jianmu
/
lammps
forked from lijiext/lammps
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Add includes and linked libs for HIP_PLATFORM=nvcc 

Do not set CMAKE_CXX_COMPILER=hipcc in the NVCC case since it causes
all sorts of problems. Instead we use the regular host compiler and
set the necessary include and link paths.
This commit is contained in:
Richard Berger 2020-04-10 12:45:39 -04:00
parent 6381940b9a
commit b7a2b375c6
1 changed files with 14 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
cmake/Modules/Packages/GPU.cmake
View File

@ -214,9 +214,8 @@ elseif(GPU_API STREQUAL "HIP")
find_package(CUDA REQUIRED)
set(HIP_ARCH "sm_30" CACHE STRING "HIP primary CUDA architecture (e.g. sm_60)")
# build arch/gencode commands for nvcc based on CUDA toolkit version and use choice
# --arch translates directly instead of JIT, so this should be for the preferred or most common architecture
set(HIP_CUDA_GENCODE "-arch=${HIP_ARCH} ")
# build gencode commands for nvcc based on CUDA toolkit version and use choice
set(HIP_CUDA_GENCODE "")
# Fermi (GPU Arch 2.x) is supported by CUDA 3.2 to CUDA 8.0
if((CUDA_VERSION VERSION_GREATER "3.1") AND (CUDA_VERSION VERSION_LESS "9.0"))
string(APPEND HIP_CUDA_GENCODE "-gencode arch=compute_20,code=[sm_20,compute_20] ")
@ -287,6 +286,18 @@ elseif(GPU_API STREQUAL "HIP")
hip_add_executable(hip_get_devices ${LAMMPS_LIB_SOURCE_DIR}/gpu/geryon/ucl_get_devices.cpp)
target_compile_definitions(hip_get_devices PRIVATE -DUCL_HIP)
if(HIP_PLATFORM STREQUAL "nvcc")
target_compile_definitions(gpu PRIVATE -D__HIP_PLATFORM_NVCC__)
target_include_directories(gpu PRIVATE ${HIP_ROOT_DIR}/include)
target_include_directories(gpu PRIVATE ${CUDA_INCLUDE_DIRS})
target_link_libraries(gpu PRIVATE ${CUDA_LIBRARIES} ${CUDA_CUDA_LIBRARY})
target_compile_definitions(hip_get_devices PRIVATE -D__HIP_PLATFORM_NVCC__)
target_include_directories(hip_get_devices PRIVATE ${HIP_ROOT_DIR}/include)
target_include_directories(hip_get_devices PRIVATE ${CUDA_INCLUDE_DIRS})
target_link_libraries(hip_get_devices PRIVATE ${CUDA_LIBRARIES} ${CUDA_CUDA_LIBRARY})
endif()
target_link_libraries(lammps PRIVATE gpu)
endif()