Add amdflang to CI

.github/tools/install-amd-flang.sh

@@ -0,0 +1,84 @@
+#!/bin/sh
+
+# Install amdflang
+# 
+#
+# Originally written for Squash <https://github.com/quixdb/squash> by
+# Evan Nemerson.  For documentation, bug reports, support requests,
+# etc. please use <https://github.com/nemequ/pgi-travis>.
+#
+# To the extent possible under law, the author(s) of this script have
+# waived all copyright and related or neighboring rights to this work.
+# See <https://creativecommons.org/publicdomain/zero/1.0/> for
+# details.
+
+VERSION=7.0.5
+
+TEMPORARY_FILES="${TMPDIR:-/tmp}"
+export AMDFLANG_INSTALL_DIR=$(pwd)/amdflang-install
+export AMDFLANG_SILENT=true
+while [ $# != 0 ]; do
+    case "$1" in
+    "--prefix")
+        export AMDFLANG_INSTALL_DIR="$2"; shift
+        ;;
+    "--tmpdir")
+        TEMPORARY_FILES="$2"; shift
+        ;;
+    "--verbose")
+        export AMDFLANG_SILENT=false;
+        ;;
+    "--version")
+        VERSION="$2"; shift
+        ;;
+    *)
+        echo "Unrecognized argument '$1'"
+        exit 1
+        ;;
+    esac
+    shift
+done
+
+# Example download URL for version 6.0.0
+#    https://repo.radeon.com/rocm/misc/flang/rocm-afar-7450-drop-6.0.0-ubu.tar.bz2
+
+# Get name of drop based on version
+BASENAME=$(curl -s "https://repo.radeon.com/rocm/misc/flang/" | \
+    grep -oP "rocm-afar-[1-9][0-9]*-drop-{1}.{2}.{3}" | grep "$VERSION" | sort | tail -1)
+
+# Get URL of drop
+URL_SHORT=$(curl -s "https://repo.radeon.com/rocm/misc/flang/" | \
+    grep -o "$BASENAME-ubuntu.tar.bz2" | sort | tail -1)
+URL=https://repo.radeon.com/rocm/misc/flang/${URL_SHORT}
+
+if [ ! -f "${TEMPORARY_FILES}/${URL_SHORT}" ]; then
+    echo "Downloading [${URL}]"
+    wget -q -P ${TEMPORARY_FILES} "${URL}"
+else
+    echo "Download already present in ${TEMPORARY_FILES}"
+fi
+
+if [ ! -d "${AMDFLANG_INSTALL_DIR}/${BASENAME}" ]; then
+    if [ ! -d "${AMDFLANG_INSTALL_DIR}" ]; then
+        mkdir -p ${AMDFLANG_INSTALL_DIR}
+    fi
+    tar xjf ${TEMPORARY_FILES}/${URL_SHORT} -C ${AMDFLANG_INSTALL_DIR}
+else
+    echo "Install already present in ${AMDFLANG_INSTALL_DIR}"
+fi
+
+echo "ls ${AMDFLANG_INSTALL_DIR}"
+ls ${AMDFLANG_INSTALL_DIR}
+
+cat > ${AMDFLANG_INSTALL_DIR}/env.sh << EOF
+### Variables
+export AMDFLANG_INSTALL_DIR=${AMDFLANG_INSTALL_DIR}/${BASENAME}
+export AMDFLANG_VERSION=${VERSION}
+
+### Compilers
+export PATH=\${AMDFLANG_INSTALL_DIR}/bin:\${PATH}
+export LD_LIBRARY_PATH=\${AMDFLANG_INSTALL_DIR}/lib
+export LD_LIBRARY_PATH=\${AMDFLANG_INSTALL_DIR}/llvm/lib:\$LD_LIBRARY_PATH
+EOF
+
+cat ${AMDFLANG_INSTALL_DIR}/env.sh

.github/workflows/build.yml

@@ -35,6 +35,7 @@ jobs:
           - linux clang-18
           - linux nvhpc-25.1
           - linux intel-classic
+          - linux amd-flang-7.0.5
           - macos
 
         include:
@@ -67,14 +68,22 @@ jobs:
             ctest_options: -E memory
             caching: false
 
-          - name : linux intel-classic
+          - name: linux intel-classic
             os: ubuntu-22.04
             compiler: intel-classic
             compiler_cc: icc
             compiler_cxx: icpc
             compiler_fc: ifort
             caching: true
 
