[OPENMP][NVPTX]Basic support for reductions across the teams.

[ABataev]

Summary

Added functions kmpc_nvptx_teams_reduce_nowait_simple and
kmpc_nvptx_teams_end_reduce_nowait_simple to implement basic support
for reductions across the teams.

[gtbercea]

LG

[kkwli0]

libomptarget/deviceRTLs/nvptx/src/reduction.cu

435 ↗

(On Diff #175555)

Why is global_tid being passed in but not used?

446 ↗

(On Diff #175555)

Same as above, loc and global_tid are not used in the routine?

[ABataev]

libomptarget/deviceRTLs/nvptx/src/reduction.cu

435 ↗

(On Diff #175555)

They are going to be used in future, at least in some diagnostics.

openmp/trunk/libomptarget/deviceRTLs/nvptx/src/interface.h

 EXTERN int32_t __kmpc_nvptx_teams_reduce_nowait_simple_spmd(
     int32_t global_tid, int32_t num_vars, size_t reduce_size, void *reduce_data,
     kmp_ShuffleReductFctPtr shflFct, kmp_InterWarpCopyFctPtr cpyFct,
     kmp_CopyToScratchpadFctPtr sratchFct, kmp_LoadReduceFctPtr ldFct);
 EXTERN int32_t __kmpc_nvptx_teams_reduce_nowait_simple_generic(
     int32_t global_tid, int32_t num_vars, size_t reduce_size, void *reduce_data,
     kmp_ShuffleReductFctPtr shflFct, kmp_InterWarpCopyFctPtr cpyFct,
     kmp_CopyToScratchpadFctPtr sratchFct, kmp_LoadReduceFctPtr ldFct);
+EXTERN int32_t __kmpc_nvptx_teams_reduce_nowait_simple(kmp_Ident *loc,
+                                                       int32_t global_tid,
+                                                       kmp_CriticalName *crit);
+EXTERN void __kmpc_nvptx_teams_end_reduce_nowait_simple(kmp_Ident *loc,
+                                                        int32_t global_tid,
+                                                        kmp_CriticalName *crit);
 EXTERN int32_t __kmpc_shuffle_int32(int32_t val, int16_t delta, int16_t size);
 EXTERN int64_t __kmpc_shuffle_int64(int64_t val, int16_t delta, int16_t size);
 
 // sync barrier
 EXTERN void __kmpc_barrier(kmp_Ident *loc_ref, int32_t tid);
 EXTERN void __kmpc_barrier_simple_spmd(kmp_Ident *loc_ref, int32_t tid);
 EXTERN void __kmpc_barrier_simple_generic(kmp_Ident *loc_ref, int32_t tid);
 EXTERN int32_t __kmpc_cancel_barrier(kmp_Ident *loc, int32_t global_tid);

openmp/trunk/libomptarget/deviceRTLs/nvptx/src/reduction.cu

int32_t __kmpc_nvptx_teams_reduce_nowait_simple_generic(
     kmp_ShuffleReductFctPtr shflFct, kmp_InterWarpCopyFctPtr cpyFct,
     kmp_CopyToScratchpadFctPtr scratchFct, kmp_LoadReduceFctPtr ldFct) {
   return nvptx_teams_reduce_nowait(global_tid, num_vars, reduce_size,
                                    reduce_data, shflFct, cpyFct, scratchFct,
                                    ldFct,
                                    /*isSPMDExecutionMode=*/false,
                                    /*isRuntimeUninitialized=*/true);
 }
+
+EXTERN int32_t __kmpc_nvptx_teams_reduce_nowait_simple(kmp_Ident *loc,
+                                                       int32_t global_tid,
+                                                       kmp_CriticalName *crit) {
+  if (checkSPMDMode(loc) && GetThreadIdInBlock() != 0)
+    return 0;
+  // The master thread of the team actually does the reduction.
+  while (atomicCAS((uint32_t *)crit, 0, 1))
+    ;
+  return 1;
+}
+
+EXTERN void
+__kmpc_nvptx_teams_end_reduce_nowait_simple(kmp_Ident *loc, int32_t global_tid,
+                                            kmp_CriticalName *crit) {
+  (void)atomicExch((uint32_t *)crit, 0);
+}
+