Draft: Add SW4CK kernels

src/apps/CMakeLists.txt

@@ -21,7 +21,7 @@ blt_add_library(
           DEL_DOT_VEC_2D-Cuda.cpp 
           DEL_DOT_VEC_2D-OMP.cpp 
           DEL_DOT_VEC_2D-OMPTarget.cpp 
-	  DEL_DOT_VEC_2D-Sycl.cpp 
+      	  DEL_DOT_VEC_2D-Sycl.cpp 
           DIFFUSION3DPA.cpp
           DIFFUSION3DPA-Cuda.cpp
           DIFFUSION3DPA-Hip.cpp
@@ -86,7 +86,11 @@ blt_add_library(
           PRESSURE-Cuda.cpp 
           PRESSURE-OMP.cpp 
           PRESSURE-OMPTarget.cpp 
-          PRESSURE-Sycl.cpp 
+          PRESSURE-Sycl.cpp
+          SW4CK_KERNEL_2.cpp
+          SW4CK_KERNEL_2-Seq.cpp
+          SW4CK_KERNEL_5.cpp
+          SW4CK_KERNEL_5-Seq.cpp
           VOL3D.cpp
           VOL3D-Seq.cpp
           VOL3D-Hip.cpp 

src/apps/SW4CK_KERNEL_2-Seq.cpp

@@ -0,0 +1,145 @@
+//~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~//
+// Copyright (c) 2017-23, 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 "SW4CK_KERNEL_2.hpp"
+
+#include "RAJA/RAJA.hpp"
+
+#include "AppsData.hpp"
+
+#include <iostream>
+
+namespace rajaperf
+{
+namespace apps
+{
+
+
+void SW4CK_KERNEL_2::runSeqVariant(VariantID vid, size_t RAJAPERF_UNUSED_ARG(tune_idx))
+{
+  const Index_type run_reps = getRunReps();
+
+  //To be populated later with
+
+  char op = '=';
+
+  SW4CK_KERNEL_2_DATA_SETUP;
+
+
+  switch ( vid ) {
+
+    case Base_Seq : {
+
+      startTimer();
+      for (RepIndex_type irep = 0; irep < run_reps; ++irep) {
+
+        //Reference impl
+        for(int k=kstart; k<kend+1; k++) {
+          for(int j=jstart; j<jend+1; j++) {
+            for(int i=istart; i<ilast - 1; i++) {
+
+              // 5 ops
+              SW4CK_KERNEL_2_BODY_1;
+
+              // pp derivative (u)
+              // 53 ops, tot=58
+              SW4CK_KERNEL_2_BODY_2;
+
+              // qq derivative (u)
+              // 43 ops, tot=101
+              SW4CK_KERNEL_2_BODY_3;
+
+              // rr derivative (u)
+              // 5*11+14+14=83 ops, tot=184
+              SW4CK_KERNEL_2_BODY_4;
+
+              // rr derivative (v)
+              // 42 ops, tot=226
+              SW4CK_KERNEL_2_BODY_5;
+
+              // rr derivative (w)
+              // 43 ops, tot=269
+              SW4CK_KERNEL_2_BODY_6;
+
+              // pq-derivatives
+              // 38 ops, tot=307
+              SW4CK_KERNEL_2_BODY_7;
+
+              // qp-derivatives
+              // 38 ops, tot=345
+              SW4CK_KERNEL_2_BODY_8;
+
+              // pr-derivatives
+              // 130 ops., tot=475
+              SW4CK_KERNEL_2_BODY_9;
+
+              // rp derivatives
+              // 130 ops, tot=605
+              SW4CK_KERNEL_2_BODY_10;
+
+              // qr derivatives
+              // 82 ops, tot=687
+              SW4CK_KERNEL_2_BODY_11;
+
+              // rq derivatives
+              // 82 ops, tot=769
+              SW4CK_KERNEL_2_BODY_12;
+
+              // 4 ops, tot=773
+              SW4CK_KERNEL_2_BODY_13;
+
+            }
+          }
+        }
+
+
+
+      }
+      stopTimer();
+
+      break;
+    }
+
+#if defined(RUN_RAJA_SEQ)
+    case Lambda_Seq : {
+
+      startTimer();
+      for (RepIndex_type irep = 0; irep < run_reps; ++irep) {
+
+        //Lambda impl
+
+      }
+      stopTimer();
+
+      break;
+    }
+
+    case RAJA_Seq : {
+
+      startTimer();
+      for (RepIndex_type irep = 0; irep < run_reps; ++irep) {
+
+        //RAJA impl
+
+      }
+      stopTimer();
+
+      break;
+    }
+#endif // RUN_RAJA_SEQ
+
+    default : {
+      getCout() << "\n  SW4CK_KERNEL_2 : Unknown variant id = " << vid << std::endl;
+    }
+
+  }
+
+}
+
+} // end namespace apps
+} // end namespace rajaperf

