Merge branch 'develop' into option_for_vec

CHANGELOG.md

@@ -4,6 +4,7 @@ All notable changes to this project will be documented in this file.
 ## Unreleased
 ### Added
 - Ability to build Kokkos and RAJA versions against existing packages.
+- Thrust managed memory.
 
 ### Changed
 - RAJA CUDA CMake build issues resolved.

src/raja/model.cmake

@@ -8,6 +8,8 @@ register_flag_optional(RAJA_IN_TREE
          Make sure to use the release version of RAJA or clone RAJA recursively with submodules.
          Remember to append RAJA specific flags as well, for example:
              -DRAJA_IN_TREE=... -DENABLE_OPENMP=ON -DENABLE_CUDA=ON ...
+         For RAJA >= v2022.03.0, remember to use the RAJA prefixed CMake options:
+             -DRAJA_IN_TREE=... -DRAJA_ENABLE_OPENMP=ON -DRAJA_ENABLE_CUDA=ON ...
          See https://github.com/LLNL/RAJA/blob/08cbbafd2d21589ebf341f7275c229412d0fe903/CMakeLists.txt#L44 for all available options
 " "")
 
@@ -20,7 +22,7 @@ register_flag_optional(TARGET
         CPU)
 
 register_flag_optional(CUDA_TOOLKIT_ROOT_DIR
-        "[TARGET==NVIDIA only] Path to the CUDA toolkit directory (e.g `/opt/cuda-11.2`) if the ENABLE_CUDA flag is specified for RAJA" "")
+        "[TARGET==NVIDIA only] Path to the CUDA toolkit directory (e.g `/opt/cuda-11.2`) if the RAJA_ENABLE_CUDA or ENABLE_CUDA flag is specified for RAJA" "")
 
 # XXX CMake 3.18 supports CMAKE_CUDA_ARCHITECTURES/CUDA_ARCHITECTURES but we support older CMakes
 register_flag_optional(CUDA_ARCH
@@ -58,7 +60,20 @@ macro(setup)
         set(ENABLE_BENCHMARKS OFF CACHE BOOL "")
         set(ENABLE_CUDA ${ENABLE_CUDA} CACHE BOOL "" FORCE)
 
-        if (ENABLE_CUDA)
+        # RAJA >= v2022.03.0 switched to prefixed variables, we keep the legacy ones for backwards compatibiity
+        set(RAJA_ENABLE_TESTS OFF CACHE BOOL "")
+        set(RAJA_ENABLE_EXAMPLES OFF CACHE BOOL "")
+        set(RAJA_ENABLE_REPRODUCERS OFF CACHE BOOL "")
+        set(RAJA_ENABLE_EXERCISES OFF CACHE BOOL "")
+        set(RAJA_ENABLE_DOCUMENTATION OFF CACHE BOOL "")
+        set(RAJA_ENABLE_BENCHMARKS OFF CACHE BOOL "")
+        set(RAJA_ENABLE_CUDA ${RAJA_ENABLE_CUDA} CACHE BOOL "" FORCE)
+
+        if (ENABLE_CUDA OR RAJA_ENABLE_CUDA)
+
+            # RAJA still needs ENABLE_CUDA for internal use, so if either is on, assert both.
+            set(RAJA_ENABLE_CUDA ON)
+            set(ENABLE_CUDA ON)
 
             # XXX CMake 3.18 supports CMAKE_CUDA_ARCHITECTURES/CUDA_ARCHITECTURES but we support older CMakes
             if(POLICY CMP0104)
@@ -70,6 +85,10 @@ macro(setup)
             set(CMAKE_CUDA_FLAGS ${CMAKE_CUDA_FLAGS} "-forward-unknown-to-host-compiler -extended-lambda -arch=${CUDA_ARCH}" ${CUDA_EXTRA_FLAGS})
             list(APPEND CMAKE_CUDA_FLAGS)
 
+            # See https://github.com/LLNL/RAJA/pull/1302
+            # And https://github.com/LLNL/RAJA/pull/1339
+            set(RAJA_ENABLE_VECTORIZATION OFF CACHE BOOL "")
+
             message(STATUS "NVCC flags: ${CMAKE_CUDA_FLAGS}")
         endif ()
 

src/thrust/ThrustStream.h

@@ -8,7 +8,11 @@
 
 #include <iostream>
 #include <vector>
+#if defined(MANAGED)
+#include <thrust/universal_vector.h>
+#else
 #include <thrust/device_vector.h>
+#endif
 
 #include "Stream.h"
 
@@ -21,9 +25,15 @@ class ThrustStream : public Stream<T>
     // Size of arrays
     int array_size;
 
+  #if defined(MANAGED)
+    thrust::universtal_vector<T> a;
+    thrust::universtal_vector<T> b;
+    thrust::universtal_vector<T> c;
+  #else
     thrust::device_vector<T> a;
     thrust::device_vector<T> b;
     thrust::device_vector<T> c;
+  #endif
 
   public:
     ThrustStream(const int, int);

src/thrust/model.cmake

@@ -18,6 +18,9 @@ register_flag_optional(BACKEND
         "
         "CUDA")
 
+      register_flag_optional(MANAGED "Enabled managed memory mode."
+        "OFF")
+
 register_flag_optional(CMAKE_CUDA_COMPILER
         "[THRUST_IMPL==CUDA] Path to the CUDA nvcc compiler"
         "")
@@ -34,6 +37,9 @@ register_flag_optional(CUDA_EXTRA_FLAGS
 
 macro(setup)
     set(CMAKE_CXX_STANDARD 14)
+    if (MANAGED)
+      register_definitions(MANAGED)
+    endif ()
 
     if (${THRUST_IMPL} STREQUAL "CUDA")
 
@@ -91,4 +97,4 @@ macro(setup)
 endmacro()
 
 
- 
+