Adding parameter to install.sh to build for cuda backend

.gitignore

@@ -27,3 +27,6 @@
 *.exe
 *.out
 *.app
+
+# Tool meta-data
+.vscode

CMakeLists.txt

@@ -60,11 +60,13 @@ option( BUILD_SHARED_LIBS "Build hipBLAS as a shared library" ON )
 
 # Find HCC/HIP dependencies
 if( CMAKE_CXX_COMPILER MATCHES ".*/hcc$" )
+  message( STATUS "Building for ROCm backend" )
   find_package( hcc REQUIRED CONFIG PATHS /opt/rocm )
 elseif( CMAKE_CXX_COMPILER MATCHES ".*/hipcc$" )
+  message( STATUS "Building for ROCm backend" )
   find_package( hip REQUIRED CONFIG PATHS /opt/rocm )
 else( )
-  message( AUTHOR_WARNING "hipblas library probably won't build without either the hcc or hipcc compilers. Set the CXX environment variable to specify your desired compiler" )
+  message( STATUS "Building for CUDA backend" )
 endif( )
 
 # CMake list of machine targets

clients/gtest/CMakeLists.txt

@@ -30,7 +30,7 @@ if( NOT TARGET hipblas )
   find_package( hipblas CONFIG PATHS /opt/rocm/hipblas )
 
   if( NOT hipblas_FOUND )
-    message( FATAL_ERROR "rocBLAS is a required dependency and is not found; try adding hipblas path to CMAKE_PREFIX_PATH")
+    message( FATAL_ERROR "hipBLAS is a required dependency and is not found; try adding hipblas path to CMAKE_PREFIX_PATH")
   endif( )
 endif( )
 
@@ -57,7 +57,18 @@ set( hipblas_benchmark_common
   ../common/hipblas_template_specialization.cpp
 )
 
-add_executable( hipblas-test ${hipblas_test_source} ${hipblas_benchmark_common} )
+if( CMAKE_CXX_COMPILER MATCHES ".*/hcc$" OR CMAKE_CXX_COMPILER MATCHES ".*/hipcc$" )
+  # ROCm path
+  add_executable( hipblas-test ${hipblas_test_source} ${hipblas_benchmark_common} )
+else( )
+  # Nvidia path
+  # If HiP is not installed in default path, user can append to CMAKE_MODULE_PATH
+  list( APPEND CMAKE_MODULE_PATH /opt/rocm/hip/cmake )
+  find_package( HIP REQUIRED ) # Using module mode;
+  find_package( CUDA REQUIRED ) # Using module mode; packaged in cmake
+
+  add_executable( hipblas-test ${hipblas_test_source} ${hipblas_benchmark_common} )
+endif( )
 
 set( THREADS_PREFER_PTHREAD_FLAG ON )
 find_package( Threads REQUIRED )
@@ -89,14 +100,16 @@ endif( )
 
 target_link_libraries( hipblas-test PRIVATE cblas lapack ${GTEST_LIBRARIES} ${Boost_LIBRARIES} )
 
-if( CMAKE_CXX_COMPILER MATCHES ".*/hcc$" )
+if( CMAKE_CXX_COMPILER MATCHES ".*/hcc$" OR CMAKE_CXX_COMPILER MATCHES ".*/hipcc$" )
   # Remove following when hcc is fixed; hcc emits following spurious warning ROCm v1.6.1
   # "clang-5.0: warning: argument unused during compilation: '-isystem /opt/rocm/include'"
   target_compile_options( hipblas-test PRIVATE -Wno-unused-command-line-argument )
 
   foreach( target ${AMDGPU_TARGETS} )
     target_link_libraries( hipblas-test PRIVATE --amdgpu-target=${target} )
   endforeach( )
+else( )
+  target_link_libraries( hipblas-test PRIVATE ${CUDA_LIBRARIES} )
 endif( )
 
 set_target_properties( hipblas-test PROPERTIES DEBUG_POSTFIX "-d" CXX_EXTENSIONS NO )

install.sh

@@ -38,6 +38,7 @@ function display_help()
   echo "    [-i|--install] install after build"
   echo "    [-d|--dependencies] install build dependencies"
   echo "    [-c|--clients] build library clients too (combines with -i & -d)"
+  echo "    [--cuda] build library for cuda backend"
 }
 