+          - name: linux amd-flang-7.0.5
+            os: ubuntu-24.04
+            compiler: amd-flang-7.0.5
+            compiler_cc: amdclang
+            compiler_cxx: amdclang++
+            compiler_fc: amdflang
+            cmake_options: -DCMAKE_Fortran_FLAGS_INIT='-fPIC -fdynamic-heap-array' -DCMAKE_CXX_FLAGS_INIT=-fPIC -DCMAKE_C_FLAGS_INIT=-fPIC -DENABLE_MPI=OFF -DOpenMP_Fortran_FLAGS=-fopenmp -DOpenMP_CXX_FLAGS=-fopenmp -DOpenMP_C_FLAGS=-fopenmp
+
           - name: macos
             # Xcode compiler requires empty environment variables, so we pass null (~) here
             os: macos-13
@@ -150,7 +159,17 @@ jobs:
         printenv >> $GITHUB_ENV
         echo "CACHE_SUFFIX=$CC-$($CC -dumpversion)" >> $GITHUB_ENV
 
+    - name: Install AMD flang compiler
+      if: contains( matrix.compiler, 'amd-flang' )
+      shell: bash -eux {0}
+      run: |
+        ${ECTRANS_TOOLS}/install-amd-flang.sh --prefix /opt/amd-flang
+        source /opt/amd-flang/env.sh
+        echo "${AMDFLANG_INSTALL_DIR}/bin"                   >> $GITHUB_PATH
+        echo "LD_LIBRARY_PATH=${AMDFLANG_INSTALL_DIR}/lib:${AMDFLANG_INSTALL_DIR}/llvm/lib" >> $GITHUB_ENV
+
     - name: Install MPI
+      if: ${{ !contains( matrix.compiler, 'amd-flang' ) }}
       shell: bash -eux {0}
       run: |
         FCFLAGS=-fPIC CFLAGS=-fPIC FFLAGS=-fPIC ${ECTRANS_TOOLS}/install-mpi.sh --mpi openmpi --prefix  ${DEPS_DIR}/openmpi
@@ -187,8 +206,8 @@ jobs:
           ecmwf-ifs/fiat
         dependency_branch: develop
         dependency_cmake_options: |
-          ecmwf-ifs/fiat: "-G Ninja -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} -DENABLE_TESTS=OFF -DENABLE_MPI=ON"
-        cmake_options:    "-G Ninja -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ${{ matrix.cmake_options }} -DENABLE_MPI=ON -DENABLE_FFTW=ON -DENABLE_ETRANS=ON"
+          ecmwf-ifs/fiat: "-G Ninja -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} -DENABLE_TESTS=OFF -DENABLE_MPI=ON ${{ matrix.cmake_options }}"
+        cmake_options:    "-G Ninja -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} -DENABLE_MPI=ON -DENABLE_FFTW=ON -DENABLE_ETRANS=ON ${{ matrix.cmake_options }}"
         ctest_options: "${{ matrix.ctest_options }}"
 
     - name: Verify tools

tests/test-install.sh.in

@@ -49,6 +49,10 @@ cmake $SOURCE \
 
 make VERBOSE=1
 
+if [[ "@CMAKE_Fortran_COMPILER_ID@" = "LLVMFlang" ]]; then
+  ulimit -s unlimited
+fi
+
 if [ -f bin/main_dp ]; then
   bin/main_dp
 fi

tests/trans/test_dirtrans_adjoint.F90

@@ -162,6 +162,11 @@ PROGRAM TEST_DIRTRANS_ADJOINT
 CALL DIST_SPEC(PSPECG=ZVORXG, KFDISTG=JP_NUMLEVELS_G, KFROM=ITOSP, PSPEC=ZVORX, KVSET=IVSET)
 CALL DIST_SPEC(PSPECG=ZDIVXG, KFDISTG=JP_NUMLEVELS_G, KFROM=ITOSP, PSPEC=ZDIVX, KVSET=IVSET)
 
+! We don't need these anymore
+DEALLOCATE(ZSPECXG)
+DEALLOCATE(ZVORXG)
+DEALLOCATE(ZDIVXG)
+
 ! ===== Allocate and initialize gridpoint data =====
 ALLOCATE(ZGXG(IGPTOTG,3*JP_NUMLEVELS_G))
 ALLOCATE(ZGX(JPPROMA,3*JP_NUMLEVELS_G,IGPBLKS))
@@ -187,6 +192,11 @@ PROGRAM TEST_DIRTRANS_ADJOINT
 ADJ_VALUE_1 = ADJ_VALUE_1 + SCALPRODSP(ZVORX, ZVORY)
 ADJ_VALUE_1 = ADJ_VALUE_1 + SCALPRODSP(ZDIVX, ZDIVY)
 
