From 99ae590e4dd23fe155ba290fdad5d16cf68d84f2 Mon Sep 17 00:00:00 2001
From: fabrizio22 <ferrariruffino.fz@gmail.com>
Date: Fri, 22 Oct 2021 13:44:53 +0200
Subject: [PATCH] beef_acc - first step

---
 XClib/beefun.c              |  9 +++++++--
 XClib/qe_drivers_gga.f90    | 17 +++++++++++++----
 XClib/xc_beef_interface.f90 |  1 +
 3 files changed, 21 insertions(+), 6 deletions(-)

diff --git a/XClib/beefun.c b/XClib/beefun.c
index b592be90d..d10b58b3d 100755
--- a/XClib/beefun.c
+++ b/XClib/beefun.c
@@ -10,6 +10,7 @@
 #include "pbecor.h"
 
 // evaluate bee exchange energy and its derivatives de/drho and ( de/d|grad rho| ) / |grad rho|
+#pragma acc routine seq
 void beefx_(double *r, double *g, double *e, double *dr, double *dg, int *addlda)
 {
     double s2,t,r43,r83,s,sx,dx,fx,dl,dfx;
@@ -17,7 +18,11 @@ void beefx_(double *r, double *g, double *e, double *dr, double *dg, int *addlda
     const int i1=1;
     const int i2=1;
 
-    switch(beeftype) {
+    *e=6.;
+    *dr=7.;
+    *dg=8.;
+
+/*    switch(beeftype) {
     case 0: //BEEF-vdW xc    
     r43 = pow(*r, 4./3.);
     r83 = r43*r43;
@@ -62,7 +67,7 @@ void beefx_(double *r, double *g, double *e, double *dr, double *dg, int *addlda
     }
     
     break;
-    }
+    } */
 }
 
 
diff --git a/XClib/qe_drivers_gga.f90 b/XClib/qe_drivers_gga.f90
index 0bddeb8cf..93fbf2001 100644
--- a/XClib/qe_drivers_gga.f90
+++ b/XClib/qe_drivers_gga.f90
@@ -1106,6 +1106,8 @@ SUBROUTINE gcxc_beef( length, rho_in, grho_in, sx_out, sc_out, v1x_out, &
   !---------------------------------------------------------------------
   !! Driver for BEEF gga xc terms. Unpolarized case.
   !
+  ! ****note: to compile add -ta=tesla::cc60 to both f90 and cc flags
+  
   USE beef_interface, ONLY: beefx, beeflocalcorr
   !
   IMPLICIT NONE
@@ -1126,6 +1128,10 @@ SUBROUTINE gcxc_beef( length, rho_in, grho_in, sx_out, sc_out, v1x_out, &
   !
   IF (igcx == 43 .OR. igcc == 14) THEN
     !
+#if defined(_OPENACC)
+    !$acc data copyin(rho_in,grho_in), copyout(sx_out,sc_out,v1x_out,v2x_out,v1c_out,v2c_out)
+    !$acc parallel loop
+#endif
     DO ir = 1, length
       grho = grho_in(ir)
       IF ( rho_in(ir) <= rho_threshold_gga .OR. grho <= grho_threshold_gga ) THEN
@@ -1149,12 +1155,15 @@ SUBROUTINE gcxc_beef( length, rho_in, grho_in, sx_out, sc_out, v1x_out, &
       IF ( igcc == 14 ) THEN
         ! last parameter 0 means: do not add lda contributions
         ! espresso will do that itself
-        CALL beeflocalcorr( rho, grho, sc, v1c, v2c, 0 )
-        sc_out(ir) = sc
-        v1c_out(ir) = v1c
-        v2c_out(ir) = v2c
+!        CALL beeflocalcorr( rho, grho, sc, v1c, v2c, 0 )
+!        sc_out(ir) = sc
+!        v1c_out(ir) = v1c
+!        v2c_out(ir) = v2c
       ENDIF
     ENDDO
+#if defined(_OPENACC)
+    !$acc end data
+#endif
     !
   ENDIF
   !
diff --git a/XClib/xc_beef_interface.f90 b/XClib/xc_beef_interface.f90
index b5ac12cda..0c429de5e 100644
--- a/XClib/xc_beef_interface.f90
+++ b/XClib/xc_beef_interface.f90
@@ -22,6 +22,7 @@ MODULE beef_interface
     INTERFACE
     !
     SUBROUTINE beefx( r, g, e, dr, dg, addlda ) BIND(C, NAME="beefx_")
+    !$acc routine seq
     USE iso_c_binding
         REAL (C_DOUBLE)            :: r, g, e, dr, dg
         INTEGER(C_INT), INTENT(IN) :: addlda