From 5bebcea5b407d6a48992b4f18232868c99535948 Mon Sep 17 00:00:00 2001
From: fabrizio22 <ferrariruffino.fz@gmail.com>
Date: Fri, 3 May 2024 14:31:30 +0200
Subject: [PATCH] Vpaw_acc - offload switch removed in radial_gradient

---
 Modules/radial_gradients.f90 | 18 +-----------------
 PW/src/paw_onecenter.f90     | 13 ++++++++++---
 2 files changed, 11 insertions(+), 20 deletions(-)

diff --git a/Modules/radial_gradients.f90 b/Modules/radial_gradients.f90
index 293ab0155..c166507b2 100644
--- a/Modules/radial_gradients.f90
+++ b/Modules/radial_gradients.f90
@@ -6,7 +6,7 @@
 ! or http://www.gnu.org/copyleft/gpl.txt .
 !
 !------------------------------------------------------------------------------
-SUBROUTINE radial_gradient( f, gf, r, mesh, iflag, offld_ )
+SUBROUTINE radial_gradient( f, gf, r, mesh, iflag )
   !---------------------------------------------------------------------------
   !! This subroutine calculates the derivative with respect to r of a
   !! radial function defined on the mesh r.
@@ -23,34 +23,22 @@ SUBROUTINE radial_gradient( f, gf, r, mesh, iflag, offld_ )
   INTEGER, INTENT(IN) :: mesh, iflag
   REAL(DP), INTENT(IN) :: f(mesh), r(mesh)
   REAL(DP), INTENT(OUT) :: gf(mesh)
-  LOGICAL, INTENT(IN), OPTIONAL :: offld_
   !
   INTEGER :: i, j, k, imin, npoint
   REAL(DP) :: delta, b(5), faux(6), raux(6)
-  LOGICAL :: offld
   !
   ! ... This formula is used in the all-electron case.
   !
-  offld = .FALSE.
-  IF (PRESENT(offld_)) offld = offld_
-  !
   IF (iflag==0) THEN
      !
-     !$acc data copyin(f,r) copyout(gf) if(offld)
-     !
-     !$acc parallel loop if(offld)
      DO i = 2, mesh-1
         gf(i) = ( (r(i+1)-r(i))**2*(f(i-1)-f(i)) &
                    -(r(i-1)-r(i))**2*(f(i+1)-f(i)) ) &
                    / ((r(i+1)-r(i))*(r(i-1)-r(i))*(r(i+1)-r(i-1)))
      ENDDO
      !
-     !$acc parallel if(offld)
      gf(mesh) = 0.0_DP
      gf(1) = gf(2) + (gf(3)-gf(2)) * (r(1)-r(2)) / (r(3)-r(2))
-     !$acc end parallel
-     !
-     !$acc end data
      !
   ELSE
      !
@@ -64,8 +52,6 @@ SUBROUTINE radial_gradient( f, gf, r, mesh, iflag, offld_ )
      ! ... and this formula coincides with the previous one.
      ! ... (ADC 08/2007)
      !
-     !$acc update self(f) if(offld)
-     !
      delta = 0.00001_DP
      imin = 1
      !
@@ -126,8 +112,6 @@ SUBROUTINE radial_gradient( f, gf, r, mesh, iflag, offld_ )
        gf(i) = b(1)+r(i)*(b(2)+r(i)*(b(3)+r(i)*b(4)))
     ENDDO
     !
-    !$acc update device(gf) if(offld)
-    !
   ENDIF
   !
   RETURN
diff --git a/PW/src/paw_onecenter.f90 b/PW/src/paw_onecenter.f90
index 99ae7ef46..25b8787b7 100644
--- a/PW/src/paw_onecenter.f90
+++ b/PW/src/paw_onecenter.f90
@@ -1004,12 +1004,15 @@ MODULE paw_onecenter
       ENDDO
     ENDDO
     !
+    !$acc update self(F_lm(:,1:1,:,:))
+    !
     ! ... Compute partial radial derivative d/dr
     DO is = 1, nspin_gga
       DO lm = 1, lmaxq_out**2
         ! ... Derive along \hat{r} (F already contains a r**2 factor, otherwise
         ! ... it may be better to expand (1/r**2) d(A*r**2)/dr = dA/dr + 2A/r)
-        CALL radial_gradient( F_lm(1:i%m,1,lm,is), aux, g(i%t)%r, i%m, radial_grad_style, .TRUE. )
+        CALL radial_gradient( F_lm(1:i%m,1,lm,is), aux, g(i%t)%r, i%m, radial_grad_style )
+        !$acc update device(aux)
         !
         ! ... Sum it in the divergence: it is already in the right Y_lm form
         !
@@ -1093,11 +1096,13 @@ MODULE paw_onecenter
         ENDDO
       ENDDO
       !
+      !$acc update self(aux)
       DO ix = 1, nx_loc
         ixs = (ix-1)*i%m+1
         ixe = ix*i%m
-        CALL radial_gradient( aux(ixs:ixe), aux2(ixs:ixe), g(i%t)%r, i%m, radial_grad_style, .TRUE. )
+        CALL radial_gradient( aux(ixs:ixe), aux2(ixs:ixe), g(i%t)%r, i%m, radial_grad_style )
       ENDDO
+      !$acc update device(aux2)
       !
       !$acc parallel loop collapse(2) present(rad(i%t:i%t),g(i%t:i%t))
       DO ix = 1, nx_loc
@@ -1368,7 +1373,9 @@ MODULE paw_onecenter
   !--------------------------------------------------------------------------------
   SUBROUTINE PAW_rad2lm( i, F_rad, F_lm, lmax_loc, nspin )
     !------------------------------------------------------------------------------
-    !! gpu double
+    !! Computes:
+    !! \[ F_{lm}(r) = \int d \Omega\ F(r,\text{th},\text{ph})\ Y_{lm}(\text{th},
+    !! \text{ph}) \]
     !
     IMPLICIT NONE
     !