Merge pull request #23 from LLNL/feature/chen59/update-raja-v0.14.0

Update to RAJA v0.14.1
LLNL · Nov 15, 2021 · 9853d2e · 9853d2e
2 parents cbd91a5 + 724f49d
commit 9853d2e
Show file tree

Hide file tree

Showing 117 changed files with 1,058 additions and 2,139 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -19,9 +19,9 @@ option(ENABLE_LULESH_TWO "Build LULESH v2.0" On)
 option(ENABLE_KRIPKE "Build Kripke" On)
 option(ENABLE_COMD "Build CoMD" On)
 
-set (ENABLE_TESTS Off CACHE Bool "Don't build tests")
-set (ENABLE_EXAMPLES Off CACHE Bool "Don't build examples")
-set (ENABLE_EXERCISES Off CACHE Bool "Don't build exercises")
+set (ENABLE_TESTS Off CACHE BOOL "Don't build tests")
+set (ENABLE_EXAMPLES Off CACHE BOOL "Don't build examples")
+set (ENABLE_EXERCISES Off CACHE BOOL "Don't build exercises")
 
 if (PROXIES_ENABLE_WARNINGS)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -Wextra -Werror")

diff --git a/CoMD/CMakeLists.txt b/CoMD/CMakeLists.txt
@@ -6,11 +6,11 @@ project(COMD LANGUAGES CXX)
 # Initialize the BLT build system
 #
 
-set(ENABLE_GTEST Off CACHE Bool "")
-set(ENABLE_TESTS Off CACHE Bool "")
-set(ENABLE_TBB Off CACHE Bool "")
-set(ENABLE_EXAMPLES Off CACHE Bool "")
-set(ENABLE_DOCUMENTATION Off CACHE Bool "")
+set(ENABLE_GTEST Off CACHE BOOL "")
+set(ENABLE_TESTS Off CACHE BOOL "")
+set(ENABLE_TBB Off CACHE BOOL "")
+set(ENABLE_EXAMPLES Off CACHE BOOL "")
+set(ENABLE_DOCUMENTATION Off CACHE BOOL "")
 
 #
 # Configure RAJA

diff --git a/CoMD/src-threaded/CoMD.cpp b/CoMD/src-threaded/CoMD.cpp
@@ -235,12 +235,16 @@ SimFlat* initSimulation(Command cmd)
    // create lattice with desired temperature and displacement.
    createFccLattice(cmd.nx, cmd.ny, cmd.nz, latticeConstant, sim);
 
-   RAJA::resources::Resource defHostRes{RAJA::resources::Host::get_default()};
+#ifdef ENABLE_CUDA
+   RAJA::resources::Resource erasedRes{RAJA::resources::Cuda()};
+#else
+   RAJA::resources::Resource erasedRes{RAJA::resources::Host::get_default()};
+#endif
 
    /* Create Total IndexSets */
-   sim->isTotal = new RAJA::TypedIndexSet<RAJA::RangeSegment>() ;
+   sim->isTotal = new RAJA::TypedIndexSet<RAJA::TypedRangeSegment<int>>() ;
    for (int i=0; i<sim->boxes->nTotalBoxes; ++i) {
-     RAJA::RangeSegment myseg(i*MAXATOMS,i*MAXATOMS + sim->boxes->nAtoms[i]);
+     RAJA::TypedRangeSegment<int> myseg(i*MAXATOMS,i*MAXATOMS + sim->boxes->nAtoms[i]);
      sim->isTotal->push_back( myseg ) ;
    }
 
@@ -256,21 +260,21 @@ SimFlat* initSimulation(Command cmd)
            tmpBox[tmpCount++] = sim->boxes->nbrBoxes[i][j] ;
          }
        }
-         sim->boxes->nbrSegments[i] = new RAJA::TypedListSegment<int>(tmpBox, tmpCount, defHostRes);
+         sim->boxes->nbrSegments[i] = new RAJA::TypedListSegment<int>(tmpBox, tmpCount, erasedRes);
      }
      else {
-       sim->boxes->nbrSegments[i] = new RAJA::TypedListSegment<int>(sim->boxes->nbrBoxes[i], 27, defHostRes);
+       sim->boxes->nbrSegments[i] = new RAJA::TypedListSegment<int>(sim->boxes->nbrBoxes[i], 27, erasedRes);
      }
    }
 
