Merge branch 'nwchemgit:master' into remove-bgj-crap

nwchemgit · May 16, 2021 · e069175 · e069175
2 parents 168a31c + 16a03cb
commit e069175
Show file tree

Hide file tree

Showing 81 changed files with 1,286 additions and 172 deletions.
diff --git a/.github/workflows/github_actions.yml b/.github/workflows/github_actions.yml
@@ -93,6 +93,22 @@ jobs:
             nwchem_modules: "nwdft solvation driver"
             fc: gfortran-10
             use_simint: 1
+          - os: ubuntu-20.04
+            experimental: true
+            mpi_impl: mpich
+            armci_network: mpi-ts
+            nwchem_modules: "nwdft solvation driver"
+            fc: nvfortran
+            cc: gcc
+            use_simint: 1
+          - os: ubuntu-20.04
+            experimental: true
+            mpi_impl: intel
+            armci_network: mpi-ts
+            nwchem_modules: "nwdft solvation driver"
+            fc: ifort
+            cc: icc
+            use_simint: 1
           - os: ubuntu-20.04
             experimental: true 
             mpi_impl: mpich
@@ -166,7 +182,7 @@ jobs:
         with:
           path: |
             ~/cache
-          key: ${{ matrix.os }}-${{ matrix.mpi_impl}}-${{ matrix.fc}}-xcode${{ matrix.xcode_version}}-simint${{ matrix.use_simint}}-blas${{ matrix.blas}}-simd${{ steps.get-simd.outputs.simd }}-nwchemcache-v004
+          key: ${{ matrix.os }}-${{ matrix.mpi_impl}}-${{ matrix.fc}}-xcode${{ matrix.xcode_version}}-simint${{ matrix.use_simint}}-blas${{ matrix.blas}}-simd${{ steps.get-simd.outputs.simd }}-nwchemcache-v005
       - name: build environment
         run: |
           pwd

