1 //===---- Parallelism.cpp - OpenMP GPU parallel implementation ---- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // Parallel implementation in the GPU. Here is the pattern:
10 //
11 //    while (not finished) {
12 //
13 //    if (master) {
14 //      sequential code, decide which par loop to do, or if finished
15 //     __kmpc_kernel_prepare_parallel() // exec by master only
16 //    }
17 //    syncthreads // A
18 //    __kmpc_kernel_parallel() // exec by all
19 //    if (this thread is included in the parallel) {
20 //      switch () for all parallel loops
21 //      __kmpc_kernel_end_parallel() // exec only by threads in parallel
22 //    }
23 //
24 //
25 //    The reason we don't exec end_parallel for the threads not included
26 //    in the parallel loop is that for each barrier in the parallel
27 //    region, these non-included threads will cycle through the
28 //    syncthread A. Thus they must preserve their current threadId that
29 //    is larger than thread in team.
30 //
31 //    To make a long story short...
32 //
33 //===----------------------------------------------------------------------===//
34 
35 #include "Debug.h"
36 #include "Interface.h"
37 #include "Mapping.h"
38 #include "State.h"
39 #include "Synchronization.h"
40 #include "Types.h"
41 #include "Utils.h"
42 
43 using namespace _OMP;
44 
45 #pragma omp declare target
46 
47 namespace {
48 
49 uint32_t determineNumberOfThreads(int32_t NumThreadsClause) {
50   uint32_t NThreadsICV =
51       NumThreadsClause != -1 ? NumThreadsClause : icv::NThreads;
52   uint32_t NumThreads = mapping::getBlockSize();
53 
54   if (NThreadsICV != 0 && NThreadsICV < NumThreads)
55     NumThreads = NThreadsICV;
56 
57   // Round down to a multiple of WARPSIZE since it is legal to do so in OpenMP.
58   if (NumThreads < mapping::getWarpSize())
59     NumThreads = 1;
60   else
61     NumThreads = (NumThreads & ~((uint32_t)mapping::getWarpSize() - 1));
62 
63   return NumThreads;
64 }
65 
66 // Invoke an outlined parallel function unwrapping arguments (up to 32).
67 void invokeMicrotask(int32_t global_tid, int32_t bound_tid, void *fn,
68                      void **args, int64_t nargs) {
69   switch (nargs) {
70 #include "generated_microtask_cases.gen"
71   default:
72     PRINT("Too many arguments in kmp_invoke_microtask, aborting execution.\n");
73     __builtin_trap();
74   }
75 }
76 
77 } // namespace
78 
79 extern "C" {
80 
81 void __kmpc_parallel_51(IdentTy *ident, int32_t, int32_t if_expr,
82                         int32_t num_threads, int proc_bind, void *fn,
83                         void *wrapper_fn, void **args, int64_t nargs) {
84 
85   uint32_t TId = mapping::getThreadIdInBlock();
86   // Handle the serialized case first, same for SPMD/non-SPMD.
87   if (OMP_UNLIKELY(!if_expr || icv::Level)) {
88     state::enterDataEnvironment();
89     ++icv::Level;
90     invokeMicrotask(TId, 0, fn, args, nargs);
91     state::exitDataEnvironment();
92     return;
93   }
94 
95   uint32_t NumThreads = determineNumberOfThreads(num_threads);
96   if (mapping::isSPMDMode()) {
97     // Avoid the race between the read of the `icv::Level` above and the write
98     // below by synchronizing all threads here.
99     synchronize::threadsAligned();
100     {
101       // Note that the order here is important. `icv::Level` has to be updated
102       // last or the other updates will cause a thread specific state to be
103       // created.
104       state::ValueRAII ParallelTeamSizeRAII(state::ParallelTeamSize, NumThreads,
105                                             1u, TId == 0);
106       state::ValueRAII ActiveLevelRAII(icv::ActiveLevel, 1u, 0u, TId == 0);
107       state::ValueRAII LevelRAII(icv::Level, 1u, 0u, TId == 0);
108 
109       // Synchronize all threads after the main thread (TId == 0) set up the
110       // team state properly.
