Backup

.github/workflows/osx.yml

@@ -103,6 +103,7 @@ jobs:
           -DKokkosKernels_INST_OFFSET_SIZE_T=ON \
           -DKokkosKernels_ENABLE_TPL_CUSPARSE=OFF \
           -DKokkosKernels_ENABLE_TPL_CUBLAS=OFF \
+          -DKokkosKernels_ENABLE_TPL_CUSOLVER=OFF \
           ..
 
       - name: build_kokkos_kernels

CMakeLists.txt

@@ -190,7 +190,7 @@ ELSE()
     "ALL"
     STRING
     "A list of components to enable in testing and building"
-    VALID_ENTRIES BATCHED BLAS LAPACK GRAPH SPARSE ALL
+    VALID_ENTRIES BATCHED BLAS LAPACK CUSOLVER ROCSOLVER GRAPH SPARSE ALL
   )
 
   # ==================================================================
@@ -242,6 +242,8 @@ ELSE()
   MESSAGE("   BATCHED:   ${KokkosKernels_ENABLE_COMPONENT_BATCHED}")
   MESSAGE("   BLAS:      ${KokkosKernels_ENABLE_COMPONENT_BLAS}")
   MESSAGE("   LAPACK:    ${KokkosKernels_ENABLE_COMPONENT_LAPACK}")
+  MESSAGE("   CUSOLVER:  ${KokkosKernels_ENABLE_COMPONENT_CUSOLVER}")
+  MESSAGE("   ROCSOLVER: ${KokkosKernels_ENABLE_COMPONENT_ROCSOLVER}")
   MESSAGE("   GRAPH:     ${KokkosKernels_ENABLE_COMPONENT_GRAPH}")
   MESSAGE("   SPARSE:    ${KokkosKernels_ENABLE_COMPONENT_SPARSE}")
   MESSAGE("   ODE:       ${KokkosKernels_ENABLE_COMPONENT_ODE}")
@@ -289,6 +291,12 @@ ELSE()
   IF (KokkosKernels_ENABLE_COMPONENT_LAPACK)
     INCLUDE(lapack/CMakeLists.txt)
   ENDIF()
+  IF (KokkosKernels_ENABLE_COMPONENT_CUSOLVER)
+    INCLUDE(lapack/CMakeLists.txt)
+  ENDIF()
+  IF (KokkosKernels_ENABLE_COMPONENT_ROCSOLVER)
+    INCLUDE(lapack/CMakeLists.txt)
+  ENDIF()
   IF (KokkosKernels_ENABLE_COMPONENT_GRAPH)
     INCLUDE(graph/CMakeLists.txt)
   ENDIF()
@@ -371,8 +379,10 @@ ELSE()
   KOKKOSKERNELS_LINK_TPL(kokkoskernels PUBLIC CHOLMOD)
   KOKKOSKERNELS_LINK_TPL(kokkoskernels PUBLIC MKL)
   KOKKOSKERNELS_LINK_TPL(kokkoskernels PUBLIC CUBLAS)
+  KOKKOSKERNELS_LINK_TPL(kokkoskernels PUBLIC CUSOLVER)
   KOKKOSKERNELS_LINK_TPL(kokkoskernels PUBLIC CUSPARSE)
   KOKKOSKERNELS_LINK_TPL(kokkoskernels PUBLIC ROCBLAS)
+  KOKKOSKERNELS_LINK_TPL(kokkoskernels PUBLIC ROCSOLVER)
   KOKKOSKERNELS_LINK_TPL(kokkoskernels PUBLIC ROCSPARSE)
   KOKKOSKERNELS_LINK_TPL(kokkoskernels PUBLIC METIS)
   KOKKOSKERNELS_LINK_TPL(kokkoskernels PUBLIC ARMPL)
@@ -410,6 +420,12 @@ ELSE()
   IF (KokkosKernels_ENABLE_COMPONENT_LAPACK)
     KOKKOSKERNELS_ADD_TEST_DIRECTORIES(lapack/unit_test)
   ENDIF()
+  IF (KokkosKernels_ENABLE_COMPONENT_CUSOLVER)
+    KOKKOSKERNELS_ADD_TEST_DIRECTORIES(lapack/unit_test)
+  ENDIF()
+  IF (KokkosKernels_ENABLE_COMPONENT_ROCSOLVER)
+    KOKKOSKERNELS_ADD_TEST_DIRECTORIES(lapack/unit_test)
+  ENDIF()
   IF (KokkosKernels_ENABLE_COMPONENT_GRAPH)
     KOKKOSKERNELS_ADD_TEST_DIRECTORIES(graph/unit_test)
   ENDIF()

