);
275   CGBuilderTy CtorBuilder(CGM, Context);
276 
277   CtorBuilder.SetInsertPoint(CtorEntryBB);
278 
279   // For each GPU binary, register it with the CUDA runtime and store returned
280   // handle in a global variable and save the handle in GpuBinaryHandles vector
281   // to be cleaned up in destructor on exit. Then associate all known kernels
282   // with the GPU binary handle so CUDA runtime can figure out what to call on
283   // the GPU side.
284   for (const std::string &GpuBinaryFileName :
285        CGM.getCodeGenOpts().CudaGpuBinaryFileNames) {
286     llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> GpuBinaryOrErr =
287         llvm::MemoryBuffer::getFileOrSTDIN(GpuBinaryFileName);
288     if (std::error_code EC = GpuBinaryOrErr.getError()) {
289       CGM.getDiags().Report(diag::err_cannot_open_file) << GpuBinaryFileName
290                                                         << EC.message();
291       continue;
292     }
293 
294     const char *FatbinConstantName =
295         CGM.getTriple().isMacOSX() ? "__NV_CUDA,__nv_fatbin" : ".nv_fatbin";
296     // NVIDIA's cuobjdump looks for fatbins in this section.
297     const char *FatbinSectionName =
298         CGM.getTriple().isMacOSX() ? "__NV_CUDA,__fatbin" : ".nvFatBinSegment";
299 
300     // Create initialized wrapper structure that points to the loaded GPU binary
301     ConstantInitBuilder Builder(CGM);
302     auto Values = Builder.beginStruct(FatbinWrapperTy);
303     // Fatbin wrapper magic.
304     Values.addInt(IntTy, 0x466243b1);
305     // Fatbin version.
306     Values.addInt(IntTy, 1);
307     // Data.
308     Values.add(makeConstantString(GpuBinaryOrErr.get()->getBuffer(),
309                                   "", FatbinConstantName, 8));
310     // Unused in fatbin v1.
311     Values.add(llvm::ConstantPointerNull::get(VoidPtrTy));
312     llvm::GlobalVariable *FatbinWrapper =
313       Values.finishAndCreateGlobal("__cuda_fatbin_wrapper",
314                                    CGM.getPointerAlign(),
315                                    /*constant*/ true);
316     FatbinWrapper->setSection(FatbinSectionName);
317 
318     // GpuBinaryHandle = __cudaRegisterFatBinary(&FatbinWrapper);
319     llvm::CallInst *RegisterFatbinCall = CtorBuilder.CreateCall(
320         RegisterFatbinFunc,
321         CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy));
322     llvm::GlobalVariable *GpuBinaryHandle = new llvm::GlobalVariable(
323         TheModule, VoidPtrPtrTy, false, llvm::GlobalValue::InternalLinkage,
324         llvm::ConstantPointerNull::get(VoidPtrPtrTy), "__cuda_gpubin_handle");
325     CtorBuilder.CreateAlignedStore(RegisterFatbinCall, GpuBinaryHandle,
326                                    CGM.getPointerAlign());
327 
328     // Call __cuda_register_globals(GpuBinaryHandle);
329     if (RegisterGlobalsFunc)
330       CtorBuilder.CreateCall(RegisterGlobalsFunc, RegisterFatbinCall);
331 
332     // Save GpuBinaryHandle so we can unregister it in destructor.
333     GpuBinaryHandles.push_back(GpuBinaryHandle);
334   }
335 
336   CtorBuilder.CreateRetVoid();
337   return ModuleCtorFunc;
338 }
339 
340 /// Creates a global destructor function that unregisters all GPU code blobs
341 /// registered by constructor.
342 /// \code
343 /// void __cuda_module_dtor(void*) {
344 ///     __cudaUnregisterFatBinary(Handle0);
345 ///     ...
346 ///     __cudaUnregisterFatBinary(HandleN);
347 /// }
348 /// \endcode
349 llvm::Function *CGNVCUDARuntime::makeModuleDtorFunction() {
350   // No need for destructor if we don't have handles to unregister.
351   if (GpuBinaryHandles.empty())
352     return nullptr;
353 
354   // void __cudaUnregisterFatBinary(void ** handle);
355   llvm::Constant *UnregisterFatbinFunc = CGM.CreateRuntimeFunction(
356       llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false),
357       "__cudaUnregisterFatBinary");
358 
359   llvm::Function *ModuleDtorFunc = llvm::Function::Create(
360       llvm::FunctionType::get(VoidTy, VoidPtrTy, false),
361       llvm::GlobalValue::InternalLinkage, "__cuda_module_dtor", &TheModule);
362   llvm::BasicBlock *DtorEntryBB =
363       llvm::BasicBlock::Create(Context, "entry", ModuleDtorFunc);
364   CGBuilderTy DtorBuilder(CGM, Context);
365   DtorBuilder.SetInsertPoint(DtorEntryBB);
366 
367   for (llvm::GlobalVariable *GpuBinaryHandle : GpuBinaryHandles) {
368     auto HandleValue =
369       DtorBuilder.CreateAlignedLoad(GpuBinaryHandle, CGM.getPointerAlign());
370     DtorBuilder.CreateCall(UnregisterFatbinFunc, HandleValue);
371   }
372 
373   DtorBuilder.CreateRetVoid();
374   return ModuleDtorFunc;
375 }
376 
377 CGCUDARuntime *CodeGen::CreateNVCUDARuntime(CodeGenModule &CGM) {
378   return new CGNVCUDARuntime(CGM);
379 }
380