PTEROSOR/QuantumPackage
10
0
Fork 0
mirror of https://github.com/QuantumPackage/qp2.git synced 2024-12-22 12:23:43 +01:00
Code Releases Activity

Use less memory in Cholesky

Browse Source
This commit is contained in:
Anthony Scemama 2024-06-17 14:57:48 +02:00
parent e876f635d6
commit f671c669f8
2 changed files with 119 additions and 280 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

336
src/ao_two_e_ints/cholesky.irp.f
View File

@ -31,14 +31,14 @@ END_PROVIDER
integer*8 :: ndim8
integer :: rank
double precision :: tau, tau2
double precision, pointer :: L(:,:)
double precision, pointer :: L(:,:), Delta(:,:)
double precision :: s
double precision :: dscale, dscale_tmp
double precision, allocatable :: D(:), Ltmp_p(:,:), Ltmp_q(:,:), D_sorted(:), Delta_col(:), Delta(:,:)
double precision, allocatable :: D(:), Ltmp_p(:,:), Ltmp_q(:,:), D_sorted(:), Delta_col(:)
integer, allocatable :: addr1(:), addr2(:)
integer*8, allocatable :: Lset(:), Dset(:), addr3(:)
integer*8, allocatable :: Lset(:), Dset(:)
logical, allocatable :: computed(:)
integer :: i,j,k,m,p,q, dj, p2, q2, ii, jj
@ -64,11 +64,8 @@ END_PROVIDER
type(c_ptr) :: c_pointer(2)
integer :: fd(2)
logical :: delta_on_disk
integer :: dgemm_block_size, nqq
double precision, allocatable :: dgemm_buffer1(:,:), dgemm_buffer2(:,:), dgemm_buffer3(:,:)
PROVIDE nproc
PROVIDE nproc ao_cholesky_threshold do_direct_integrals qp_max_mem
PROVIDE nucl_coord ao_two_e_integral_schwartz
call set_multiple_levels_omp(.False.)
@ -88,19 +85,8 @@ END_PROVIDER
else
PROVIDE nucl_coord ao_two_e_integral_schwartz
call set_multiple_levels_omp(.False.)
call resident_memory(mem0)
rank_max = min(ndim8,(qp_max_mem*1024_8*1024_8*1024_8/8_8)/ndim8)
call mmap(trim(ezfio_work_dir)//'cholesky_ao_tmp', (/ ndim8, rank_max /), 8, fd(1), .False., .True., c_pointer(1))
call c_f_pointer(c_pointer(1), L, (/ ndim8, rank_max /))
! Deleting the file while it is open makes the file invisible on the filesystem,
! and automatically deleted, even if the program crashes
iunit = getUnitAndOpen(trim(ezfio_work_dir)//'cholesky_ao_tmp', 'R')
close(iunit,status='delete')
if (do_direct_integrals) then
if (ao_two_e_integral(1,1,1,1) < huge(1.d0)) then
! Trigger providers inside ao_two_e_integral
@ -113,8 +99,12 @@ END_PROVIDER
tau = ao_cholesky_threshold
tau2 = tau*tau
mem = 6.d0 * memory_of_double8(ndim8) + 6.d0 * memory_of_int8(ndim8)
call check_mem(mem, irp_here)
rank = 0
allocate( D(ndim8), Lset(ndim8), Dset(ndim8), D_sorted(ndim8))
allocate( addr1(ndim8), addr2(ndim8), Delta_col(ndim8) )
call resident_memory(mem0)
call print_memory_usage()
@ -127,46 +117,35 @@ END_PROVIDER
print *, '============ ============='
rank = 0
allocate( D(ndim8), Lset(ndim8), Dset(ndim8), D_sorted(ndim8))
allocate( addr1(ndim8), addr2(ndim8), addr3(ndim8) )
!print *, 'allocate : (D(ndim8))', memory_of_int8(ndim8)
!print *, 'allocate : (Lset(ndim8))', memory_of_int8(ndim8)
!print *, 'allocate : (Dset(ndim8))', memory_of_int8(ndim8)
!print *, 'allocate : (4,addr(ndim8))', memory_of_int8(4_8*ndim8)
! 1.
