1 //===----- CGOpenMPRuntimeNVPTX.h - Interface to OpenMP NVPTX Runtimes ----===//
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 OpenMP runtime code generation specialized to NVPTX
11 // targets.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #ifndef LLVM_CLANG_LIB_CODEGEN_CGOPENMPRUNTIMENVPTX_H
16 #define LLVM_CLANG_LIB_CODEGEN_CGOPENMPRUNTIMENVPTX_H
17 
18 #include "CGOpenMPRuntime.h"
19 #include "CodeGenFunction.h"
20 #include "clang/AST/StmtOpenMP.h"
21 #include "llvm/IR/CallSite.h"
22 
23 namespace clang {
24 namespace CodeGen {
25 
26 class CGOpenMPRuntimeNVPTX : public CGOpenMPRuntime {
27 public:
28   class EntryFunctionState {
29   public:
30     llvm::BasicBlock *ExitBB;
31 
EntryFunctionState()32     EntryFunctionState() : ExitBB(nullptr){};
33   };
34 
35   class WorkerFunctionState {
36   public:
37     llvm::Function *WorkerFn;
38     const CGFunctionInfo *CGFI;
39 
40     WorkerFunctionState(CodeGenModule &CGM);
41 
42   private:
43     void createWorkerFunction(CodeGenModule &CGM);
44   };
45 
46   /// \brief Helper for target entry function. Guide the master and worker
47   /// threads to their respective locations.
48   void emitEntryHeader(CodeGenFunction &CGF, EntryFunctionState &EST,
49                        WorkerFunctionState &WST);
50 
51   /// \brief Signal termination of OMP execution.
52   void emitEntryFooter(CodeGenFunction &CGF, EntryFunctionState &EST);
53 
54 private:
55   //
56   // NVPTX calls.
57   //
58 
59   /// \brief Get the GPU warp size.
60   llvm::Value *getNVPTXWarpSize(CodeGenFunction &CGF);
61 
62   /// \brief Get the id of the current thread on the GPU.
63   llvm::Value *getNVPTXThreadID(CodeGenFunction &CGF);
64 
65   // \brief Get the maximum number of threads in a block of the GPU.
66   llvm::Value *getNVPTXNumThreads(CodeGenFunction &CGF);
67 
68   /// \brief Get barrier to synchronize all threads in a block.
69   void getNVPTXCTABarrier(CodeGenFunction &CGF);
70 
71   // \brief Synchronize all GPU threads in a block.
72   void syncCTAThreads(CodeGenFunction &CGF);
73 
74   //
75   // OMP calls.
76   //
77 
78   /// \brief Get the thread id of the OMP master thread.
79   /// The master thread id is the first thread (lane) of the last warp in the
80   /// GPU block.  Warp size is assumed to be some power of 2.
81   /// Thread id is 0 indexed.
82   /// E.g: If NumThreads is 33, master id is 32.
83   ///      If NumThreads is 64, master id is 32.
84   ///      If NumThreads is 1024, master id is 992.
85   llvm::Value *getMasterThreadID(CodeGenFunction &CGF);
86 
87   //
88   // Private state and methods.
89   //
90 
91   // Master-worker control state.
92   // Number of requested OMP threads in parallel region.
93   llvm::GlobalVariable *ActiveWorkers;
94   // Outlined function for the workers to execute.
95   llvm::GlobalVariable *WorkID;
96 
97   /// \brief Initialize master-worker control state.
98   void initializeEnvironment();
99 
100   /// \brief Emit the worker function for the current target region.
101   void emitWorkerFunction(WorkerFunctionState &WST);
102 
103   /// \brief Helper for worker function. Emit body of worker loop.
104   void emitWorkerLoop(CodeGenFunction &CGF, WorkerFunctionState &WST);
105 
106   /// \brief Returns specified OpenMP runtime function for the current OpenMP
107   /// implementation.  Specialized for the NVPTX device.
108   /// \param Function OpenMP runtime function.
109   /// \return Specified function.
110   llvm::Constant *createNVPTXRuntimeFunction(unsigned Function);
111 
112   //
113   // Base class overrides.
114   //
115 
116   /// \brief Creates offloading entry for the provided entry ID \a ID,
117   /// address \a Addr and size \a Size.
118   void createOffloadEntry(llvm::Constant *ID, llvm::Constant *Addr,
119                           uint64_t Size) override;
120 
121   /// \brief Emit outlined function for 'target' directive on the NVPTX
122   /// device.
123   /// \param D Directive to emit.
124   /// \param ParentName Name of the function that encloses the target region.
125   /// \param OutlinedFn Outlined function value to be defined by this call.
126   /// \param OutlinedFnID Outlined function ID value to be defined by this call.
127   /// \param IsOffloadEntry True if the outlined function is an offload entry.
128   /// An outlined function may not be an entry if, e.g. the if clause always
129   /// evaluates to false.
130   void emitTargetOutlinedFunction(const OMPExecutableDirective &D,
131                                   StringRef ParentName,
132                                   llvm::Function *&OutlinedFn,
133                                   llvm::Constant *&OutlinedFnID,
134                                   bool IsOffloadEntry,
135                                   const RegionCodeGenTy &CodeGen) override;
136 
137 public:
138   explicit CGOpenMPRuntimeNVPTX(CodeGenModule &CGM);
139 
140   /// \brief This function ought to emit, in the general case, a call to
141   // the openmp runtime kmpc_push_num_teams. In NVPTX backend it is not needed
142   // as these numbers are obtained through the PTX grid and block configuration.
143   /// \param NumTeams An integer expression of teams.
144   /// \param ThreadLimit An integer expression of threads.
145   void emitNumTeamsClause(CodeGenFunction &CGF, const Expr *NumTeams,
146                           const Expr *ThreadLimit, SourceLocation Loc) override;
147 
148   /// \brief Emits inlined function for the specified OpenMP parallel
149   //  directive but an inlined function for teams.
150   /// \a D. This outlined function has type void(*)(kmp_int32 *ThreadID,
151   /// kmp_int32 BoundID, struct context_vars*).
152   /// \param D OpenMP directive.
153   /// \param ThreadIDVar Variable for thread id in the current OpenMP region.
154   /// \param InnermostKind Kind of innermost directive (for simple directives it
155   /// is a directive itself, for combined - its innermost directive).
156   /// \param CodeGen Code generation sequence for the \a D directive.
157   llvm::Value *
158   emitParallelOrTeamsOutlinedFunction(const OMPExecutableDirective &D,
159                                       const VarDecl *ThreadIDVar,
160                                       OpenMPDirectiveKind InnermostKind,
161                                       const RegionCodeGenTy &CodeGen) override;
162 
163   /// \brief Emits code for teams call of the \a OutlinedFn with
164   /// variables captured in a record which address is stored in \a
165   /// CapturedStruct.
166   /// \param OutlinedFn Outlined function to be run by team masters. Type of
167   /// this function is void(*)(kmp_int32 *, kmp_int32, struct context_vars*).
168   /// \param CapturedVars A pointer to the record with the references to
169   /// variables used in \a OutlinedFn function.
170   ///
171   void emitTeamsCall(CodeGenFunction &CGF, const OMPExecutableDirective &D,
172                      SourceLocation Loc, llvm::Value *OutlinedFn,
173                      ArrayRef<llvm::Value *> CapturedVars) override;
174 };
175 
176 } // CodeGen namespace.
177 } // clang namespace.
178 
179 #endif // LLVM_CLANG_LIB_CODEGEN_CGOPENMPRUNTIMENVPTX_H
180