Support HIP/ROCm backends for GPUs (#101)

* Add hydrogen error handling mechanisms * new cuda management infrastructure * everything in rocm compiles i think. linker issues pending * remove override decoration from Element/BlockMatrix functions * patch for finding rocblas; not sure if this is strictly necessary any more * forward kernel arguments by reference * a few tweaks to the CMakeLists * Make sure ROCm and CUDA aren't enabled at the same time. * correct a discrepancy in hipMemcpy2DAsync semantics * clean up HAVE_CUDA macro usage; streamline copy syntax * use nonblocking stream; clean up the mempool * straggler HAVE_CUDA use in include tree * preprocessor macro cleanup in blaslike tests * Remove debugging print statements * add short-circuit returns to copy/fill routines when size is zero * some cleanup * fix some new rocm issues * update aluminum version number * update version number * remove some unneeded CMake * revert changes related to the hip override bug * add support for hipCUB and generalize cublas tensor option * fix annoying clang warnings (that GCC _should_ throw, too, but it doesn't) * address some review comments * fix use of streams that should have been SyncInfos * Clean up device library functions * cleanup timer nonsense in Gemm test * fix some hipCUB linkage * Apply suggestions from code review Co-authored-by: Tim Moon <moon13@llnl.gov> * Apply suggestions from code review Co-authored-by: Tim Moon <moon13@llnl.gov> * remove unneeded metafunction. DiHydrogen has a cleaner implementation anyway.
LLNL · Jun 5, 2020 · d2feee8 · d2feee8
1 parent 48de387
commit d2feee8
Show file tree

Hide file tree

Showing 179 changed files with 4,733 additions and 1,917 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -53,8 +53,8 @@ endif (__GIT_EXECUTABLE)
 
 # This must be set because version tags
 set(HYDROGEN_VERSION_MAJOR 1)
-set(HYDROGEN_VERSION_MINOR 3)
-set(HYDROGEN_VERSION_PATCH 4)
+set(HYDROGEN_VERSION_MINOR 4)
+set(HYDROGEN_VERSION_PATCH 0)
 set(HYDROGEN_VERSION_MAJOR_MINOR
     "${HYDROGEN_VERSION_MAJOR}.${HYDROGEN_VERSION_MINOR}")
 set(HYDROGEN_VERSION
@@ -154,20 +154,32 @@ option(Hydrogen_ENABLE_CUDA
   "Search for CUDA support and enable related features if found."
   OFF)
 
-if (Hydrogen_ENABLE_CUDA)
+option(Hydrogen_ENABLE_ROCM
+  "Search for ROCm/HIP support and enable related features if found."
+  OFF)
+
+if (Hydrogen_ENABLE_CUDA OR Hydrogen_ENABLE_ROCM)
   option(Hydrogen_ENABLE_CUB
     "Search for CUB support and enable related features if found."
     ON)
 
-  option(Hydrogen_ENABLE_CUBLAS_TENSOR_MATH
-    "Use the cuBLAS tensor operation math."
+  option(Hydrogen_ENABLE_GPU_TENSOR_MATH
+    "Use the GPU tensor operations when available."
     OFF)
 
   option(Hydrogen_ENABLE_GPU_FP16
     "Enable FP16 arithmetic in GPU code."
     ON)
 endif ()
 
+if (Hydrogen_ENABLE_ROCM AND Hydrogen_ENABLE_CUDA)
+  message(FATAL_ERROR
+    "ROCm and CUDA code paths are mutually exclusive. "
+    "Please enable the one that corresponds to your hardware. "
+    "If you have mixed hardware, please contact the Hydrogen developers "
+    "as this would be of great interest.")
+endif ()
+
 #
 # MEMORY-RELATED OPTIONS
 #
@@ -334,8 +346,8 @@ if (Hydrogen_ENABLE_CUDA)
     find_package(CUDA REQUIRED) # Enable all the macros
     find_package(NVML REQUIRED)
 
