Convert OpenMP target variants of reduction kernels to new interface.

Note: A RAJA fix is needed to have this compile.
LLNL · Oct 15, 2024 · bd434a5 · bd434a5
1 parent 7696a9f
commit bd434a5
Show file tree

Hide file tree

Showing 8 changed files with 30 additions and 17 deletions.
diff --git a/src/algorithm/ATOMIC.hpp b/src/algorithm/ATOMIC.hpp
@@ -74,6 +74,7 @@ class ATOMIC : public KernelBase
   void setOpenMPTuningDefinitions(VariantID vid);
   void setCudaTuningDefinitions(VariantID vid);
   void setHipTuningDefinitions(VariantID vid);
+  void setOpenMPTargetTuningDefinitions(VariantID vid);
 
   template < size_t replication >
   void runSeqVariantReplicate(VariantID vid);

diff --git a/src/algorithm/REDUCE_SUM-OMPTarget.cpp b/src/algorithm/REDUCE_SUM-OMPTarget.cpp
@@ -64,7 +64,8 @@ void REDUCE_SUM::runOpenMPTargetVariant(VariantID vid, size_t RAJAPERF_UNUSED_AR
       RAJA::forall<RAJA::omp_target_parallel_for_exec<threads_per_team>>(
         RAJA::RangeSegment(ibegin, iend),
         RAJA::expt::Reduce<RAJA::operators::plus>(&tsum),
-        [=] (Index_type i, Real_type& sum) {
+        [=] (Index_type i,
+          RAJA::expt::ValOp<Real_type, RAJA::operators::plus>& sum) {
           REDUCE_SUM_BODY;
         }
       );

diff --git a/src/basic/PI_REDUCE-OMPTarget.cpp b/src/basic/PI_REDUCE-OMPTarget.cpp
@@ -64,7 +64,8 @@ void PI_REDUCE::runOpenMPTargetVariant(VariantID vid, size_t RAJAPERF_UNUSED_ARG
       RAJA::forall<RAJA::omp_target_parallel_for_exec<threads_per_team>>(
         RAJA::RangeSegment(ibegin, iend),
         RAJA::expt::Reduce<RAJA::operators::plus>(&tpi),
-        [=] (Index_type i, Real_type& pi) {
+        [=] (Index_type i,
+          RAJA::expt::ValOp<Real_type, RAJA::operators::plus>& pi) {
           PI_REDUCE_BODY;
         }
       );

diff --git a/src/basic/REDUCE3_INT-OMPTarget.cpp b/src/basic/REDUCE3_INT-OMPTarget.cpp
@@ -74,8 +74,11 @@ void REDUCE3_INT::runOpenMPTargetVariant(VariantID vid, size_t RAJAPERF_UNUSED_A
         RAJA::expt::Reduce<RAJA::operators::plus>(&tvsum),
         RAJA::expt::Reduce<RAJA::operators::minimum>(&tvmin),
         RAJA::expt::Reduce<RAJA::operators::maximum>(&tvmax),
-        [=](Index_type i, Int_type& vsum, Int_type& vmin, Int_type& vmax) {
-          REDUCE3_INT_BODY;
+        [=](Index_type i,
+          RAJA::expt::ValOp<Int_type, RAJA::operators::plus>& vsum,
+          RAJA::expt::ValOp<Int_type, RAJA::operators::minimum>& vmin,
+          RAJA::expt::ValOp<Int_type, RAJA::operators::maximum>& vmax) {
+          REDUCE3_INT_BODY_RAJA;
         }
       );
 

