Merge branch 'ICLDisco:master' into update/parsec623

CMakeLists.txt

@@ -16,6 +16,10 @@ set(DPLASMA_VERSION "${DPLASMA_VERSION_MAJOR}.${DPLASMA_VERSION_MINOR}")
 
 ############################################################################
 # CMake Policies Tuning
+if(POLICY CMP0144)
+  # CMP0144: find_package uses upper-case <PACKAGENAME>_ROOT variables in addition to <PackageName>_ROOT
+  cmake_policy(SET CMP0144 NEW)
+endif(POLICY CMP0144)
 set(CMAKE_NO_SYSTEM_FROM_IMPORTED True)
 
 ############################################################################
@@ -231,12 +235,30 @@ endif(NOT TARGET PaRSEC::parsec AND NOT TARGET PaRSEC::parsec_ptgpp)
 
 ############################################################################
 # Resume configuring dplasma
-option(DPLASMA_HAVE_CUDA "Use CUDA to accelerate DPLASMA routines" ${PARSEC_HAVE_CUDA})
-if(DPLASMA_HAVE_CUDA)
+option(DPLASMA_GPU_WITH_CUDA "Use CUDA to accelerate DPLASMA routines" ${PARSEC_HAVE_CUDA})
+if(DPLASMA_GPU_WITH_CUDA)
+  if(NOT PARSEC_HAVE_CUDA)
+    message(FATAL_ERROR "CUDA support for DPLASMA requested, but detected PaRSEC does not support it")
+  endif()
   message(STATUS "CUDA support for DPLASMA enabled")
   if(NOT TARGET CUDA::cusolver)
     find_package(CUDAToolkit REQUIRED)
   endif(NOT TARGET CUDA::cusolver)
+  set(DPLASMA_HAVE_CUDA ${PARSEC_HAVE_CUDA} CACHE BOOL "True if DPLASMA provide support for CUDA")
+endif()
+option(DPLASMA_GPU_WITH_HIP "Use HIP to accelerate DPLASMA routines" ${PARSEC_HAVE_HIP})
+if(DPLASMA_GPU_WITH_HIP)
+  if(NOT PARSEC_HAVE_HIP)
+    message(FATAL_ERROR "HIP support for DPLASMA requested, but detected PaRSEC does not support it")
+  endif()
+  message(STATUS "HIP support for DPLASMA enabled")
+  # This is kinda ugly but the PATH and HINTS don't get transmitted to sub-dependents
+  set(CMAKE_SYSTEM_PREFIX_PATH_save ${CMAKE_SYSTEM_PREFIX_PATH})
+  list(APPEND CMAKE_SYSTEM_PREFIX_PATH /opt/rocm)
+  find_package(hipblas REQUIRED)
+  find_package(rocsolver REQUIRED)
+  set(CMAKE_SYSTEM_PREFIX_PATH ${CMAKE_SYSTEM_PREFIX_PATH_save})
+  set(DPLASMA_HAVE_HIP ${PARSEC_HAVE_HIP} CACHE BOOL "True if DPLASMA provide support for HIP")
 endif()
 
 ############################################################################

cmake_modules/FindLAPACKE.cmake

@@ -284,7 +284,7 @@ foreach(_comp ${LAPACKE_FIND_COMPONENTS})
   elseif(_comp STREQUAL "BLAS")
     if(NOT BLAS_FOUND)
       _find_library_with_header(${_comp} "" blas refblas)
-      set(BLA_VENDOR CACHE "Generic")
+      set(BLA_VENDOR "Generic" CACHE STRING "BLAS vendor")
     else()
       set(LAPACKE_BLAS_FOUND 1)
       set(LAPACKE_BLAS_LIB_FOUND 1)

configure

@@ -132,6 +132,9 @@ cat <<EOF
                 compile kernels optimized for the CUDA SM model x, y and z
                 where x,y,z are two digit numbers representing a valid CUDA architecture (e.g. 35,37,60) (default=autodetect)
 