+! We don't need these anymore
+DEALLOCATE(ZSPECY)
+DEALLOCATE(ZVORY)
+DEALLOCATE(ZDIVY)
+
 ! ===== Compute dirtrans_adj and gather result on proc 1 =====
 ! i.e. dirtrans_adj(rspscalarx, rspvorx, rspdivx) = rgpy
 
@@ -196,13 +206,16 @@ PROGRAM TEST_DIRTRANS_ADJOINT
 ! ===== Compute: adj_value2 = <rgpx, dirtrans_adj(rspscalarx, rspvorx, rspdivx)> =====
 ! i.e. adj_value2 = <rgpy, rgpx>
 
-ADJ_VALUE_2 = SCALPRODGP(ZGY, ZGX)
+ZGX = ZGX * ZGY
+CALL GATH_GRID(ZGXG, JPPROMA, 3*JP_NUMLEVELS_G, ITOGP, PGP=ZGX)
 
 ! Only task 1 should perform the correctness check
 IF (IMYPROC == 1) THEN
   ! ===== Compare inner products =====
   ! i.e. <dirtrans(rgpx), (rspscalarx, rspvorx, rspdivx)> == <rgpx, dirtrans_adj(rspscalarx, rspvorx, rspdivx)>
 
+  ADJ_VALUE_2 = SUM(ZGXG)
+
   ZRELATIVE_ERROR = ABS(ADJ_VALUE_1 - ADJ_VALUE_2) / ABS(ADJ_VALUE_1)
 
   WRITE(IOUT, '(A,1E30.15)') '<Fx,y>  = ', ADJ_VALUE_1
@@ -232,39 +245,6 @@ PROGRAM TEST_DIRTRANS_ADJOINT
 
 CONTAINS
 
-FUNCTION SCALPRODGP(RGP1, RGP2) RESULT(RSC)
-
-  ! Scalar product in spectral space
-  REAL(KIND=JPRB) :: RGP1(JPPROMA,3*JP_NUMLEVELS_G,IGPBLKS), RGP2(JPPROMA,3*JP_NUMLEVELS_G,IGPBLKS)
-  REAL(KIND=JPRB) :: RSC
-
-  INTEGER(KIND=JPIM) :: JLEV, JKGLO, IEND, IBL, JROF
-  REAL(KIND=JPRB) :: RGP(JPPROMA,3*JP_NUMLEVELS_G,IGPBLKS), RGPG(IGPTOTG,3*JP_NUMLEVELS_G)
-
-  RSC = 0.0_JPRB
-
-  !$OMP PARALLEL DO SCHEDULE(STATIC,1) PRIVATE(JLEV,JKGLO,IEND,IBL,JROF)
-  DO JLEV = 1, 3 * JP_NUMLEVELS_G
-    DO JKGLO = 1, IGPTOT, JPPROMA
-      IEND = MIN(JPPROMA, IGPTOT - JKGLO + 1)
-      IBL  = (JKGLO - 1) / JPPROMA+1
-      DO JROF = 1, IEND
-        RGP(JROF,JLEV,IBL) = RGP1(JROF,JLEV,IBL) * RGP2(JROF,JLEV,IBL)
-      ENDDO
-    ENDDO
-  ENDDO
-  !$OMP END PARALLEL DO
-
-  CALL GATH_GRID(RGPG, JPPROMA, 3*JP_NUMLEVELS_G, ITOGP, PGP=RGP)
-
-  IF(IMYPROC == 1) THEN
-    RSC = SUM(RGPG)
-  ELSE
-    RSC = 0.0_JPRB
-  ENDIF
-
-END FUNCTION SCALPRODGP
-
 FUNCTION SCALPRODSP(PSP1,PSP2) RESULT(PSC)
 
   ! Scalar product in spectral space
@@ -340,4 +320,4 @@ FUNCTION DETECT_MPIRUN() RESULT(LMPI_REQUIRED)
   ENDIF
 END FUNCTION
 
