LCPQ
/
qmcchem
mirror of https://gitlab.com/scemama/qmcchem.git
10
1
Fork 1
Code Releases Activity

Prefetch in sparse-full-mv

This commit is contained in:
Anthony Scemama 2016-06-03 13:51:01 +02:00
parent 0e243a8afc
commit 3ccee1b3f0
1 changed files with 34 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

52
src/mo.irp.f
View File

@ -76,7 +76,6 @@ BEGIN_PROVIDER [ real, mo_coef, (ao_num_8,mo_num_8) ]
f = 1./mo_scale
do j=1,mo_num
!DIR$ VECTOR ALIGNED
!DIR$ LOOP COUNT (2000)
do i=1,ao_num_8
mo_coef(i,j) *= f
enddo
@ -218,7 +217,6 @@ END_PROVIDER
cycle
endif
r_inv = nucl_elec_dist_inv(k,i)
!DIR$ LOOP COUNT (500)
do j=1,mo_num
mo_value_transp(j,i) = mo_value_transp(j,i) + nucl_fitcusp_param(1,j,k) +&
r * (nucl_fitcusp_param(2,j,k) + &
@ -250,7 +248,6 @@ END_PROVIDER
! Scale off-diagonal elements
t = prepare_walkers_t
do i=1,mo_num
!DIR$ LOOP COUNT (100)
do j=1,elec_alpha_num
if (i /= j) then
mo_value_transp(i,j) *= t
@ -260,7 +257,6 @@ END_PROVIDER
mo_lapl_transp(i,j) *= t
endif
enddo
!DIR$ LOOP COUNT (100)
do j=1,elec_beta_num
if (i /= j) then
mo_value_transp(i,j+elec_alpha_num) *= t
@ -284,7 +280,6 @@ END_PROVIDER
enddo
END_PROVIDER
BEGIN_PROVIDER [ real, mo_value, (elec_num_8,mo_num) ]
implicit none
BEGIN_DOC
@ -342,10 +337,27 @@ BEGIN_PROVIDER [ double precision, mo_grad_lapl, (4,elec_num,mo_num) ]
integer :: i,j
do j=1,mo_num
do i=1,elec_num
mo_grad_lapl(1,i,j) = mo_grad_x(i,j)
mo_grad_lapl(2,i,j) = mo_grad_y(i,j)
mo_grad_lapl(3,i,j) = mo_grad_z(i,j)
mo_grad_lapl(4,i,j) = mo_lapl (i,j)
mo_grad_lapl(1,i,j) = mo_grad_transp_x(j,i)
mo_grad_lapl(2,i,j) = mo_grad_transp_y(j,i)
mo_grad_lapl(3,i,j) = mo_grad_transp_z(j,i)
mo_grad_lapl(4,i,j) = mo_lapl_transp (j,i)
enddo
enddo
END_PROVIDER
BEGIN_PROVIDER [ double precision, mo_grad_lapl_transp, (4,mo_num,elec_num) ]
implicit none
BEGIN_DOC
! Gradients and laplacian
END_DOC
integer :: i,j
do i=1,elec_num
do j=1,mo_num
mo_grad_lapl_transp(1,j,i) = mo_grad_transp_x(j,i)
mo_grad_lapl_transp(2,j,i) = mo_grad_transp_y(j,i)
mo_grad_lapl_transp(3,j,i) = mo_grad_transp_z(j,i)
mo_grad_lapl_transp(4,j,i) = mo_lapl_transp (j,i)
enddo
enddo
@ -419,7 +431,6 @@ BEGIN_PROVIDER [ double precision , mo_value_at_nucl, (mo_num_8,nucl_num) ]
PROVIDE ao_value_p
!DIR$ LOOP COUNT (2000)
do i=1,ao_num
if (ao_nucl(i) /= k) then
ao_value_at_nucl_no_S(i) = ao_value_p(i)
@ -461,7 +472,6 @@ END_PROVIDER
point(3) = nucl_coord(k,3)+ nucl_fitcusp_radius(k)
TOUCH point
!DIR$ LOOP COUNT (2000)
do j=1,ao_num
ao_value_at_fitcusp_radius(j,k) = ao_value_p(j)
ao_grad_at_fitcusp_radius(j,k) = ao_grad_p(j,3)
@ -615,7 +625,6 @@ BEGIN_PROVIDER [ real, nucl_fitcusp_param, (4,mo_num,nucl_num) ]
cycle
endif
R = nucl_fitcusp_radius(k)
!DIR$ LOOP COUNT (500)
do i=1,mo_num
double precision :: lap_phi, grad_phi, phi, eta
lap_phi = mo_lapl_at_fitcusp_radius(i,k)
@ -728,22 +737,29 @@ subroutine sparse_full_mv(A,LDA, &
do k=0,LDA-1,$IRP_ALIGN/4
!DIR$ VECTOR ALIGNED
do j=1,$IRP_ALIGN/4
IRP_IF NO_PREFETCH
IRP_ELSE
call MM_PREFETCH (A(j+k,indices(kao+4)),3)
call MM_PREFETCH (A(j+k,indices(kao+5)),3)
call MM_PREFETCH (A(j+k,indices(kao+6)),3)
call MM_PREFETCH (A(j+k,indices(kao+7)),3)
IRP_ENDIF
C1(j+k) = C1(j+k) + A(j+k,k_vec(1))*d11 + A(j+k,k_vec(2))*d21&
+ A(j+k,k_vec(3))*d31 + A(j+k,k_vec(4))*d41
enddo
!DIR$ VECTOR ALIGNED
do j=1,$IRP_ALIGN/4
C2(j+k) = C2(j+k) + A(j+k,k_vec(1))*d12 + A(j+k,k_vec(2))*d22&
+ A(j+k,k_vec(3))*d32 + A(j+k,k_vec(4))*d42
enddo
!DIR$ VECTOR ALIGNED
do j=1,$IRP_ALIGN/4
C3(j+k) = C3(j+k) + A(j+k,k_vec(1))*d13 + A(j+k,k_vec(2))*d23&
+ A(j+k,k_vec(3))*d33 + A(j+k,k_vec(4))*d43
C4(j+k) = C4(j+k) + A(j+k,k_vec(1))*d14 + A(j+k,k_vec(2))*d24&
+ A(j+k,k_vec(3))*d34 + A(j+k,k_vec(4))*d44
enddo
!DIR$ VECTOR ALIGNED
do j=1,$IRP_ALIGN/4
C4(j+k) = C4(j+k) + A(j+k,k_vec(1))*d14 + A(j+k,k_vec(2))*d24&
+ A(j+k,k_vec(3))*d34 + A(j+k,k_vec(4))*d44
C5(j+k) = C5(j+k) + A(j+k,k_vec(1))*d15 + A(j+k,k_vec(2))*d25&
+ A(j+k,k_vec(3))*d35 + A(j+k,k_vec(4))*d45
enddo