Limited-area transforms on GPU

CMakeLists.txt

@@ -163,12 +163,13 @@ ecbuild_add_option( FEATURE GPU_STATIC
 
 ecbuild_add_option( FEATURE ETRANS
                     DEFAULT OFF
+					CONDITION NOT HAVE_GPU_GRAPHS_FFT
                     DESCRIPTION "Include Limited-Area-Model Transforms" )
 
 
 ecbuild_add_option( FEATURE ECTRANS4PY
                     DEFAULT OFF
-		    CONDITION HAVE_ETRANS AND HAVE_DOUBLE_PRECISION
+		            CONDITION HAVE_ETRANS AND HAVE_DOUBLE_PRECISION
                     DESCRIPTION "Compile ectrans4py interface routines for python binding w/ ctypesForFortran" )
 
 

src/etrans/CMakeLists.txt

@@ -1,5 +1,3 @@
-
-
 # (C) Copyright 2020- ECMWF.
 #
 # This software is licensed under the terms of the Apache Licence Version 2.0
@@ -66,7 +64,7 @@ function(generate_backend_includes)
     get_filename_component(outfile_name_we ${file_i} NAME_WE)
     get_filename_component(outfile_ext     ${file_i} EXT)
     get_filename_component(outfile_dir     ${file_i} DIRECTORY)
-    if (${file_i} IN_LIST ectrans_common_includes)
+    if (${file_i} IN_LIST ectrans_lam_common_includes)
       configure_file(${_PAR_INCLUDE_DIRECTORY}/${file_i} ${destination}/${outfile_name})
     else()
       set(outfile "${destination}/${outfile_name_we}_${backend}${outfile_ext}")
@@ -94,61 +92,13 @@ function(generate_backend_includes)
   target_include_directories(${_PAR_TARGET} INTERFACE $<BUILD_INTERFACE:${destination}>)
 endfunction(generate_backend_includes)
 
-
-
-
-
-# TODO: move precision-independent files to common
-#add_subdirectory( common )
+add_subdirectory( common )
 
 if( HAVE_CPU)
   add_subdirectory( cpu )
 endif()
 
 # placeholder
-#if( HAVE_GPU )
-#  add_subdirectory( gpu )
-#endif()
-
-
-if (FALSE)
-# original cmake file for etrans; keeping it for reference, but should be cleaned later
-message(FATAL_ERROR "Hold it right there!")
-
-# build list of sources to add to trans library 
-# (using CMAKE_CURRENT_SOURCE_DIR is necessary because sources are in a different directory than the target library (trans_${prec})
-ecbuild_list_add_pattern( LIST etrans_src
-					  GLOB
-							${CMAKE_CURRENT_SOURCE_DIR}/biper/internal/*
-							${CMAKE_CURRENT_SOURCE_DIR}/biper/external/*
-							${CMAKE_CURRENT_SOURCE_DIR}/etrans/internal/*
-							${CMAKE_CURRENT_SOURCE_DIR}/etrans/external/*
-					  QUIET
-					)
-
-# dummies to be able to loop over precisions
-set( HAVE_dp ${HAVE_DOUBLE_PRECISION} )
-set( HAVE_sp ${HAVE_SINGLE_PRECISION} )
-
-# loop over precisions
-foreach( prec sp dp )
-  if( HAVE_${prec} )
-	# add sources
-	target_sources(trans_${prec} PRIVATE ${etrans_src})
-	# add include directories
-	target_include_directories(trans_${prec}
-	    PUBLIC
-			$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/biper/include>
-			$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/etrans/include>
-    )
-  endif()
-endforeach()
-
-# install headers
-file( GLOB etrans_interface biper/include/* etrans/include/*)
-install(
-  FILES        ${etrans_interface}
-  DESTINATION  include/ectrans
-)
-
-endif()
+if( HAVE_GPU )
+  add_subdirectory( gpu )
+endif()