diff --git a/contrib/getfiles.nwchem b/contrib/getfiles.nwchem
@@ -0,0 +1,15 @@
+#!/usr/bin/env bash
+#
+# $Id$
+cd $NWCHEM_TOP/src/nwpw/nwpwlib/nwpwxc/
+rm -f dftd3.tgz
+wget https://www.chemie.uni-bonn.de/pctc/mulliken-center/software/dft-d3/dftd3.tgz
+cd $NWCHEM_TOP/src/libext/openblas
+VERSION=0.3.13
+rm -rf OpenBLAS*gz
+curl -L https://github.com/xianyi/OpenBLAS/archive/v${VERSION}.tar.gz -o OpenBLAS-${VERSION}.tar.gz
+cd $NWCHEM_TOP/src/libext/scalapack
+COMMIT=bc6cad585362aa58e05186bb85d4b619080c45a9
+rm -f scalapack-$COMMIT.zip
+curl -L https://github.com/Reference-ScaLAPACK/scalapack/archive/$COMMIT.zip -o scalapack-$COMMIT.zip
+
diff --git a/src/NWints/simint/libsimint_source/build_simint.sh b/src/NWints/simint/libsimint_source/build_simint.sh
@@ -12,6 +12,16 @@ if  [ -z "$(command -v python3)" ]; then
     echo please install python3
     exit 1
 fi
+if  [ -z "$(command -v curl)" ] && [ -z "$(command -v wget)" ]; then
+    echo curl and wget not installed
+    echo please install curl or wget
+    exit 1
+fi
+if  [ -z "$(command -v patch)" ]; then
+    echo patch not installed
+    echo please install patch
+    exit 1
+fi
 UNAME_S=$(uname -s)
 if [[ ${UNAME_S} == Linux ]]; then
     CPU_FLAGS=$(cat /proc/cpuinfo | grep flags |tail -n 1)
@@ -41,6 +51,19 @@ else
     VEC=scalar
 fi
 echo VEC $VEC
+if [[ "${VEC}" == "avx512" ]]; then
+if [[   -z "${CC}" ]]; then
+    CC=cc
+fi
+let GCCVERSIONGT5=$(expr `${CC} -dumpversion | cut -f1 -d.` \> 5)
+    if [[ ${GCCVERSIONGT5} != 1 ]]; then
+	echo
+	echo you have gcc version $(${CC} -dumpversion | cut -f1 -d.)
+	echo gcc version 6 and later needed for skylake
+	echo
+	exit 1
+    fi
+fi
 SRC_HOME=`pwd`
 DERIV=1
 if [[  -z "${SIMINT_MAXAM}" ]]; then
@@ -50,8 +73,16 @@ fi
 PERMUTE_SLOW=${SIMINT_MAXAM}
 GITHUB_USERID=edoapra
 rm -rf simint.l${SIMINT_MAXAM}_p${PERMUTE_SLOW}_d${DERIVE}* *-chem-simint-generator-?????? simint-chem-simint-generator.tar.gz simint_lib
-curl -L https://github.com/${GITHUB_USERID}/simint-generator/tarball/master -o simint-chem-simint-generator.tar.gz
-#curl -LJ https://github.com/simint-chem/simint-generator/tarball/master -o simint-chem-simint-generator.tar.gz
+
+GITHUB_URL=https://github.com/${GITHUB_USERID}/simint-generator/tarball/master
+#GITHUB_URL=https://github.com/simint-chem/simint-generator/tarball/master
+TAR_NAME=simint-chem-simint-generator.tar.gz
+if  [ ! -z "$(command -v curl)" ] ; then
+    curl -L "${GITHUB_URL}" -o "${TAR_NAME}"
+else
+    wget -O "${TAR_NAME}" "${GITHUB_URL}"
+fi
+
 tar xzf simint-chem-simint-generator.tar.gz
 cd *-simint-generator-???????
 rm -f generator_types.patch
@@ -112,7 +143,10 @@ if [[ ${CMAKE_VER} -lt 3 ]]; then
     echo define the CMAKE env. variable
     exit 1
 fi
-$CMAKE ../
+if [[ -z "${SIMINT_BUILD_TYPE}" ]]; then
+    SIMINT_BUILD_TYPE=Release
+fi
+$CMAKE  -DCMAKE_BUILD_TYPE="${SIMINT_BUILD_TYPE}"  ../
 make -j2
 cd ..
 #./create.py -g build/generator/ostei -l 6 -p 4 -d 1 simint.l6_p4_d1
@@ -167,9 +201,10 @@ elif  [ ${FC} == xlf ] || [ ${FC} == xlf_r ] || [ ${FC} == xlf90 ]|| [ ${FC} ==
     Fortran_FLAGS=" -qintsize=8 -qextname -qpreprocess"
 elif  [ ${FC} == ifort ]; then
     Fortran_FLAGS="-i8 -fpp"
-fi
-if [[ -z "${SIMINT_BUILD_TYPE}" ]]; then
-    SIMINT_BUILD_TYPE=Release
+elif  [ ${FC} == nvfortran ] || [ ${FC} == pgf90 ] ; then
+    Fortran_FLAGS="-i8 -cpp"
+    CC=gcc
+    CXX=g++
 fi
 echo Fortran_FLAGS equal "$Fortran_FLAGS"
 FC="${FC}" CXX="${CXX}" $CMAKE \