-    if (Hydrogen_ENABLE_CUBLAS_TENSOR_MATH)
-      set(HYDROGEN_CUBLAS_USE_TENSOR_OP_MATH TRUE)
+    if (Hydrogen_ENABLE_GPU_TENSOR_MATH)
+      set(HYDROGEN_GPU_USE_TENSOR_OP_MATH TRUE)
     endif ()
 
     if (Hydrogen_ENABLE_GPU_FP16)
@@ -387,38 +399,64 @@ if (Hydrogen_ENABLE_CUDA)
     set(HYDROGEN_HAVE_CUDA FALSE)
 
   endif ()
-
 endif (Hydrogen_ENABLE_CUDA)
 
-set(HYDROGEN_HAVE_GPU ${HYDROGEN_HAVE_CUDA})
+if (Hydrogen_ENABLE_ROCM)
+  set(CMAKE_MODULE_PATH "/opt/rocm/hip/cmake" ${CMAKE_MODULE_PATH})
+  find_package(HIP REQUIRED)
+
+  if (Hydrogen_ENABLE_CUB)
+    set(CMAKE_PREFIX_PATH "/opt/rocm/hip" ${CMAKE_PREFIX_PATH})
+    set(HIP_FOUND FALSE)
+    find_package(HIP CONFIG REQUIRED)
+    find_package(rocPRIM REQUIRED)
+    find_package(hipCUB REQUIRED)
+    set(HYDROGEN_HAVE_CUB TRUE)
+  else ()
+    set(HYDROGEN_HAVE_CUB FALSE)
+  endif ()
+
+  if (HIP_FOUND)
+    set(CMAKE_CXX_EXTENSIONS FALSE)
+    find_package(ROCBLAS REQUIRED)
+    set(HYDROGEN_HAVE_ROCM TRUE)
+    message(STATUS "Found ROCm/HIP toolchain. Using HIP/ROCm.")
+  else ()
+    message(FATAL_ERROR "ROCm requested but not found.")
+  endif ()
+endif (Hydrogen_ENABLE_ROCM)
+
+if (HYDROGEN_HAVE_CUDA OR HYDROGEN_HAVE_ROCM)
+  set(HYDROGEN_HAVE_GPU TRUE)
+endif ()
 
 if (Hydrogen_ENABLE_ALUMINUM)
-  find_package(Aluminum 0.3.0 NO_MODULE
+  find_package(Aluminum 0.4.0 NO_MODULE
     HINTS ${Aluminum_DIR} ${ALUMINUM_DIR} ${AL_DIR}
     $ENV{Aluminum_DIR} $ENV{ALUMINUM_DIR} $ENV{AL_DIR}
     PATH_SUFFIXES lib64/cmake/aluminum lib/cmake/aluminum
     NO_DEFAULT_PATH)
   if (NOT Aluminum_FOUND)
-    find_package(Aluminum 0.3.0 NO_MODULE)
+    find_package(Aluminum 0.4.0 NO_MODULE)
   endif ()
 
   if (Aluminum_FOUND)
     set(HYDROGEN_HAVE_ALUMINUM TRUE)
     message(STATUS "Found Aluminum: ${Aluminum_DIR}")
 
-    if (HYDROGEN_HAVE_CUDA AND AL_HAS_NCCL)
+    if (HYDROGEN_HAVE_GPU AND AL_HAS_NCCL)
       set(HYDROGEN_HAVE_NCCL2 TRUE)
       message(STATUS "Aluminum detected with NCCL2 backend support.")
     else ()
       set(HYDROGEN_HAVE_NCCL2 FALSE)
-    endif (HYDROGEN_HAVE_CUDA AND AL_HAS_NCCL)
+    endif (HYDROGEN_HAVE_GPU AND AL_HAS_NCCL)
 