+--with-hip[=DIR]
+                use the AMD RoCM accelerator libray [installed in DIR] (default=autodetect)
+
 
 Some influential environment variables:
   CC          C compiler command
@@ -296,6 +299,11 @@ while [ "x$1" != x ]; do
         --with-cuda-sm-targets) with_cuda_sm_targets=yes; shift;;
         --without-cuda-sm-targets) with_cuda_sm_targets=no; shift;;
 
+# RoCM options
+        --with-hip=*) with_hip="${1#*=}"; shift;;
+        --with-hip) with_hip=yes; shift;;
+        --without-hip) with_hip=no; shift;;
+
 # Python options
         --with-python=*) with_python="${1#*=}"; shift;;
         --with-python) with_python=yes; shift;;
@@ -399,12 +407,12 @@ _EOF
     mkdir -p "$NATIVE_DIR" && pushd "$NATIVE_DIR"
     rm -rf CMakeCache.txt CMakeFiles
 
-    # Disable MPI, CUDA, HWLOC when creating the build-tools
+    # Disable MPI, GPU, HWLOC when creating the build-tools
     local NATIVE_MPI="-DPARSEC_DIST_WITH_MPI=OFF"
-    local NATIVE_CUDA="-DPARSEC_GPU_WITH_CUDA=OFF"
+    local NATIVE_GPU="-DPARSEC_GPU_WITH_CUDA=OFF -DPARSEC_GPU_WITH_HIP=OFF"
     local NATIVE_HWLOC=""
     local NATIVE_COMPILERS="-DSUPPORT_FORTRAN=OFF"
-    local NATIVE_OPTS="-DBUILD_TESTING=OFF -DBUILD_TOOLS=ON -DBUILD_PARSEC=ON -DCMAKE_INSTALL_PREFIX=$NATIVE_PREFIX $NATIVE_MPI $NATIVE_CUDA $NATIVE_HWLOC $NATIVE_COMPILERS"
+    local NATIVE_OPTS="-DBUILD_TESTING=OFF -DBUILD_TOOLS=ON -DBUILD_PARSEC=ON -DCMAKE_INSTALL_PREFIX=$NATIVE_PREFIX $NATIVE_MPI $NATIVE_GPU $NATIVE_HWLOC $NATIVE_COMPILERS"
 
     set_cmake_executable #may have been changed in the platform file
     echo "CC=\"${NATIVE_CC}\" CFLAGS=\"${NATIVE_CFLAGS}\" CXX=\"${NATIVE_CXX}\" CXXFLAGS=\"${NATIVE_CXXFLAGS}\" LDFLAGS=\"${NATIVE_LDFLAGS}\" ${cmake_executable} -G\"${cmake_generator}\" ${NATIVE_OPTS} ${PARSEC_TOOLCHAIN_OPTIONS} $(for i in "$@"; do printf ' %q' "$i"; done) ${srcdir}"
@@ -621,6 +629,12 @@ x) ;;
 *) CMAKE_DEFINES+=" -DCUDA_SM_TARGETS='${with_cuda_sm_targets/,/;}'";;
 esac
 
+case x$with_hip in
+xno) CMAKE_DEFINES+=" -DPARSEC_GPU_WITH_HIP=OFF -DDPLASMA_GPU_WITH_HIP=OFF";;
+xyes) CMAKE_DEFINES+=" -DPARSEC_GPU_WITH_HIP=ON -DDPLASMA_GPU_WITH_HIP=ON";;
+x) ;;
+*) CMAKE_DEFINES+=" -DPARSEC_GPU_WITH_HIP=ON -DHIP_ROOT=$(printf %q "$with_hip") -DDPLASMA_GPU_WITH_HIP=ON";;
+esac
 
 case x$with_python in
 xno) echo >&2 "Python is required. Please provide a path to the python executable."; exit 3;;

contrib/platforms/macosx

@@ -12,6 +12,27 @@ for this reason the Apple VecLib remains the default option.
 EOF
 with_blas=Apple
 