    /* Create Local IndexSet View */
-   sim->isLocal = new RAJA::TypedIndexSet<RAJA::RangeSegment>() ;
+   sim->isLocal = new RAJA::TypedIndexSet<RAJA::TypedRangeSegment<int>>() ;
    for(int i = 0; i < sim->boxes->nLocalBoxes; ++i) {
-     RAJA::RangeSegment myseg(i*MAXATOMS, i*MAXATOMS + sim->boxes->nAtoms[i]);
+     RAJA::TypedRangeSegment<int> myseg(i*MAXATOMS, i*MAXATOMS + sim->boxes->nAtoms[i]);
      sim->isLocal->push_back( myseg ) ;
    }
 
-   sim->isLocalSegment = new RAJA::RangeSegment(0, sim->boxes->nLocalBoxes);
+   sim->isLocalSegment = new RAJA::TypedRangeSegment<int>(0, sim->boxes->nLocalBoxes);
 
    setTemperature(sim, cmd.temperature);
 
@@ -412,7 +416,7 @@ void sumAtoms(SimFlat* s)
    */
   RAJA::kernel<redistributeKernel>(
   RAJA::make_tuple(
-    RAJA::RangeSegment(0, nLocalBoxes)),
+    RAJA::TypedRangeSegment<int>(0, nLocalBoxes)),
     [=] RAJA_DEVICE (int i)
     {
       nLocalReduce += s->boxes->nAtoms[i];

diff --git a/CoMD/src-threaded/CoMDTypes.h b/CoMD/src-threaded/CoMDTypes.h
@@ -90,9 +90,9 @@ typedef struct SimFlatSt
 
    HaloExchange* atomExchange;
 
-   RAJA::TypedIndexSet<RAJA::RangeSegment> *isTotal ;
-   RAJA::TypedIndexSet<RAJA::RangeSegment> *isLocal ;
-   RAJA::RangeSegment *isLocalSegment ;
+   RAJA::TypedIndexSet<RAJA::TypedRangeSegment<int>> *isTotal ;
+   RAJA::TypedIndexSet<RAJA::TypedRangeSegment<int>> *isLocal ;
+   RAJA::TypedRangeSegment<int> *isLocalSegment ;
 
 } SimFlat;
 

diff --git a/CoMD/src-threaded/eam.cpp b/CoMD/src-threaded/eam.cpp
@@ -241,8 +241,8 @@ int eamForce(SimFlat* s)
 
    RAJA::kernel<atomWorkKernel>(
            RAJA::make_tuple(
-           RAJA::RangeSegment(0, s->boxes->nLocalBoxes),
-           RAJA::RangeSegment(0, MAXATOMS) ),
+           RAJA::TypedRangeSegment<int>(0, s->boxes->nLocalBoxes),
+           RAJA::TypedRangeSegment<int>(0, MAXATOMS) ),
            [=] COMD_DEVICE (int iBox, int iOffLocal) {
 
               const int nIBox = s->boxes->nAtoms[iBox];
@@ -271,9 +271,9 @@ int eamForce(SimFlat* s)
    RAJA::kernel<forcePolicyKernel>(
      RAJA::make_tuple(
        *s->isLocalSegment,                // local boxes
-       RAJA::RangeSegment(0,27),          // 27 neighbor boxes
-       RAJA::RangeSegment(0, MAXATOMS),   // atoms i in local box
-       RAJA::RangeSegment(0, MAXATOMS) ), // atoms j in neighbor box
+       RAJA::TypedRangeSegment<int>(0,27),          // 27 neighbor boxes
+       RAJA::TypedRangeSegment<int>(0, MAXATOMS),   // atoms i in local box
+       RAJA::TypedRangeSegment<int>(0, MAXATOMS) ), // atoms j in neighbor box
      [=] COMD_DEVICE (int iBox, int nghb, int iOff, int jOff) {
 