batched/KokkosBatched_Util.hpp

@@ -21,6 +21,7 @@
 // no experimental name space guard for trilinos
 #define __KOKKOSBATCHED_PROMOTION__ 1
 
+#include <iostream>
 #include <iomanip>
 #include <random>
 #include <string>

batched/dense/impl/KokkosBatched_HostLevel_Gemm_Serial_Impl.hpp

@@ -16,6 +16,7 @@
 #ifndef __KOKKOSBATCHED_HOSTLEVEL_GEMM_SERIAL_IMPL_HPP__
 #define __KOKKOSBATCHED_HOSTLEVEL_GEMM_SERIAL_IMPL_HPP__
 #include "KokkosBatched_Gemm_Decl.hpp"
+#include <iostream>
 
 namespace KokkosBatched {
 namespace Impl {
@@ -181,4 +182,4 @@ class BatchedSerialGemm {
 };
 }  // namespace Impl
 }  // namespace KokkosBatched
-#endif
+#endif

cm_generate_makefile.bash

@@ -177,8 +177,10 @@ get_kernels_tpls_list() {
   KOKKOSKERNELS_USER_TPL_PATH_CMD=
   KOKKOSKERNELS_USER_TPL_LIBNAME_CMD=
   CUBLAS_DEFAULT=OFF
+  CUSOLVER_DEFAULT=OFF
   CUSPARSE_DEFAULT=OFF
   ROCBLAS_DEFAULT=OFF
+  ROCSOLVER_DEFAULT=OFF
   ROCSPARSE_DEFAULT=OFF
   PARSE_TPLS_LIST=$(echo $KOKKOSKERNELS_TPLS | tr "," "\n")
   for TPLS_ in $PARSE_TPLS_LIST
@@ -188,12 +190,18 @@ get_kernels_tpls_list() {
     if [ "$UC_TPLS" == "CUBLAS" ]; then
       CUBLAS_DEFAULT=ON
     fi
+    if [ "$UC_TPLS" == "CUSOLVER" ]; then
+      CUSOLVER_DEFAULT=ON
+    fi
     if [ "$UC_TPLS" == "CUSPARSE" ]; then
       CUSPARSE_DEFAULT=ON
     fi
     if [ "$UC_TPLS" == "ROCBLAS" ]; then
       ROCBLAS_DEFAULT=ON
     fi
+    if [ "$UC_TPLS" == "ROCSOLVER" ]; then
+      ROCSOLVER_DEFAULT=ON
+    fi
     if [ "$UC_TPLS" == "ROCSPARSE" ]; then
       ROCSPARSE_DEFAULT=ON
     fi
@@ -221,12 +229,18 @@ get_kernels_tpls_list() {
   if [ "$CUBLAS_DEFAULT" == "OFF" ]; then
     KOKKOSKERNELS_TPLS_CMD="-DKokkosKernels_ENABLE_TPL_CUBLAS=OFF ${KOKKOSKERNELS_TPLS_CMD}"
   fi
+  if [ "$CUSOLVER_DEFAULT" == "OFF" ]; then
+    KOKKOSKERNELS_TPLS_CMD="-DKokkosKernels_ENABLE_TPL_CUSOLVER=OFF ${KOKKOSKERNELS_TPLS_CMD}"
+  fi
   if [ "$CUSPARSE_DEFAULT" == "OFF" ]; then
     KOKKOSKERNELS_TPLS_CMD="-DKokkosKernels_ENABLE_TPL_CUSPARSE=OFF ${KOKKOSKERNELS_TPLS_CMD}"
   fi
   if [ "$ROCBLAS_DEFAULT" == "OFF" ]; then
     KOKKOSKERNELS_TPLS_CMD="-DKokkosKernels_ENABLE_TPL_ROCBLAS=OFF ${KOKKOSKERNELS_TPLS_CMD}"
   fi
+  if [ "$ROCSOLVER_DEFAULT" == "OFF" ]; then
+    KOKKOSKERNELS_TPLS_CMD="-DKokkosKernels_ENABLE_TPL_ROCSOLVER=OFF ${KOKKOSKERNELS_TPLS_CMD}"
+  fi
   if [ "$ROCSPARSE_DEFAULT" == "OFF" ]; then
     KOKKOSKERNELS_TPLS_CMD="-DKokkosKernels_ENABLE_TPL_ROCSPARSE=OFF ${KOKKOSKERNELS_TPLS_CMD}"
   fi
@@ -345,7 +359,7 @@ display_help_text() {
       echo "--with-tpls=[TPLS]:           Set tpls to be instantiated (Proper support requies that appropriate compiler and device must be enabled)."
       echo "                              This may require providing paths and the library name if using custom installs not on a default path"
       echo "                              that CMake searches"
-      echo "                                Options: blas, mkl, cublas, cusparse, magma, armpl, rocblas, rocsparse"
+      echo "                                Options: blas, mkl, cublas, cusolver, cusparse, magma, armpl, rocblas, rocsolver, rocsparse"
       echo "--user-blas-path=[PATH]:      Set path to location of user-specified BLAS library."
       echo "--user-blas-lib=[LIB]:        Library name of desired BLAS install."
       echo "                                Example: For the typical \"libblas.a\" provide \"blas\""

