Add DAXPY_ATOMIC, formatting, in REDUCE3, Kokkos::min, Kokkos::max

LLNL · Feb 7, 2022 · 1320300 · 1320300
1 parent 0d8b139
commit 1320300
Show file tree

Hide file tree

Showing 6 changed files with 112 additions and 62 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -155,7 +155,6 @@ Kokkos_ENABLE_HIP variable to ON")
 
   if(ENABLE_TARGET_OPENMP)
     set(Kokkos_ENABLE_OPENMPTARGET ON CACHE BOOL "Docstring")
-    set(RAJA_ENABLE_TARGET_OPENMP ON CACHE BOOL "Docstring")
     if(NOT CMAKE_BUILD_TYPE MATCHES Debug)
       if(NOT EXPERIMENTAL_BUILD)
       message(FATAL_ERROR "Kokkos builds with OpenMPTarget require a Debug build to succeed at the moment. Rebuild with CMAKE_BUILD_TYPE=Debug. If you're a compiler developer, rebuild with -DEXPERIMENTAL_BUILD=ON")

diff --git a/src/basic-kokkos/CMakeLists.txt b/src/basic-kokkos/CMakeLists.txt
@@ -21,5 +21,6 @@ blt_add_library(
           NESTED_INIT-Kokkos.cpp
           REDUCE3_INT-Kokkos.cpp
           TRAP_INT-Kokkos.cpp
+          DAXPY_ATOMIC-Kokkos.cpp
   DEPENDS_ON common ${RAJA_PERFSUITE_DEPENDS}
   )
diff --git a/src/basic-kokkos/DAXPY_ATOMIC-Kokkos.cpp b/src/basic-kokkos/DAXPY_ATOMIC-Kokkos.cpp
@@ -0,0 +1,67 @@
+//~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~//
+// Copyright (c) 2017-21, Lawrence Livermore National Security, LLC
+// and RAJA Performance Suite project contributors.
+// See the RAJAPerf/LICENSE file for details.
+//
+// SPDX-License-Identifier: (BSD-3-Clause)
+//~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~//
+
+#include "DAXPY_ATOMIC.hpp"
+#include "common/KokkosViewUtils.hpp"
+#include <iostream>
+
+// Delete me
+// For de-bugging:
+#include "RAJA/RAJA.hpp"
+
+namespace rajaperf {
+namespace basic {
+
+void DAXPY_ATOMIC::runKokkosVariant(VariantID vid) {
+
+  const Index_type run_reps = getRunReps();
+  const Index_type ibegin = 0;
+  const Index_type iend = getActualProblemSize();
+
+  DAXPY_ATOMIC_DATA_SETUP;
+  //
+  // Kokkos Views to wrap pointers declared in DAXPY_ATOMIC.hpp
+  //
+
+  auto x_view = getViewFromPointer(x, iend);
+  auto y_view = getViewFromPointer(y, iend);
+
+  switch (vid) {
+
+  case Kokkos_Lambda: {
+
+    Kokkos::fence();
+    startTimer();
+
+    for (RepIndex_type irep = 0; irep < run_reps; ++irep) {
+
+      Kokkos::parallel_for(
+          "DAXPY_ATOMIC_Kokkos Kokkos_Lambda",
+          Kokkos::RangePolicy<Kokkos::DefaultExecutionSpace>(ibegin, iend),
+          KOKKOS_LAMBDA(Index_type i) {
+            Kokkos::atomic_add(&y_view[i], a * x_view[i]);
+          });
+    }
+
+    Kokkos::fence();
+    stopTimer();
+
+    break;
+  }
+
+  default: {
+    getCout() << "\n  DAXPY_ATOMIC : Unknown variant id = " << vid << std::endl;
+  }
+  }
+
+  moveDataToHostFromKokkosView(x, x_view, iend);
+  moveDataToHostFromKokkosView(y, y_view, iend);
+}
+
+} // end namespace basic
+} // end namespace rajaperf
diff --git a/src/basic-kokkos/REDUCE3_INT-Kokkos.cpp b/src/basic-kokkos/REDUCE3_INT-Kokkos.cpp
@@ -33,16 +33,11 @@ void REDUCE3_INT::runKokkosVariant(VariantID vid) {
     Kokkos::fence();
     startTimer();
     for (RepIndex_type irep = 0; irep < run_reps; ++irep) {
-      // The values below are initilized elsewhere by RPS
-      // These variables were declared to Kokkos-ify the parallel_reduce
-      // construct:
 