-    if (HYDROGEN_HAVE_CUDA AND AL_HAS_MPI_CUDA)
+    if (HYDROGEN_HAVE_GPU AND AL_HAS_MPI_CUDA)
       set(HYDROGEN_HAVE_AL_MPI_CUDA TRUE)
       message(STATUS "Aluminum detected with MPI-CUDA backend support.")
     else ()
       set(HYDROGEN_HAVE_AL_MPI_CUDA FALSE)
-    endif (HYDROGEN_HAVE_CUDA AND AL_HAS_MPI_CUDA)
+    endif (HYDROGEN_HAVE_GPU AND AL_HAS_MPI_CUDA)
   else ()
     set(HYDROGEN_HAVE_ALUMINUM FALSE)
     set(HYDROGEN_HAVE_NCCL2 FALSE)
@@ -497,7 +535,12 @@ configure_file("${PROJECT_SOURCE_DIR}/cmake/configure_files/hydrogen_config.h.in
 configure_file("${PROJECT_SOURCE_DIR}/doxy/Doxyfile.in"
   "${PROJECT_BINARY_DIR}/doxy/Doxyfile")
 
-add_library(Hydrogen_CXX "${HYDROGEN_SOURCES}" "${HYDROGEN_HEADERS}")
+if (HYDROGEN_HAVE_ROCM)
+  hip_add_library(Hydrogen_CXX "${HYDROGEN_SOURCES}" "${HYDROGEN_HEADERS}")
+else ()
+  add_library(Hydrogen_CXX "${HYDROGEN_SOURCES}" "${HYDROGEN_HEADERS}")
+endif ()
+
 target_include_directories(Hydrogen_CXX PUBLIC
   $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include>
   $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include/El>
@@ -509,40 +552,60 @@ target_include_directories(Hydrogen_CXX PUBLIC
 # be forced to build with that (even though they maybe should)...
 target_compile_options(Hydrogen_CXX PRIVATE ${EXTRA_CXX_FLAGS})
 
-target_link_libraries(Hydrogen_CXX PUBLIC ${Aluminum_LIBRARIES})
-target_link_libraries(Hydrogen_CXX PUBLIC ${HALF_LIBRARIES})
-
-if (TARGET OpenMP::OpenMP_CXX)
-  target_link_libraries(Hydrogen_CXX PUBLIC OpenMP::OpenMP_CXX)
-endif ()
-target_link_libraries(Hydrogen_CXX PUBLIC MPI::MPI_CXX)
-target_link_libraries(Hydrogen_CXX PUBLIC LAPACK::lapack)
-target_link_libraries(Hydrogen_CXX PUBLIC EP::extended_precision)
-
-target_link_libraries(Hydrogen_CXX PUBLIC ${VTUNE_LIBRARIES})
-target_link_libraries(Hydrogen_CXX PUBLIC ${NVTX_LIBRARIES})
-if (HYDROGEN_HAVE_CUDA)
-  target_link_libraries(Hydrogen_CXX PUBLIC cuda::toolkit)
-endif ()
+target_link_libraries(
+  Hydrogen_CXX PUBLIC
+  ${Aluminum_LIBRARIES}
+  ${HALF_LIBRARIES}
+  ${VTUNE_LIBRARIES}
+  ${NVTX_LIBRARIES}
+  ${ROCBLAS_LIBRARIES}
+  $<TARGET_NAME_IF_EXISTS:OpenMP::OpenMP_CXX>
+  $<TARGET_NAME_IF_EXISTS:MPI::MPI_CXX>
+  $<TARGET_NAME_IF_EXISTS:LAPACK::lapack>
+  $<TARGET_NAME_IF_EXISTS:EP::extended_precision>
+  $<TARGET_NAME_IF_EXISTS:cuda::toolkit>
+  $<TARGET_NAME_IF_EXISTS:hip::rocprim_hip>
+  $<TARGET_NAME_IF_EXISTS:hip::hipcub>
+  )
 
