LCPQ/qp2
9
1
Fork 0
mirror of https://github.com/QuantumPackage/qp2.git synced 2024-11-09 06:53:38 +01:00
Code Releases Activity

merged
Some checks failed
continuous-integration/drone/push Build is failing
Details

Browse Source
This commit is contained in:
eginer 2024-06-02 19:18:19 +02:00
commit 1b48f30c27
3 changed files with 251 additions and 80 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

63
config/gfortran_debug_mkl.cfg Normal file
View File

@ -0,0 +1,63 @@
# Common flags
##############
#
# -ffree-line-length-none : Needed for IRPF90 which produces long lines
# -lblas -llapack : Link with libblas and liblapack libraries provided by the system
# -I . : Include the curent directory (Mandatory)
#
# --ninja : Allow the utilisation of ninja. (Mandatory)
# --align=32 : Align all provided arrays on a 32-byte boundary
#
#
[COMMON]
FC : gfortran -g -ffree-line-length-none -I . -fPIC -std=legacy
LAPACK_LIB : -I${MKLROOT}/include -L${MKLROOT}/lib/intel64 -Wl,--no-as-needed -lmkl_gf_lp64 -lmkl_core -lpthread -lm -ldl -lmkl_gnu_thread -lgomp -fopenmp
IRPF90 : irpf90
IRPF90_FLAGS : --ninja --align=32 --assert -DSET_NESTED
# Global options
################
#
# 1 : Activate
# 0 : Deactivate
#
[OPTION]
MODE : DEBUG ; [ OPT | PROFILE | DEBUG ] : Chooses the section below
CACHE : 0 ; Enable cache_compile.py
OPENMP : 1 ; Append OpenMP flags
# Optimization flags
####################
#
# -Ofast : Disregard strict standards compliance. Enables all -O3 optimizations.
# It also enables optimizations that are not valid
# for all standard-compliant programs. It turns on
# -ffast-math and the Fortran-specific
# -fno-protect-parens and -fstack-arrays.
[OPT]
FCFLAGS : -Ofast
# Profiling flags
#################
#
[PROFILE]
FC : -p -g
FCFLAGS : -Ofast
# Debugging flags
#################
#
# -fcheck=all : Checks uninitialized variables, array subscripts, etc...
# -g : Extra debugging information
#
[DEBUG]
#FCFLAGS : -g -msse4.2 -fcheck=all -Waliasing -Wampersand -Wconversion -Wsurprising -Wintrinsics-std -Wno-tabs -Wintrinsic-shadow -Wline-truncation -Wreal-q-constant -Wuninitialized -fbacktrace -ffpe-trap=zero,overflow,underflow -finit-real=nan
FCFLAGS : -g -mavx -fcheck=all -Waliasing -Wampersand -Wconversion -Wsurprising -Wintrinsics-std -Wno-tabs -Wintrinsic-shadow -Wline-truncation -Wreal-q-constant -Wuninitialized -fbacktrace -ffpe-trap=zero,overflow -finit-real=nan
# OpenMP flags
#################
#
[OPENMP]
FC : -fopenmp
IRPF90_FLAGS : --openmp

216
src/ao_two_e_ints/cholesky.irp.f
View File

@ -22,41 +22,51 @@ END_PROVIDER
! <ij|kl> = (ik|jl) = sum_a (ik|a).(a|jl)
!
! Last dimension of cholesky_ao is cholesky_ao_num
!