+# OpenBLAS on OSX is not in system directories,
+# lets make sure it gets auto-detected, if requested with with_blas
+if [ -d /opt/homebrew/opt/openblas/lib/cmake/OpenBLAS ]; then # brew (arm)
+  ENVVARS+=" CMAKE_PREFIX_PATH+=':/opt/homebrew/opt/openblas/lib'"
+elif [ -d /usr/local/opt/openblas/lib/cmake/OpenBLAS ]; then # brew (intel)
+  ENVVARS+=" CMAKE_PREFIX_PATH+=':/usr/local/opt/openblas/lib'"
+fi
+if [ -d /opt/local/lib/cmake/OpenBLAS ]; then # macports
+  ENVVARS+=" CMAKE_PREFIX_PATH+=':/opt/local/lib'"
+fi
+if [ -d /opt/sw/lib/cmake/OpenBLAS ]; then # fink (10.15..)
+  ENVVARS+=" CMAKE_PREFIX_PATH+=':/opt/sw/lib'"
+elif [ -d /sw/lib/cmake/OpenBLAS ]; then # fink (..10.14)
+  ENVVARS+=" CMAKE_PREFIX_PATH+=':/sw/lib'"
+fi
+
+#####################################################################
+# Everything below is imported from the parsec contrib plaform file
+# If you need to modify something below, it probably needs modified
+# in parsec first, and backported here second
+
 # if icc/CLang is not set, detect the matching gcc/gfortran pair
 # OS X has no fortran by default, and flavor varies between fink, port, self installed, etc
 # Lets try to guess and use the same version for all 3 compilers
@@ -30,24 +51,25 @@ if [ "x$enable_fortran" != xno -a ! -x "$(command -v "$CC")" ]; then
   done
 fi
 
-# We have Brew OpenBLAS, lets make sure it gets auto-detected, if requested
-if [ -d /usr/local/opt/openblas/lib ]; then
-  ENVVARS+=" CMAKE_PREFIX_PATH+=':/usr/local/opt/openblas/lib'"
-fi
-
 # OS-X 12.2 provides Bison 2.3, we need Bison 3 or better
-# Try to get the 'brew' Bison if installed
-if [ -d /usr/local/opt/bison ]; then
-  ENVVARS+=" BISON_ROOT=${BISON_ROOT:-/usr/local/opt/bison}"
+local bison_version=$(expr "$(bison --version)" : "bison.*\([0-9]\)\.[0-9]")
+# do not override valid version in PATH or user choice
+if [ "$bison_version" -lt 3 -a -z "$BISON_ROOT" ]; then
+  # Try to find Bison if installed with brew, ports, fink
+  if [ -d /opt/homebrew/opt/bison ]; then # brew (arm)
+    bison_dir=/opt/homebrew/opt/bison
+  elif [ -d /usr/local/opt/bison ]; then # brew (intel)
+    bison_dir=/usr/local/opt/bison
+  elif [ -x /opt/local/bin/bison ]; then # macports
+    bison_dir=/opt/local
+  elif [ -x /opt/sw/bin/bison ]; then # fink (10.15..)
+    bison_dir=/opt/sw
+  elif [ -x /sw/bin/bison ]; then # fink (..10.14)
+    bison_dir=/sw
+  fi
+  if [ -n "${bison_dir}" ]; then
+    echo "Default Bison is too old, will search in ${bison_dir}; override by setting BISON_ROOT"
+    ENVVARS+=" BISON_ROOT=${bison_dir}"
+  fi
 fi
-# Try to get the 'MacPort' Bison if installed
-if [ -x /opt/local/bin/bison ]; then
-  ENVVARS+=" BISON_ROOT=${BISON_ROOT:-/opt/local}"
-fi
-# Try to get the 'Fink' Bison if installed
-if [ -x /sw/bin/bison ]; then
-  ENVVARS+=" BISON_ROOT=${BISON_ROOT:-/sw}"
-fi
-# If Bison still not found, please set BISON_ROOT by hand
-
 

