Cusolver (#653)

This PR adds the option of using cuSolver.

Co-authored-by: Paul Mullowney <Paul.Mullowney@nrel.gov>

src/CMakeLists.txt

@@ -108,6 +108,7 @@ option(HYPRE_ENABLE_UNIFIED_MEMORY   "Use unified memory for allocating the memo
 # CUDA options
 option(HYPRE_ENABLE_CUDA_STREAMS     "Use CUDA streams" ON)
 option(HYPRE_ENABLE_CUSPARSE         "Use cuSPARSE" ON)
+option(HYPRE_ENABLE_CUSOLVER         "Use cuSOLVER" OFF)
 option(HYPRE_ENABLE_DEVICE_POOL      "Use device memory pool" OFF)
 option(HYPRE_ENABLE_CUBLAS           "Use cuBLAS" OFF)
 option(HYPRE_ENABLE_CURAND           "Use cuRAND" ON)

src/config/HYPRE_config.h.cmake.in

@@ -82,6 +82,9 @@
 /* Use cuSPARSE */
 #cmakedefine HYPRE_USING_CUSPARSE 1
 
+/* Use cuSOLVER */
+#cmakedefine HYPRE_USING_CUSOLVER 1
+
 /* Use device memory pool */
 #cmakedefine HYPRE_USING_DEVICE_POOL 1
 

src/config/HYPRE_config.h.in

@@ -172,6 +172,9 @@
 /* Define to 1 if using cuSPARSE */
 #undef HYPRE_USING_CUSPARSE
 
+/* Define to 1 if using cuSOLVER */
+#undef HYPRE_USING_CUSOLVER
+
 /* Define to 1 if using device async malloc */
 #undef HYPRE_USING_DEVICE_MALLOC_ASYNC
 

src/config/cmake/HYPRE_SetupCUDAToolkit.cmake

@@ -50,6 +50,16 @@ if (CMAKE_VERSION VERSION_LESS 3.17)
     endif (HYPRE_SHARED)
   endif (HYPRE_ENABLE_CUBLAS)
 
+  if (HYPRE_ENABLE_CUSOLVER)
+    set(HYPRE_USING_CUSOLVER ON CACHE BOOL "" FORCE)
+    if (HYPRE_SHARED)
+      list(APPEND EXPORT_INTERFACE_CUDA_LIBS ${CUDA_cusolver_LIBRARY})
+    else ()
+      list(APPEND EXPORT_INTERFACE_CUDA_LIBS
+        ${CUDA_TOOLKIT_ROOT_DIR}/lib64/libcusolver_static.a)
+    endif ()
+  endif ()
+
   if (NOT HYPRE_SHARED)
     list(APPEND EXPORT_INTERFACE_CUDA_LIBS
       ${CUDA_TOOLKIT_ROOT_DIR}/lib64/libculibos.a)
@@ -101,6 +111,15 @@ else()
     endif (HYPRE_SHARED)
   endif (HYPRE_ENABLE_CUBLAS)
 
+  if (HYPRE_ENABLE_CUSOLVER)
+    set(HYPRE_USING_CUSOLVER ON CACHE BOOL "" FORCE)
+    if (HYPRE_CUDA_TOOLKIT_STATIC)
+      list(APPEND EXPORT_INTERFACE_CUDA_LIBS CUDA::cusolver_static)
+    else ()
+      list(APPEND EXPORT_INTERFACE_CUDA_LIBS CUDA::cusolver)
+    endif ()
+  endif ()
+
   if (HYPRE_CUDA_TOOLKIT_STATIC)
     list(APPEND EXPORT_INTERFACE_CUDA_LIBS CUDA::culibos)
   endif (HYPRE_CUDA_TOOLKIT_STATIC)

src/config/configure.in

@@ -175,6 +175,7 @@ hypre_using_cuda_streams=no
 hypre_using_cusparse=yes
 hypre_using_cublas=yes
 hypre_using_curand=yes
+hypre_using_cusolver=no
 hypre_using_device_pool=no
 hypre_using_device_malloc_async=no
 hypre_using_umpire=no
@@ -432,6 +433,17 @@ AS_HELP_STRING([--enable-cusparse],
 [hypre_using_cusparse=yes]
 )
 
+AC_ARG_ENABLE(cusolver,
+AS_HELP_STRING([--enable-cusolver],
+               [Use cusolver (default is NO).]),
+[case "${enableval}" in
+    yes) hypre_using_cusolver=yes ;;
+    no)  hypre_using_cusolver=no ;;
+    *)   hypre_using_cusolver=yes ;;
+ esac],
+[hypre_using_cusolver=no]
+)
+
 AC_ARG_ENABLE(device-memory-pool,
 AS_HELP_STRING([--enable-device-memory-pool],
                [Use device pooling allocator (default is NO).]),
@@ -2270,6 +2282,12 @@ then
       HYPRE_CUDA_LIBS+=" -lcurand"
    fi
 
+   if test "$hypre_using_cusolver" = "yes"
+   then
+      AC_DEFINE(HYPRE_USING_CUSOLVER, 1, [Define to 1 if using cuSolver])
+      HYPRE_CUDA_LIBS+=" -lcusolver"
+   fi
+
    if test "$hypre_using_device_pool" = "yes"
    then
       AC_DEFINE(HYPRE_USING_DEVICE_POOL, 1, [Define to 1 if using device pooling allocator])

src/configure

@@ -784,6 +784,7 @@ enable_fortran
 enable_unified_memory
 enable_cuda_streams
 enable_cusparse
+enable_cusolver
 enable_device_memory_pool
 enable_device_malloc_async
 enable_cublas
@@ -1517,6 +1518,7 @@ Optional Features:
                           (default is NO).
   --enable-cuda-streams   Use CUDA streams (default is YES).
   --enable-cusparse       Use cuSPARSE (default is YES).
+  --enable-cusolver       Use cuSOLVER (default is NO).
   --enable-device-memory-pool
                           Use device pooling allocator (default is NO).
   --enable-device-malloc-async
@@ -2787,6 +2789,7 @@ hypre_using_cuda_streams=no
 hypre_using_cusparse=yes
 hypre_using_cublas=yes
 hypre_using_curand=yes
+hypre_using_cusolver=no
 hypre_using_device_pool=no
 hypre_using_device_malloc_async=no
 hypre_using_umpire=no
@@ -3155,6 +3158,19 @@ else
 fi
 
 
+# Check whether --enable-cusolver was given.
+if test "${enable_cusolver+set}" = set; then :
+  enableval=$enable_cusolver; case "${enableval}" in
+    yes) hypre_using_cusolver=yes ;;
+    no)  hypre_using_cusolver=no ;;
+    *)   hypre_using_cusolver=yes ;;
+ esac
+else
+  hypre_using_cusolver=no
+
+fi
+
+
 # Check whether --enable-device-malloc-async was given.
 if test "${enable_device_malloc_async+set}" = set; then :
   enableval=$enable_device_malloc_async; case "${enableval}" in
@@ -9195,6 +9211,14 @@ $as_echo "#define HYPRE_USING_CURAND 1" >>confdefs.h
       HYPRE_CUDA_LIBS+=" -lcurand"
    fi
 
+   if test "$hypre_using_cusolver" = "yes"
+   then
+
+$as_echo "#define HYPRE_USING_CUSOLVER 1" >>confdefs.h
+
+       HYPRE_CUDA_LIBS+=" -lcusolver"
+   fi
+
    if test "$hypre_using_device_pool" = "yes"
    then