Homec4science
Diffusion miniapp_gbs_gradients (test_is_and_use_device_ptr) Browse

R11986/test_gbs_gradients.F90
f16fd9b8f371test_is_and_use_device_…

Last Change
Raw File

test_gbs_gradients.F90
View Options

program main
use space_grid
use gradients
use prec_const
use iso_fortran_env
#ifdef CUDA
use gradients_cuda
#endif
implicit none
! INPUT: Number of cells in each direction
integer :: nx=1000,ny=1000,nz=4
!
real(dp), dimension(:,:,:), allocatable :: fin,fout ! array in input and array in output (computed from different gradients)
real(dp), dimension(:,:,:), allocatable :: fout_omp ! same but computed with openmp
integer :: i,j,k
integer(di) :: startc, endc ! For timing
real(dp) :: rate ! For timing
#ifdef CUDA
real(dp), pointer :: fout_cuda(:,:,:)
real(dp), pointer :: fin_cuda(:,:,:)
#endif
call system_clock(count_rate=rate)
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
! Compute mesh
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
call compute_mesh(nx,ny,nz)
!
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
! Initialize coefficient for gradients (as in gbs)
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
call intialize_coefficients
!
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
! allocate arrays (as in gbs)
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
call gbs_allocate(fin,iysg,iyeg,ixsg,ixeg,izsg,izeg)
call gbs_allocate(fout,iysg,iyeg,ixsg,ixeg,izsg,izeg)
call gbs_allocate(fout_omp,iysg,iyeg,ixsg,ixeg,izsg,izeg)
call gbs_allocate(gradpar_y_n,iysg,iyeg,ixsg,ixeg)
call gbs_allocate(gradpar_x_n,iysg,iyeg,ixsg,ixeg)
!
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
! initialize input and output arrays
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
call init_array(fin)
call init_array(fout)
call init_array(fout_omp)
call init_array2d(gradpar_x_n)
call init_array2d(gradpar_y_n)
!$omp target enter data map (to: fin, gradpar_x_n,gradpar_y_n)
! !$acc enter data copyin(fin, gradpar_x_n,gradpar_y_n)
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
! Test gradients with OpenMP and OpenACC
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
call system_clock(startc)
call gradz_n2n_fd4(fin(:,:,:),fout(:,:,:))
call system_clock(endc)
write(*,*) "timing for gradz_n2n_fd4 CPU: ", real(endc-startc, dp)/rate*1000, "ms"
call system_clock(startc)
call gradz_n2n_fd4_omp(fin(:,:,:),fout_omp(:,:,:))
call system_clock(endc)
write(*,*) "timing for gradz_n2n_fd4 DEVICE: ", real(endc-startc, dp)/rate*1000, "ms"
write(*,*) 'gradz_n2n_fd4 L2_norm and maxval:', &
norm2(fout(iys:iye,ixs:ixe,izs:ize) - fout_omp(iys:iye,ixs:ixe,izs:ize)), &
maxval(fout(iys:iye,ixs:ixe,izs:ize) - fout_omp(iys:iye,ixs:ixe,izs:ize))
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
call init_array(fout)
call init_array(fout_omp)
call system_clock(startc)
CALL gradz_v2n_fd4(fin(:,:,:),fout(:,:,:))
call system_clock(endc)
write(*,*) "timing for gradz_v2n_fd4 CPU: ", real(endc-startc, dp)/rate*1000, "ms"
call system_clock(startc)
CALL gradz_v2n_fd4_omp(fin(:,:,:),fout_omp(:,:,:))
call system_clock(endc)
write(*,*) "timing for gradz_v2n_fd4 DEVICE: ", real(endc-startc, dp)/rate*1000, "ms"
write(*,*) 'gradz_v2n_fd4 L2_norm and maxval:', &
norm2(fout(iys:iye,ixs:ixe,izs:ize) - fout_omp(iys:iye,ixs:ixe,izs:ize)), &
maxval(fout(iys:iye,ixs:ixe,izs:ize) - fout_omp(iys:iye,ixs:ixe,izs:ize))