 # Add the CXX library to "Hydrogen"
 set(HYDROGEN_LIBRARIES Hydrogen_CXX)
 
 if (HYDROGEN_HAVE_CUDA)
-  add_library(Hydrogen_CUDA "${HYDROGEN_CUDA_SOURCES}")
+  add_library(Hydrogen_CUDA "${HYDROGEN_GPU_SOURCES}")
   target_include_directories(Hydrogen_CUDA PUBLIC
     $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include>
     $<BUILD_INTERFACE:${PROJECT_BINARY_DIR}/include>
     $<INSTALL_INTERFACE:include>)
 
-  target_link_libraries(Hydrogen_CUDA PUBLIC ${HALF_LIBRARIES})
-  target_link_libraries(Hydrogen_CUDA PUBLIC ${NVTX_LIBRARIES})
-  target_link_libraries(Hydrogen_CUDA PUBLIC cuda::toolkit)
+  target_link_libraries(
+    Hydrogen_CUDA PUBLIC
+    ${HALF_LIBRARIES}
+    ${NVTX_LIBRARIES}
+    $<TARGET_NAME_IF_EXISTS:cuda::toolkit>
+    )
 
   target_link_libraries(Hydrogen_CXX PUBLIC Hydrogen_CUDA)
   list(APPEND HYDROGEN_LIBRARIES Hydrogen_CUDA)
 endif ()
 
+if (HYDROGEN_HAVE_ROCM)
+  hip_add_library(Hydrogen_ROCM STATIC "${HYDROGEN_GPU_SOURCES}")
+  target_include_directories(Hydrogen_ROCM PUBLIC
+    $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include>
+    $<BUILD_INTERFACE:${PROJECT_BINARY_DIR}/include>
+    $<INSTALL_INTERFACE:include>
+    )
+
+  target_link_libraries(Hydrogen_ROCM PUBLIC
+    ${HALF_LIBRARIES}
+    ${ROCBLAS_LIBRARIES}
+    )
+
+  #set_target_properties(Hydrogen_ROCM PROPERTIES LINKER_LANGUAGE CXX)
+  list(APPEND HYDROGEN_LIBRARIES Hydrogen_ROCM)
+endif ()
+
 # Setup the tests
 if (Hydrogen_ENABLE_TESTING OR Hydrogen_ENABLE_UNIT_TESTS)
   include(CTest)

diff --git a/cmake/configure_files/HydrogenConfig.cmake.in b/cmake/configure_files/HydrogenConfig.cmake.in
@@ -18,11 +18,14 @@ set(HYDROGEN_MPI_CXX_COMPILER "@MPI_CXX_COMPILER@")
 set(MPI_CXX_COMPILER "${HYDROGEN_MPI_CXX_COMPILER}"
   CACHE FILEPATH "The MPI CXX compiler.")
 
-set(_OpenMP_DIR "@OpenMP_DIR@")
-if (NOT OpenMP_DIR)
-  set(OpenMP_DIR "${_OpenMP_DIR}")
-endif ()
-include (FindAndVerifyOpenMP)
+set(_HYDROGEN_HAVE_OPENMP @EL_HAVE_OPENMP@)
+if (_HYDROGEN_HAVE_OPENMP)
+  set(_OpenMP_DIR "@OpenMP_DIR@")
+  if (NOT OpenMP_DIR)
+    set(OpenMP_DIR "${_OpenMP_DIR}")
+  endif ()
+  include (FindAndVerifyOpenMP)
+endif (_HYDROGEN_HAVE_OPENMP)
 # FIXME: I should do verification to make sure all found features are
 #   the same.
 include (FindAndVerifyMPI)
@@ -33,14 +36,14 @@ set(_HYDROGEN_HAVE_NCCL2 @HYDROGEN_HAVE_NCCL2@)
 set(_HYDROGEN_HAVE_AL_MPI_CUDA @HYDROGEN_HAVE_AL_MPI_CUDA@)
 if (_HYDROGEN_HAVE_ALUMINUM)
   if (NOT Aluminum_FOUND)
-    find_package(Aluminum 0.3.0 NO_MODULE QUIET
+    find_package(Aluminum 0.4.0 NO_MODULE QUIET
       HINTS ${Aluminum_DIR} ${ALUMINUM_DIR} ${AL_DIR}
       $ENV{Aluminum_DIR} $ENV{ALUMINUM_DIR} $ENV{AL_DIR}
       PATH_SUFFIXES lib64/cmake/aluminum lib/cmake/aluminum
       NO_DEFAULT_PATH)
     if (NOT Aluminum_FOUND)
       set(Aluminum_DIR "@Aluminum_DIR@")
-      find_package(Aluminum 0.3.0 NO_MODULE REQUIRED)
+      find_package(Aluminum 0.4.0 NO_MODULE REQUIRED)
     endif ()
   endif ()
 