src/etrans/common/CMakeLists.txt

@@ -0,0 +1,44 @@
+# (C) Copyright 2020- ECMWF.
+#
+# This software is licensed under the terms of the Apache Licence Version 2.0
+# which can be obtained at http://www.apache.org/licenses/LICENSE-2.0.
+# In applying this licence, ECMWF does not waive the privileges and immunities
+# granted to it by virtue of its status as an intergovernmental organisation
+# nor does it submit to any jurisdiction.
+
+## Sources which are precision independent can go into a common library
+list( APPEND ectrans_lam_common_src
+  internal/tpmald_distr.F90
+  internal/tpmald_dim.F90
+  internal/esetup_dims_mod.F90
+  internal/esetup_geom_mod.F90
+  internal/suemplat_mod.F90
+  internal/suemplatb_mod.F90
+  internal/ellips.F90
+)
+list( APPEND ectrans_lam_common_includes
+)
+
+ecbuild_add_library(
+  TARGET           ectrans_lam_common
+  LINKER_LANGUAGE  Fortran
+  SOURCES          ${ectrans_lam_common_src}
+  PUBLIC_LIBS      fiat ectrans_common
+  PRIVATE_LIBS     ${LAPACK_LIBRARIES}
+  PUBLIC_INCLUDES  $<INSTALL_INTERFACE:include/ectrans>
+                   $<INSTALL_INTERFACE:include>
+                   $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/src/trans/include/ectrans>
+                   $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/src/trans/include>
+)
+ectrans_target_fortran_module_directory(
+  TARGET            ectrans_lam_common
+  MODULE_DIRECTORY  ${CMAKE_BINARY_DIR}/module/ectrans
+  INSTALL_DIRECTORY module/ectrans
+)
+
+if( HAVE_OMP )
+  ecbuild_debug("target_link_libraries( ectrans_common PRIVATE OpenMP::OpenMP_Fortran )")
+  target_link_libraries( ectrans_lam_common PRIVATE OpenMP::OpenMP_Fortran )
+endif()
+
+set( ectrans_lam_common_includes ${ectrans_lam_common_includes} PARENT_SCOPE )

src/etrans/cpu/internal/ellips.F90 → src/etrans/common/internal/ellips.F90

@@ -11,7 +11,7 @@
 
 ! Jan-2011 P. Marguinaud Interface to thread-safe FA
 SUBROUTINE ELLIPS (KSMAX,KMSMAX,KNTMP,KMTMP)
-USE PARKIND1, ONLY : JPRD, JPIM
+USE EC_PARKIND, ONLY : JPRD, JPIM
 USE YOMHOOK , ONLY : LHOOK, DR_HOOK, JPHOOK
 IMPLICIT NONE
 !

src/etrans/cpu/internal/esetup_dims_mod.F90 → src/etrans/common/internal/esetup_dims_mod.F90

@@ -13,7 +13,7 @@ MODULE ESETUP_DIMS_MOD
 CONTAINS
 SUBROUTINE ESETUP_DIMS
 
-USE PARKIND1  ,ONLY : JPIM     ,JPRB
+USE EC_PARKIND  ,ONLY : JPIM
 USE YOMHOOK   ,ONLY : LHOOK,   DR_HOOK, JPHOOK
 
 USE TPM_DIM         ,ONLY : R

src/etrans/cpu/internal/esetup_geom_mod.F90 → src/etrans/common/internal/esetup_geom_mod.F90

@@ -13,7 +13,7 @@ MODULE ESETUP_GEOM_MOD
 CONTAINS
 SUBROUTINE ESETUP_GEOM
 
-USE PARKIND1  ,ONLY : JPIM     ,JPRB
+USE EC_PARKIND  ,ONLY : JPIM
 USE YOMHOOK   ,ONLY : LHOOK,   DR_HOOK, JPHOOK
 
 USE TPM_GEN         ,ONLY : NOUT, NPRINTLEV

src/etrans/cpu/internal/suemplat_mod.F90 → src/etrans/common/internal/suemplat_mod.F90

@@ -93,7 +93,7 @@ SUBROUTINE SUEMPLAT(KDGL,KPROC,KPROCA,KMYSETA,LDSPLIT,LDEQ_REGIONS,&
 !        R. El Khatib 09-Aug-2013 Allow LEQ_REGIONS
 !     ------------------------------------------------------------------
 