k=0
i8=0
do j=1,ao_num
do i=1,ao_num
k = k+1
addr1(k) = i
addr2(k) = j
addr3(k) = (i-1)*ao_num + j
i8 = i8+1
addr1(i8) = i
addr2(i8) = j
enddo
enddo
if (do_direct_integrals) then
!$OMP PARALLEL DO DEFAULT(SHARED) PRIVATE(i8) SCHEDULE(dynamic,16)
!$OMP PARALLEL DO DEFAULT(SHARED) PRIVATE(i8) SCHEDULE(dynamic,21)
do i8=ndim8,1,-1
D(i8) = ao_two_e_integral(addr1(i8), addr2(i8), &
addr1(i8), addr2(i8))
enddo
!$OMP END PARALLEL DO
else
!$OMP PARALLEL DO DEFAULT(SHARED) PRIVATE(i8) SCHEDULE(dynamic,16)
!$OMP PARALLEL DO DEFAULT(SHARED) PRIVATE(i8) SCHEDULE(dynamic,21)
do i8=ndim8,1,-1
D(i8) = get_ao_two_e_integral(addr1(i8), addr1(i8), &
addr2(i8), addr2(i8), &
ao_integrals_map)
addr2(i8), addr2(i8), ao_integrals_map)
enddo
!$OMP END PARALLEL DO
endif
D_sorted(:) = -D(:)
call dsort_noidx_big(D_sorted,ndim8)
D_sorted(:) = dabs(D_sorted(:))
D_sorted(:) = -D_sorted(:)
Dmax = D_sorted(1)
! 2.
@ -174,12 +153,24 @@ END_PROVIDER
dscale_tmp = dscale*dscale*Dmax
np8=0_8
do p8=1,ndim8
if ( dscale_tmp*D(p8) > tau2 ) then
if ( dscale_tmp*D(p8) >= tau2 ) then
np8 = np8+1_8
Lset(np8) = p8
endif
enddo
np = np8
if (np <= 0) stop 'np<=0'
if (np > ndim8) stop 'np>ndim8'
rank_max = min(np,20*elec_num*elec_num)
call mmap(trim(ezfio_work_dir)//'cholesky_ao_tmp', (/ ndim8, rank_max /), 8, fd(1), .False., .True., c_pointer(1))
call c_f_pointer(c_pointer(1), L, (/ ndim8, rank_max /))
! Deleting the file while it is open makes the file invisible on the filesystem,
! and automatically deleted, even if the program crashes
iunit = getUnitAndOpen(trim(ezfio_work_dir)//'cholesky_ao_tmp', 'R')
close(iunit,status='delete')
! 3.
N = 0
@ -187,85 +178,59 @@ END_PROVIDER
! 4.
i = 0
mem = memory_of_double(np) & ! Delta(np,nq)
+ (np+1)*memory_of_double(block_size) ! Ltmp_p(np,block_size) + Ltmp_q(nq,block_size)
! call check_mem(mem)
! 5.
do while ( (Dmax > tau).and.(rank*1_8 < min(ndim8,rank_max)) )
do while ( (Dmax > tau).and.(np > 0) )
! a.
i = i+1
! Inrease s until the arrays fit in memory
s = 0.01d0
block_size = max(N,24)
! Determine nq so that Delta fits in memory
s = 0.1d0
Dmin = max(s*Dmax,tau)
do nq=2,np-1
if (D_sorted(nq) < Dmin) exit
enddo
do while (.True.)
! b.
Dmin = max(s*Dmax,tau)
mem = mem0 &
+ np*memory_of_double(nq) & ! Delta(np,nq)
+ (np+nq)*memory_of_double(block_size) & ! Ltmp_p(np,block_size) + Ltmp_q(nq,block_size)
+ memory_of_int(nq) ! computed(nq)
! c.
