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
24using namespace clang;
25using namespace CodeGen;
26
27namespace {
28
29class CGNVCUDARuntime : public CGCUDARuntime {
30
31private:
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
68public:
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
84CGNVCUDARuntime::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
99llvm::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
107llvm::Constant *CGNVCUDARuntime::getLaunchFn() const {
108  // cudaError_t cudaLaunch(char *)
109  return CGM.CreateRuntimeFunction(
110      llvm::FunctionType::get(IntTy, CharPtrTy, false), "cudaLaunch");
111}
112
113void CGNVCUDARuntime::emitDeviceStub(CodeGenFunction &CGF,
114                                     FunctionArgList &Args) {
115  EmittedKernels.push_back(CGF.CurFn);
116  emitDeviceStubBody(CGF, Args);
117}
118
119void CGNVCUDARuntime::emitDeviceStubBody(CodeGenFunction &CGF,
120                                         FunctionArgList &Args) {
121  // Build the argument value list and the argument stack struct type.
122  SmallVector<llvm::Value *, 16> ArgValues;
123  std::vector<llvm::Type *> ArgTypes;
124  for (FunctionArgList::const_iterator I = Args.begin(), E = Args.end();
125       I != E; ++I) {
126    llvm::Value *V = CGF.GetAddrOfLocalVar(*I).getPointer();
127    ArgValues.push_back(V);
128    assert(isa<llvm::PointerType>(V->getType()) && "Arg type not PointerType");
129    ArgTypes.push_back(cast<llvm::PointerType>(V->getType())->getElementType());
130  }
131  llvm::StructType *ArgStackTy = llvm::StructType::get(Context, ArgTypes);
132
133  llvm::BasicBlock *EndBlock = CGF.createBasicBlock("setup.end");
134
135  // Emit the calls to cudaSetupArgument
136  llvm::Constant *cudaSetupArgFn = getSetupArgumentFn();
137  for (unsigned I = 0, E = Args.size(); I != E; ++I) {
138    llvm::Value *Args[3];
139    llvm::BasicBlock *NextBlock = CGF.createBasicBlock("setup.next");
140    Args[0] = CGF.Builder.CreatePointerCast(ArgValues[I], VoidPtrTy);
141    Args[1] = CGF.Builder.CreateIntCast(
142        llvm::ConstantExpr::getSizeOf(ArgTypes[I]),
143        SizeTy, false);
144    Args[2] = CGF.Builder.CreateIntCast(
145        llvm::ConstantExpr::getOffsetOf(ArgStackTy, I),
146        SizeTy, false);
147    llvm::CallSite CS = CGF.EmitRuntimeCallOrInvoke(cudaSetupArgFn, Args);
148    llvm::Constant *Zero = llvm::ConstantInt::get(IntTy, 0);
149    llvm::Value *CSZero = CGF.Builder.CreateICmpEQ(CS.getInstruction(), Zero);
150    CGF.Builder.CreateCondBr(CSZero, NextBlock, EndBlock);
151    CGF.EmitBlock(NextBlock);
152  }
153
154  // Emit the call to cudaLaunch
155  llvm::Constant *cudaLaunchFn = getLaunchFn();
156  llvm::Value *Arg = CGF.Builder.CreatePointerCast(CGF.CurFn, CharPtrTy);
157  CGF.EmitRuntimeCallOrInvoke(cudaLaunchFn, Arg);
158  CGF.EmitBranch(EndBlock);
159
160  CGF.EmitBlock(EndBlock);
161}
162
163/// Creates a function that sets up state on the host side for CUDA objects that
164/// have a presence on both the host and device sides. Specifically, registers
165/// the host side of kernel functions and device global variables with the CUDA
166/// runtime.
167/// \code
168/// void __cuda_register_globals(void** GpuBinaryHandle) {
169///    __cudaRegisterFunction(GpuBinaryHandle,Kernel0,...);
170///    ...
171///    __cudaRegisterFunction(GpuBinaryHandle,KernelM,...);
172///    __cudaRegisterVar(GpuBinaryHandle, GlobalVar0, ...);
173///    ...
174///    __cudaRegisterVar(GpuBinaryHandle, GlobalVarN, ...);
175/// }
176/// \endcode
177llvm::Function *CGNVCUDARuntime::makeRegisterGlobalsFn() {
178  // No need to register anything
179  if (EmittedKernels.empty() && DeviceVars.empty())
180    return nullptr;
181
182  llvm::Function *RegisterKernelsFunc = llvm::Function::Create(
183      llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false),
184      llvm::GlobalValue::InternalLinkage, "__cuda_register_globals", &TheModule);
185  llvm::BasicBlock *EntryBB =
186      llvm::BasicBlock::Create(Context, "entry", RegisterKernelsFunc);
187  CGBuilderTy Builder(CGM, Context);
188  Builder.SetInsertPoint(EntryBB);
189
190  // void __cudaRegisterFunction(void **, const char *, char *, const char *,
191  //                             int, uint3*, uint3*, dim3*, dim3*, int*)
192  llvm::Type *RegisterFuncParams[] = {
193      VoidPtrPtrTy, CharPtrTy, CharPtrTy, CharPtrTy, IntTy,
194      VoidPtrTy,    VoidPtrTy, VoidPtrTy, VoidPtrTy, IntTy->getPointerTo()};
195  llvm::Constant *RegisterFunc = CGM.CreateRuntimeFunction(
196      llvm::FunctionType::get(IntTy, RegisterFuncParams, false),
197      "__cudaRegisterFunction");
198
199  // Extract GpuBinaryHandle passed as the first argument passed to
200  // __cuda_register_globals() and generate __cudaRegisterFunction() call for
201  // each emitted kernel.