-USE PARKIND1  ,ONLY : JPIM     ,JPRB, JPRD
+USE EC_PARKIND  ,ONLY : JPIM ,JPRD
 USE YOMHOOK   ,ONLY : LHOOK,   DR_HOOK, JPHOOK
 
 USE TPM_GEN         ,ONLY : NOUT, NPRINTLEV

src/etrans/cpu/internal/suemplatb_mod.F90 → src/etrans/common/internal/suemplatb_mod.F90

@@ -74,7 +74,7 @@ SUBROUTINE SUEMPLATB(KDGSA,KDGL,KPROCA,KLOENG,LDSPLIT,&
 !        A.Bogatchev   21-Sep-2010 phasing CY37
 !     ------------------------------------------------------------------
 
-USE PARKIND1  ,ONLY : JPIM     ,JPRB, JPRD
+USE EC_PARKIND  ,ONLY : JPIM, JPRD
 USE YOMHOOK   ,ONLY : LHOOK,   DR_HOOK, JPHOOK
 
 USE ABORT_TRANS_MOD ,ONLY : ABORT_TRANS

src/etrans/cpu/internal/tpmald_dim.F90 → src/etrans/common/internal/tpmald_dim.F90

@@ -13,7 +13,7 @@ MODULE TPMALD_DIM
 
 ! Module for dimensions.
 
-USE PARKIND1  ,ONLY : JPIM     ,JPRB
+USE EC_PARKIND  ,ONLY : JPIM, JPIB
 
 IMPLICIT NONE
 
@@ -26,6 +26,13 @@ MODULE TPMALD_DIM
 INTEGER(KIND=JPIM) :: NDGLSUR       ! Number of rows of latitudes+...
 INTEGER(KIND=JPIM) :: NMSMAX        ! Zonal truncation
 INTEGER(KIND=JPIM) :: NDGUX         ! Number of rows in zone C+I
+
+! arguments to pass to EXECUTE_FFT: kept here to make sure their addresses are constant (necessary for cuda graphs)
+INTEGER(KIND=JPIM) :: NLOENS_LON(1)
+INTEGER(KIND=JPIB) :: NOFFSETS_LON(2)
+INTEGER(KIND=JPIM) :: NLOENS_LAT(1)
+INTEGER(KIND=JPIB) :: NOFFSETS_LAT(2)
+
 END TYPE ALDDIM_TYPE
 
 TYPE(ALDDIM_TYPE),ALLOCATABLE,TARGET :: ALDDIM_RESOL(:)

src/etrans/cpu/internal/tpmald_distr.F90 → src/etrans/common/internal/tpmald_distr.F90

@@ -13,7 +13,7 @@ MODULE TPMALD_DISTR
 
 ! Module for distributed memory environment.
 
-USE PARKIND1  ,ONLY : JPIM     ,JPRB
+USE EC_PARKIND  ,ONLY : JPIM
 
 IMPLICIT NONE
 

src/etrans/cpu/CMakeLists.txt

@@ -27,7 +27,7 @@ function(generate_backend_sources)
     internal/*.F90
     external/*.F90
     biper/internal/*.F90
-          biper/external/*.F90
+    biper/external/*.F90
     QUIET
   )
 
@@ -50,20 +50,20 @@ set( BUILD_INTERFACE_INCLUDE_DIR ${CMAKE_BINARY_DIR}/include/ectrans )
 foreach( prec dp sp )
   if( HAVE_${prec} )
 