src/apps/SW4CK_KERNEL_2.cpp

@@ -0,0 +1,95 @@
+//~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~//
+// Copyright (c) 2017-23, 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 "SW4CK_KERNEL_2.hpp"
+
+#include "RAJA/RAJA.hpp"
+
+#include "AppsData.hpp"
+#include "common/DataUtils.hpp"
+
+#include <cmath>
+
+
+namespace rajaperf
+{
+namespace apps
+{
+
+
+SW4CK_KERNEL_2::SW4CK_KERNEL_2(const RunParams& params)
+  : KernelBase(rajaperf::Apps_SW4CK_KERNEL_2, params)
+{
+  setDefaultProblemSize(100*100*100);  // See rzmax in ADomain struct
+  setDefaultReps(100);
+
+  Index_type rzmax = std::cbrt(getTargetProblemSize())+1;
+  //m_domain = new ADomain(rzmax, /* ndims = */ 3);
+
+  //m_array_length = m_domain->nnalls;
+
+  //setActualProblemSize( m_domain->lpz+1 - m_domain->fpz );
+
+  //setItsPerRep( m_domain->lpz+1 - m_domain->fpz );
+  setKernelsPerRep(1);
+  // touched data size, not actual number of stores and loads
+  //  setBytesPerRep( (1*sizeof(Real_type) + 0*sizeof(Real_type)) * getItsPerRep() +
+  //(0*sizeof(Real_type) + 3*sizeof(Real_type)) * (getItsPerRep() + 1+m_domain->jp+m_domain->kp) );
+
+  //setFLOPsPerRep(72 * (m_domain->lpz+1 - m_domain->fpz));
+
+  checksum_scale_factor = 0.001 *
+              ( static_cast<Checksum_type>(getDefaultProblemSize()) /
+                                           getActualProblemSize() );
+
+  setUsesFeature(Teams);
+
+  //Goal is to get the following three variants right first
+  setVariantDefined( Base_Seq );
+  setVariantDefined( Lambda_Seq );
+  setVariantDefined( RAJA_Seq );
+
+  /*
+  setVariantDefined( Base_OpenMP );
+  setVariantDefined( Lambda_OpenMP );
+  setVariantDefined( RAJA_OpenMP );
+
+  setVariantDefined( Base_OpenMPTarget );
+  setVariantDefined( RAJA_OpenMPTarget );
+
+  setVariantDefined( Base_CUDA );
+  setVariantDefined( RAJA_CUDA );
+
+  setVariantDefined( Base_HIP );
+  setVariantDefined( RAJA_HIP );
+  */
+}
+
+SW4CK_KERNEL_2::~SW4CK_KERNEL_2()
+{
+  //  delete m_domain;
+}
+
+void SW4CK_KERNEL_2::setUp(VariantID vid, size_t RAJAPERF_UNUSED_ARG(tune_idx))
+{
+
+}
+
+void SW4CK_KERNEL_2::updateChecksum(VariantID vid, size_t tune_idx)
+{
+  //checksum[vid][tune_idx] += calcChecksum(m_vol, m_array_length, checksum_scale_factor );
+}
+
+void SW4CK_KERNEL_2::tearDown(VariantID vid, size_t RAJAPERF_UNUSED_ARG(tune_idx))
+{
+  (void) vid;
+
+}
+
+} // end namespace apps
+} // end namespace rajaperf