@@ -56,6 +59,16 @@ if (_HYDROGEN_HAVE_ALUMINUM)
   endif ()
 endif (_HYDROGEN_HAVE_ALUMINUM)
 
+# ROCm
+set(_HYDROGEN_HAVE_ROCM @HYDROGEN_HAVE_ROCM@)
+if (_HYDROGEN_HAVE_ROCM)
+  find_package(HIP REQUIRED)
+  find_package(ROCBLAS REQUIRED)
+
+  # query this beforehand, to set to what it was?
+  set(CMAKE_CXX_EXTENSIONS FALSE)
+endif (_HYDROGEN_HAVE_ROCM)
+
 # CUDA!
 set(_HYDROGEN_HAVE_CUDA @HYDROGEN_HAVE_CUDA@)
 set(_HYDROGEN_HAVE_CUB @HYDROGEN_HAVE_CUB@)

diff --git a/cmake/configure_files/HydrogenConfigVersion.cmake.in b/cmake/configure_files/HydrogenConfigVersion.cmake.in
@@ -7,13 +7,15 @@
 # [0.87 1.0.0)
 # [1.0.0 1.1.0)
 # [1.1.0 1.2.0)
-# [1.2.0 ???)
+# [1.2.0 1.3.0)
+# [1.3.0 1.4.0)
+# [1.4.0 ???)
 #
 # IMPORTANT: IF YOU MAKE A BREAKING CHANGE TO HYDROGEN, THE UPDATE
 # MUST BE GIVEN A NEW VERSION NUMBER, WHICH THEN MUST BE APPENDED TO
 # THIS LIST.
 
-set(_version_compat_ranges 0.0.0 0.87.0 1.0.0 1.1.0 1.2.0)
+set(_version_compat_ranges 0.0.0 0.87.0 1.0.0 1.1.0 1.2.0 1.3.0 1.4.0)
 
 # This is the version that has been installed.
 set(PACKAGE_VERSION "@HYDROGEN_VERSION@")

diff --git a/cmake/configure_files/hydrogen_config.h.in b/cmake/configure_files/hydrogen_config.h.in
@@ -33,16 +33,18 @@
 #cmakedefine HYDROGEN_HAVE_MKL
 #cmakedefine HYDROGEN_HAVE_MKL_GEMMT
 
+#cmakedefine HYDROGEN_HAVE_GPU
+
 // CUDA stuff
 #cmakedefine HYDROGEN_HAVE_CUDA
-#cmakedefine HYDROGEN_HAVE_CUB
-#cmakedefine HYDROGEN_CUBLAS_USE_TENSOR_OP_MATH
+
+// ROCm stuff
+#cmakedefine HYDROGEN_HAVE_ROCM
 
 // General GPU stuff
-#ifdef HYDROGEN_HAVE_CUDA
-#define HYDROGEN_HAVE_GPU
+#cmakedefine HYDROGEN_HAVE_CUB
+#cmakedefine HYDROGEN_GPU_USE_TENSOR_OP_MATH
 #cmakedefine HYDROGEN_GPU_USE_FP16
-#endif // HYDROGEN_HAVE_CUDA
 