! https://mogp-emulator.readthedocs.io/en/latest/methods/proc/ProcPivotedCholesky.html
! https://doi.org/10.1016/j.apnum.2011.10.001 : Page 4, Algorithm 1
END_DOC
integer :: rank, ndim
double precision :: tau
integer*8 :: ndim8
integer :: rank
double precision :: tau, tau2
double precision, pointer :: L(:,:), L_old(:,:)
double precision :: s
double precision, parameter :: dscale = 1.d0
double precision :: dscale, dscale_tmp
double precision, allocatable :: D(:), Delta(:,:), Ltmp_p(:,:), Ltmp_q(:,:)
integer, allocatable :: Lset(:), Dset(:), addr(:,:)
integer, allocatable :: addr1(:), addr2(:)
integer*8, allocatable :: Lset(:), Dset(:), addr3(:)
logical, allocatable :: computed(:)
integer :: i,j,k,m,p,q, qj, dj, p2, q2
integer :: N, np, nq
integer :: i,j,k,m,p,q, dj, p2, q2
integer*8 :: i8, j8, p8, qj8
integer :: N, np, nq, npmax
double precision :: Dmax, Dmin, Qmax, f
double precision, external :: get_ao_two_e_integral
logical, external :: ao_two_e_integral_zero
double precision, external :: ao_two_e_integral
integer :: block_size, iblock, ierr
integer :: block_size, iblock
double precision :: mem
double precision, external :: memory_of_double, memory_of_int
double precision, external :: memory_of_double8, memory_of_int8
integer, external :: getUnitAndOpen
integer :: iunit
integer :: iunit, ierr
ndim8 = ao_num*ao_num*1_8
double precision :: wall0,wall1
call wall_time(wall0)
ndim = ao_num*ao_num
deallocate(cholesky_ao)
! TODO : Save L() to disk
if (read_ao_cholesky) then
print *, 'Reading Cholesky vectors from disk...'
iunit = getUnitAndOpen(trim(ezfio_work_dir)//'cholesky_ao', 'R')
@ -81,13 +91,15 @@ END_PROVIDER
endif
tau = ao_cholesky_threshold
tau2 = tau*tau
mem = 6.d0 * memory_of_double(ndim) + 6.d0 * memory_of_int(ndim)
mem = 6.d0 * memory_of_double8(ndim8) + 6.d0 * memory_of_int8(ndim8)
call check_mem(mem, irp_here)
call print_memory_usage()
allocate(L(ndim,1))
allocate(L(ndim8,1))
!print *, 'allocate : (L(ndim8,1))', memory_of_double8(ndim8)
print *, ''
print *, 'Cholesky decomposition of AO integrals'
@ -100,32 +112,36 @@ END_PROVIDER
rank = 0
allocate( D(ndim), Lset(ndim), Dset(ndim) )
allocate( addr(3,ndim) )
allocate( D(ndim8), Lset(ndim8), Dset(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
do j=1,ao_num
do i=1,ao_num
k = k+1
addr(1,k) = i
addr(2,k) = j
addr(3,k) = (i-1)*ao_num + j
addr1(k) = i
addr2(k) = j
addr3(k) = (i-1)*ao_num + j
enddo
enddo
if (do_direct_integrals) then
!$OMP PARALLEL DO DEFAULT(SHARED) PRIVATE(i) SCHEDULE(guided)
do i=1,ndim
D(i) = ao_two_e_integral(addr(1,i), addr(2,i), &
addr(1,i), addr(2,i))
!$OMP PARALLEL DO DEFAULT(SHARED) PRIVATE(i8) SCHEDULE(dynamic,16)
do i8=1,ndim8
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(i) SCHEDULE(guided)
do i=1,ndim
D(i) = get_ao_two_e_integral(addr(1,i), addr(1,i), &
addr(2,i), addr(2,i), &
!$OMP PARALLEL DO DEFAULT(SHARED) PRIVATE(i8) SCHEDULE(dynamic,16)
do i8=1,ndim8
D(i8) = get_ao_two_e_integral(addr1(i8), addr1(i8), &
addr2(i8), addr2(i8), &
ao_integrals_map)
enddo
!$OMP END PARALLEL DO
@ -134,12 +150,26 @@ END_PROVIDER
Dmax = maxval(D)
! 2.
npmax = huge(1_4)*1_8
np = npmax
dscale = 1.d0
dscale_tmp = Dmax
do while (np == npmax)
np=0
do p=1,ndim
if ( dscale*dscale*Dmax*D(p) > tau*tau ) then
do p8=1,ndim8
if ( dscale_tmp*D(p8) > tau2 ) then
np = np+1
Lset(np) = p
Lset(np) = p8
if (np == npmax) then
! Overflow detected
dscale = dscale*0.1d0
dscale_tmp = dscale*dscale*Dmax
!print *, 'Overflow detected '
!print *, 'dscale = ', dscale
exit
endif
endif
enddo
enddo
! 3.