-// If the RAJA OPENMP TARGET OPTION IS NOT DEFINED
-//#ifndef RAJA_ENABLE_TARGET_OPENMP
       Int_type max_value = m_vmax_init;
       Int_type min_value = m_vmin_init;
       Int_type sum = m_vsum_init;
-
+      // ADL: argument-dependent look up here
       parallel_reduce(
           "REDUCE3-Kokkos Kokkos_Lambda",
           Kokkos::RangePolicy<Kokkos::DefaultExecutionSpace>(ibegin, iend),
@@ -58,10 +53,10 @@ void REDUCE3_INT::runKokkosVariant(VariantID vid) {
           Kokkos::Max<Int_type>(max_value), Kokkos::Min<Int_type>(min_value),
           sum);
       m_vsum += static_cast<Int_type>(sum);
-      m_vmin = Kokkos::Experimental::min(m_vmin, static_cast<Int_type>(min_value));
-      m_vmax = Kokkos::Experimental::max(m_vmax, static_cast<Int_type>(max_value));
-
-//#endif
+      m_vmin =
+          Kokkos::Experimental::min(m_vmin, static_cast<Int_type>(min_value));
+      m_vmax =
+          Kokkos::Experimental::max(m_vmax, static_cast<Int_type>(max_value));
     }
     Kokkos::fence();
     stopTimer();

diff --git a/src/basic/DAXPY_ATOMIC.cpp b/src/basic/DAXPY_ATOMIC.cpp
@@ -12,66 +12,60 @@
 
 #include "common/DataUtils.hpp"
 
-namespace rajaperf
-{
-namespace basic
-{
+namespace rajaperf {
+namespace basic {
 
-
-DAXPY_ATOMIC::DAXPY_ATOMIC(const RunParams& params)
-  : KernelBase(rajaperf::Basic_DAXPY_ATOMIC, params)
-{
+DAXPY_ATOMIC::DAXPY_ATOMIC(const RunParams &params)
+    : KernelBase(rajaperf::Basic_DAXPY_ATOMIC, params) {
   setDefaultProblemSize(1000000);
   setDefaultReps(500);
 
-  setActualProblemSize( getTargetProblemSize() );
+  setActualProblemSize(getTargetProblemSize());
 
-  setItsPerRep( getActualProblemSize() );
+  setItsPerRep(getActualProblemSize());
   setKernelsPerRep(1);
-  setBytesPerRep( (1*sizeof(Real_type) + 2*sizeof(Real_type)) * getActualProblemSize() );
+  setBytesPerRep((1 * sizeof(Real_type) + 2 * sizeof(Real_type)) *
+                 getActualProblemSize());
   setFLOPsPerRep(2 * getActualProblemSize());
 
   setUsesFeature(Forall);
 
-  setVariantDefined( Base_Seq );
-  setVariantDefined( Lambda_Seq );
-  setVariantDefined( RAJA_Seq );
+  setVariantDefined(Base_Seq);
+  setVariantDefined(Lambda_Seq);
+  setVariantDefined(RAJA_Seq);
 
-  setVariantDefined( Base_OpenMP );
-  setVariantDefined( Lambda_OpenMP );
-  setVariantDefined( RAJA_OpenMP );
+  setVariantDefined(Base_OpenMP);
+  setVariantDefined(Lambda_OpenMP);
+  setVariantDefined(RAJA_OpenMP);
 
