XClib - beef gnu+acc compatibility and small cleanup

PW/src/v_of_rho.f90

@@ -254,8 +254,8 @@ SUBROUTINE v_xc_meta( rho, rho_core, rhog_core, etxc, vtxc, v, kedtaur )
     CALL xc_metagcx( dfftp%nnr, 1, np, rho%of_r, grho, tau, ex, ec, &
                      v1x, v2x, v3x, v1c, v2c, v3c, gpu_args_=.TRUE. )
     !
-    !$acc parallel loop reduction(+:etxc) reduction(+:vtxc) reduction(-:rhoneg1) &
-    !$acc&              reduction(-:rhoneg2) present(rho)
+    !$acc parallel loop reduction(+:etxc,vtxc) reduction(-:rhoneg1,rhoneg2) &
+    !$acc&              present(rho)
     DO k = 1, dfftp%nnr
        !
        v(k,1) = (v1x(k,1)+v1c(k,1)) * e2
@@ -290,8 +290,7 @@ SUBROUTINE v_xc_meta( rho, rho_core, rhog_core, etxc, vtxc, v, kedtaur )
     !
     ! ... first term of the gradient correction : D(rho*Exc)/D(rho)
     !
-    !$acc parallel loop reduction(+:etxc) reduction(+:vtxc) reduction(-:rhoneg1) &
-    !$acc&              reduction(-:rhoneg2)
+    !$acc parallel loop reduction(+:etxc,vtxc) reduction(-:rhoneg1,rhoneg2)
     DO k = 1, dfftp%nnr
        !
        v(k,1) = (v1x(k,1) + v1c(k,1)) * e2
@@ -474,8 +473,7 @@ SUBROUTINE v_xc( rho, rho_core, rhog_core, etxc, vtxc, v )
      !
      CALL xc( dfftp%nnr, 1, 1, rho%of_r, ex, ec, vx, vc, gpu_args_=.TRUE. )
      !
-     !$acc parallel loop reduction(+:etxc) reduction(+:vtxc) reduction(-:rhoneg1) &
-     !$acc&              present(rho)
+     !$acc parallel loop reduction(+:etxc,vtxc) reduction(-:rhoneg1) present(rho)
      DO ir = 1, dfftp%nnr
         v(ir,1) = e2*( vx(ir,1) + vc(ir,1) )
         etxc = etxc + e2*( ex(ir) + ec(ir) )*rho%of_r(ir,1)
@@ -490,8 +488,8 @@ SUBROUTINE v_xc( rho, rho_core, rhog_core, etxc, vtxc, v )
      !
      CALL xc( dfftp%nnr, 2, 2, rho%of_r, ex, ec, vx, vc, gpu_args_=.TRUE. )
      !
-     !$acc parallel loop reduction(+:etxc) reduction(+:vtxc) reduction(-:rhoneg1) &
-     !$acc&              reduction(-:rhoneg2) present(rho)
+     !$acc parallel loop reduction(+:etxc,vtxc) reduction(-:rhoneg1,rhoneg2) &
+     !$acc&              present(rho)
      DO ir = 1, dfftp%nnr
         v(ir,1) = e2*( vx(ir,1) + vc(ir,1) )
         v(ir,2) = e2*( vx(ir,2) + vc(ir,2) )
@@ -511,8 +509,8 @@ SUBROUTINE v_xc( rho, rho_core, rhog_core, etxc, vtxc, v )
       !
       CALL xc( dfftp%nnr, 4, 2, rho%of_r, ex, ec, vx, vc, gpu_args_=.TRUE. )
       !
-      !$acc parallel loop reduction(+:etxc) reduction(+:vtxc) reduction(-:rhoneg1) &
-      !$acc&              reduction(+:rhoneg2) present(rho)
+      !$acc parallel loop reduction(+:etxc,vtxc) reduction(-:rhoneg1,rhoneg2) &
+      !$acc&              present(rho)
       DO ir = 1, dfftp%nnr
          arho = ABS( rho%of_r(ir,1) )
          IF ( arho < vanishing_charge ) THEN

XClib/beefleg.h

@@ -8,10 +8,10 @@
 
 #define nmax 30
 
-#pragma acc routine (dgemv_) seq
 extern void dgemv_(const char *, const int *, const int *, const double *,
     double *, const int *, double *, const int *, const double *, double *,
     const int *);
+#pragma acc routine (dgemv_) seq
 
 #pragma acc routine seq
 double ddot1(double v[], double u[], int n)