@ -149,13 +179,14 @@ END_PROVIDER
i = 0
! 5.
do while ( (Dmax > tau).and.(rank < ndim) )
do while ( (Dmax > tau).and.(rank < min(ndim8,huge(1_4))) )
! a.
i = i+1
s = 0.01d0
! Inrease s until the arrays fit in memory
s = 0.01d0
block_size = max(N,24)
do while (.True.)
! b.
@ -170,10 +201,11 @@ END_PROVIDER
endif
enddo
call total_memory(mem)
mem = mem &
+ np*memory_of_double(nq) &! Delta(np,nq)
+ (rank+nq)* memory_of_double(ndim) &! L(ndim,rank+nq)
+ (rank+nq)* memory_of_double8(ndim8) &! L(ndim8,rank+nq)
+ (np+nq)*memory_of_double(block_size) ! Ltmp_p(np,block_size) + Ltmp_q(nq,block_size)
if (mem > qp_max_mem) then
@ -184,6 +216,9 @@ END_PROVIDER
if ((s > 1.d0).or.(nq == 0)) then
call print_memory_usage()
print *, 'Required peak memory: ', mem, 'Gb'
call total_memory(mem)
print *, 'Already used memory: ', mem, 'Gb'
print *, 'Not enough memory. Reduce cholesky threshold'
stop -1
endif
@ -191,20 +226,21 @@ END_PROVIDER
enddo
! d., e.
block_size = max(N,24)
L_old => L
allocate(L(ndim,rank+nq), stat=ierr)
allocate(L(ndim8,rank+nq), stat=ierr)
!print *, 'allocate : L(ndim8,rank+nq)', memory_of_double8(ndim8*(rank+nq))
if (ierr /= 0) then
call print_memory_usage()
print *, irp_here, ': allocation failed : (L(ndim,rank+nq))'
print *, irp_here, ': allocation failed : (L(ndim8,rank+nq))'
stop -1
endif
!$OMP PARALLEL DO PRIVATE(k,j)
!$OMP PARALLEL DO PRIVATE(k,j8)
do k=1,rank
do j=1,ndim
L(j,k) = L_old(j,k)
do j8=1,ndim8
L(j8,k) = L_old(j8,k)
enddo
enddo
!$OMP END PARALLEL DO
@ -212,6 +248,8 @@ END_PROVIDER
deallocate(L_old)
allocate(Delta(np,nq), stat=ierr)
!print *, 'allocate : Delta(np,nq)', memory_of_double8(np*nq*1_8)
if (ierr /= 0) then
call print_memory_usage()
print *, irp_here, ': allocation failed : (Delta(np,nq))'
@ -219,6 +257,8 @@ END_PROVIDER
endif
allocate(Ltmp_p(np,block_size), stat=ierr)
!print *, 'allocate : Ltmp_p(np,block_size)', memory_of_double8(np*block_size*1_8), np, block_size
if (ierr /= 0) then
call print_memory_usage()
print *, irp_here, ': allocation failed : (Ltmp_p(np,block_size))'
@ -226,6 +266,8 @@ END_PROVIDER
endif
allocate(Ltmp_q(nq,block_size), stat=ierr)
!print *, 'allocate : Ltmp_q(nq,block_size)', memory_of_double8(nq*block_size*1_8), nq, block_size
if (ierr /= 0) then
call print_memory_usage()
print *, irp_here, ': allocation failed : (Ltmp_q(nq,block_size))'
@ -234,32 +276,47 @@ END_PROVIDER
allocate(computed(nq))
!print *, 'allocate : computed(nq)', memory_of_int(nq)
!$OMP PARALLEL DEFAULT(SHARED) PRIVATE(m,k,p,q,j)
!print *, 'N, rank, block_size', N, rank, block_size
!print *, 'p1'
!$OMP PARALLEL DEFAULT(SHARED) PRIVATE(k,p,q,j)
!print *, 'computed'
!$OMP DO
do q=1,nq
computed(q) = .False.