111       synchronize::threadsAligned();
112 
113       ASSERT(state::ParallelTeamSize == NumThreads);
114       ASSERT(icv::ActiveLevel == 1u);
115       ASSERT(icv::Level == 1u);
116 
117       if (TId < NumThreads)
118         invokeMicrotask(TId, 0, fn, args, nargs);
119 
120       // Synchronize all threads at the end of a parallel region.
121       synchronize::threadsAligned();
122     }
123 
124     ASSERT(state::ParallelTeamSize == 1u);
125     ASSERT(icv::ActiveLevel == 0u);
126     ASSERT(icv::Level == 0u);
127     return;
128   }
129 
130   // We do *not* create a new data environment because all threads in the team
131   // that are active are now running this parallel region. They share the
132   // TeamState, which has an increase level-var and potentially active-level
133   // set, but they do not have individual ThreadStates yet. If they ever
134   // modify the ICVs beyond this point a ThreadStates will be allocated.
135 
136   bool IsActiveParallelRegion = NumThreads > 1;
137   if (!IsActiveParallelRegion) {
138     state::ValueRAII LevelRAII(icv::Level, 1u, 0u, true);
139     invokeMicrotask(TId, 0, fn, args, nargs);
140     return;
141   }
142 
143   void **GlobalArgs = nullptr;
144   if (nargs) {
145     __kmpc_begin_sharing_variables(&GlobalArgs, nargs);
146 #pragma unroll
147     for (int I = 0; I < nargs; I++)
148       GlobalArgs[I] = args[I];
149   }
150 
151   {
152     // Note that the order here is important. `icv::Level` has to be updated
153     // last or the other updates will cause a thread specific state to be
154     // created.
155     state::ValueRAII ParallelTeamSizeRAII(state::ParallelTeamSize, NumThreads,
156                                           1u, true);
157     state::ValueRAII ParallelRegionFnRAII(state::ParallelRegionFn, wrapper_fn,
158                                           (void *)nullptr, true);
159     state::ValueRAII ActiveLevelRAII(icv::ActiveLevel, 1u, 0u, true);
160     state::ValueRAII LevelRAII(icv::Level, 1u, 0u, true);
161 
162     // Master signals work to activate workers.
163     synchronize::threads();
164     // Master waits for workers to signal.
165     synchronize::threads();
166   }
167 
168   if (nargs)
169     __kmpc_end_sharing_variables();
170 }
171 
172 __attribute__((noinline)) bool
173 __kmpc_kernel_parallel(ParallelRegionFnTy *WorkFn) {
174   // Work function and arguments for L1 parallel region.
175   *WorkFn = state::ParallelRegionFn;
176 
177   // If this is the termination signal from the master, quit early.
178   if (!*WorkFn)
179     return false;
180 
181   // Set to true for workers participating in the parallel region.
182   uint32_t TId = mapping::getThreadIdInBlock();
183   bool ThreadIsActive = TId < state::ParallelTeamSize;
184   return ThreadIsActive;
185 }
186 
187 __attribute__((noinline)) void __kmpc_kernel_end_parallel() {
188   // In case we have modified an ICV for this thread before a ThreadState was
189   // created. We drop it now to not contaminate the next parallel region.
190   ASSERT(!mapping::isSPMDMode());
191   uint32_t TId = mapping::getThreadIdInBlock();
192   state::resetStateForThread(TId);
193   ASSERT(!mapping::isSPMDMode());
194 }
195 
196 uint16_t __kmpc_parallel_level(IdentTy *, uint32_t) { return omp_get_level(); }
197 
198 int32_t __kmpc_global_thread_num(IdentTy *) { return omp_get_thread_num(); }
199 
200 void __kmpc_push_num_threads(IdentTy *, int32_t, int32_t NumThreads) {
201   icv::NThreads = NumThreads;
202 }
203 
204 void __kmpc_push_num_teams(IdentTy *loc, int32_t tid, int32_t num_teams,
205                            int32_t thread_limit) {}
206 
207 void __kmpc_push_proc_bind(IdentTy *loc, uint32_t tid, int proc_bind) {}
208 }
209 
210 #pragma omp end declare target
211