LCPQ
/
quantum_package
mirror of https://github.com/LCPQ/quantum_package
10
0
Fork 0
Code Issues Releases Wiki Activity

Merge branch 'feature/4idx' of lpqlx139:~/quantum_package

This commit is contained in:
Anthony Scemama 2017-09-26 09:25:41 +02:00
parent 813d8d5e0f 42c7cf31b7
commit 2f82cb4ad7
9 changed files with 509 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
install/scripts/install_gpi2.sh
View File

@ -6,9 +6,9 @@ GPI_OPTIONS=--with-ethernet
function _install()
{
cd gpi2
cd _build/gpi2
./install.sh -p $QP_ROOT $GPI_OPTIONS
cp src/GASPI.f90 $QP_ROOT/src/plugins/GPI2/
cp src/GASPI.f90 $QP_ROOT/plugins/GPI2/
return 0
}

44
plugins/FourIdx/four_idx.irp.f Normal file
View File

@ -0,0 +1,44 @@
program FourIdx
use map_module
implicit none
BEGIN_DOC
! Performs a four index transformation of the two-electron integrals
END_DOC
type(map_type) :: test_map
integer(key_kind) :: key_max
integer(map_size_kind) :: sze
call bielec_integrals_index(ao_num,ao_num,ao_num,ao_num,key_max)
sze = key_max
call map_init(test_map,sze)
call four_index_transform(ao_integrals_map,test_map, &
mo_coef, size(mo_coef,1), &
1, 1, 1, 1, ao_num, ao_num, ao_num, ao_num, &
1, 1, 1, 1, mo_tot_num, mo_tot_num, mo_tot_num, mo_tot_num)
integer :: i,j,k,l
real(integral_kind) :: integral1, integral2
provide mo_bielec_integrals_in_map
do i=1,mo_tot_num
do j=1,mo_tot_num
do k=1,mo_tot_num
do l=1,mo_tot_num
call bielec_integrals_index(i,j,k,l,key_max)
call map_get(test_map,key_max,integral1)
call map_get(mo_integrals_map,key_max,integral2)
if (dabs(integral2) >=1.d-10 ) then
if (dabs(integral1 / integral2 -1.d0) > .001d0) then
print *, i,j,k,l
print *, integral1, integral2
print *, ''
endif
endif
enddo
enddo
enddo
enddo
end

