stress_acc - gg array on acc

Modules/recvec.f90

@@ -120,7 +120,7 @@
           ALLOCATE( igtongl_d(ngm) )
           ALLOCATE( gl_d(ngm) )
        ENDIF  
-       !$acc enter data create( mill(1:3, 1:ngm), g(1:3, 1:ngm) ) 
+       !$acc enter data create( mill(1:3,1:ngm), g(1:3,1:ngm), gg(1:ngm) ) 
        !
        RETURN 
        !
@@ -142,7 +142,10 @@
        END IF
        !
        !  
-       IF( ALLOCATED( gg ) )     DEALLOCATE( gg )
+       IF( ALLOCATED( gg ) ) THEN
+!$acc    exit data delete(gg)
+       DEALLOCATE( gg )
+       END IF
        IF( ALLOCATED( g ) )  THEN 
 !$acc    exit data delete(g) 
          DEALLOCATE( g )

PW/src/Coul_cut_2D.f90

@@ -795,7 +795,7 @@ SUBROUTINE cutoff_stres_sigmaewa( alpha, sdewald, sigmaewa )
   !
   sdewald = 0._DP
   !
-  !$acc parallel loop copyin(g,gg,cutoff_2D,tau,zv,ityp) &
+  !$acc parallel loop copyin(cutoff_2D,tau,zv,ityp) &
   !$acc& reduction(+:sigma11,sigma21,sigma22,sigma31,sigma32, &
   !$acc&             sigma33)
   DO ng = gstart, ngm

PW/src/init_run.f90

@@ -98,7 +98,7 @@ SUBROUTINE init_run()
      gg_d   = gg
   END IF
 #endif
-  !$acc update device(mill, g)
+  !$acc update device(mill, g, gg)
   !
   IF (do_comp_esm) CALL esm_init(.NOT. lrism)
   !

PW/src/read_file_new.f90

@@ -230,10 +230,10 @@ SUBROUTINE post_xml_init (  )
   g_d    = g
   gg_d   = gg
 #endif
-  !$acc update device(mill, g)
+  !$acc update device(mill, g, gg)
   !
   CALL ggens( dffts, gamma_only, at, g, gg, mill, gcutms, ngms ) 
-  CALL gshells ( lmovecell ) 
+  CALL gshells ( lmovecell )
   !
   IF (do_comp_esm) CALL esm_init()
   IF (do_cutoff_2D) CALL cutoff_fact()

PW/src/stres_ewa.f90

@@ -87,7 +87,7 @@ SUBROUTINE stres_ewa( alat, nat, ntyp, ityp, zv, at, bg, tau, &
   COMPLEX(DP) :: rhostar
   REAL(DP) :: sigma11, sigma21, sigma22, sigma31, sigma32, sigma33
   !
-  !$acc data present_or_copyin( g, gg )
+  !$acc data present( g, gg )
   !
   tpiba2 = (tpi / alat)**2
   sigmaewa(:,:) = 0.d0
@@ -109,23 +109,22 @@ SUBROUTINE stres_ewa( alat, nat, ntyp, ityp, zv, at, bg, tau, &
   !
   IF (upperbound > 1d-7) GOTO 12
   !
-  ! G-space sum here
-  !
-  ! Determine if this processor contains G=0 and set the constant term
-  !
+  ! ... Determine if this processor contains G=0 and set the constant term
+  !     sdewald is the diagonal term
   IF (gstart == 2) THEN
      sdewald = tpi * e2 / 4.d0 / alpha * (charge / omega)**2
   ELSE
      sdewald = 0.d0
   ENDIF
   !
-  ! sdewald is the diagonal term
   IF (gamma_only) THEN
     fact = 2.d0
   ELSE
     fact = 1.d0
   ENDIF
   !
+  ! ... G-space sum here below
+  !
   IF (do_cutoff_2D) THEN 
      !
      CALL cutoff_stres_sigmaewa( alpha, sdewald, sigmaewa )
@@ -189,14 +188,14 @@ SUBROUTINE stres_ewa( alat, nat, ntyp, ityp, zv, at, bg, tau, &
      sigmaewa(l,l) = sigmaewa(l,l) + sdewald
   ENDDO
   !
-  ! R-space sum here (see ewald.f90 for details on parallelization)
+  ! ... R-space sum here (see ewald.f90 for details on parallelization)
   !
   CALL block_distribute( nat, me_bgrp, nproc_bgrp, na_s, na_e, mykey )
   !
   IF ( mykey == 0 ) THEN
      rmax = 4.0d0 / SQRT(alpha) / alat
      !
-     ! with this choice terms up to ZiZj*erfc(5) are counted (erfc(5)=2x10^-1
+     ! ... with this choice terms up to ZiZj*erfc(5) are counted (erfc(5)=2x10^-1
      !
 !$omp parallel do default(none) shared(na_s, na_e, nat, tau, rmax, at, bg, alat, ityp, alpha, omega, zv)&
 !$omp                          &private(nb, dtau, r, r2, nrm, nr, rr, fac, l, m)&
@@ -205,7 +204,7 @@ SUBROUTINE stres_ewa( alat, nat, ntyp, ityp, zv, at, bg, tau, &
         DO nb = 1, nat
            dtau(:) = tau(:,na) - tau(:,nb)
            !
-           !     generates nearest-neighbors shells r(i)=R(i)-dtau(i)
+           ! ... generates nearest-neighbors shells r(i)=R(i)-dtau(i)
            !
            CALL rgen( dtau, rmax, mxr, at, bg, r, r2, nrm )
            !

PW/src/stress.f90

@@ -70,6 +70,10 @@ SUBROUTINE stress( sigma )
   !
   CALL start_clock( 'stress' )
   !
+  ! --------------- ... provisional ... ---------------
+  !$acc update device( g, gg )
+  ! -------------------------------------------
+  !
   !   contribution from local potential
   !
   IF (.NOT. use_gpu) CALL stres_loc( sigmaloc )
@@ -88,13 +92,13 @@ SUBROUTINE stress( sigma )
   ELSE
      IF (.NOT. use_gpu) CALL stres_har( sigmahar )
      IF (      use_gpu) CALL stres_har_gpu( sigmahar )
-  END IF
+  ENDIF
   !
   !  xc contribution (diagonal)
   !
   sigmaxc(:,:) = 0.d0
   DO l = 1, 3
-     sigmaxc (l, l) = - (etxc - vtxc) / omega
+     sigmaxc(l,l) = - (etxc - vtxc) / omega
   ENDDO
   !
   !  xc contribution: add gradient corrections (non diagonal)