scnc
/
quantum-espresso
mirror of https://gitlab.com/QEF/q-e.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

complying to triplet notation in openacc statements

This commit is contained in:
Pietro Delugas 2021-12-08 22:45:56 +01:00
parent ff90d73b73
commit 7d75915c51
9 changed files with 10 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
CPV/src/gram.f90
View File

@ -5,22 +5,7 @@
! in the root directory of the present distribution,
! or http://www.gnu.org/copyleft/gpl.txt .
!
#if defined (_OPENACC)
#ifndef __OPENACC
#define __OPENACC
#endif
#endif
#if defined (__OPENACC)
#define DEV_ACC !$acc
#define DEV_OMP !!!
#define START_WSHARE DEV_ACC kernels
#define END_WSHARE DEV_ACC end kernels
#else
#define DEV_ACC !!!
#define DEV_OMP !$omp
#define START_WSHARE DEV_OMP workshare
#define END_WSHARE DEV_OMP end workshare
#endif
#include <cpv_device_macros.h>
!-------------------------------------------------------------------------
SUBROUTINE gram_bgrp( betae, bec_bgrp, nkbx, cp_bgrp, ngwx )
!-----------------------------------------------------------------------

2
CPV/src/potentials.f90
View File

@ -357,7 +357,7 @@
!
DEV_ACC data present(rhoeg, rhops, mill,g ) copy(fion) create(rp(s_ngm_)) copyin(sfac, screen_coul, gg, vps, ityp,ei1, ei2, ei3)
DEV_ACC data present(rhoeg, rhops, mill,g ) copy(fion) create(rp(1:s_ngm_)) copyin(sfac, screen_coul, gg, vps, ityp,ei1, ei2, ei3)
!
DEV_OMP parallel default(none) &
DEV_OMP shared(gstart, dffts,sfac, rhops, screen_coul, rhoeg, nsp, gg, tpiba2, tpiba, mill, g, &

2
Modules/fft_rho.f90
View File

@ -123,7 +123,7 @@ CONTAINS
nspin= SIZE(rhor_d, 2)
!
ALLOCATE( psi_d(desc%nnr) )
!$acc data create( psi_d(desc%nnr) )
!$acc data create( psi_d(1:desc%nnr) )
!$acc host_data use_device( psi_d )
IF( nspin == 1 ) THEN
iss=1

2
Modules/recvec.f90
View File

@ -120,7 +120,7 @@
ALLOCATE( igtongl_d(ngm) )
ALLOCATE( gl_d(ngm) )
ENDIF
!$acc enter data create( mill(3, ngm), g(3, ngm) )
!$acc enter data create( mill(1:3, 1:ngm), g(1:3, 1:ngm) )
!
RETURN
!

2
PW/src/allocate_wfc.f90
View File

@ -75,7 +75,7 @@ SUBROUTINE allocate_wfc_k()
!
ALLOCATE( vkb(npwx,nkb) )
#if defined __CUDA
!$acc enter data create(vkb(npwx,nkb) )
!$acc enter data create(vkb(1:npwx,1:nkb) )
#endif
!
! g2kin contains the kinetic energy \hbar^2(k+G)^2/2m

4
PW/src/force_us_gpu.f90
View File

@ -103,7 +103,7 @@ SUBROUTINE force_us_gpu( forcenl )
CALL using_evc_d(0)
DO ipol = 1, 3
!
!$acc data present(vkb(:,:), vkb1(npwx,nkb), igk_k(:,:))
!$acc data present(vkb(:,:), vkb1(1:npwx,1:nkb), igk_k(:,:))
!$acc parallel loop collapse(2)
DO jkb = 1, nkb
DO ig = 1, npw
@ -112,7 +112,7 @@ SUBROUTINE force_us_gpu( forcenl )
ENDDO
!$acc end data
!
!$acc data present(vkb1(npwx,nkb))
!$acc data present(vkb1(1:npwx,1:nkb))
!$acc host_data use_device(vkb1)
CALL calbec_gpu ( npw, vkb1, evc_d, dbecp_d )
!$acc end host_data

2
PW/src/init_us_2.f90
View File

@ -46,7 +46,7 @@ CONTAINS
!
if(use_gpu.and.run_on_gpu) then
!
!$acc data present(igk_(npw_), mill(:,:), g(:,:), vkb_(npwx,nkb), eigts1(:,:), eigts2(:,:), eigts3(:,:))
!$acc data present(igk_(1:npw_), mill(:,:), g(:,:), vkb_(1:npwx,1:nkb), eigts1(:,:), eigts2(:,:), eigts3(:,:))
!$acc host_data use_device(eigts1, eigts2, eigts3, mill, g, igk_, vkb_)
CALL init_us_2_base_gpu(npw_, npwx, igk_, q_, nat, tau, ityp, tpiba, omega,&
dfftp%nr1, dfftp%nr2, dfftp%nr3, eigts1, eigts2, eigts3, mill, g,&

2
PW/src/pwcom.f90
View File

@ -90,7 +90,7 @@ CONTAINS
INTEGER :: ik
!
IF (.NOT.ALLOCATED(igk_k)) ALLOCATE( igk_k(npwx,nks) )
!$acc enter data create(igk_k(npwx,nks))
!$acc enter data create(igk_k(1:npwx,1:nks))
!
IF (.NOT.ALLOCATED(ngk)) ALLOCATE( ngk(nks) )
!

2
dft-d3/core.f90
View File

@ -2905,7 +2905,7 @@ contains
IF ( mykey == 0 ) THEN
na_smax = max(3,na_s)
!$acc data copyin(xyz(3,n),iz(n),cc6ab(n*n),lat(3,3),r0ab(max_elem,max_elem))
!$acc data copyin(xyz(1:3,1:n),iz(1:n),cc6ab(1:n*n),lat(1:3,1:3),r0ab(1:max_elem,1:max_elem))
!$acc kernels vector_length(32)
!$acc loop collapse(2) gang private(ijvec1,ijvec2,ijvec3, ikvec1,ikvec2,ikvec3, jkvec1,jkvec2,jkvec3, c9, r0ij,r0ik,r0jk, &
!$acc& repmin1,repmin2,repmin3, repmax1,repmax2,repmax3, jtau1,jtau2,jtau3, &