149
plugins/FourIdx/four_index.irp.f Normal file
View File

@ -0,0 +1,149 @@
subroutine four_index_transform(map_a,map_c,matrix_B,LDB, &
i_start, j_start, k_start, l_start, &
i_end , j_end , k_end , l_end , &
a_start, b_start, c_start, d_start, &
a_end , b_end , c_end , d_end )
implicit none
use map_module
BEGIN_DOC
! Performs a four-index transformation of map_a(N^4) into map_c(M^4) using b(NxM)
! C_{abcd} = \sum_{ijkl} A_{ijkl}.B_{ia}.B_{jb}.B_{kc}.B_{ld}
! Loops run over *_start->*_end
END_DOC
type(map_type), intent(in) :: map_a
type(map_type), intent(inout) :: map_c
integer, intent(in) :: LDB
double precision, intent(in) :: matrix_B(LDB,*)
integer, intent(in) :: i_start, j_start, k_start, l_start
integer, intent(in) :: i_end , j_end , k_end , l_end
integer, intent(in) :: a_start, b_start, c_start, d_start
integer, intent(in) :: a_end , b_end , c_end , d_end
double precision, allocatable :: T(:,:,:), U(:,:,:), V(:,:,:)
integer :: i_max, j_max, k_max, l_max
integer :: i_min, j_min, k_min, l_min
integer :: i, j, k, l
integer :: a, b, c, d
double precision, external :: get_ao_bielec_integral
integer(key_kind) :: idx
real(integral_kind) :: tmp
integer(key_kind), allocatable :: key(:)
real(integral_kind), allocatable :: value(:)
ASSERT (k_start == i_start)
ASSERT (l_start == j_start)
ASSERT (a_start == c_start)
ASSERT (b_start == d_start)
i_min = min(i_start,a_start)
i_max = max(i_end ,a_end )
j_min = min(j_start,b_start)
j_max = max(j_end ,b_end )
k_min = min(k_start,c_start)
k_max = max(k_end ,c_end )
l_min = min(l_start,d_start)
l_max = max(l_end ,d_end )
ASSERT (0 < i_max)
ASSERT (0 < j_max)
ASSERT (0 < k_max)
ASSERT (0 < l_max)
ASSERT (LDB >= i_max)
ASSERT (LDB >= j_max)
ASSERT (LDB >= k_max)
ASSERT (LDB >= l_max)
!$OMP PARALLEL DEFAULT(PRIVATE) SHARED( &
!$OMP a_start,a_end,b_start,b_end,c_start,c_end,d_start,d_end,&
!$OMP i_start,i_end,j_start,j_end,k_start,k_end,l_start,l_end,&
!$OMP i_min,i_max,j_min,j_max,k_min,k_max,l_min,l_max, &
!$OMP map_a,map_c,matrix_B)
allocate( key(i_max*j_max*k_max), value(i_max*j_max*k_max) )
allocate( U(a_start:a_end, c_start:c_end, b_start:b_end) )
!$OMP DO SCHEDULE(static,1)
do d=d_start,d_end
U = 0.d0
do l=1,l_end
if (dabs(matrix_B(l,d)) < 1.d-10) then
cycle
endif
print *, d, l
allocate( T(i_start:i_end, k_start:k_end, j_start:j_end), &
V(a_start:a_end, k_start:k_end, j_start:j_end) )
do k=k_start,k_end
do j=j_start,j_end
do i=i_start,k
call bielec_integrals_index(i,j,k,l,idx)
call map_get(map_a,idx,tmp)
T(i, k,j) = tmp
enddo
enddo
enddo
do j=j_start,j_end
do k=k_start,k_end
do i=k+1,i_end
T(i, k,j) = T(k, i,j)
enddo
enddo
enddo
call DGEMM('T','N', (a_end-a_start+1), &
(k_end-k_start+1)*(j_end-j_start+1), &
(i_end-i_start+1), 1.d0, &
matrix_B(i_start,a_start), size(matrix_B,1), &
T(i_start,k_start,j_start), size(T,1), 0.d0, &
V(a_start,k_start,j_start), size(V, 1) )
deallocate(T)
allocate( T(a_start:a_end, k_start:k_end, b_start:d) )
call DGEMM('N','N', (a_end-a_start+1)*(k_end-k_start+1), &
(d-b_start+1), &
(j_end-j_start+1), 1.d0, &
V(a_start,k_start,j_start), size(V,1)*size(V,2), &
matrix_B(j_start,b_start), size(matrix_B,1),0.d0, &
T(a_start,k_start,b_start), size(T,1)*size(T,2) )
deallocate(V)
do b=b_start,d
call DGEMM('N','N', (b-a_start+1), (c_end-c_start+1), &
(k_end-k_start+1), matrix_B(l, d), &
T(a_start,k_start,b), size(T,1), &
matrix_B(k_start,k_start), size(matrix_B,1), 1.d0, &
U(a_start,c_start,b), size(U,1) )
enddo
deallocate(T)
enddo
idx = 0_8
do b=b_start,d
do c=c_start,c_end
do a=a_start,min(b,c)
if (dabs(U(a,c,b)) < 1.d-15) then
cycle
endif
idx = idx+1_8
call bielec_integrals_index(a,b,c,d,key(idx))
value(idx) = U(a,c,b)
enddo
enddo
enddo
!$OMP CRITICAL
call map_append(map_c, key, value, idx)
call map_sort(map_c)
call map_unique(map_c)
!$OMP END CRITICAL
enddo
!$OMP END DO
deallocate(key,value)
!$OMP END PARALLEL
end

