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