Merge branch 'release/1.5.0'

* release/1.5.0: (86 commits)
  Version 1.5.0
  Workaround for supporting GPU-aware MPI on Cray/AMD platforms
  Workaround for ASYNC statements in TRMTOL for AMD GPUs
  Pin buffers in benchmark (#158)
  Remove default values for TPM_GEN variables (#159)
  Fix case inconsistencies
  Add option to not do in-place FFT
  Search for OpenACC only if explicitly enabled
  Use Fortran linker for ectrans_test_setup_trans0
  Update CMakeLists.txt
  Fix tabs
  Fix typo in TRGTOL
  Update src/trans/gpu/algor/hicfft.hip.cpp
  Update paths to HIP includes to support ROCm > 6
  add gpu graphs define
  Fixup diagnostic prints missing since #141
  add ungrouepd gemms
  Make default for --niter-warmup 3 if not specified, experimentally a better value
  Add --niter-warmup command line option to be able to configure warmup iterations
  Remove spurious print
  ...

.github/workflows/build.yml

@@ -124,6 +124,25 @@ jobs:
         path: ${{ env.DEPS_DIR }}
         key: deps-${{ matrix.os }}-${{ matrix.compiler }}-${{ matrix.build_type }}-${{ env.CACHE_SUFFIX }}
 
+    # Free up disk space for nvhpc
+    - name: Free Disk Space (Ubuntu)
+      uses: jlumbroso/free-disk-space@main
+      if: contains( matrix.compiler, 'nvhpc' )
+      continue-on-error: true
+      with:
+        # this might remove tools that are actually needed,
+        # if set to "true" but frees about 6 GB
+        tool-cache: false
+
+        # all of these default to true, but feel free to set to
+        # "false" if necessary for your workflow
+        android: true
+        dotnet: true
+        haskell: true
+        large-packages: true
+        docker-images: true
+        swap-storage: true
+
     - name: Install NVHPC compiler
       if: contains( matrix.compiler, 'nvhpc' )
       shell: bash -eux {0}
@@ -171,13 +190,9 @@ jobs:
         recreate_cache: ${{ matrix.caching == false }}
         dependencies: |
           ecmwf/ecbuild
-          ecmwf/eckit
-          ecmwf/fckit
           ecmwf-ifs/fiat
         dependency_branch: develop
         dependency_cmake_options: |
-          ecmwf/eckit:    "-G Ninja -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} -DENABLE_TESTS=OFF -DENABLE_ECKIT_CMD=OFF -DENABLE_ECKIT_SQL=OFF -DENABLE_MPI=ON -DENABLE_OMP=OFF"
-          ecmwf/fckit:    "-G Ninja -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} -DENABLE_TESTS=OFF"
           ecmwf-ifs/fiat: "-G Ninja -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} -DENABLE_TESTS=OFF -DENABLE_MPI=ON"
         cmake_options:    "-G Ninja -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ${{ matrix.cmake_options }} -DENABLE_MPI=ON -DENABLE_FFTW=ON"
         ctest_options: "${{ matrix.ctest_options }}"

CMakeLists.txt

@@ -34,14 +34,16 @@ ecbuild_add_option( FEATURE OMP
                     DESCRIPTION "Support for OpenMP shared memory parallelism"
                     REQUIRED_PACKAGES "OpenMP COMPONENTS Fortran" )
 
-if( ${CMAKE_VERSION} VERSION_LESS "3.25" AND (NOT DEFINED ENABLE_ACC OR ENABLE_ACC ) )
-  # See https://gitlab.kitware.com/cmake/cmake/-/issues/23691, fixed in CMake 3.25
-  # (TL;DR: FindOpenACC sets OpenACC_<LANG>_FOUND correctly but does not set
-  #  OpenACC_FOUND unless all three C, CXX, and Fortran have been found - even if
-  #  only one language has been requested via COMPONENTS)
-  find_package( OpenACC COMPONENTS Fortran )
-  if( OpenACC_Fortran_FOUND )
-    set( OpenACC_FOUND ON )
+if( ${CMAKE_VERSION} VERSION_LESS "3.25" )
+  if ( ECTRANS_ENABLE_ACC OR (NOT DEFINED ECTRANS_ENABLE_ACC AND ENABLE_ACC) )
+    # See https://gitlab.kitware.com/cmake/cmake/-/issues/23691, fixed in CMake 3.25
+    # (TL;DR: FindOpenACC sets OpenACC_<LANG>_FOUND correctly but does not set
+    #  OpenACC_FOUND unless all three C, CXX, and Fortran have been found - even if
+    #  only one language has been requested via COMPONENTS)
+    find_package( OpenACC COMPONENTS Fortran )
+    if( OpenACC_Fortran_FOUND )
+      set( OpenACC_FOUND ON )
+    endif()
   endif()
 endif()
 