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
call init_array(fout)
call init_array(fout_omp)
call system_clock(startc)
CALL grady_v2n_fd4(fin(:,:,:),fout(:,:,:))
call system_clock(endc)
write(*,*) "timing for grady_v2n_fd4 CPU: ", real(endc-startc, dp)/rate*1000, "ms"
call system_clock(startc)
CALL grady_v2n_fd4_omp(fin(:,:,:),fout_omp(:,:,:))
call system_clock(endc)
write(*,*) "timing for grady_v2n_fd4 DEVICE: ", real(endc-startc, dp)/rate*1000, "ms"
write(*,*) 'grady_v2n_fd4 L2_norm and maxval:', &
norm2(fout(iys:iye,ixs:ixe,izs:ize) - fout_omp(iys:iye,ixs:ixe,izs:ize)), &
maxval(fout(iys:iye,ixs:ixe,izs:ize) - fout_omp(iys:iye,ixs:ixe,izs:ize))
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
call init_array(fout)
call init_array(fout_omp)
call system_clock(startc)
CALL interp_v2n_fd4(fin(:,:,:),fout(:,:,:))
call system_clock(endc)
write(*,*) "timing for interp_v2n_fd4 CPU: ", real(endc-startc, dp)/rate*1000, "ms"
call system_clock(startc)
CALL interp_v2n_fd4_omp(fin(:,:,:),fout_omp(:,:,:))
call system_clock(endc)
write(*,*) "timing for interp_v2n_fd4 DEVICE: ", real(endc-startc, dp)/rate*1000, "ms"
write(*,*) 'interp_v2n_fd4 L2_norm and maxval:', &
norm2(fout(iys:iye,ixs:ixe,izs:ize) - fout_omp(iys:iye,ixs:ixe,izs:ize)), &
maxval(fout(iys:iye,ixs:ixe,izs:ize) - fout_omp(iys:iye,ixs:ixe,izs:ize))
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
call init_array(fout)
call init_array(fout_omp)
call system_clock(startc)
CALL gradx_n2n_fd4(fin(:,:,:),fout(:,:,:))
call system_clock(endc)
write(*,*) "timing for gradx_n2n_fd4 CPU: ", real(endc-startc, dp)/rate*1000, "ms"
call system_clock(startc)
CALL gradx_n2n_fd4_omp(fin(:,:,:),fout_omp(:,:,:))
call system_clock(endc)
write(*,*) "timing for gradx_n2n_fd4 DEVICE: ", real(endc-startc, dp)/rate*1000, "ms"
write(*,*) 'gradx_n2n_fd4 L2_norm and maxval:', &
norm2(fout(iys:iye,ixs:ixe,izs:ize) - fout_omp(iys:iye,ixs:ixe,izs:ize)), &
maxval(fout(iys:iye,ixs:ixe,izs:ize) - fout_omp(iys:iye,ixs:ixe,izs:ize))
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
#ifndef CUDA
call init_array(fout)
call init_array(fout_omp)
call system_clock(startc)
CALL gradpar_v2n_fd4(fin(:,:,:),fout(:,:,:))
call system_clock(endc)
write(*,*) "timing for gradpar_v2n_fd4 CPU: ", real(endc-startc, dp)/rate*1000, "ms"
call system_clock(startc)
CALL gradpar_v2n_fd4_omp(fin(:,:,:),fout_omp(:,:,:))
call system_clock(endc)
write(*,*) "timing for gradpar_v2n_fd4 DEVICE: ", real(endc-startc, dp)/rate*1000, "ms"
write(*,*) 'gradpar_v2n_fd4 L2_norm and maxval:', &
norm2(fout(iys:iye,ixs:ixe,izs:ize) - fout_omp(iys:iye,ixs:ixe,izs:ize)), &
maxval(fout(iys:iye,ixs:ixe,izs:ize) - fout_omp(iys:iye,ixs:ixe,izs:ize))
#endif
#ifdef CUDA
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
! Test gradients with CUDA C
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
call gbs_allocate_cuda(fout_cuda,iysg,iyeg,ixsg,ixeg,izsg,izeg)
call gbs_allocate_cuda(fin_cuda,iysg,iyeg,ixsg,ixeg,izsg,izeg)
call gbs_allocate_cuda(gradpar_y_n_cuda,iysg,iyeg,ixsg,ixeg)
call gbs_allocate_cuda(gradpar_x_n_cuda,iysg,iyeg,ixsg,ixeg)
call init_array2d(gradpar_x_n_cuda)
call init_array2d(gradpar_y_n_cuda)
call init_which_grad
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
call init_array(fin_cuda)
call init_array(fout_cuda)
call init_array(fout)
call system_clock(startc)
call gradz_n2n_fd4(fin(:,:,:),fout(:,:,:))
