@@ -176,109 +176,11 @@ int32_t __kmpc_nvptx_parallel_reduce_nowait_v2(IdentTy *Loc,
176176 return nvptx_parallel_reduce_nowait (reduce_data, shflFct, cpyFct);
177177}
178178
179- // / Mostly like _v2 but with the builtin assumption that we have less than
180- // / num_of_records (by default 1024) teams.
181- int32_t __kmpc_nvptx_teams_reduce_nowait_v3 (
182- IdentTy *Loc, void *__restrict__ GlobalBuffer, uint32_t num_of_records,
183- uint64_t reduce_data_size, void *reduce_data, ShuffleReductFnTy shflFct,
184- InterWarpCopyFnTy cpyFct, ListGlobalFnTy lgcpyFct, ListGlobalFnTy lgredFct,
185- ListGlobalFnTy glcpyFct, ListGlobalFnTy glredFct) {
186- // Terminate all threads in non-SPMD mode except for the main thread.
187- uint32_t ThreadId = mapping::getThreadIdInBlock ();
188- if (mapping::isGenericMode ()) {
189- if (!mapping::isMainThreadInGenericMode ())
190- return 0 ;
191- ThreadId = 0 ;
192- }
193-
194- uint32_t &Cnt = state::getKernelLaunchEnvironment ().ReductionCnt ;
195-
196- // In non-generic mode all workers participate in the teams reduction.
197- // In generic mode only the team main participates in the teams
198- // reduction because the workers are waiting for parallel work.
199- uint32_t NumThreads = omp_get_num_threads ();
200- uint32_t TeamId = omp_get_team_num ();
201- uint32_t NumTeams = omp_get_num_teams ();
202- static unsigned SHARED (ChunkTeamCount);
203-
204- // Block progress for teams greater than the current upper
205- // limit. We always only allow a number of teams less or equal
206- // to the number of slots in the buffer.
207- bool IsMain = (ThreadId == 0 );
208-
209- if (IsMain) {
210- lgcpyFct (GlobalBuffer, TeamId, reduce_data);
211-
212- // Propagate the memory writes above to the world.
213- fence::kernel (atomic::release);
214-
215- // Increment team counter.
216- // This counter is incremented by all teams in the current
217- // BUFFER_SIZE chunk.
218- ChunkTeamCount = atomic::inc (&Cnt, NumTeams, atomic::acq_rel,
219- atomic::MemScopeTy::device);
220- }
221-
222- // Synchronize in SPMD mode as in generic mode all but 1 threads are in the
223- // state machine.
224- if (mapping::isSPMDMode ())
225- synchronize::threadsAligned (atomic::acq_rel);
226-
227- // Each thread will have a local struct containing the values to be
228- // reduced:
229- // 1. do reduction within each warp.
230- // 2. do reduction across warps.
231- // 3. write the final result to the main reduction variable
232- // by returning 1 in the thread holding the reduction result.
233-
234- // Check if this is the very last team.
235- if (ChunkTeamCount != NumTeams - 1 )
236- return 0 ;
237-
238- // Last team processing.
239- NumThreads = roundToWarpsize (kmpcMin (NumThreads, NumTeams));
240- if (ThreadId >= NumThreads)
241- return 0 ;
242-
243- // Ensure we see the global memory writes by other teams
244- fence::kernel (atomic::aquire);
245-
246- // Load from buffer and reduce.
247- glcpyFct (GlobalBuffer, ThreadId, reduce_data);
248- for (uint32_t i = NumThreads + ThreadId; i < NumTeams; i += NumThreads)
249- glredFct (GlobalBuffer, i, reduce_data);
250-
251- // Reduce across warps to the warp main.
252- gpu_regular_warp_reduce (reduce_data, shflFct);
253-
254- uint32_t ActiveThreads = kmpcMin (NumTeams, NumThreads);
255- uint32_t WarpsNeeded =
256- (ActiveThreads + mapping::getWarpSize () - 1 ) / mapping::getWarpSize ();
257- // Gather all the reduced values from each warp
258- // to the first warp.
259- cpyFct (reduce_data, WarpsNeeded);
260-
261- if (mapping::getWarpIdInBlock () == 0 )
262- gpu_irregular_warp_reduce (reduce_data, shflFct, WarpsNeeded, ThreadId);
263-
264- return IsMain;
265- }
266-
267179int32_t __kmpc_nvptx_teams_reduce_nowait_v2 (
268- IdentTy *Loc, void *GlobalBuffer, uint32_t num_of_records,
269- uint64_t reduce_data_size, void *reduce_data, ShuffleReductFnTy shflFct,
270- InterWarpCopyFnTy cpyFct, ListGlobalFnTy lgcpyFct, ListGlobalFnTy lgredFct,
271- ListGlobalFnTy glcpyFct, ListGlobalFnTy glredFct) {
272- // The first check is a compile time constant, the second one a runtime check.
273- // If the first one succeeds we will use the specialized version.
274- if ((state::getKernelEnvironment ().Configuration .MaxTeams >= 0 &&
275- state::getKernelEnvironment ().Configuration .MaxTeams <= num_of_records &&
276- num_of_records == 1024 ) ||
277- (omp_get_num_teams () <= num_of_records))
278- return __kmpc_nvptx_teams_reduce_nowait_v3 (
279- Loc, GlobalBuffer, num_of_records, reduce_data_size, reduce_data,
280- shflFct, cpyFct, lgcpyFct, lgredFct, glcpyFct, glredFct);
281-
180+ IdentTy *Loc, int32_t TId, void *GlobalBuffer, uint32_t num_of_records,
181+ void *reduce_data, ShuffleReductFnTy shflFct, InterWarpCopyFnTy cpyFct,
182+ ListGlobalFnTy lgcpyFct, ListGlobalFnTy lgredFct, ListGlobalFnTy glcpyFct,
183+ ListGlobalFnTy glredFct) {
282184 // Terminate all threads in non-SPMD mode except for the master thread.
283185 uint32_t ThreadId = mapping::getThreadIdInBlock ();
284186 if (mapping::isGenericMode ()) {
0 commit comments