enddo
!$OMP ENDDO NOWAIT
!print *, 'Delta'
!$OMP DO
do q=1,nq
do j=1,np
Delta(j,q) = 0.d0
enddo
computed(q) = .False.
enddo
!$OMP ENDDO NOWAIT
!$OMP DO
!print *, 'Ltmp_p'
do k=1,N
!$OMP DO
do p=1,np
Ltmp_p(p,k) = L(Lset(p),k)
enddo
!$OMP END DO NOWAIT
!$OMP DO
do q=1,nq
Ltmp_q(q,k) = L(Dset(q),k)
enddo
enddo
!$OMP END DO NOWAIT
enddo
!$OMP BARRIER
!$OMP END PARALLEL
!print *, 'p2', np, nq, N
if (N>0) then
call dgemm('N','T', np, nq, N, -1.d0, &
Ltmp_p, np, Ltmp_q, nq, 1.d0, Delta, np)
@ -276,54 +333,67 @@ END_PROVIDER
iblock = 0
do j=1,nq
if ( (Qmax <= Dmin).or.(N+j > ndim) ) exit
if ( (Qmax <= Dmin).or.(N+j*1_8 > ndim8) ) exit
! i.
rank = N+j
if (iblock == block_size) then
!print *, 'dgemm'
call dgemm('N','T',np,nq,block_size,-1.d0, &
Ltmp_p, np, Ltmp_q, nq, 1.d0, Delta, np)
iblock = 0
endif
! ii.
do dj=1,nq
qj = Dset(dj)
if (D(qj) == Qmax) then
qj8 = Dset(dj)
if (D(qj8) == Qmax) then
exit
endif
enddo
L(1:ndim, rank) = 0.d0
do i8=1,ndim8
L(i8, rank) = 0.d0
enddo
if (.not.computed(dj)) then
m = dj
!$OMP PARALLEL DO PRIVATE(k) SCHEDULE(guided)
do k=np,1,-1
if (.not.ao_two_e_integral_zero( addr(1,Lset(k)), addr(1,Dset(m)),&
addr(2,Lset(k)), addr(2,Dset(m)) ) ) then
if (do_direct_integrals) then
!$OMP PARALLEL DO PRIVATE(k) SCHEDULE(dynamic,16)
do k=np,1,-1
if (.not.ao_two_e_integral_zero( addr1(Lset(k)), addr1(Dset(m)),&
addr2(Lset(k)), addr2(Dset(m)) ) ) then
Delta(k,m) = Delta(k,m) + &
ao_two_e_integral(addr(1,Lset(k)), addr(2,Lset(k)),&
addr(1,Dset(m)), addr(2,Dset(m)))
else
Delta(k,m) = Delta(k,m) + &
get_ao_two_e_integral( addr(1,Lset(k)), addr(1,Dset(m)),&
addr(2,Lset(k)), addr(2,Dset(m)), ao_integrals_map)
endif
ao_two_e_integral(addr1(Lset(k)), addr2(Lset(k)),&
addr1(Dset(m)), addr2(Dset(m)))
endif
enddo
!$OMP END PARALLEL DO
else
!$OMP PARALLEL DO PRIVATE(k) SCHEDULE(dynamic,16)
do k=np,1,-1
if (.not.ao_two_e_integral_zero( addr1(Lset(k)), addr1(Dset(m)),&
addr2(Lset(k)), addr2(Dset(m)) ) ) then
Delta(k,m) = Delta(k,m) + &
get_ao_two_e_integral( addr1(Lset(k)), addr1(Dset(m)),&
addr2(Lset(k)), addr2(Dset(m)), ao_integrals_map)
endif
enddo
!$OMP END PARALLEL DO
endif
computed(dj) = .True.