call system_clock(endc)
write(*,*) "timing for gradz_n2n_fd4 CPU: ", real(endc-startc, dp)/rate*1000, "ms"
call system_clock(startc)
call gradz_n2n_cuda(fin_cuda(:,:,:),fout_cuda(:,:,:))
call system_clock(endc)
write(*,*) "timing for gradz_n2n_fd4 DEVICE: ", real(endc-startc, dp)/rate*1000, "ms"
write(*,*) 'gradz_n2n_fd4 with CUDA L2_norm and maxval:', &
norm2(fout(iys:iye,ixs:ixe,izs:ize) - fout_cuda(iys:iye,ixs:ixe,izs:ize)), &
maxval(fout(iys:iye,ixs:ixe,izs:ize) - fout_cuda(iys:iye,ixs:ixe,izs:ize))
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
call init_array(fin_cuda)
call init_array(fout_cuda)
call init_array(fout)
call system_clock(startc)
call gradz_v2n_fd4(fin(:,:,:),fout(:,:,:))
call system_clock(endc)
write(*,*) "timing for gradz_v2n_fd4 CPU: ", real(endc-startc, dp)/rate*1000, "ms"
call system_clock(startc)
call gradz_v2n_cuda(fin_cuda(:,:,:),fout_cuda(:,:,:))
call system_clock(endc)
write(*,*) "timing for gradz_v2n_fd4 DEVICE: ", real(endc-startc, dp)/rate*1000, "ms"
write(*,*) 'gradz_v2n_fd4 with CUDA L2_norm and maxval:', &
norm2(fout(iys:iye,ixs:ixe,izs:ize) - fout_cuda(iys:iye,ixs:ixe,izs:ize)), &
maxval(fout(iys:iye,ixs:ixe,izs:ize) - fout_cuda(iys:iye,ixs:ixe,izs:ize))
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
call init_array(fin_cuda)
call init_array(fout_cuda)
call init_array(fout)
call system_clock(startc)
call grady_v2n_fd4(fin(:,:,:),fout(:,:,:))
call system_clock(endc)
write(*,*) "timing for grady_v2n_fd4 CPU: ", real(endc-startc, dp)/rate*1000, "ms"
call system_clock(startc)
call grady_v2n_cuda(fin_cuda(:,:,:),fout_cuda(:,:,:))
call system_clock(endc)
write(*,*) "timing for grady_v2n_fd4 DEVICE: ", real(endc-startc, dp)/rate*1000, "ms"
write(*,*) 'grady_v2n_fd4 with CUDA L2_norm and maxval:', &
norm2(fout(iys:iye,ixs:ixe,izs:ize) - fout_cuda(iys:iye,ixs:ixe,izs:ize)), &
maxval(fout(iys:iye,ixs:ixe,izs:ize) - fout_cuda(iys:iye,ixs:ixe,izs:ize))
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
call init_array(fin_cuda)
call init_array(fout_cuda)
call init_array(fout)
call system_clock(startc)
call interp_v2n_fd4(fin(:,:,:),fout(:,:,:))
call system_clock(endc)
write(*,*) "timing for interp_v2n CPU: ", real(endc-startc, dp)/rate*1000, "ms"
call system_clock(startc)
call interp_v2n_fd4cuda(fin_cuda(:,:,:),fout_cuda(:,:,:))
call system_clock(endc)
write(*,*) "timing for interp_v2n DEVICE: ", real(endc-startc, dp)/rate*1000, "ms"
write(*,*) 'interp_v2n with CUDA L2_norm and maxval:', &
norm2(fout(iys:iye,ixs:ixe,izs:ize) - fout_cuda(iys:iye,ixs:ixe,izs:ize)), &
maxval(fout(iys:iye,ixs:ixe,izs:ize) - fout_cuda(iys:iye,ixs:ixe,izs:ize))
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
call init_array(fin_cuda)
call init_array(fout_cuda)
call init_array(fout)
call system_clock(startc)
call gradpar_v2n_fd4(fin(:,:,:),fout(:,:,:))
call system_clock(endc)
write(*,*) "timing for gradpar_v2n CPU: ", real(endc-startc, dp)/rate*1000, "ms"
call system_clock(startc)
call gradpar_v2n_cuda(fin_cuda(:,:,:),fout_cuda(:,:,:))
call system_clock(endc)
write(*,*) "timing for gradpar_v2n DEVICE: ", real(endc-startc, dp)/rate*1000, "ms"
write(*,*) 'gradpar_v2n with CUDA L2_norm and maxval:', &
norm2(fout(iys:iye,ixs:ixe,izs:ize) - fout_cuda(iys:iye,ixs:ixe,izs:ize)), &
maxval(fout(iys:iye,ixs:ixe,izs:ize) - fout_cuda(iys:iye,ixs:ixe,izs:ize))
#endif
print*, "END TESTS"
end program main
c4science · Help