if (mem > qp_max_mem*0.5d0) then
nq = nq/2
else
exit
endif
enddo
if (nq <= 0) then
print *, nq
stop 'bug in cholesky: nq <= 0'
endif
Dmin = D_sorted(nq)
nq=0
do p=1,np
if ( D(Lset(p)) > Dmin ) then
if ( D(Lset(p)) >= Dmin ) then
nq = nq+1
Dset(nq) = Lset(p)
endif
enddo
mem = mem0 &
+ np*memory_of_double(nq)
!print *, 'mem = ', mem
if (mem > qp_max_mem/2) then
s = s*2.d0
else
exit
endif
if ((s > 1.d0).or.(nq == 0)) then
call print_memory_usage()
print *, 'Required peak memory: ', mem, 'Gb'
call resident_memory(mem)
print *, 'Already used memory: ', mem, 'Gb'
print *, 'Not enough memory. Reduce cholesky threshold'
stop -1
endif
if (s > 0.3d0) then
exit
endif
enddo
! d., e.
mem = mem0 &
+ memory_of_int(nq) &! computed(nq)
+ np*memory_of_int(nq) &! computed(nq)
+ memory_of_double(np) &! Delta_col(np)
+ 7*memory_of_double8(ndim8) &! D, Lset, Dset, D_sorted, addr[1-3]
+ np*memory_of_double(nq) &! Delta(np,nq)
+ (np+nq)*memory_of_double(block_size) ! Ltmp_p(np,block_size) + Ltmp_q(nq,block_size)
if (1.1*mem > qp_max_mem) then
! call mmap(trim(ezfio_work_dir)//'cholesky_delta', (/ np*1_8, nq*1_8 /), 8, fd(2), .False., .True., c_pointer(2))
! call c_f_pointer(c_pointer(2), Delta, (/ np, nq /))
! ! Deleting the file while it is open makes the file invisible on the filesystem,
! ! and automatically deleted, even if the program crashes
iunit = getUnitAndOpen(trim(ezfio_work_dir)//'cholesky_delta', 'R')
close(iunit,status='delete')
open(unit=iunit, access='DIRECT', form='UNFORMATTED', RECL=(np+1)*8, &
ASYNCHRONOUS='YES', file=trim(ezfio_work_dir)//'cholesky_delta')
delta_on_disk = .True.
else
allocate(Delta(np,nq))
delta_on_disk = .False.
endif
!print *, delta_on_disk
allocate(Delta_col(np))
allocate(Ltmp_p(np,block_size), stat=ierr)
!print *, 'allocate : Ltmp_p(np,block_size)', memory_of_double8(np*block_size*1_8), np, block_size
if (ierr /= 0) then
call print_memory_usage()
@ -274,7 +239,6 @@ END_PROVIDER
endif
allocate(Ltmp_q(nq,block_size), stat=ierr)
!print *, 'allocate : Ltmp_q(nq,block_size)', memory_of_double8(nq*block_size*1_8), nq, block_size
if (ierr /= 0) then
call print_memory_usage()
@ -287,7 +251,6 @@ END_PROVIDER
computed(:) = .False.
!print *, 'N, rank, block_size', N, rank, block_size
!$OMP PARALLEL DEFAULT(SHARED) PRIVATE(k,p,q)
do k=1,N
!$OMP DO
@ -305,67 +268,6 @@ END_PROVIDER
!$OMP BARRIER
!$OMP END PARALLEL
PROVIDE nproc
if (delta_on_disk) then
dgemm_block_size = nproc*4
allocate (dgemm_buffer1(np,dgemm_block_size))
allocate (dgemm_buffer3(np,dgemm_block_size))
allocate (dgemm_buffer2(dgemm_block_size,N))
if (N>0) then
! Blocking improves I/O performance
do jj=1,nq,dgemm_block_size
nqq = min(nq, jj+dgemm_block_size-1) - jj + 1
!$OMP PARALLEL DO DEFAULT(SHARED) PRIVATE(q,ii)
do ii=1,N
do q=jj,jj+nqq-1
dgemm_buffer2(q-jj+1,ii) = Ltmp_q(q,ii)
enddo
enddo
!$OMP END PARALLEL DO
print *, np, nq, jj, nqq, N
call dgemm('N', 'T', np, nqq, N, -1.d0, &
Ltmp_p, np, dgemm_buffer2, dgemm_block_size, 0.d0, dgemm_buffer1, np)
wait(iunit)
dgemm_buffer3(:,:) = dgemm_buffer1(:,:)
! !$OMP PARALLEL DO DEFAULT(SHARED) PRIVATE(q)
do q=jj,jj+nqq-1
write(iunit, ASYNCHRONOUS='YES', rec=q) dgemm_buffer3(1:np, q-jj+1)
enddo
! !$OMP END PARALLEL DO
enddo
print *, 'ok1'
else
dgemm_buffer1(1:np,1) = 0.d0
! !$OMP PARALLEL DO DEFAULT(SHARED) PRIVATE(q)
do q=1,nq
write(iunit, ASYNCHRONOUS='YES', rec=q) dgemm_buffer1(1:np, 1)
enddo
! !$OMP END PARALLEL DO
endif
deallocate(dgemm_buffer1, dgemm_buffer2)
if (delta_on_disk) wait(iunit)
deallocate(dgemm_buffer3)
else
if (N>0) then
call dgemm('N', 'T', np, nq, N, -1.d0, &
@ -381,8 +283,6 @@ print *, 'ok1'
endif
endif
! f.