endif
iblock = iblock+1
!print *, iblock
do p=1,np
Ltmp_p(p,iblock) = Delta(p,dj)
enddo
! iv.
if (iblock > 1) then
!print *, 'dgemv', iblock
call dgemv('N', np, iblock-1, -1.d0, Ltmp_p, np, Ltmp_q(dj,1), nq, 1.d0,&
Ltmp_p(1,iblock), 1)
endif
@ -331,7 +401,8 @@ END_PROVIDER
! iii.
f = 1.d0/dsqrt(Qmax)
!$OMP PARALLEL PRIVATE(m,p,q,k) DEFAULT(shared)
!print *, 'p4'
!$OMP PARALLEL PRIVATE(p,q) DEFAULT(shared)
!$OMP DO
do p=1,np
Ltmp_p(p,iblock) = Ltmp_p(p,iblock) * f
@ -345,7 +416,6 @@ END_PROVIDER
Ltmp_q(q,iblock) = L(Dset(q), rank)
enddo
!$OMP END DO
!$OMP END PARALLEL
Qmax = D(Dset(1))
@ -357,10 +427,10 @@ END_PROVIDER
print '(I10, 4X, ES12.3)', rank, Qmax
deallocate(computed)
deallocate(Delta)
deallocate(Ltmp_p)
deallocate(Ltmp_q)
deallocate(computed)
deallocate(Delta)
! i.
N = rank
@ -371,17 +441,33 @@ END_PROVIDER
Dmax = max(Dmax, D(Lset(p)))
enddo
np = npmax
dscale = 1.d0
dscale_tmp = Dmax
do while (np == npmax)
np=0
do p=1,ndim
if ( dscale*dscale*Dmax*D(p) > tau*tau ) then
do p8=1,ndim8
if ( dscale_tmp*D(p8) > tau2 ) then
np = np+1
Lset(np) = p
Lset(np) = p8
if (np == npmax) then
! Overflow detected
dscale = dscale*0.5d0
dscale_tmp = dscale*dscale*Dmax
!print *, 'Overflow detected '
!print *, 'dscale = ', dscale
exit
endif
endif
enddo
enddo
enddo
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()
print *, irp_here, ': Allocation failed'
@ -389,7 +475,9 @@ END_PROVIDER
endif
!$OMP PARALLEL DO PRIVATE(k)
do k=1,rank
call dcopy(ndim, L(1,k), 1, cholesky_ao(1,1,k), 1)
do j=1,ao_num
cholesky_ao(1:ao_num,j,k) = L((j-1)*ao_num+1:j*ao_num,k)
enddo
enddo
!$OMP END PARALLEL DO
deallocate(L)

20
src/utils/memory.irp.f
View File

@ -79,6 +79,26 @@ IRP_ENDIF
call unlock_io()
end function
double precision function memory_of_double8(n)
implicit none
BEGIN_DOC
! Computes the memory required for n double precision elements in gigabytes.
END_DOC
integer*8, intent(in) :: n
double precision, parameter :: f = 8.d0 / (1024.d0*1024.d0*1024.d0)
memory_of_double8 = dble(n) * f
end function
double precision function memory_of_int8(n)
implicit none
BEGIN_DOC
! Computes the memory required for n double precision elements in gigabytes.
END_DOC
integer*8, intent(in) :: n
double precision, parameter :: f = 4.d0 / (1024.d0*1024.d0*1024.d0)
memory_of_int8 = dble(n) * f
end function
double precision function memory_of_double(n)
implicit none
BEGIN_DOC