diff --git a/src/NWints/simint/source/nwcsim_facef90.F b/src/NWints/simint/source/nwcsim_facef90.F
@@ -102,6 +102,7 @@ subroutine nwcsim_init(rtdb,nbas,bases,num_der)
      &              call errquit(pname//'Exiting ',8, BASIS_ERR)
                if (.not. geom_cent_get(geom, iat, tag,
      &              coord, q))call errquit
+     &              (pname//'Exiting ',9, GEOM_ERR)
                nwcsim_noshell(bas)=nwcsim_noshell(bas)+1
                call simint_initialize_shell(
      S              smnt_sh(nwcsim_noshell(bas),bas))
@@ -124,10 +125,8 @@ subroutine nwcsim_init(rtdb,nbas,bases,num_der)
 c
 c     memory allocation
 c
-         isz_2e4c = max(isz_2e4c,
-     S        simint_eri_worksize(num_der, max_ang))
          mem_2e4c = max(mem_2e4c,
-     S        simint_eri_workmem(num_der, max_ang))
+     S        simint_eri_worksize(num_der, max_ang))
       enddo ! basis loop
       endif
 c
@@ -136,16 +135,16 @@ subroutine nwcsim_init(rtdb,nbas,bases,num_der)
 c     SIMINT_PRIM_SCREEN_STAT needs 4 more doubles
 c      isz_2e4c        = isz_2e4c + 4
 c      
-      call util_align(isz_2e4c,SIMINT_SIMD_LEN)
-      call util_align(mem_2e4c,SIMINT_SIMD_LEN)
 c
-      iszb_2e4c=isz_2e4c
+c      iszb_2e4c=isz_2e4c
 
       if(num_der.eq.1) then
          memb_2e4c = mem_2e4c + mem_2e4c/5
       else
          memb_2e4c = mem_2e4c + mem_2e4c/10 ! +10% to be safe
       endif
+      call util_align(isz_2e4c,SIMINT_SIMD_LEN)
+      call util_align(mem_2e4c,SIMINT_SIMD_LEN)
       call util_align(memb_2e4c,SIMINT_SIMD_LEN)
       mem_2e3c = mem_2e4c
       mem_2e2c = mem_2e4c

diff --git a/src/basis/bas_input.F b/src/basis/bas_input.F
@@ -358,7 +358,7 @@ subroutine bas_input_body(basis, osegment, oIs_rel, oHas_Star)
       integer nltypes           ! No. of known angular momentum types
       integer nsptypes          ! No. of known sp type shells
       integer nopts             ! No. of options
-      parameter (nltypes = 7, nsptypes = 3, nopts = 6)
+      parameter (nltypes = 8, nsptypes = 3, nopts = 6)
       character*1 ltypes(nltypes)
       character*2 sptypes(nsptypes)
       character*8 opts(nopts)
@@ -374,7 +374,7 @@ subroutine bas_input_body(basis, osegment, oIs_rel, oHas_Star)
 cc AJL/Begin/SPIN-POLARISED ECPs
       integer channel           ! Both (Default)=0; Alpha=1; Beta=2
 cc AJL/End
-      data ltypes /'s','p','d','f','g','h','i'/
+      data ltypes /'s','p','d','f','g','h','i','k'/
       data sptypes / 'sp', 'l ', 'ul'/
       data spvalues/  -1 , -1 ,   -1 /
       data opts / 'except', 'library', 'file', 'rel', 'nelec' ,

diff --git a/src/basis/basis.F b/src/basis/basis.F
@@ -658,7 +658,7 @@ logical function bas_print(basisin)
       integer atn, len_tag, len_ele
       character*2 symbol
       character*16 element
-      character*3 ctype(0:6),cltype(2)
+      character*3 ctype(0:7),cltype(2)
       character*3 shell_type
 *. . . . . . . . . . . ! Room for tag+space+(+element+) = 16+1+1+16+1
       character*35 buffer  
@@ -683,6 +683,7 @@ logical function bas_print(basisin)
       ctype(4)='G'
       ctype(5)='H'
       ctype(6)='I'
+      ctype(7)='K'
       cltype(1)='SP'
       cltype(2)='SPD'
       bas_print = .true.

diff --git a/src/ccsd/GNUmakefile b/src/ccsd/GNUmakefile
@@ -109,6 +109,12 @@ ifeq ($(HAVE_SET_GA_PROPERTY),Y)
       LIB_DEFINES += -DHAVE_SET_GA_PROPERTY
 endif
 