254
plugins/GPI2/broadcast.irp.f Normal file
View File

@ -0,0 +1,254 @@
subroutine broadcast_wf(energy)
implicit none
BEGIN_DOC
! Segment corresponding to the wave function. This is segment 0.
END_DOC
use bitmasks
use GASPI
use ISO_C_BINDING
double precision, intent(inout) :: energy(N_states)
integer(gaspi_return_t) :: res
if (is_gaspi_master) then
call broadcast_wf_put(energy)
else
call broadcast_wf_get(energy)
endif
res = gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK)
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_barrier failed"
stop -1
end if
integer(gaspi_segment_id_t) :: seg_id
do seg_id=0,3
res = gaspi_segment_delete(seg_id)
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_segment_delete failed", seg_id
stop -1
end if
end do
end
subroutine broadcast_wf_put(energy)
implicit none
BEGIN_DOC
! Initiates the broadcast of the wave function
END_DOC
use bitmasks
use GASPI
use ISO_C_BINDING
double precision, intent(in) :: energy(N_states)
integer(gaspi_segment_id_t) :: seg_id
integer(gaspi_alloc_t) :: seg_alloc_policy
integer(gaspi_size_t) :: seg_size(0:3)
type(c_ptr) :: seg_ptr(0:3)
integer, pointer :: params_int(:) ! Segment 0
double precision, pointer :: psi_coef_tmp(:,:) ! Segment 1
integer(bit_kind), pointer :: psi_det_tmp(:,:,:) ! Segment 2
double precision, pointer :: params_double(:) ! Segment 3
integer(gaspi_return_t) :: res
seg_alloc_policy = GASPI_MEM_UNINITIALIZED
seg_size(0) = 4 * 5
seg_id=0
res = gaspi_segment_create(seg_id, seg_size(seg_id), GASPI_GROUP_ALL, &
GASPI_BLOCK, seg_alloc_policy)
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_create_segment failed", gaspi_rank, seg_id
stop -1
end if
res = gaspi_segment_ptr(seg_id, seg_ptr(seg_id))
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_segment_ptr failed", gaspi_rank
stop -1
end if
call c_f_pointer(seg_ptr(0), params_int, shape=(/ 5 /))
params_int(1) = N_states
params_int(2) = N_det
params_int(3) = psi_det_size
res = gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK)
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_barrier failed", gaspi_rank
stop -1
end if
seg_size(1) = 8 * psi_det_size * N_states
seg_size(2) = bit_kind * psi_det_size * 2 * N_int
seg_size(3) = 8 * N_states
do seg_id=1, 3
res = gaspi_segment_create(seg_id, seg_size(seg_id), GASPI_GROUP_ALL, &
GASPI_BLOCK, seg_alloc_policy)
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_create_segment failed", gaspi_rank, seg_id
stop -1
end if
res = gaspi_segment_ptr(seg_id, seg_ptr(seg_id))
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_segment_ptr failed", gaspi_rank
stop -1
end if
end do
call c_f_pointer(seg_ptr(1), psi_coef_tmp, shape=shape(psi_coef))
call c_f_pointer(seg_ptr(2), psi_det_tmp, shape=shape(psi_det))
call c_f_pointer(seg_ptr(3), params_double, shape=(/ N_states /))
psi_coef_tmp = psi_coef
psi_det_tmp = psi_det
params_double = energy
res = gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK)
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_barrier failed", gaspi_rank
stop -1
end if
end
subroutine broadcast_wf_get(energy)
implicit none
BEGIN_DOC
! Gets the broadcasted wave function
END_DOC
use bitmasks
use GASPI
use ISO_C_BINDING
double precision, intent(out) :: energy(N_states)
integer(gaspi_segment_id_t) :: seg_id
integer(gaspi_alloc_t) :: seg_alloc_policy
integer(gaspi_size_t) :: seg_size(0:3)
type(c_ptr) :: seg_ptr(0:3)
integer, pointer :: params_int(:) ! Segment 0
double precision, pointer :: psi_coef_tmp(:,:) ! Segment 1
integer(bit_kind), pointer :: psi_det_tmp(:,:,:) ! Segment 2
double precision, pointer :: params_double(:) ! Segment 3
integer(gaspi_return_t) :: res
seg_alloc_policy = GASPI_MEM_UNINITIALIZED
seg_size(0) = 4 * 5
seg_id=0
res = gaspi_segment_create(seg_id, seg_size(seg_id), GASPI_GROUP_ALL,&
GASPI_BLOCK, seg_alloc_policy)
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_create_segment failed"
stop -1
end if
res = gaspi_segment_ptr(seg_id, seg_ptr(seg_id))
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_segment_ptr failed"
stop -1
end if
res = gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK)
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_barrier failed"
stop -1
end if
integer(gaspi_offset_t) :: localOff, remoteOff
integer(gaspi_rank_t) :: remoteRank
integer(gaspi_queue_id_t) :: queue
localOff = 0
remoteRank = 0
queue = 0
res = gaspi_read(seg_id, localOff, remoteRank, &
seg_id, remoteOff, seg_size(seg_id), queue, GASPI_BLOCK)
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_read failed"
stop -1
end if
res = gaspi_wait(queue, GASPI_BLOCK)
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_wait failed"
stop -1
end if
call c_f_pointer(seg_ptr(0), params_int, shape=shape( (/ 5 /) ))
N_states = params_int(1)
N_det = params_int(2)
psi_det_size = params_int(3)
TOUCH N_states N_det psi_det_size
seg_size(1) = 8 * psi_det_size * N_states
seg_size(2) = bit_kind * psi_det_size * 2 * N_int
seg_size(3) = 8 * N_states
do seg_id=1, 3
res = gaspi_segment_create(seg_id, seg_size(seg_id), GASPI_GROUP_ALL, &
GASPI_BLOCK, seg_alloc_policy)
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_create_segment failed"
stop -1
end if
res = gaspi_segment_ptr(seg_id, seg_ptr(seg_id))
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_segment_ptr failed"
stop -1
end if
end do
res = gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK)
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_barrier failed"
stop -1
end if
do seg_id=1, 3
res = gaspi_read(seg_id, localOff, remoteRank, &
seg_id, remoteOff, seg_size(seg_id), queue, GASPI_BLOCK)
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_read failed"
stop -1
end if
res = gaspi_wait(queue, GASPI_BLOCK)
if(res .ne. GASPI_SUCCESS) then
write(*,*) "gaspi_wait failed"
stop -1
end if
end do
call c_f_pointer(seg_ptr(1), psi_coef_tmp, shape=shape(psi_coef))
call c_f_pointer(seg_ptr(2), psi_det_tmp, shape=shape(psi_det))
call c_f_pointer(seg_ptr(3), params_double, shape=shape(energy))
psi_coef = psi_coef_tmp
psi_det = psi_det_tmp
energy = params_double
end