@@ -104,8 +106,8 @@ endif()
 
 ecbuild_add_option( FEATURE GPU
                     DEFAULT OFF
-                    DESCRIPTION "Compile GPU version of ectrans (Requires OpenACC or sufficient OpenMP offloading support and MPI)"
-                    CONDITION (HAVE_HIP OR HAVE_CUDA) AND (HAVE_ACC OR HAVE_OMP) AND HAVE_MPI )
+                    DESCRIPTION "Compile GPU version of ectrans (Requires OpenACC or sufficient OpenMP offloading support)"
+                    CONDITION (HAVE_HIP OR HAVE_CUDA) AND (HAVE_ACC OR HAVE_OMP) )
 
 if( HAVE_GPU )
   if( HAVE_ACC )
@@ -140,6 +142,11 @@ ecbuild_add_option( FEATURE GPU_GRAPHS_GEMM
                     CONDITION HAVE_GPU
                     DESCRIPTION "Enable graph-based optimisation of Legendre transform GEMM kernel" )
 
+	    ecbuild_add_option( FEATURE GPU_GRAPHS_FFT
+                    DEFAULT ON
+                    CONDITION HAVE_GPU
+		    DESCRIPTION "Enable graph-based optimisation of FFT kernels" )
+
 if( BUILD_SHARED_LIBS )
   set( GPU_STATIC_DEFAULT OFF )
 else()

README.md

@@ -82,6 +82,7 @@ Extra options can be added to the `cmake` command to control the build:
 Specific extra options exist for GPU installation:
  - `-DENABLE_GPU_AWARE_MPI=<ON|OFF>`    default=OF
  - `-DENABLE_GPU_GRAPHS_GEMM=<ON|OFF>`  default=ON
+ - `-DENABLE_GPU_GRAPHS_FFT=<ON|OFF>`   default=ON
  - `-DENABLE_CUTLASS=<ON|OFF>`          default=OFF
  - `-DENABLE_3XTF32=<ON|OFF>`           default=OFF
 

VERSION

@@ -1 +1 @@
-1.4.0
+1.5.0

cmake/ectrans_compile_options.cmake

@@ -44,8 +44,8 @@ if( NOT DEFINED ECTRANS_HAVE_CONTIGUOUS_ISSUE )
 endif()
 
 macro( ectrans_add_compile_options )
-  set( options  )
-  set( single_value_args FLAGS )
+  set( options NOFAIL )
+  set( single_value_args FLAGS)
   set( multi_value_args SOURCES )
   cmake_parse_arguments( _PAR "${options}" "${single_value_args}" "${multi_value_args}"  ${_FIRST_ARG} ${ARGN} )
   if(_PAR_UNPARSED_ARGUMENTS)
@@ -59,7 +59,7 @@ macro( ectrans_add_compile_options )
   endif()
   foreach( _file ${_PAR_SOURCES} )
     ecbuild_warn("Adding custom compile flags for file ${_file} : [${_PAR_FLAGS}]")
-    if( NOT EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${_file} )
+    if( NOT EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${_file} AND NOT _PAR_NOFAIL)
         ecbuild_error("${_file} does not exist")
     endif()
     set_source_files_properties( ${_file} PROPERTIES COMPILE_FLAGS "${_PAR_FLAGS}" )

src/programs/CMakeLists.txt

@@ -36,9 +36,17 @@ foreach( program ectrans-benchmark )
                                   parkind_${prec}
                                   trans_gpu_${prec}
                                   OpenACC::OpenACC_Fortran
-       )
-       target_compile_definitions(${program}-gpu-${prec} PRIVATE VERSION="gpu")
-       endif()
+                                DEFINITIONS
+                                  VERSION="gpu"
+                                  $<$<COMPILE_LANG_AND_ID:Fortran,NVHPC>:USE_PINNED>
+        )
+        target_compile_options(${program}-gpu-${prec} PUBLIC
+          $<$<COMPILE_LANG_AND_ID:Fortran,NVHPC>:-cuda>
+        )
+        target_link_options(${program}-gpu-${prec} PUBLIC
+          $<$<LINK_LANG_AND_ID:Fortran,NVHPC>:-cuda>
+        )
+      endif()
     endforeach( prec )
   endif( HAVE_GPU )
 endforeach( program )