Qmax = D(Dset(1))
do q=1,nq
@ -395,63 +295,19 @@ print *, 'ok1'
do j=1,nq
if ( (Qmax <= Dmin).or.(N+j*1_8 > ndim8) ) exit
! i.
rank = N+j
if (rank == rank_max) then
print *, 'cholesky: rank_max reached'
exit
endif
if (iblock == block_size) then
if (delta_on_disk) then
! Blocking improves I/O performance
dgemm_block_size = nproc*4
allocate (dgemm_buffer1(np,dgemm_block_size))
allocate (dgemm_buffer2(dgemm_block_size,block_size))
do jj=1,nq,dgemm_block_size
nqq = min(nq, jj+dgemm_block_size-1) - jj + 1
!$OMP PARALLEL DEFAULT(SHARED) PRIVATE(q,ii)
!$OMP DO
do ii=1,block_size
do q=jj,jj+nqq-1
dgemm_buffer2(q-jj+1,ii) = Ltmp_q(q,ii)
enddo
enddo
!$OMP END DO
!$OMP END PARALLEL
! !$OMP DO
do q=jj,jj+nqq-1
read(iunit, rec=q) dgemm_buffer1(1:np, q-jj+1)
enddo
! !$OMP END DO
print *, np, nq, jj, nqq, block_size
call dgemm('N', 'T', np, nqq, block_size, -1.d0, &
Ltmp_p(1,1), np, dgemm_buffer2, dgemm_block_size, 1.d0, dgemm_buffer1, np)
wait(iunit)
dgemm_buffer3 = dgemm_buffer1
! !$OMP PARALLEL DO DEFAULT(SHARED) PRIVATE(q)
do q=jj,jj+nqq-1
write(iunit, ASYNCHRONOUS='YES', rec=q) dgemm_buffer3(:, q-jj+1)
enddo
! !$OMP END PARALLEL DO
enddo
print *, 'ok'
deallocate(dgemm_buffer1, dgemm_buffer2)
wait(iunit)
deallocate(dgemm_buffer3)
else
call dgemm('N','T',np,nq,block_size,-1.d0, &
Ltmp_p, np, Ltmp_q, nq, 1.d0, Delta, np)
endif
iblock = 0
endif
@ -469,63 +325,41 @@ print *, 'ok'
enddo
iblock = iblock+1
if (delta_on_disk) then
read(iunit,rec=dj) Ltmp_p(1:np,iblock)
else
!$OMP PARALLEL DO PRIVATE(p)
do p=1,np
Ltmp_p(p,iblock) = Delta(p,dj)
enddo
!$OMP END PARALLEL DO
endif
if (.not.computed(dj)) then
m = dj
if (do_direct_integrals) then
!$OMP PARALLEL DO PRIVATE(k) SCHEDULE(dynamic,21)
do k=1,np
Delta_col(k) = 0.d0
if (.not.ao_two_e_integral_zero( addr1(Lset(k)), addr1(Dset(m)),&
addr2(Lset(k)), addr2(Dset(m)) ) ) then
Delta_col(k) = &
ao_two_e_integral(addr1(Lset(k)), addr2(Lset(k)),&
addr1(Dset(m)), addr2(Dset(m)))
else
Delta_col(k) = 0.d0
endif
enddo
!$OMP END PARALLEL DO
else
PROVIDE ao_integrals_map
!$OMP PARALLEL DO PRIVATE(k) SCHEDULE(dynamic,21)
do k=1,np
Delta_col(k) = 0.d0
if (.not.ao_two_e_integral_zero( addr1(Lset(k)), addr1(Dset(m)),&
addr2(Lset(k)), addr2(Dset(m)) ) ) then
Delta_col(k) = &
get_ao_two_e_integral( addr1(Lset(k)), addr1(Dset(m)),&
addr2(Lset(k)), addr2(Dset(m)), ao_integrals_map)
else