cmake/KokkosKernels_config.h.in

@@ -117,6 +117,8 @@
 #cmakedefine KOKKOSKERNELS_ENABLE_TPL_CUSPARSE
 /* CUBLAS */
 #cmakedefine KOKKOSKERNELS_ENABLE_TPL_CUBLAS
+/* CUSOLVER */
+#cmakedefine KOKKOSKERNELS_ENABLE_TPL_CUSOLVER
 /* MAGMA */
 #cmakedefine KOKKOSKERNELS_ENABLE_TPL_MAGMA
 /* SuperLU */
@@ -135,6 +137,8 @@
 #cmakedefine ARMPL_BUILD @ARMPL_BUILD@
 /* ROCBLAS */
 #cmakedefine KOKKOSKERNELS_ENABLE_TPL_ROCBLAS
+/* ROCSOLVER */
+#cmakedefine KOKKOSKERNELS_ENABLE_TPL_ROCSOLVER
 /* ROCSPARSE */
 #cmakedefine KOKKOSKERNELS_ENABLE_TPL_ROCSPARSE
 

cmake/Modules/FindTPLCUSOLVER.cmake

@@ -0,0 +1,18 @@
+FIND_PACKAGE(CUDA)
+
+INCLUDE(FindPackageHandleStandardArgs)
+IF (NOT CUDA_FOUND)
+  #Important note here: this find Module is named TPLCUSOLVER
+  #The eventual target is named CUSOLVER. To avoid naming conflicts
+  #the find module is called TPLCUSOLVER. This call will cause
+  #the find_package call to fail in a "standard" CMake way
+  FIND_PACKAGE_HANDLE_STANDARD_ARGS(TPLCUSOLVER REQUIRED_VARS CUDA_FOUND)
+ELSE()
+  #The libraries might be empty - OR they might explicitly be not found
+  IF("${CUDA_CUSOLVER_LIBRARIES}" MATCHES "NOTFOUND")
+    FIND_PACKAGE_HANDLE_STANDARD_ARGS(TPLCUSOLVER REQUIRED_VARS CUDA_CUSOLVER_LIBRARIES)
+  ELSE()
+    KOKKOSKERNELS_CREATE_IMPORTED_TPL(CUSOLVER INTERFACE
+      LINK_LIBRARIES "${CUDA_CUSOLVER_LIBRARIES}")
+  ENDIF()
+ENDIF()

