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

opt[overhead,acc]: present and deviceptr moved outside the loop 

This way the compiler checks only once if data are on the GPU
Reduces wait status in openacc regions
This commit is contained in:
Laura Bellentani 2024-08-22 13:23:26 +02:00
parent d3e2b4d3b0
commit 31de484af1
1 changed files with 3 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
PW/src/add_vuspsi_gpu.f90
View File

@ -240,6 +240,7 @@ SUBROUTINE add_vuspsi_gpu( lda, n, m, hpsi_d )
becp_k = becp%k
!$acc end kernels
!
!$acc data present(deeq) deviceptr(deeaux_d)
DO nt = 1, ntyp
!
IF ( nh(nt) == 0 ) CYCLE
@ -255,7 +256,7 @@ SUBROUTINE add_vuspsi_gpu( lda, n, m, hpsi_d )
!
!deeaux_d(:,:) = CMPLX(deeq(1:nh(nt),1:nh(nt),na,current_spin), 0.0_dp, KIND=dp )
!
!$acc parallel loop collapse(2) present(deeq)
!$acc parallel loop collapse(2)
DO j = 1, nhnt
DO k = 1, nhnt
deeaux_d(k,j) = CMPLX(deeq(k,j,na,current_spin), 0.0_dp, KIND=DP )
@ -273,6 +274,7 @@ SUBROUTINE add_vuspsi_gpu( lda, n, m, hpsi_d )
END DO
!
END DO
!$acc end data
CALL dev_buf%release_buffer(deeaux_d, ierr) ! DEALLOCATE (deeaux_d)
!
!$acc host_data use_device(vkb)