Delta_col(k) = 0.d0
endif
enddo
!$OMP END PARALLEL DO
endif
if (delta_on_disk) then
!$OMP PARALLEL DO PRIVATE(p)
do p=1,np
Ltmp_p(p,iblock) = Ltmp_p(p,iblock) + Delta_col(p)
enddo
!$OMP END PARALLEL DO
write(iunit, ASYNCHRONOUS='YES', rec=dj) Ltmp_p(1:np,iblock)
else
!$OMP PARALLEL DO PRIVATE(p)
do p=1,np
Ltmp_p(p,iblock) = Ltmp_p(p,iblock) + Delta_col(p)
@ -533,8 +367,6 @@ print *, 'ok'
enddo
!$OMP END PARALLEL DO
endif
computed(dj) = .True.
endif
@ -572,30 +404,26 @@ print *, 'ok'
print '(I10, 4X, ES12.3)', rank, Qmax
deallocate(Delta_col)
deallocate(Ltmp_p)
deallocate(Ltmp_q)
deallocate(computed)
if (delta_on_disk) then
close(iunit, status='delete')
else
deallocate(Delta)
endif
! i.
N = rank
! j.
Dmax = D(Lset(1))
do p=1,np
Dmax = max(Dmax, D(Lset(p)))
enddo
D_sorted(:) = -D(:)
call dsort_noidx_big(D_sorted,ndim8)
D_sorted(:) = -D_sorted(:)
Dmax = D_sorted(1)
dscale = 1.d0
dscale_tmp = dscale*dscale*Dmax
np8=0_8
do p8=1,ndim8
if ( dscale_tmp*D(p8) > tau2 ) then
if ( dscale_tmp*D(p8) >= tau2 ) then
np8 = np8+1_8
Lset(np8) = p8
endif
@ -609,7 +437,6 @@ print *, 'ok'
print *, ''
allocate(cholesky_ao(ao_num,ao_num,rank), stat=ierr)
!print *, 'allocate : cholesky_ao(ao_num,ao_num,rank)', memory_of_double8(ao_num*ao_num*rank*1_8)
if (ierr /= 0) then
call print_memory_usage()
@ -621,7 +448,7 @@ print *, 'ok'
!$OMP PARALLEL DO PRIVATE(k,j)
do k=1,rank
do j=1,ao_num
cholesky_ao(1:ao_num,j,k) = L((j-1)*ao_num+1:j*ao_num,k)
cholesky_ao(1:ao_num,j,k) = L((j-1_8)*ao_num+1_8:1_8*j*ao_num,k)
enddo
enddo
!$OMP END PARALLEL DO
@ -646,5 +473,6 @@ print *, 'ok'
call wall_time(wall1)
print*,'Time to provide AO cholesky vectors = ',(wall1-wall0)/60.d0, ' min'
END_PROVIDER

19
src/hartree_fock/fock_matrix_hf.irp.f
View File

@ -194,14 +194,25 @@ END_PROVIDER
endif
double precision :: rss
double precision :: rss, mem0, mem
double precision :: memory_of_double
integer :: iblock
integer, parameter :: block_size = 32
integer :: block_size
call resident_memory(mem0)
block_size = 1024
rss = memory_of_double(2.d0*ao_num*ao_num)
do
mem = mem0 + block_size*rss
if ( (block_size < 2).or.(mem < qp_max_mem) ) exit
block_size = block_size/2
enddo
call check_mem(block_size*rss, irp_here)
rss = memory_of_double(ao_num*ao_num)
call check_mem(2.d0*block_size*rss, irp_here)
allocate(X2(ao_num,ao_num,block_size,2))
allocate(X3(ao_num,block_size,ao_num,2))