cmake/Modules/FindTPLROCSOLVER.cmake

@@ -0,0 +1,12 @@
+FIND_PACKAGE(ROCSOLVER)
+if(TARGET roc::rocsolver)
+## MPL: 12/29/2022: Variable TPL_ROCSOLVER_IMPORTED_NAME follows the requested convention
+## of KokkosKernel (method kokkoskernels_import_tpl of kokkoskernels_tpls.cmake)
+  SET(TPL_ROCSOLVER_IMPORTED_NAME roc::rocsolver)
+  SET(TPL_IMPORTED_NAME roc::rocsolver)
+## MPL: 12/29/2022: A target comming from a TPL must follows the requested convention
+## of KokkosKernel (method kokkoskernels_link_tpl of kokkoskernels_tpls.cmake)
+  ADD_LIBRARY(KokkosKernels::ROCSOLVER ALIAS roc::rocsolver)
+ELSE()
+  MESSAGE(FATAL_ERROR "Package ROCSOLVER requested but not found")
+ENDIF()

cmake/kokkoskernels_tpls.cmake

@@ -456,27 +456,35 @@ ENDIF()
 
 KOKKOSKERNELS_ADD_OPTION(NO_DEFAULT_CUDA_TPLS OFF BOOL "Whether CUDA TPLs should be enabled by default. Default: OFF")
 SET(CUBLAS_DEFAULT   ${KOKKOS_ENABLE_CUDA})
+SET(CUSOLVER_DEFAULT ${KOKKOS_ENABLE_CUDA})
 SET(CUSPARSE_DEFAULT ${KOKKOS_ENABLE_CUDA})
 IF(KOKKOSKERNELS_NO_DEFAULT_CUDA_TPLS)
   SET(CUBLAS_DEFAULT   OFF)
+  SET(CUSOLVER_DEFAULT OFF)
   SET(CUSPARSE_DEFAULT OFF)
 ENDIF()
 KOKKOSKERNELS_ADD_TPL_OPTION(CUBLAS   ${CUBLAS_DEFAULT}     "Whether to enable CUBLAS"
   DEFAULT_DOCSTRING "ON if CUDA-enabled Kokkos, otherwise OFF")
+KOKKOSKERNELS_ADD_TPL_OPTION(CUSOLVER ${CUSOLVER_DEFAULT}   "Whether to enable CUSOLVER"
+  DEFAULT_DOCSTRING "ON if CUDA-enabled Kokkos, otherwise OFF")
 KOKKOSKERNELS_ADD_TPL_OPTION(CUSPARSE ${CUSPARSE_DEFAULT}   "Whether to enable CUSPARSE"
   DEFAULT_DOCSTRING "ON if CUDA-enabled Kokkos, otherwise OFF")
 
 KOKKOSKERNELS_ADD_OPTION(NO_DEFAULT_ROCM_TPLS OFF BOOL "Whether ROCM TPLs should be enabled by default. Default: OFF")
 # Unlike CUDA, ROCm does not automatically install these TPLs
 SET(ROCBLAS_DEFAULT   OFF)
+SET(ROCSOLVER_DEFAULT OFF)
 SET(ROCSPARSE_DEFAULT OFF)
 # Since the default is OFF we do not really need this piece of logic here.
 # IF(KOKKOSKERNELS_NO_DEFAULT_ROCM_TPLS)
 #   SET(ROCBLAS_DEFAULT   OFF)
+#   SET(ROCSOLVER_DEFAULT OFF)
 #   SET(ROCSPARSE_DEFAULT OFF)
 # ENDIF()
 KOKKOSKERNELS_ADD_TPL_OPTION(ROCBLAS   ${ROCBLAS_DEFAULT}     "Whether to enable ROCBLAS"
   DEFAULT_DOCSTRING "ON if HIP-enabled Kokkos, otherwise OFF")
+KOKKOSKERNELS_ADD_TPL_OPTION(ROCSOLVER ${ROCSOLVER_DEFAULT}   "Whether to enable ROCSOLVER"
+  DEFAULT_DOCSTRING "ON if HIP-enabled Kokkos, otherwise OFF")
 KOKKOSKERNELS_ADD_TPL_OPTION(ROCSPARSE ${ROCSPARSE_DEFAULT}   "Whether to enable ROCSPARSE"
   DEFAULT_DOCSTRING "ON if HIP-enabled Kokkos, otherwise OFF")
 