-END PROGRAM TEST_DIRTRANS_ADJOINT
+END PROGRAM TEST_DIRTRANS_ADJOINT

tests/trans/test_invtrans_adjoint.F90

@@ -162,6 +162,11 @@ PROGRAM TEST_INVTRANS_ADJOINT
 CALL DIST_SPEC(PSPECG=ZVORXG, KFDISTG=JP_NUMLEVELS_G, KFROM=ITOSP, PSPEC=ZVORX, KVSET=IVSET)
 CALL DIST_SPEC(PSPECG=ZDIVXG, KFDISTG=JP_NUMLEVELS_G, KFROM=ITOSP, PSPEC=ZDIVX, KVSET=IVSET)
 
+! We don't need these anymore
+DEALLOCATE(ZSPECXG)
+DEALLOCATE(ZVORXG)
+DEALLOCATE(ZDIVXG)
+
 ! ===== Allocate and initialize gridpoint data =====
 ALLOCATE(ZGXG(IGPTOTG,3*JP_NUMLEVELS_G))
 ALLOCATE(ZGX(JPPROMA,3*JP_NUMLEVELS_G,IGPBLKS))
@@ -183,7 +188,8 @@ PROGRAM TEST_INVTRANS_ADJOINT
 ! ===== Compute: adj_value2 = <invtrans(rspscalarx, rspvorx, rspdivx), rgpx> =====
 ! i.e. adj_value2 = <rgpy, rgpx>
 
-ADJ_VALUE_1 = SCALPRODGP(ZGY, ZGX)
+ZGY = ZGX * ZGY
+CALL GATH_GRID(ZGXG, JPPROMA, 3*JP_NUMLEVELS_G, ITOGP, PGP=ZGY)
 
 ! ===== Compute adjoint invtrans and gather result on proc 1 =====
 ! i.e. invtrans_adj(rgpx) = (rspscalary, rspvory, rspdivy)
@@ -211,6 +217,8 @@ PROGRAM TEST_INVTRANS_ADJOINT
   ! ===== Compare inner products =====
   ! i.e. <invtrans_adj(rgpx), (rspscalarx, rspvorx, rspdivx)> == <rgpx, invtrans(rspscalarx, rspvorx, rspdivx)>
 
+  ADJ_VALUE_1 = SUM(ZGXG)
+
   ZRELATIVE_ERROR = ABS(ADJ_VALUE_1 - ADJ_VALUE_2) / ABS(ADJ_VALUE_1)
 
   WRITE(IOUT, '(A,1E30.15)') '<Fx,y>  = ', ADJ_VALUE_1
@@ -240,39 +248,6 @@ PROGRAM TEST_INVTRANS_ADJOINT
 
 CONTAINS
 
-FUNCTION SCALPRODGP(RGP1, RGP2) RESULT(RSC)
-
-  ! Scalar product in spectral space
-  REAL(KIND=JPRB) :: RGP1(JPPROMA,3*JP_NUMLEVELS_G,IGPBLKS), RGP2(JPPROMA,3*JP_NUMLEVELS_G,IGPBLKS)
-  REAL(KIND=JPRB) :: RSC
-
-  INTEGER(KIND=JPIM) :: JLEV, JKGLO, IEND, IBL, JROF
-  REAL(KIND=JPRB) :: RGP(JPPROMA,3*JP_NUMLEVELS_G,IGPBLKS), RGPG(IGPTOTG,3*JP_NUMLEVELS_G)
-
-  RSC = 0.0_JPRB
-
-  !$OMP PARALLEL DO SCHEDULE(STATIC,1) PRIVATE(JLEV,JKGLO,IEND,IBL,JROF)
-  DO JLEV = 1, 3 * JP_NUMLEVELS_G
-    DO JKGLO = 1, IGPTOT, JPPROMA
-      IEND = MIN(JPPROMA, IGPTOT - JKGLO + 1)
-      IBL  = (JKGLO - 1) / JPPROMA+1
-      DO JROF = 1, IEND
-        RGP(JROF,JLEV,IBL) = RGP1(JROF,JLEV,IBL) * RGP2(JROF,JLEV,IBL)
-      ENDDO
-    ENDDO
-  ENDDO
-  !$OMP END PARALLEL DO
-
-  CALL GATH_GRID(RGPG, JPPROMA, 3*JP_NUMLEVELS_G, ITOGP, PGP=RGP)
-
-  IF(IMYPROC == 1) THEN
-    RSC = SUM(RGPG)
-  ELSE
-    RSC = 0.0_JPRB
-  ENDIF
-
-END FUNCTION SCALPRODGP
-
 FUNCTION SCALPRODSP(PSP1,PSP2) RESULT(PSC)
 
   ! Scalar product in spectral space