        const int jBox = s->boxes->nbrBoxes[iBox][nghb];
@@ -338,8 +338,8 @@ int eamForce(SimFlat* s)
    profileStart(forceFunctionTimer);
    RAJA::kernel<atomWorkKernel>(
    RAJA::make_tuple(
-     RAJA::RangeSegment(0, s->boxes->nLocalBoxes),
-     RAJA::RangeSegment(0, MAXATOMS) ),
+     RAJA::TypedRangeSegment<int>(0, s->boxes->nLocalBoxes),
+     RAJA::TypedRangeSegment<int>(0, MAXATOMS) ),
    [=] COMD_DEVICE (int iBox, int iOffLocal) {
 
         const int nIBox = s->boxes->nAtoms[iBox];
@@ -368,9 +368,9 @@ int eamForce(SimFlat* s)
    RAJA::kernel<forcePolicyKernel>(
      RAJA::make_tuple(
        *s->isLocalSegment,                // local boxes
-       RAJA::RangeSegment(0,27),          // 27 neighbor boxes
-       RAJA::RangeSegment(0, MAXATOMS),   // atoms i in local box
-       RAJA::RangeSegment(0, MAXATOMS) ), // atoms j in neighbor box
+       RAJA::TypedRangeSegment<int>(0,27),          // 27 neighbor boxes
+       RAJA::TypedRangeSegment<int>(0, MAXATOMS),   // atoms i in local box
+       RAJA::TypedRangeSegment<int>(0, MAXATOMS) ), // atoms j in neighbor box
      [=] COMD_DEVICE (int iBox, int nghb, int iOff, int jOff) {
 
        const int jBox = s->boxes->nbrBoxes[iBox][nghb];

diff --git a/CoMD/src-threaded/haloExchange.cpp b/CoMD/src-threaded/haloExchange.cpp
@@ -443,7 +443,7 @@ int loadAtomsBuffer(void* vparms, void* data, int face, char* charBuf)
    // TODO: Try to optimize this.
    RAJA::kernel<redistributeKernel>(
      RAJA::make_tuple(
-     RAJA::RangeSegment(0, nCells)),
+     RAJA::TypedRangeSegment<int>(0, nCells)),
    [=] RAJA_DEVICE (int iCell) {
        SimFlat* s = (SimFlat*) data;
        int* offsetBuf = (int*)charBuf;

diff --git a/CoMD/src-threaded/linkCells.cpp b/CoMD/src-threaded/linkCells.cpp
@@ -125,7 +125,7 @@ LinkCell* initLinkCells(const Domain* domain, real_t cutoff)
    // Adding nbrSegments
    ll->nbrSegments = (RAJA::TypedListSegment<int> **)comdMalloc(ll->nTotalBoxes*sizeof(RAJA::TypedListSegment<int>*));
 
-   ll->neighbors = (RAJA::TypedIndexSet<RAJA::RangeSegment> **)comdMalloc(ll->nTotalBoxes*sizeof(RAJA::TypedIndexSet<RAJA::RangeSegment> *));
+   ll->neighbors = (RAJA::TypedIndexSet<RAJA::TypedRangeSegment<int>> **)comdMalloc(ll->nTotalBoxes*sizeof(RAJA::TypedIndexSet<RAJA::TypedRangeSegment<int>> *));
 
    return ll;
 }
@@ -407,7 +407,7 @@ void emptyHaloCells(LinkCell* boxes)
 {
    RAJA::kernel<redistributeKernel>(
    RAJA::make_tuple(
-                    RAJA::RangeSegment(boxes->nLocalBoxes, boxes->nTotalBoxes)),
+                    RAJA::TypedRangeSegment<int>(boxes->nLocalBoxes, boxes->nTotalBoxes)),
    [=] RAJA_DEVICE (int ii) {
      boxes->nAtoms[ii] = 0;
    });
@@ -524,7 +524,7 @@ void unloadAtomsBuffer(void* vparms, void* data, int face, int bufSize, char* ch
 