@@ -508,6 +516,7 @@ IF (NOT KOKKOSKERNELS_HAS_TRILINOS)
   KOKKOSKERNELS_IMPORT_TPL(LAPACK)
   KOKKOSKERNELS_IMPORT_TPL(MKL)
   KOKKOSKERNELS_IMPORT_TPL(CUBLAS)
+  KOKKOSKERNELS_IMPORT_TPL(CUSOLVER)
   KOKKOSKERNELS_IMPORT_TPL(CUSPARSE)
   KOKKOSKERNELS_IMPORT_TPL(CBLAS)
   KOKKOSKERNELS_IMPORT_TPL(LAPACKE)
@@ -517,6 +526,7 @@ IF (NOT KOKKOSKERNELS_HAS_TRILINOS)
   KOKKOSKERNELS_IMPORT_TPL(ARMPL)
   KOKKOSKERNELS_IMPORT_TPL(MAGMA)
   KOKKOSKERNELS_IMPORT_TPL(ROCBLAS)
+  KOKKOSKERNELS_IMPORT_TPL(ROCSOLVER)
   KOKKOSKERNELS_IMPORT_TPL(ROCSPARSE)
 ELSE ()
   IF (Trilinos_ENABLE_SuperLU5_API)

example/half/xpy.cpp

@@ -109,4 +109,4 @@ int main(int argc, char **argv) {
   do_xpy<Kokkos::Experimental::bhalf_t, DeviceType, LayoutType>(n, time_only);
   Kokkos::finalize();
   return 0;
-}
+}

lapack/CMakeLists.txt

@@ -8,11 +8,11 @@ LIST(APPEND KK_INCLUDE_DIRS ${CMAKE_CURRENT_SOURCE_DIR}/lapack/tpls)
 KOKKOSKERNELS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_BINARY_DIR}/lapack)
 KOKKOSKERNELS_INCLUDE_DIRECTORIES(REQUIRED_DURING_INSTALLATION_TESTING ${CMAKE_CURRENT_SOURCE_DIR}/lapack)
 
-#######################
-#                     #
+#########################
+#                       #
 # Logic for LAPACK TPLs #
-#                     #
-#######################
+#                       #
+#########################
 
 #Include LAPACK, Lapack host wrapper
 IF (KOKKOSKERNELS_ENABLE_TPL_LAPACK OR KOKKOSKERNELS_ENABLE_TPL_MKL OR KOKKOSKERNELS_ENABLE_TPL_ARMPL)
@@ -28,14 +28,14 @@ IF (KOKKOSKERNELS_ENABLE_TPL_LAPACK OR KOKKOSKERNELS_ENABLE_TPL_MKL OR KOKKOSKER
 ENDIF()
 
 # Include cuda lapack TPL source file
-IF (KOKKOSKERNELS_ENABLE_TPL_CULAPACK)
+IF (KOKKOSKERNELS_ENABLE_TPL_CUSOLVER)
   LIST(APPEND SOURCES
    lapack/tpls/KokkosLapack_Cuda_tpl.cpp
   )
 ENDIF()
 
 # Include rocm lapack TPL source file
-IF (KOKKOSKERNELS_ENABLE_TPL_ROCLAPACK)
+IF (KOKKOSKERNELS_ENABLE_TPL_ROCSOLVER)
   LIST(APPEND SOURCES
    lapack/tpls/KokkosLapack_Rocm_tpl.cpp
   )