Fix build issues for V100 target

Author: fluidnumerics-joe

CMakeLists.txt

@@ -28,10 +28,9 @@ cmake_minimum_required(VERSION 3.21)
 cmake_policy(VERSION 3.21...3.27)
 
 # C Language is needed in order to verify Fortran compiler is C-interoperable
-# HIP language is needed to properly find "hip" package
 project(SELF VERSION 1.0.0
 	DESCRIPTION "Spectral Element Library in Fortran"
-	LANGUAGES Fortran C HIP)
+	LANGUAGES Fortran C)
 
 option(SELF_ENABLE_MULTITHREADING "Option to enable CPU multithreading for `do concurrent` loop blocks."  OFF)
 option(SELF_ENABLE_TESTING "Option to enable build of tests. (Default On)"  ON)
@@ -47,13 +46,6 @@ if(SELF_ENABLE_MULTITHREADING)
     set(SELF_MULITHREADING_NTHREADS "4" CACHE STRING "Number of threads to use for `do concurrent` loop blocks. This option is only used with GNU compilers. Other compilers use OMP_NUM_THREADS environment variable at runtime.")
 endif()
 
-if(NOT DEFINED ROCM_PATH)
-    if(NOT DEFINED ENV{ROCM_PATH})
-        set(ROCM_PATH "/opt/rocm/" CACHE PATH "Path to which ROCm has been installed")
-    else()
-        set(ROCM_PATH $ENV{ROCM_PATH} CACHE PATH "Path to which ROCm has been installed")
-    endif()
-endif()
 
 
 # Fortran compiler requirements
@@ -202,25 +194,29 @@ if(SELF_ENABLE_GPU)
             message( FATAL_ERROR "MPI installation is not GPU-aware" )
         endif()
     else()
-        # CUDA (Optional)
-        find_package(cuda)
-        if(cuda_FOUND)
-            if(MPI_HAS_QUERY_CUDA_SUPPORT)
-                message("-- CUDA found. Enabling CUDA language.")
-                enable_language(CUDA)
-                set( CMAKE_Fortran_FLAGS "${CMAKE_Fortran_FLAGS} -DHAVE_CUDA" )
-                set( CMAKE_Fortran_FLAGS_DEBUG "${CMAKE_Fortran_FLAGS_DEBUG} -DHAVE_CUDA" )
-                set( CMAKE_Fortran_FLAGS_COVERAGE "${CMAKE_Fortran_FLAGS_COVERAGE} -DHAVE_CUDA")
-                set( CMAKE_Fortran_FLAGS_PROFILE "${CMAKE_Fortran_FLAGS_PROFILE} -DHAVE_CUDA")
-                set( CMAKE_Fortran_FLAGS_RELEASE "${CMAKE_Fortran_FLAGS_RELEASE} -DHAVE_CUDA" )
-            
-    
-                # TO DO - need cuda libraries and hipblas libraries
-            else()
-                message( FATAL_ERROR "MPI installation is not GPU-aware" )
+        # CUDA
+        find_package(CUDAToolkit REQUIRED)
+        message("-- CUDA found. Enabling CUDA language.")
+        enable_language(CUDA)
+        if(MPI_HAS_QUERY_CUDA_SUPPORT)
+            set( CMAKE_Fortran_FLAGS "${CMAKE_Fortran_FLAGS} -DHAVE_CUDA" )
+            set( CMAKE_Fortran_FLAGS_DEBUG "${CMAKE_Fortran_FLAGS_DEBUG} -DHAVE_CUDA" )
+            set( CMAKE_Fortran_FLAGS_COVERAGE "${CMAKE_Fortran_FLAGS_COVERAGE} -DHAVE_CUDA")
+            set( CMAKE_Fortran_FLAGS_PROFILE "${CMAKE_Fortran_FLAGS_PROFILE} -DHAVE_CUDA")
+            set( CMAKE_Fortran_FLAGS_RELEASE "${CMAKE_Fortran_FLAGS_RELEASE} -DHAVE_CUDA" )
+
+            if(SELF_ENABLE_DOUBLE_PRECISION)
+                set( CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -DDOUBLE_PRECISION" )
+                set( CMAKE_CUDA_FLAGS_DEBUG "${CMAKE_CUDA_FLAGS_DEBUG} -DDOUBLE_PRECISION" )
+                set( CMAKE_CUDA_FLAGS_COVERAGE "${CMAKE_CUDA_FLAGS_COVERAGE} -DDOUBLE_PRECISION")
+                set( CMAKE_CUDA_FLAGS_PROFILE "${CMAKE_CUDA_FLAGS_PROFILE} -DDOUBLE_PRECISION")
+                set( CMAKE_CUDA_FLAGS_RELEASE "${CMAKE_CUDA_FLAGS_RELEASE} -DDOUBLE_PRECISION" )
             endif()
+
+            set( BACKEND_LIBRARIES CUDA::cuda_driver CUDA::cudart CUDA::cublas)
+
         else()
-            message( FATAL_ERROR "Enabling GPU support requires either HIP or CUDA." )
+            message( FATAL_ERROR "MPI installation is not GPU-aware" )
         endif()
     endif()
 endif()

src/CMakeLists.txt

@@ -29,16 +29,10 @@ file(GLOB SELF_FSRC "${CMAKE_CURRENT_SOURCE_DIR}/*.f*")
 if(SELF_ENABLE_GPU)
     file(GLOB SELF_BACKEND_FSRC "${CMAKE_CURRENT_SOURCE_DIR}/gpu/*.f*")
     file(GLOB SELF_BACKEND_CPPSRC "${CMAKE_CURRENT_SOURCE_DIR}/gpu/*.cpp*")
-    # Note : [email protected] (Oct. 1 2024)
-    # Ultimately, we want to be able to use the language support for HIP/CUDA
-    # rather than bringing in HIP/CUDA through `find_package`. At the moment
-    # we are doing a hack overrided the CXX compiler with either hipcc or nvcc
-    # The reason we're doing it this way (hacky) at the moment is that we get
-    # segmentation faults on our AMD GPU tests when using the HIP language support
-    # via Cmake, for some yet unknown reason.
+
     if(hip_FOUND)
       set_source_files_properties(${SELF_BACKEND_CPPSRC} PROPERTIES LANGUAGE HIP)
-    elseif(cuda_FOUND)
+    else()
       set_source_files_properties(${SELF_BACKEND_CPPSRC} PROPERTIES LANGUAGE CUDA)
     endif()
 else()

src/gpu/SELF_GPU_Macros.h

@@ -31,6 +31,7 @@ static void check(const hipError_t err, const char *const file, const int line)
 #else
 
 #include <cuda_runtime.h>
+#include <stdint.h> // required to provide uint32_t
 
 #define hipLaunchKernelGGL(F,G,B,M,S,...) F<<<G,B,M,S>>>(__VA_ARGS__)