202  llvm::Argument &GpuBinaryHandlePtr = *RegisterKernelsFunc->arg_begin();
203  for (llvm::Function *Kernel : EmittedKernels) {
204    llvm::Constant *KernelName = makeConstantString(Kernel->getName());
205    llvm::Constant *NullPtr = llvm::ConstantPointerNull::get(VoidPtrTy);
206    llvm::Value *Args[] = {
207        &GpuBinaryHandlePtr, Builder.CreateBitCast(Kernel, VoidPtrTy),
208        KernelName, KernelName, llvm::ConstantInt::get(IntTy, -1), NullPtr,
209        NullPtr, NullPtr, NullPtr,
210        llvm::ConstantPointerNull::get(IntTy->getPointerTo())};
211    Builder.CreateCall(RegisterFunc, Args);
212  }
213
214  // void __cudaRegisterVar(void **, char *, char *, const char *,
215  //                        int, int, int, int)
216  llvm::Type *RegisterVarParams[] = {VoidPtrPtrTy, CharPtrTy, CharPtrTy,
217                                     CharPtrTy,    IntTy,     IntTy,
218                                     IntTy,        IntTy};
219  llvm::Constant *RegisterVar = CGM.CreateRuntimeFunction(
220      llvm::FunctionType::get(IntTy, RegisterVarParams, false),
221      "__cudaRegisterVar");
222  for (auto &Pair : DeviceVars) {
223    llvm::GlobalVariable *Var = Pair.first;
224    unsigned Flags = Pair.second;
225    llvm::Constant *VarName = makeConstantString(Var->getName());
226    uint64_t VarSize =
227        CGM.getDataLayout().getTypeAllocSize(Var->getValueType());
228    llvm::Value *Args[] = {
229        &GpuBinaryHandlePtr,
230        Builder.CreateBitCast(Var, VoidPtrTy),
231        VarName,
232        VarName,
233        llvm::ConstantInt::get(IntTy, (Flags & ExternDeviceVar) ? 1 : 0),
234        llvm::ConstantInt::get(IntTy, VarSize),
235        llvm::ConstantInt::get(IntTy, (Flags & ConstantDeviceVar) ? 1 : 0),
236        llvm::ConstantInt::get(IntTy, 0)};
237    Builder.CreateCall(RegisterVar, Args);
238  }
239
240  Builder.CreateRetVoid();
241  return RegisterKernelsFunc;
242}
243
244/// Creates a global constructor function for the module:
245/// \code
246/// void __cuda_module_ctor(void*) {
247///     Handle0 = __cudaRegisterFatBinary(GpuBinaryBlob0);
248///     __cuda_register_globals(Handle0);
249///     ...
250///     HandleN = __cudaRegisterFatBinary(GpuBinaryBlobN);
251///     __cuda_register_globals(HandleN);
252/// }
253/// \endcode
254llvm::Function *CGNVCUDARuntime::makeModuleCtorFunction() {
255  // No need to generate ctors/dtors if there are no GPU binaries.
256  if (CGM.getCodeGenOpts().CudaGpuBinaryFileNames.empty())
257    return nullptr;
258
259  // void __cuda_register_globals(void* handle);
260  llvm::Function *RegisterGlobalsFunc = makeRegisterGlobalsFn();
261  // void ** __cudaRegisterFatBinary(void *);
262  llvm::Constant *RegisterFatbinFunc = CGM.CreateRuntimeFunction(
263      llvm::FunctionType::get(VoidPtrPtrTy, VoidPtrTy, false),
264      "__cudaRegisterFatBinary");
265  // struct { int magic, int version, void * gpu_binary, void * dont_care };
266  llvm::StructType *FatbinWrapperTy =
267      llvm::StructType::get(IntTy, IntTy, VoidPtrTy, VoidPtrTy, nullptr);
268
269  llvm::Function *ModuleCtorFunc = llvm::Function::Create(
270      llvm::FunctionType::get(VoidTy, VoidPtrTy, false),
271      llvm::GlobalValue::InternalLinkage, "__cuda_module_ctor", &TheModule);
272  llvm::BasicBlock *CtorEntryBB =
273      llvm::BasicBlock::Create(Context, "entry", ModuleCtorFunc);
274  CGBuilderTy CtorBuilder(CGM, Context);
275
276  CtorBuilder.SetInsertPoint(CtorEntryBB);
277
278  // For each GPU binary, register it with the CUDA runtime and store returned
279  // handle in a global variable and save the handle in GpuBinaryHandles vector
280  // to be cleaned up in destructor on exit. Then associate all known kernels
281  // with the GPU binary handle so CUDA runtime can figure out what to call on
282  // the GPU side.
283  for (const std::string &GpuBinaryFileName :
284       CGM.getCodeGenOpts().CudaGpuBinaryFileNames) {
285    llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> GpuBinaryOrErr =
286        llvm::MemoryBuffer::getFileOrSTDIN(GpuBinaryFileName);
287    if (std::error_code EC = GpuBinaryOrErr.getError()) {
288      CGM.getDiags().Report(diag::err_cannot_open_file) << GpuBinaryFileName
289                                                        << EC.message();
290      continue;
291    }
292
293    // Create initialized wrapper structure that points to the loaded GPU binary
294    llvm::Constant *Values[] = {
295        llvm::ConstantInt::get(IntTy, 0x466243b1), // Fatbin wrapper magic.
296        llvm::ConstantInt::get(IntTy, 1),          // Fatbin version.
297        makeConstantString(GpuBinaryOrErr.get()->getBuffer(), "", 16), // Data.
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
337llvm::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
365CGCUDARuntime *CodeGen::CreateNVCUDARuntime(CodeGenModule &CGM) {
366  return new CGNVCUDARuntime(CGM);
367}
368