-  setVariantDefined( Base_OpenMPTarget );
-  setVariantDefined( RAJA_OpenMPTarget );
+  setVariantDefined(Base_OpenMPTarget);
+  setVariantDefined(RAJA_OpenMPTarget);
 
-  setVariantDefined( Base_CUDA );
-  setVariantDefined( Lambda_CUDA );
-  setVariantDefined( RAJA_CUDA );
+  setVariantDefined(Base_CUDA);
+  setVariantDefined(Lambda_CUDA);
+  setVariantDefined(RAJA_CUDA);
 
-  setVariantDefined( Base_HIP );
-  setVariantDefined( Lambda_HIP );
-  setVariantDefined( RAJA_HIP );
-}
+  setVariantDefined(Base_HIP);
+  setVariantDefined(Lambda_HIP);
+  setVariantDefined(RAJA_HIP);
 
-DAXPY_ATOMIC::~DAXPY_ATOMIC()
-{
+  setVariantDefined(Kokkos_Lambda);
 }
 
-void DAXPY_ATOMIC::setUp(VariantID vid)
-{
+DAXPY_ATOMIC::~DAXPY_ATOMIC() {}
+
+void DAXPY_ATOMIC::setUp(VariantID vid) {
   allocAndInitDataConst(m_y, getActualProblemSize(), 0.0, vid);
   allocAndInitData(m_x, getActualProblemSize(), vid);
   initData(m_a);
 }
 
-void DAXPY_ATOMIC::updateChecksum(VariantID vid)
-{
+void DAXPY_ATOMIC::updateChecksum(VariantID vid) {
   checksum[vid] += calcChecksum(m_y, getActualProblemSize());
 }
 
-void DAXPY_ATOMIC::tearDown(VariantID vid)
-{
-  (void) vid;
+void DAXPY_ATOMIC::tearDown(VariantID vid) {
+  (void)vid;
   deallocData(m_x);
   deallocData(m_y);
 }

diff --git a/src/basic/DAXPY_ATOMIC.hpp b/src/basic/DAXPY_ATOMIC.hpp
@@ -17,32 +17,25 @@
 #ifndef RAJAPerf_Basic_DAXPY_ATOMIC_HPP
 #define RAJAPerf_Basic_DAXPY_ATOMIC_HPP
 
-#define DAXPY_ATOMIC_DATA_SETUP \
-  Real_ptr x = m_x; \
-  Real_ptr y = m_y; \
+#define DAXPY_ATOMIC_DATA_SETUP                                                \
+  Real_ptr x = m_x;                                                            \
+  Real_ptr y = m_y;                                                            \
   Real_type a = m_a;
 
-#define DAXPY_ATOMIC_BODY  \
-  y[i] += a * x[i] ;
-
-#define DAXPY_ATOMIC_RAJA_BODY(policy)  \
-  RAJA::atomicAdd<policy>(&y[i], a * x[i]);
+#define DAXPY_ATOMIC_BODY y[i] += a * x[i];
 
+#define DAXPY_ATOMIC_RAJA_BODY(policy) RAJA::atomicAdd<policy>(&y[i], a * x[i]);
 
 #include "common/KernelBase.hpp"
 
-namespace rajaperf
-{
+namespace rajaperf {
 class RunParams;
 
-namespace basic
-{
+namespace basic {
 
-class DAXPY_ATOMIC : public KernelBase
-{
+class DAXPY_ATOMIC : public KernelBase {
 public:
-
-  DAXPY_ATOMIC(const RunParams& params);
+  DAXPY_ATOMIC(const RunParams &params);
 
   ~DAXPY_ATOMIC();
 
@@ -55,6 +48,7 @@ class DAXPY_ATOMIC : public KernelBase
   void runCudaVariant(VariantID vid);
   void runHipVariant(VariantID vid);
   void runOpenMPTargetVariant(VariantID vid);
+  void runKokkosVariant(VariantID vid);
 
 private:
   Real_ptr m_x;