   RAJA::kernel<redistributeKernel>(
   RAJA::make_tuple(
-    RAJA::RangeSegment(0, nCells)),
+    RAJA::TypedRangeSegment<int>(0, nCells)),
   [=] RAJA_DEVICE (int iCell) {
     AtomMsg* buf = (AtomMsg*)(charBuf + size);
     int* offsets = (int*)(charBuf);

diff --git a/CoMD/src-threaded/linkCells.h b/CoMD/src-threaded/linkCells.h
@@ -31,7 +31,7 @@ typedef struct LinkCellSt
    int* nAtoms;         //!< total number of atoms in each box
    int** nbrBoxes;      //!< neighbor boxes for each box
 
-  RAJA::TypedIndexSet<RAJA::RangeSegment> ** neighbors ; //!< neighbor IndexSets (Slices) for each box
+  RAJA::TypedIndexSet<RAJA::TypedRangeSegment<int>> ** neighbors ; //!< neighbor IndexSets (Slices) for each box
   RAJA::TypedListSegment<int> ** nbrSegments ; //!< neighbor list per box, same indexing as boxes
 
 

diff --git a/CoMD/src-threaded/ljForce.cpp b/CoMD/src-threaded/ljForce.cpp
@@ -160,8 +160,8 @@ int ljForce(SimFlat* s)
    profileStart(forceZeroingTimer);
   RAJA::kernel<atomWorkKernel>(
   RAJA::make_tuple(
-    RAJA::RangeSegment(0, s->boxes->nLocalBoxes),
-    RAJA::RangeSegment(0, MAXATOMS) ),
+    RAJA::TypedRangeSegment<int>(0, s->boxes->nLocalBoxes),
+    RAJA::TypedRangeSegment<int>(0, MAXATOMS) ),
     [=] COMD_DEVICE (int iBox, int iOffLocal) {
       const int nIBox = s->boxes->nAtoms[iBox];
       if(iOffLocal < nIBox) {
@@ -181,9 +181,9 @@ int ljForce(SimFlat* s)
      RAJA::kernel<forcePolicyKernel>(
        RAJA::make_tuple(
          *s->isLocalSegment,                // local boxes
-         RAJA::RangeSegment(0,27),          // 27 neighbor boxes
-         RAJA::RangeSegment(0, MAXATOMS),   // atoms i in local box
-         RAJA::RangeSegment(0, MAXATOMS) ), // atoms j in neighbor box
+         RAJA::TypedRangeSegment<int>(0,27),          // 27 neighbor boxes
+         RAJA::TypedRangeSegment<int>(0, MAXATOMS),   // atoms i in local box
+         RAJA::TypedRangeSegment<int>(0, MAXATOMS) ), // atoms j in neighbor box
        [=] COMD_DEVICE (int iBoxID, int nghb, int iOff, int jOff) {
          const int nLocalBoxes = s->boxes->nLocalBoxes;
          const int nIBox = s->boxes->nAtoms[iBoxID];

diff --git a/CoMD/src-threaded/timestep.cpp b/CoMD/src-threaded/timestep.cpp
@@ -8,8 +8,8 @@
 #include "parallel.h"
 #include "performanceTimers.h"
 
-static void advanceVelocity(SimFlat* s, RAJA::TypedIndexSet<RAJA::RangeSegment> *extent, real_t dt);
-static void advancePosition(SimFlat* s, RAJA::TypedIndexSet<RAJA::RangeSegment> *extent, real_t dt);
+static void advanceVelocity(SimFlat* s, RAJA::TypedIndexSet<RAJA::TypedRangeSegment<int>> *extent, real_t dt);
+static void advancePosition(SimFlat* s, RAJA::TypedIndexSet<RAJA::TypedRangeSegment<int>> *extent, real_t dt);
 
 extern "C" int sortAtomsById(const void* a, const void* b);
 
