Add kernel fusing using RAJA

config/SAMRAI_config.h.cmake.in

@@ -358,7 +358,9 @@
 /* Configure for compiling on BGL family of machines */
 #undef __BGL_FAMILY__
 
-
+#ifdef HAVE_RAJA
+#define HAVE_KERNEL_FUSER
+#endif
 
 namespace SAMRAI {
    static const unsigned short MAX_DIM_VAL = SAMRAI_MAXIMUM_DIMENSION;

source/SAMRAI/hier/PatchData.C

@@ -31,6 +31,33 @@ PatchData::~PatchData()
 {
 }
 
+void
+PatchData::copy(
+   const PatchData& src,
+   const BoxOverlap& overlap,
+   tbox::KernelFuser& fuser)
+{
+   copy(src, overlap);
+}
+
+void
+PatchData::packStream(
+   tbox::MessageStream& stream,
+   const BoxOverlap& overlap,
+   tbox::KernelFuser& fuser)
+{
+   packStream(stream, overlap);
+}
+
+void
+PatchData::unpackStream(
+   tbox::MessageStream& stream,
+   const BoxOverlap& overlap,
+   tbox::KernelFuser& fuser)
+{
+   unpackStream(stream, overlap);
+}
+
 /*
  *************************************************************************
  *

source/SAMRAI/hier/PatchData.h

@@ -20,6 +20,15 @@
 #include "SAMRAI/tbox/Utilities.h"
 
 namespace SAMRAI {
+
+/*
+ * Forward declaration of KernelFuser class - required here because it sucks in
+ * RAJA and requires CUDA.
+ */
+namespace tbox {
+class KernelFuser;
+}
+
 namespace hier {
 
 /**
@@ -160,6 +169,12 @@ class PatchData
       const PatchData& src,
       const BoxOverlap& overlap) = 0;
 
+   virtual void
+   copy(
+      const PatchData& src,
+      const BoxOverlap& overlap,
+      tbox::KernelFuser& fuser);
+
    /**
     * Copy data from the source into the destination using the designated
     * overlap descriptor.  The overlap description will have been computed
@@ -206,6 +221,19 @@ class PatchData
       tbox::MessageStream& stream,
       const BoxOverlap& overlap) const = 0;
 
+   /**
+    * Pack data lying on the specified index set into the output stream using
+    * the given KernelFuser. The default implementation of this method will
+    * call packStream without the fuser argument. See the abstract stream
+    * virtual base class for more information about the packing operators
+    * defined for streams.
+    */
+   virtual void
+   packStream(
+      tbox::MessageStream& stream,
+      const BoxOverlap& overlap,
+      tbox::KernelFuser& fuser);
+
    /**
     * Unpack data from the message stream into the specified index set.
     * See the abstract stream virtual base class for more information about
@@ -216,6 +244,19 @@ class PatchData
       tbox::MessageStream& stream,
       const BoxOverlap& overlap) = 0;
 
+   /**
+    * Unpack data from the message stream into the specified index set using
+    * the given KernelFuser. The default implementation of this method will
+    * call unpackStream without the fuser argument. See the abstract stream
+    * virtual base class for more information about the packing operators
+    * defined for streams.
+    */
+   virtual void
+   unpackStream(
+      tbox::MessageStream& stream,
+      const BoxOverlap& overlap,
+      tbox::KernelFuser& fuser);
+
    /**
     * Checks that class version and restart file version are equal.  If so,
     * reads in the data members common to all patch data types from restart

source/SAMRAI/pdat/CMakeLists.txt

@@ -339,6 +339,10 @@ target_include_directories(
   $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/source>
   $<INSTALL_INTERFACE:include>)
 
+blt_print_target_properties(TARGET SAMRAI_pdat)
+blt_print_target_properties(TARGET raja)
+blt_print_target_properties(TARGET RAJA)
+
 install(TARGETS SAMRAI_pdat
   EXPORT SAMRAITargets
   ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}

source/SAMRAI/tbox/CMakeLists.txt

@@ -58,6 +58,7 @@ set ( tbox_headers
   TimerManager.h
   Tracer.h
   Transaction.h
+  TransactionFuseable.h
   Utilities.h)
 
 set_source_files_properties(
@@ -115,6 +116,7 @@ set (tbox_sources
   TimerManager.C
   Tracer.C
   Transaction.C
+  TransactionFuseable.C
   Utilities.C)
 
 if (ENABLE_HDF5)
@@ -146,9 +148,11 @@ if (ENABLE_RAJA)
 endif ()
 
 if (ENABLE_CUDA)
-  set(cuda_sources Schedule.C)
+  set(cuda_sources TransactionFuseable.C Schedule.C)
   set_source_files_properties(${cuda_sources} PROPERTIES LANGUAGE CUDA)
 
+  set (tbox_depends ${tbox_depends} cuda)
+
   if (ENABLE_NVTX_REGIONS)
     find_package(CUDA REQUIRED)
 
@@ -176,6 +180,9 @@ target_include_directories( SAMRAI_tbox
   $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/source>
   $<INSTALL_INTERFACE:include>)
 
+blt_print_target_properties(
+  TARGET SAMRAI_tbox)
+
 
 install(TARGETS SAMRAI_tbox
   EXPORT SAMRAITargets

source/SAMRAI/tbox/ExecutionPolicy.h

@@ -112,6 +112,11 @@ struct policy_traits<policy::parallel> {
    >;
 
    using ReductionPolicy = RAJA::cuda_reduce;
+
+   using WorkGroupPolicy = RAJA::WorkGroupPolicy<
+      RAJA::cuda_work_async<1024>, 
+      RAJA::unordered_cuda_loop_y_block_iter_x_threadblock_average,
+      RAJA::constant_stride_array_of_objects>;
 };
 
 #else

source/SAMRAI/tbox/KernelFuser.h

@@ -0,0 +1,50 @@
+#ifndef included_tbox_KernelFuser
+#define included_tbox_KernelFuser
+
+#include "SAMRAI/tbox/ExecutionPolicy.h"
+#include "SAMRAI/tbox/AllocatorDatabase.h"
+
+// #include "RAJA/RAJA.hpp"
+
+namespace SAMRAI {
+namespace tbox {
+
+class KernelFuser
+{
+public:
+  // KernelFuser() :
+  //   d_workpool(AllocatorDatabase::getDatabase()->getStreamAllocator())
+  // {}
+
+  template<typename Kernel>
+  void enqueue(int begin, int end, Kernel&& kernel) {
+    //d_workpool.enqueue(RAJA::RangeSegment(begin, end), std::forward<Kernel>(kernel));
+  }
+
+  void launch()
+  {
+    // d_workgroup = d_workpool.instantiate();
+    // d_worksite = d_workgroup.run();
+  }
+
+private:
+#ifdef HAVE_UMPIRE
+  using Allocator = umpire::TypedAllocator<char>;
+#else
+  using Allocator = ResourceAllocator;
+#endif
+
+  // using Policy = typename tbox::detail::policy_traits< tbox::policy::parallel >::WorkGroupPolicy;
+  // using WorkPool  = RAJA::WorkPool <Policy, int, RAJA::xargs<>, Allocator>;
+  // using WorkGroup = RAJA::WorkGroup<Policy, int, RAJA::xargs<>, Allocator>;
+  // using WorkSite  = RAJA::WorkSite <Policy, int, RAJA::xargs<>, Allocator>;
+
+  // WorkPool d_workpool;
+  // WorkGroup d_workgroup;
+  // WorkSite d_worksite;
+};
+
+}
+}
+
+#endif