[OPENMP][NVPTX]Fix parallel level counter in non-SPMD mode.

Summary:
In non-SPMD mode we may end up with the divergent threads when trying to
increment/decrement parallel level counter. It may lead to incorrect
calculations of the parallel level and wrong results when threads are
divergent. We need to reconverge the threads before trying to modify the
parallel level counter.

Reviewers: grokos, jdoerfert

Subscribers: guansong, openmp-commits, caomhin, kkwli0

Tags: #openmp

Differential Revision: https://reviews.llvm.org/D66802

llvm-svn: 370803

4 files changed, 45 insertions, 16 deletions

diff --git a/openmp/libomptarget/deviceRTLs/nvptx/src/parallel.cu b/openmp/libomptarget/deviceRTLs/nvptx/src/parallel.cu
index d8049d9aaf4..5db443c3b33 100644
--- a/openmp/libomptarget/deviceRTLs/nvptx/src/parallel.cu
+++ b/openmp/libomptarget/deviceRTLs/nvptx/src/parallel.cu
@@ -311,7 +311,16 @@ EXTERN bool __kmpc_kernel_parallel(void **WorkFn,
           (int)newTaskDescr->ThreadId(), (int)nThreads);
 
     isActive = true;
-    IncParallelLevel(threadsInTeam != 1);
+    // Reconverge the threads at the end of the parallel region to correctly
+    // handle parallel levels.
+    // In Cuda9+ in non-SPMD mode we have either 1 worker thread or the whole
+    // warp. If only 1 thread is active, not need to reconverge the threads.
+    // If we have the whole warp, reconverge all the threads in the warp before
+    // actually trying to change the parallel level. Otherwise, parallel level
+    // can be changed incorrectly because of threads divergence.
+    bool IsActiveParallelRegion = threadsInTeam != 1;
+    IncParallelLevel(IsActiveParallelRegion,
+                     IsActiveParallelRegion ? 0xFFFFFFFF : 1u);
   }
 
   return isActive;
@@ -329,7 +338,16 @@ EXTERN void __kmpc_kernel_end_parallel() {
   omptarget_nvptx_threadPrivateContext->SetTopLevelTaskDescr(
       threadId, currTaskDescr->GetPrevTaskDescr());
 
-  DecParallelLevel(threadsInTeam != 1);
+  // Reconverge the threads at the end of the parallel region to correctly
+  // handle parallel levels.
+  // In Cuda9+ in non-SPMD mode we have either 1 worker thread or the whole
+  // warp. If only 1 thread is active, not need to reconverge the threads.
+  // If we have the whole warp, reconverge all the threads in the warp before
+  // actually trying to change the parallel level. Otherwise, parallel level can
+  // be changed incorrectly because of threads divergence.
+    bool IsActiveParallelRegion = threadsInTeam != 1;
+    DecParallelLevel(IsActiveParallelRegion,
+                     IsActiveParallelRegion ? 0xFFFFFFFF : 1u);
 }
 
 ////////////////////////////////////////////////////////////////////////////////
@@ -339,7 +357,7 @@ EXTERN void __kmpc_kernel_end_parallel() {
 EXTERN void __kmpc_serialized_parallel(kmp_Ident *loc, uint32_t global_tid) {
   PRINT0(LD_IO, "call to __kmpc_serialized_parallel\n");
 
-  IncParallelLevel(/*ActiveParallel=*/false);
+  IncParallelLevel(/*ActiveParallel=*/false, __kmpc_impl_activemask());
 
   if (checkRuntimeUninitialized(loc)) {
     ASSERT0(LT_FUSSY, checkSPMDMode(loc),
@@ -378,7 +396,7 @@ EXTERN void __kmpc_end_serialized_parallel(kmp_Ident *loc,
                                            uint32_t global_tid) {
   PRINT0(LD_IO, "call to __kmpc_end_serialized_parallel\n");
 
-  DecParallelLevel(/*ActiveParallel=*/false);
+  DecParallelLevel(/*ActiveParallel=*/false, __kmpc_impl_activemask());
 
   if (checkRuntimeUninitialized(loc)) {
     ASSERT0(LT_FUSSY, checkSPMDMode(loc),
diff --git a/openmp/libomptarget/deviceRTLs/nvptx/src/support.h b/openmp/libomptarget/deviceRTLs/nvptx/src/support.h
index 4df75edfbaa..e10f2a19d32 100644
--- a/openmp/libomptarget/deviceRTLs/nvptx/src/support.h
+++ b/openmp/libomptarget/deviceRTLs/nvptx/src/support.h
@@ -10,6 +10,7 @@
 //
 //===----------------------------------------------------------------------===//
 