share/help-dplasma.txt

@@ -1,8 +1,9 @@
-[cu*_alloc_failed]
-There was not enough memory available on a CUDA device
+[gpu_alloc_failed]
+There was not enough memory available on a GPU device
 while trying to allocate a %s handle to manage tasks on
-this device, or another CUDA device on the node. The
+this device, or another GPU device on the node. The
 PaRSEC runtime system may be configured to reserve too
-much memory on CUDA devices. Try reducing the amount of
+much memory on GPU devices. Try reducing the amount of
 reserved memory by setting the PaRSEC MCA parameter
-'device_cuda_memory_use' to a lower value.
+'device_cuda_memory_use' (or similar for the type of
+device) to a lower value.

src/CMakeLists.txt

@@ -28,8 +28,27 @@ if( NOT DPLASMA_HAVE_COMPLEX_H )
   list(APPEND EXTRA_SOURCES complex.c)
 endif()
 if( DPLASMA_HAVE_CUDA )
-  list(APPEND EXTRA_SOURCES dplasmaaux_cuda.c)
+  list(APPEND EXTRA_SOURCES dplasmaaux_cuda.c cuda/lapack_cuda_stage_in.c)
 endif()
+if( DPLASMA_HAVE_HIP )
+  list(APPEND EXTRA_SOURCES dplasmaaux_hip.c)
+  FILE(GLOB cuda_sources cuda/[^\\.]*.[ch])
+  find_package(Perl REQUIRED)
+  find_program(HIPIFY_PERL_COMMAND NAMES hipify-perl HINTS ${HIP_BIN_INSTALL_DIR} REQUIRED)
+  foreach(cuda_file ${cuda_sources})
+    file(RELATIVE_PATH cuda_filename ${CMAKE_CURRENT_SOURCE_DIR}/cuda ${cuda_file})
+    string(REPLACE cuda hip hip_file ${cuda_filename})
+    string(PREPEND hip_file "${CMAKE_CURRENT_BINARY_DIR}/hip/")
+    add_custom_command(OUTPUT ${hip_file}
+                       DEPENDS ${cuda_file} # do not use MAIN_DEPENDENCY, that overides the default .c.o rule
+                       COMMAND ${CMAKE_COMMAND} -E copy "${cuda_file}" "${hip_file}.prehip"
+                       COMMAND ${PERL_EXECUTABLE} ${HIPIFY_PERL_COMMAND} --inplace --print-stats "${hip_file}"
+                       COMMAND ${PERL_EXECUTABLE} -i -pe "s{(cuda)}{ substr uc hip | (uc \$1 ^ \$1), 0, 3 }egi" "${hip_file}" VERBATIM) # Convert all remaining cuda/CUDA
+    if(${hip_file} MATCHES [^\\.]*.c) # do not add .h to sources
+      list(APPEND EXTRA_SOURCES ${hip_file})
+    endif()
+  endforeach()
+endif( DPLASMA_HAVE_HIP )
 
 ### Generate .c files from .jdf for all required precisions
 set(JDF
@@ -236,7 +255,9 @@ target_link_libraries(dplasma
   PaRSEC::parsec
   LAPACKE::LAPACKE
   $<$<BOOL:${DPLASMA_HAVE_CUDA}>:CUDA::cublas>
-  $<$<BOOL:${DPLASMA_HAVE_CUDA}>:CUDA::cusolver>)
+  $<$<BOOL:${DPLASMA_HAVE_CUDA}>:CUDA::cusolver>
+  $<$<BOOL:${DPLASMA_HAVE_HIP}>:roc::hipblas>
+  $<$<BOOL:${DPLASMA_HAVE_HIP}>:roc::rocsolver>)
 set_target_properties(dplasma PROPERTIES VERSION ${DPLASMA_VERSION_MAJOR}.${DPLASMA_VERSION_MINOR}
                                        SOVERSION ${DPLASMA_VERSION_MAJOR})
 

src/cuda/README.md

@@ -0,0 +1,3 @@
+This directory contains files that are automatically converted from CUDA to HIP using Hipify.
+If your file is not automatically convertible, put it somewhere else.
+