 # #################################################
@@ -46,6 +47,7 @@ function display_help()
 install_package=false
 install_dependencies=false
 build_clients=false
+build_cuda=false
 
 # #################################################
 # Parameter parsing
@@ -54,7 +56,7 @@ build_clients=false
 # check if we have a modern version of getopt that can handle whitespace and long parameters
 getopt -T
 if [[ $? -eq 4 ]]; then
-  GETOPT_PARSE=$(getopt --name "${0}" --longoptions help,install,clients,dependencies: --options hicd -- "$@")
+  GETOPT_PARSE=$(getopt --name "${0}" --longoptions help,install,clients,dependencies,cuda --options hicd -- "$@")
 else
   echo "Need a new version of getopt"
   exit 1
@@ -82,6 +84,9 @@ while true; do
     -c|--clients)
         build_clients=true
         shift ;;
+    --cuda)
+        build_cuda=true
+        shift ;;
     --) shift ; break ;;
     *)  echo "Unexpected command line parameter received; aborting";
         exit 1
@@ -103,7 +108,14 @@ rm -rf ${build_dir}
 # #################################################
 if [[ "${install_dependencies}" == true ]]; then
   # dependencies needed for hipblas and clients to build
-  library_dependencies_ubuntu=( "make" "cmake-curses-gui" "hip_hcc" "hcc" "pkg-config" )
+  library_dependencies_ubuntu=( "make" "cmake-curses-gui" "pkg-config" "hip_hcc" )
+  if [[ "${build_cuda}" == false ]]; then
+    library_dependencies_ubuntu+=( "hcc" "rocblas" )
+  else
+    # Ideally, this could be cuda-cublas-dev, but the package name has a version number in it
+    library_dependencies_ubuntu+=( "cuda" )
+  fi
+
   client_dependencies_ubuntu=( "gfortran" "libboost-program-options-dev" )
 
   sudo apt update
@@ -119,7 +131,7 @@ if [[ "${install_dependencies}" == true ]]; then
   # Dependencies required by library client apps
   if [[ "${build_clients}" == true ]]; then
     for package in "${client_dependencies_ubuntu[@]}"; do
-    if [[ $(dpkg-query --show --showformat='${db:Status-Abbrev}\n' ${package} 2> /dev/null | grep -q "ii"; echo $?) -ne 0 ]]; then
+      if [[ $(dpkg-query --show --showformat='${db:Status-Abbrev}\n' ${package} 2> /dev/null | grep -q "ii"; echo $?) -ne 0 ]]; then
         printf "\033[32mInstalling \033[33m${package}\033[32m from distro package manager\033[0m\n"
         sudo apt install -y --no-install-recommends ${package}
       fi
@@ -142,7 +154,11 @@ pushd .
 # configure
 # #################################################
   mkdir -p ${build_dir}/release && cd ${build_dir}/release
-  export CXX=/opt/rocm/bin/hcc
+
+  if [[ "${build_cuda}" == false ]]; then
+    export CXX=/opt/rocm/bin/hcc
+  fi
+
   if [[ "${build_clients}" == true ]]; then
     cmake -DBUILD_CLIENTS_TESTS=ON ../..
   else

library/src/CMakeLists.txt

@@ -66,12 +66,15 @@ else( )
 
   target_link_libraries( hipblas PRIVATE ${CUDA_CUBLAS_LIBRARIES} )
 
-  target_include_directories( hipblas PRIVATE
-    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include>
-  $<BUILD_INTERFACE:${HIP_PATH}/include>
-  $<BUILD_INTERFACE:${CUDA_INCLUDE_DIRS}> )
-
-  target_compile_definitions( hipblas PRIVATE __HIP_PLATFORM_NVCC__ )
+  target_include_directories( hipblas
+    PRIVATE
+      $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include>
+    PUBLIC
+      $<BUILD_INTERFACE:${CUDA_INCLUDE_DIRS}>
+      $<BUILD_INTERFACE:${hip_INCLUDE_DIRS}>
+    )
+
+  target_compile_definitions( hipblas PUBLIC __HIP_PLATFORM_NVCC__ )
   set_target_properties( hipblas PROPERTIES DEBUG_POSTFIX "-d" OUTPUT_NAME hipblas-nvcc )
 endif( )