XClib/qe_drivers_d_lda_lsda.f90

@@ -587,7 +587,6 @@ END SUBROUTINE dmxc_nc
 !
 !-----------------------------------------------------------------------
 FUNCTION dpz( rs, iflg )
-!$acc routine seq
   !-----------------------------------------------------------------------
   !!  Derivative of the correlation potential with respect to local density
   !!  Perdew and Zunger parameterization of the Ceperley-Alder functional.
@@ -596,6 +595,8 @@ FUNCTION dpz( rs, iflg )
   !
   IMPLICIT NONE
   !
+  !$acc routine seq
+  !
   REAL(DP), INTENT(IN) :: rs
   INTEGER,  INTENT(IN) :: iflg
   REAL(DP) :: dpz
@@ -627,7 +628,6 @@ END FUNCTION dpz
 !
 !-----------------------------------------------------------------------
 FUNCTION dpz_polarized( rs, iflg )
-!$acc routine seq
   !-----------------------------------------------------------------------
   !! Derivative of the correlation potential with respect to local density
   !! Perdew and Zunger parameterization of the Ceperley-Alder functional.  |
@@ -637,6 +637,8 @@ FUNCTION dpz_polarized( rs, iflg )
   !
   IMPLICIT NONE
   !
+  !$acc routine seq
+  !
   REAL(DP), INTENT(IN) :: rs
   INTEGER,  INTENT(IN) :: iflg
   REAL(DP) :: dpz_polarized

XClib/xc_beef_interface.f90

@@ -22,33 +22,33 @@ MODULE beef_interface
     INTERFACE
     !
     SUBROUTINE beefx( r, g, e, dr, dg, addlda ) BIND(C, NAME="beefx_")
-    !$acc routine seq
-    USE iso_c_binding
+        USE iso_c_binding
+        !$acc routine seq
         REAL (C_DOUBLE)            :: r, g, e, dr, dg
         INTEGER(C_INT), INTENT(IN) :: addlda
     END SUBROUTINE beefx
     !
     SUBROUTINE beeflocalcorr( r, g, e, dr, dg, addlda) BIND(C, NAME="beeflocalcorr_")
-    !$acc routine seq
-    USE iso_c_binding
+        USE iso_c_binding
+        !$acc routine seq
         REAL (C_DOUBLE), INTENT(INOUT) :: r, g, e, dr, dg
         INTEGER(C_INT), INTENT(IN) :: addlda
     END SUBROUTINE beeflocalcorr
     !
     SUBROUTINE beeflocalcorrspin(r, z, g, e, drup, drdown, dg, addlda) BIND(C, NAME="beeflocalcorrspin_")
-    !$acc routine seq
-    USE iso_c_binding
+        USE iso_c_binding
+        !$acc routine seq
         REAL (C_DOUBLE), INTENT(INOUT) :: r, z, g, e, drup, drdown, dg
         INTEGER(C_INT), INTENT(IN) :: addlda
     END SUBROUTINE beeflocalcorrspin
     !
     SUBROUTINE beefsetmode(mode) BIND(C, NAME="beefsetmode_")
-    USE iso_c_binding
+        USE iso_c_binding
         INTEGER(C_INT), INTENT(IN) :: mode
     END SUBROUTINE beefsetmode
     !
     SUBROUTINE beefrandinit(seed) BIND(C, NAME="beefrandinit_")
-    USE iso_c_binding
+        USE iso_c_binding
         INTEGER(C_INT), INTENT(IN) :: seed
     END SUBROUTINE beefrandinit
     !
@@ -56,7 +56,7 @@ MODULE beef_interface
     END SUBROUTINE beefrandinitdef
     !
     SUBROUTINE beefensemble(beefxc, ensemble) BIND(C, NAME="beefensemble_")
-    USE iso_c_binding
+        USE iso_c_binding
         REAL (C_DOUBLE), INTENT(INOUT) :: beefxc(*), ensemble(*)
     END SUBROUTINE beefensemble
     !
@@ -102,27 +102,27 @@ MODULE beef_interface
     ! and not in the above interfaces.
     !
     SUBROUTINE beefx( r, g, e, dr, dg, addlda )
-      !$acc routine seq
       !! Evaluate bee exchange energy and its derivatives \(d\epsilon/d\rho\) and 
       !! \( (d\epsilon/d|\nabla \rho| ) / |\nabla\rho|\).
       USE kind_l, ONLY : dp
+      !$acc routine seq
       REAL (dp) :: r, g, e, dr, dg
       INTEGER :: addlda
     END SUBROUTINE beefx
     !
     SUBROUTINE beeflocalcorr( r, g, e, dr, dg, addlda)
-      !$acc routine seq
       !! Evaluate local part of bee correlation and its derivatives \(d\epsilon/drho\)
       !! and \( (d\epsilon/d|\nabla\rho|) / |\nabla\rho| \).
       USE kind_l, ONLY : dp
+      !$acc routine seq
       REAL (dp), INTENT(INOUT) :: r, g, e, dr, dg
       INTEGER :: addlda
     END SUBROUTINE beeflocalcorr
     !
     SUBROUTINE beeflocalcorrspin(r, z, g, e, drup, drdown, dg, addlda)
-      !$acc routine seq
       !! Evaluate local part of bee correlation for spin polarized system.
       USE kind_l, ONLY : dp
+      !$acc routine seq
       REAL (dp), INTENT(INOUT) :: r, z, g, e, drup, drdown, dg
       INTEGER :: addlda
     END SUBROUTINE beeflocalcorrspin