-    generate_backend_includes(BACKEND ${prec} TARGET ectrans_etrans_${prec}_includes DESTINATION ${BUILD_INTERFACE_INCLUDE_DIR} INCLUDE_DIRECTORY ${PROJECT_SOURCE_DIR}/src/etrans/include )
-    generate_backend_sources( BACKEND ${prec} OUTPUT ectrans_etrans_${prec}_src  DESTINATION ${CMAKE_CURRENT_BINARY_DIR}/generated/ectrans_etrans_${prec})
+    generate_backend_includes(BACKEND ${prec} TARGET ectrans_lam_${prec}_includes DESTINATION ${BUILD_INTERFACE_INCLUDE_DIR} INCLUDE_DIRECTORY ${PROJECT_SOURCE_DIR}/src/etrans/include )
+    generate_backend_sources( BACKEND ${prec} OUTPUT ectrans_lam_${prec}_src  DESTINATION ${CMAKE_CURRENT_BINARY_DIR}/generated/ectrans_lam_${prec})
     ecbuild_add_library(
-      TARGET           ectrans_etrans_${prec}
+      TARGET           ectrans_lam_${prec}
       LINKER_LANGUAGE  Fortran
-      SOURCES          ${ectrans_etrans_${prec}_src}
+      SOURCES          ${ectrans_lam_${prec}_src}
       PUBLIC_INCLUDES  $<INSTALL_INTERFACE:include/ectrans>
                        $<INSTALL_INTERFACE:include>
                        $<BUILD_INTERFACE:${BUILD_INTERFACE_INCLUDE_DIR}>
-      PUBLIC_LIBS      fiat ectrans_common ectrans_${prec}_includes ectrans_${prec} ectrans_etrans_${prec}_includes
+      PUBLIC_LIBS      fiat ectrans_common ectrans_${prec}_includes ectrans_${prec} ectrans_lam_common ectrans_lam_${prec}_includes
     )
 
     ectrans_target_fortran_module_directory(
-      TARGET            ectrans_etrans_${prec}
+      TARGET            ectrans_lam_${prec}
       MODULE_DIRECTORY  ${CMAKE_BINARY_DIR}/module/ectrans
       INSTALL_DIRECTORY module/ectrans
     )
@@ -75,23 +75,23 @@ foreach( prec dp sp )
         set( FFTW_LINK PUBLIC ) # Attempt anyway to give FFTW precedence
     endif()
     ecbuild_debug("target_link_libraries( trans_${prec} ${FFTW_LINK} ${FFTW_LIBRARIES} )")
-    target_link_libraries( ectrans_etrans_${prec} ${FFTW_LINK} ${FFTW_LIBRARIES} )
-    target_include_directories( ectrans_etrans_${prec} PRIVATE ${FFTW_INCLUDE_DIRS} )
-    target_compile_definitions( ectrans_etrans_${prec} PRIVATE WITH_FFTW )
+    target_link_libraries( ectrans_lam_${prec} ${FFTW_LINK} ${FFTW_LIBRARIES} )
+    target_include_directories( ectrans_lam_${prec} PRIVATE ${FFTW_INCLUDE_DIRS} )
+    target_compile_definitions( ectrans_lam_${prec} PRIVATE WITH_FFTW )
     # daand: lam transforms don't need lapack
     #ecbuild_debug("target_link_libraries( ectrans_etrans_${prec} PRIVATE ${LAPACK_LIBRARIES} )")
     #target_link_libraries( ectrans_${prec} PRIVATE ${LAPACK_LIBRARIES} )
 
     if( HAVE_OMP )
       ecbuild_debug("target_link_libraries( ectrans_${prec} PRIVATE OpenMP::OpenMP_Fortran )")
-      target_link_libraries( ectrans_${prec} PRIVATE OpenMP::OpenMP_Fortran )
+      target_link_libraries( ectrans_lam_${prec} PRIVATE OpenMP::OpenMP_Fortran )
     endif()
 
     # This interface library is for backward compatibility, and provides the older includes
     ecbuild_add_library( TARGET etrans_${prec} TYPE INTERFACE )
     target_include_directories( etrans_${prec} INTERFACE $<BUILD_INTERFACE:${BUILD_INTERFACE_INCLUDE_DIR}/etrans_${prec}> )
     target_include_directories( etrans_${prec} INTERFACE $<INSTALL_INTERFACE:include/ectrans/etrans_${prec}> )
-    target_link_libraries( trans_${prec} INTERFACE fiat etrans_${prec} ectrans_etrans_${prec} parkind_${prec})
+	target_link_libraries( etrans_${prec} INTERFACE ectrans_lam_${prec}) # seems ok to omit fiat, parkind, trans here.
   endif()
 endforeach()