getting started with HIP support

CMakeLists.txt

@@ -130,10 +130,8 @@ add_feature_info(TBB ENABLE_TBB "Intel Thread-Building Blocks (TBB) supports pro
 option(ENABLE_CUDA "Enable use of CUDA with TiledArray" OFF)
 add_feature_info(CUDA ENABLE_CUDA "NVIDIA CUDA support for GPU")
 
-if(ENABLE_CUDA)
-  option(ENABLE_CUDA_ERROR_CHECK "TiledArray will always check errors in CUDA calls" ON)
-  add_feature_info(CUDA_ERROR_CHECK ENABLE_CUDA_ERROR_CHECK "Checks CUDA Error")
-endif()
+option(ENABLE_HIP "Enable use of HIP with TiledArray" OFF)
+add_feature_info(HIP ENABLE_HIP "AMD HIP/ROCm support for GPU")
 
 option(ENABLE_GPERFTOOLS "Enable linking with Gperftools" OFF)
 add_feature_info(GPERFTOOLS ENABLE_GPERFTOOLS "Google Performance Tools provide fast memory allocation and performance profiling")
@@ -306,10 +304,13 @@ include_directories(${PROJECT_SOURCE_DIR}/src ${PROJECT_BINARY_DIR}/src)
 add_custom_target(External-tiledarray)
 
 # required deps:
-# 1. CUDA first since others may depend on it
+# 1. derive runtime (CUDA/HIP/...) first since others may depend on it
 if(ENABLE_CUDA)
   include(external/cuda.cmake)
 endif()
+if(ENABLE_HIP)
+  include(external/hip.cmake)
+endif()
 if (TA_TTG)
   include(${PROJECT_SOURCE_DIR}/cmake/modules/FindOrFetchTTG.cmake)
 endif(TA_TTG)

INSTALL.md

@@ -329,7 +329,7 @@ Support for execution on CUDA-enabled hardware is controlled by the following va
 
 * `ENABLE_CUDA`  -- Set to `ON` to turn on CUDA support. [Default=OFF].
 * `CMAKE_CUDA_HOST_COMPILER`  -- Set to the path to the host C++ compiler to be used by CUDA compiler. CUDA compilers used to be notorious for only being able to use specific C++ host compilers, but support for more recent C++ host compilers has improved. The default is determined by the CUDA compiler and the user environment variables (`PATH` etc.).
-* `ENABLE_CUDA_ERROR_CHECK` -- Set to `ON` to turn on assertions for successful completion of calls to CUDA runtime and libraries. [Default=OFF].
+* `ENABLE_HIP`  -- Set to `ON` to turn on HIP support. [Default=OFF].
 * `LIBRETT_INSTALL_DIR` -- the installation prefix of the pre-installed LibreTT library. This should not be normally needed; it is strongly recommended to let TiledArray build and install LibreTT.
 * `UMPIRE_INSTALL_DIR` -- the installation prefix of the pre-installed Umpire library. This should not be normally needed; it is strongly recommended to let TiledArray build and install Umpire.
 

external/cuda.cmake

@@ -16,10 +16,7 @@ enable_language(CUDA)
 
 set(CUDA_FOUND TRUE)
 set(TILEDARRAY_HAS_CUDA 1 CACHE BOOL "Whether TiledArray has CUDA support")
-
-if(ENABLE_CUDA_ERROR_CHECK)
-  set (TILEDARRAY_CHECK_CUDA_ERROR 1)
-endif(ENABLE_CUDA_ERROR_CHECK)
+set(TILEDARRAY_CHECK_CUDA_ERROR 1 CACHE BOOL "Whether TiledArray will check CUDA errors")
 
 # find CUDA toolkit
 # NB CUDAToolkit does NOT have COMPONENTS

external/hip.cmake

@@ -0,0 +1,36 @@
+# cmake 3.21 introduced HIP language support
+cmake_minimum_required(VERSION 3.21.0)
+set(CMAKE_HIP_STANDARD 17)
+set(CMAKE_HIP_EXTENSIONS OFF)
+set(CMAKE_HIP_STANDARD_REQUIRED ON)
+# N.B. need relaxed constexpr for std::complex
+# see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#constexpr-functions%5B/url%5D:
+if (DEFINED CMAKE_HIP_FLAGS)
+  set(CMAKE_HIP_FLAGS "--expt-relaxed-constexpr ${CMAKE_HIPE_FLAGS}")
+else()
+  set(CMAKE_HIP_FLAGS "--expt-relaxed-constexpr")
+endif()
+enable_language(HIP)
+
+set(HIP_FOUND TRUE)
+set(TILEDARRAY_HAS_HIP 1 CACHE BOOL "Whether TiledArray has HIP support")
+set(TILEDARRAY_CHECK_HIP_ERROR 1 CACHE BOOL "Whether TiledArray will check HIP errors")
+
+# find HIP components
+find_package(hipblas REQUIRED)
+
+foreach (library hipblas)
+  if (NOT TARGET roc::${library})
+    message(FATAL_ERROR "roc::${library} not found")
+  endif()
+endforeach()
+
+##
+## Umpire
+##
+include(external/umpire.cmake)
+
+##
+## LibreTT
+##
+include(external/librett.cmake)