diff --git a/src/basic/REDUCE_STRUCT-OMPTarget.cpp b/src/basic/REDUCE_STRUCT-OMPTarget.cpp
@@ -101,10 +101,14 @@ void REDUCE_STRUCT::runOpenMPTargetVariant(VariantID vid, size_t RAJAPERF_UNUSED
           RAJA::expt::Reduce<RAJA::operators::minimum>(&tymin),
           RAJA::expt::Reduce<RAJA::operators::maximum>(&txmax),
           RAJA::expt::Reduce<RAJA::operators::maximum>(&tymax),
-          [=](Index_type i, Real_type& xsum, Real_type& ysum,
-                            Real_type& xmin, Real_type& ymin,
-                            Real_type& xmax, Real_type& ymax) {
-            REDUCE_STRUCT_BODY;
+          [=](Index_type i,
+            RAJA::expt::ValOp<Real_type, RAJA::operators::plus>& xsum,
+            RAJA::expt::ValOp<Real_type, RAJA::operators::plus>& ysum,
+            RAJA::expt::ValOp<Real_type, RAJA::operators::minimum>& xmin,
+            RAJA::expt::ValOp<Real_type, RAJA::operators::minimum>& ymin,
+            RAJA::expt::ValOp<Real_type, RAJA::operators::maximum>& xmax,
+            RAJA::expt::ValOp<Real_type, RAJA::operators::maximum>& ymax ) {
+            REDUCE_STRUCT_BODY_RAJA;
           }
         );
 

diff --git a/src/basic/TRAP_INT-OMPTarget.cpp b/src/basic/TRAP_INT-OMPTarget.cpp
@@ -71,7 +71,8 @@ void TRAP_INT::runOpenMPTargetVariant(VariantID vid, size_t RAJAPERF_UNUSED_ARG(
       RAJA::forall<RAJA::omp_target_parallel_for_exec<threads_per_team>>(
         RAJA::RangeSegment(ibegin, iend),
         RAJA::expt::Reduce<RAJA::operators::plus>(&tsumx),
-        [=] (Index_type i, Real_type& sumx) {
+        [=] (Index_type i,
+          RAJA::expt::ValOp<Real_type, RAJA::operators::plus>& sumx) {
           TRAP_INT_BODY;
         }
       );

diff --git a/src/lcals/FIRST_MIN-OMPTarget.cpp b/src/lcals/FIRST_MIN-OMPTarget.cpp
@@ -60,22 +60,23 @@ void FIRST_MIN::runOpenMPTargetVariant(VariantID vid, size_t RAJAPERF_UNUSED_ARG
 
   } else if ( vid == RAJA_OpenMPTarget ) {
 
-    using VL_TYPE = RAJA::expt::ValLoc<Real_type>;
-
     startTimer();
     for (RepIndex_type irep = 0; irep < run_reps; ++irep) {
 
-      VL_TYPE tloc(m_xmin_init, m_initloc);
+      RAJA::expt::ValLoc<Real_type, RAJA::Index_type> tminloc(m_xmin_init,
+                                                              m_initloc);
 
       RAJA::forall<RAJA::omp_target_parallel_for_exec<threads_per_team>>(
         RAJA::RangeSegment(ibegin, iend),
-        RAJA::expt::Reduce<RAJA::operators::minimum>(&tloc),
-        [=](Index_type i, VL_TYPE& loc) {
-          loc.min(x[i], i);
+        RAJA::expt::Reduce<RAJA::operators::minimum>(&tminloc),
+        [=](Index_type i,
+          RAJA::expt::ValLocOp<Real_type, Index_type,
+                               RAJA::operators::minimum>& minloc) {
+          minloc.minloc(x[i], i);
         }
       );
 
-      m_minloc = static_cast<Index_type>(tloc.getLoc());
+      m_minloc = static_cast<Index_type>(tminloc.getLoc());
 
     }
     stopTimer();

diff --git a/src/stream/DOT-OMPTarget.cpp b/src/stream/DOT-OMPTarget.cpp
@@ -68,7 +68,8 @@ void DOT::runOpenMPTargetVariant(VariantID vid, size_t RAJAPERF_UNUSED_ARG(tune_
         RAJA::forall<RAJA::omp_target_parallel_for_exec<threads_per_team>>(
           RAJA::RangeSegment(ibegin, iend),
           RAJA::expt::Reduce<RAJA::operators::plus>(&tdot),
-          [=] (Index_type i, Real_type& dot) {
+          [=] (Index_type i,
+            RAJA::expt::ValOp<Real_type, RAJA::operators::plus>& dot) {
             DOT_BODY;
           }
         );