+#include "target_impl.h"
 ////////////////////////////////////////////////////////////////////////////////
 // Execution Parameters
 ////////////////////////////////////////////////////////////////////////////////
@@ -65,8 +66,8 @@ INLINE int GetNumberOfProcsInDevice(bool isSPMDExecutionMode);
 INLINE int IsTeamMaster(int ompThreadId);
 
 // Parallel level
-INLINE void IncParallelLevel(bool ActiveParallel);
-INLINE void DecParallelLevel(bool ActiveParallel);
+INLINE void IncParallelLevel(bool ActiveParallel, __kmpc_impl_lanemask_t Mask);
+INLINE void DecParallelLevel(bool ActiveParallel, __kmpc_impl_lanemask_t Mask);
 
 ////////////////////////////////////////////////////////////////////////////////
 // Memory
diff --git a/openmp/libomptarget/deviceRTLs/nvptx/src/supporti.h b/openmp/libomptarget/deviceRTLs/nvptx/src/supporti.h
index 1a0f1a9ecd1..d4da6ad73fa 100644
--- a/openmp/libomptarget/deviceRTLs/nvptx/src/supporti.h
+++ b/openmp/libomptarget/deviceRTLs/nvptx/src/supporti.h
@@ -203,30 +203,28 @@ INLINE int IsTeamMaster(int ompThreadId) { return (ompThreadId == 0); }
 ////////////////////////////////////////////////////////////////////////////////
 // Parallel level
 
-INLINE void IncParallelLevel(bool ActiveParallel) {
-  __kmpc_impl_lanemask_t Active = __kmpc_impl_activemask();
-  __kmpc_impl_syncwarp(Active);
+INLINE void IncParallelLevel(bool ActiveParallel, __kmpc_impl_lanemask_t Mask) {
+  __kmpc_impl_syncwarp(Mask);
   __kmpc_impl_lanemask_t LaneMaskLt = __kmpc_impl_lanemask_lt();
-  unsigned Rank = __kmpc_impl_popc(Active & LaneMaskLt);
+  unsigned Rank = __kmpc_impl_popc(Mask & LaneMaskLt);
   if (Rank == 0) {
     parallelLevel[GetWarpId()] +=
         (1 + (ActiveParallel ? OMP_ACTIVE_PARALLEL_LEVEL : 0));
     __threadfence();
   }
-  __kmpc_impl_syncwarp(Active);
+  __kmpc_impl_syncwarp(Mask);
 }
 
-INLINE void DecParallelLevel(bool ActiveParallel) {
-  __kmpc_impl_lanemask_t Active = __kmpc_impl_activemask();
-  __kmpc_impl_syncwarp(Active);
+INLINE void DecParallelLevel(bool ActiveParallel, __kmpc_impl_lanemask_t Mask) {
+  __kmpc_impl_syncwarp(Mask);
   __kmpc_impl_lanemask_t LaneMaskLt = __kmpc_impl_lanemask_lt();
-  unsigned Rank = __kmpc_impl_popc(Active & LaneMaskLt);
+  unsigned Rank = __kmpc_impl_popc(Mask & LaneMaskLt);
   if (Rank == 0) {
     parallelLevel[GetWarpId()] -=
         (1 + (ActiveParallel ? OMP_ACTIVE_PARALLEL_LEVEL : 0));
     __threadfence();
   }
-  __kmpc_impl_syncwarp(Active);
+  __kmpc_impl_syncwarp(Mask);
 }
 
 ////////////////////////////////////////////////////////////////////////////////
diff --git a/openmp/libomptarget/deviceRTLs/nvptx/test/parallel/level.c b/openmp/libomptarget/deviceRTLs/nvptx/test/parallel/level.c
index edb00e09e60..33e23246456 100644
--- a/openmp/libomptarget/deviceRTLs/nvptx/test/parallel/level.c
+++ b/openmp/libomptarget/deviceRTLs/nvptx/test/parallel/level.c
@@ -135,5 +135,17 @@ int main(int argc, char *argv[]) {
     }
   }
 
+  // Check for paraller level in non-SPMD kernels.
+  level = 0;
+  #pragma omp target teams distribute num_teams(1) thread_limit(32) reduction(+:level)
+  for (int i=0; i<5032; i+=32) {
+    int ub = (i+32 > 5032) ? 5032 : i+32;
+    #pragma omp parallel for schedule(dynamic)
+    for (int j=i ; j < ub; j++) ;
+    level += omp_get_level();
+  }
+  // CHECK: Integral level = 0.
+  printf("Integral level = %d.\n", level);
+
   return 0;
 }