+ifdef USE_OPENACC_TRPDRV
+  OBJ_OPTIMIZE += ccsd_trpdrv_openacc.o
+  USES_BLAS    += ccsd_trpdrv_openacc.F
+  FOPTIONS += -acc -gpu=managed -cuda -cudalib=cublas
+endif
+
 
 ifeq ($(ARMCI_NETWORK),MPI-PR)
    LIB_DEFINES += -DACC_STRIPS

diff --git a/src/ccsd/aoccsd2.F b/src/ccsd/aoccsd2.F
@@ -16,6 +16,7 @@ subroutine aoccsd(basis,ncor,nocc,nvir,ndel,nact,nbf,maxit,
       logical oconverged, occd, use_trpdrv_nb
       logical use_trpdrv_omp, use_trpdrv_bgp2
       logical use_trpdrv_omp_mp
+      logical use_trpdrv_openacc
       logical use_trpdrv_offload
 c
 #include "ccsd_len.fh"
@@ -717,6 +718,9 @@ subroutine ccsd_iterdrv2(rtdb,basis,nsh,ncor,nocc,nvir,nact,nbf,
       if (.not. rtdb_get(rtdb, 'ccsd:use_trpdrv_omp_mp', mt_log, 1,
      1                   use_trpdrv_omp_mp))
      2    use_trpdrv_omp_mp=.false.
+      if (.not. rtdb_get(rtdb, 'ccsd:use_trpdrv_openacc', mt_log, 1,
+     1                   use_trpdrv_openacc))
+     2    use_trpdrv_openacc=.false.
       if (.not. rtdb_get(rtdb, 'ccsd:use_trpdrv_offload', mt_log, 1,
      1                   use_trpdrv_offload))
      2    use_trpdrv_offload=.false.
@@ -976,6 +980,31 @@ subroutine ccsd_iterdrv2(rtdb,basis,nsh,ncor,nocc,nvir,nact,nbf,
      $        dbl_mb(k_trp_Kka), dbl_mb(k_trp_Jij), dbl_mb(k_trp_Jkj),
      $        dbl_mb(k_trp_Kij), dbl_mb(k_trp_Kkj), dbl_mb(k_trp_Dja),
      $        dbl_mb(k_trp_Djka), dbl_mb(k_trp_Djia))
+c
+         else if (use_trpdrv_openacc) then
+#if defined(USE_OPENACC_TRPDRV)
+#ifndef USE_F90_ALLOCATABLE
+#error You must set USE_F90_ALLOCATABLE if USE_OPENACC_TRPDRV is set!
+#endif
+         if (iam.eq.0.and.oprint) then
+            write(luout,1808) nvpass,util_wallsec()
+            call util_flush(luout)
+         endif
+ 1808    format(' commencing triples evaluation - openacc version',i8,
+     I        ' at ',f20.2,' secs')
+         call ccsd_trpdrv_openacc(dbl_mb(k_t1),
+     $        f1n,f1t,f2n,f2t,f3n,f3t,f4n,f4t,
+     $        eorb,g_objo,g_objv,g_coul,g_exch,ncor,nocc,nvir,iprt,
+     $        empt(1),empt(2),oseg_lo,oseg_hi,kchunk,
+     $        dbl_mb(k_trp_Tij), dbl_mb(k_trp_Tkj), dbl_mb(k_trp_Tia),
+     $        dbl_mb(k_trp_Tka), dbl_mb(k_trp_Xia), dbl_mb(k_trp_Xka),
+     $        dbl_mb(k_trp_Jia), dbl_mb(k_trp_Jka), dbl_mb(k_trp_Kia),
+     $        dbl_mb(k_trp_Kka), dbl_mb(k_trp_Jij), dbl_mb(k_trp_Jkj),
+     $        dbl_mb(k_trp_Kij), dbl_mb(k_trp_Kkj), dbl_mb(k_trp_Dja),
+     $        dbl_mb(k_trp_Djka), dbl_mb(k_trp_Djia))
+#else
+         call errquit('aoccsd: trpdrv_openacc disabled ',0,0)
+#endif
 c
          elseif (use_trpdrv_omp_mp) then
 #ifndef TRPMIXP_OFF