@@ -67,12 +67,12 @@ void computeForce(SimFlat* s)
 }
 
 
-void advanceVelocity(SimFlat* s, RAJA::TypedIndexSet<RAJA::RangeSegment> *extent, real_t dt)
+void advanceVelocity(SimFlat* s, RAJA::TypedIndexSet<RAJA::TypedRangeSegment<int>> *extent, real_t dt)
 {
   RAJA::kernel<atomWorkKernel>(
   RAJA::make_tuple(
-    RAJA::RangeSegment(0, s->boxes->nLocalBoxes),
-    RAJA::RangeSegment(0, MAXATOMS) ),
+    RAJA::TypedRangeSegment<int>(0, s->boxes->nLocalBoxes),
+    RAJA::TypedRangeSegment<int>(0, MAXATOMS) ),
     [=] COMD_DEVICE (int iBox, int iOffLocal) {
       const int nIBox = s->boxes->nAtoms[iBox];
       if(iOffLocal < nIBox) {
@@ -87,12 +87,12 @@ void advanceVelocity(SimFlat* s, RAJA::TypedIndexSet<RAJA::RangeSegment> *extent
     } );
 }
 
-void advancePosition(SimFlat* s, RAJA::TypedIndexSet<RAJA::RangeSegment> *extent, real_t dt)
+void advancePosition(SimFlat* s, RAJA::TypedIndexSet<RAJA::TypedRangeSegment<int>> *extent, real_t dt)
 {
   RAJA::kernel<atomWorkKernel>(
   RAJA::make_tuple(
-  RAJA::RangeSegment(0, s->boxes->nLocalBoxes),
-    RAJA::RangeSegment(0, MAXATOMS) ),
+  RAJA::TypedRangeSegment<int>(0, s->boxes->nLocalBoxes),
+    RAJA::TypedRangeSegment<int>(0, MAXATOMS) ),
     [=] COMD_DEVICE (int iBox, int iOffLocal) {
       const int nIBox = s->boxes->nAtoms[iBox];
       if(iOffLocal < nIBox) {
@@ -122,8 +122,8 @@ void kineticEnergy(SimFlat* s)
 
   RAJA::kernel<atomWorkKernel>(
   RAJA::make_tuple(
-    RAJA::RangeSegment(0, s->boxes->nLocalBoxes),
-    RAJA::RangeSegment(0, MAXATOMS) ),
+    RAJA::TypedRangeSegment<int>(0, s->boxes->nLocalBoxes),
+    RAJA::TypedRangeSegment<int>(0, MAXATOMS) ),
     [=] COMD_DEVICE (int iBox, int iOffLocal) {
       const int nIBox = s->boxes->nAtoms[iBox];
       if(iOffLocal < nIBox) {
@@ -272,7 +272,7 @@ void redistributeAtoms(SimFlat* sim)
    startTimer(atomSortTimer);
    RAJA::kernel<redistributeKernel>(
    RAJA::make_tuple(
-   RAJA::RangeSegment(0, sim->boxes->nTotalBoxes)),
+   RAJA::TypedRangeSegment<int>(0, sim->boxes->nTotalBoxes)),
    [=] COMD_DEVICE (int iBox) {
      sortAtomsInCell(sim->atoms, sim->boxes, iBox);
    } );

diff --git a/blt b/blt
diff --git a/host-configs/blueos/clang_3_9_1.cmake b/host-configs/blueos/clang_3_9_1.cmake
diff --git a/host-configs/blueos/clang_4_0_0.cmake b/host-configs/blueos/clang_4_0_0.cmake
diff --git a/host-configs/blueos/clang_X.cmake b/host-configs/blueos/clang_X.cmake
@@ -0,0 +1,16 @@
+###############################################################################
+# Copyright (c) 2016-21, Lawrence Livermore National Security, LLC
+# and RAJA project contributors. See the RAJA/COPYRIGHT file for details.
+#
+# SPDX-License-Identifier: (BSD-3-Clause)
+###############################################################################
+
+set(RAJA_COMPILER "RAJA_COMPILER_CLANG" CACHE STRING "")
+
+set(CMAKE_CXX_FLAGS_RELEASE "-O3" CACHE STRING "")
+set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g" CACHE STRING "")
+set(CMAKE_CXX_FLAGS_DEBUG "-O0 -g" CACHE STRING "")
+
+set(RAJA_DATA_ALIGN 64 CACHE STRING "")
+
+set(RAJA_HOST_CONFIG_LOADED On CACHE BOOL "")
diff --git a/host-configs/blueos/clang_coral_2017_08_31.cmake b/host-configs/blueos/clang_coral_2017_08_31.cmake