 // Aluminum stuff
 #cmakedefine HYDROGEN_HAVE_ALUMINUM
@@ -62,4 +64,7 @@
 
 #cmakedefine HYDROGEN_DO_BOUNDS_CHECKING
 
+#define H_RESTRICT __restrict__
+#define H_PRETTY_FUNCTION __PRETTY_FUNCTION__
+
 #endif /* HYDROGEN_CONFIG_H */
diff --git a/cmake/modules/FindROCBLAS.cmake b/cmake/modules/FindROCBLAS.cmake
@@ -0,0 +1,46 @@
+# Find rocBLAS library and supporting header
+#
+#   rocBLAS_DIR or ROCBLAS_DIR[in]: The prefix for rocBLAS
+#
+#   ROCBLAS_INCLUDE_PATH[out,cache]: The include path for rocBLAS
+#   ROCBLAS_LIBRARY[out,cache]: The rocBLAS library
+#
+#   ROCBLAS_LIBRARIES[out]: The thing to link to for rocBLAS
+#   ROCBLAS_FOUND[out]: Variable indicating whether rocBLAS has been found
+#
+#   rocm::rocblas: Imported library for rocBLAS
+#
+
+find_path(ROCBLAS_INCLUDE_PATH rocblas.h
+  HINTS ${rocBLAS_DIR} $ENV{rocBLAS_DIR} ${ROCBLAS_DIR} $ENV{ROCBLAS_DIR}
+  PATH_SUFFIXES include
+  NO_DEFAULT_PATH
+  DOC "The rocBLAS include path.")
+find_path(ROCBLAS_INCLUDE_PATH rocblas.h)
+
+find_library(ROCBLAS_LIBRARY rocblas
+  HINTS ${rocBLAS_DIR} $ENV{rocBLAS_DIR} ${ROCBLAS_DIR} $ENV{ROCBLAS_DIR}
+  PATH_SUFFIXES lib64 lib
+  NO_DEFAULT_PATH
+  DOC "The rocBLAS library.")
+find_library(ROCBLAS_LIBRARY rocblas)
+
+# Standard handling of the package arguments
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(Rocblas
+  REQUIRED_VARS ROCBLAS_LIBRARY ROCBLAS_INCLUDE_PATH)
+
+if (NOT TARGET rocblas::rocblas)
+  add_library(rocblas::rocblas INTERFACE IMPORTED)
+endif ()
+
+if (ROCBLAS_INCLUDE_PATH AND ROCBLAS_LIBRARY)
+  set_target_properties(rocblas::rocblas PROPERTIES
+    INTERFACE_INCLUDE_DIRECTORIES
+    "${ROCBLAS_INCLUDE_PATH};/opt/rocm/hsa/include;/opt/rocm/hip/include"
+    INTERFACE_LINK_LIBRARIES "${ROCBLAS_LIBRARY}")
+endif ()
+
+set(ROCBLAS_LIBRARIES rocblas::rocblas)
+mark_as_advanced(ROCBLAS_INCLUDE_PATH)
+mark_as_advanced(ROCBLAS_LIBRARY)
diff --git a/include/El/blas_like/level1/AllReduce.hpp b/include/El/blas_like/level1/AllReduce.hpp
@@ -61,11 +61,11 @@ void AllReduce(AbstractMatrix<T>& A, mpi::Comm const& comm, mpi::Op op)
     case Device::CPU:
         AllReduce(static_cast<Matrix<T,Device::CPU>&>(A), comm, op);
         break;
-#ifdef HYDROGEN_HAVE_CUDA
+#ifdef HYDROGEN_HAVE_GPU
     case Device::GPU:
         AllReduce(static_cast<Matrix<T,Device::GPU>&>(A), comm, op);
         break;
-#endif // HYDROGEN_HAVE_CUDA
+#endif // HYDROGEN_HAVE_GPU
     default:
         LogicError("AllReduce: Bad device!");
     }