7
plugins/read_integral/read_integrals_mo.irp.f
View File

@ -1,5 +1,10 @@
program read_integrals
BEGIN_DOC
! Reads the integrals from the following files:
! - kinetic_mo
! - nuclear_mo
! - bielec_mo
END_DOC
PROVIDE ezfio_filename
call ezfio_set_integrals_monoelec_disk_access_mo_one_integrals("None")
call run

4
scripts/compilation/qp_create_ninja.py
View File

@ -36,6 +36,7 @@ except ImportError:
from qp_path import QP_ROOT, QP_SRC, QP_EZFIO
LIB = "" # join(QP_ROOT, "lib", "rdtsc.o")
GPI_LIB = join(QP_ROOT, "lib64", "libGPI2.a")
EZFIO_LIB = join(QP_ROOT, "lib", "libezfio_irp.a")
ZMQ_LIB = join(QP_ROOT, "lib", "libf77zmq.a") + " " + join(QP_ROOT, "lib", "libzmq.a") + " -lstdc++ -lrt"
ROOT_BUILD_NINJA = join(QP_ROOT, "config", "build.ninja")
@ -96,8 +97,7 @@ def ninja_create_env_variable(pwd_config_file):
l_string.append(str_)
lib_lapack = get_compilation_option(pwd_config_file, "LAPACK_LIB")
lib_gpi2 = get_compilation_option(pwd_config_file, "GPI2_LIB")
str_lib = " ".join([LIB, lib_lapack, lib_gpi2, EZFIO_LIB, ZMQ_LIB])
str_lib = " ".join([LIB, lib_lapack, GPI_LIB, EZFIO_LIB, ZMQ_LIB])
l_string.append("LIB = {0} ".format(str_lib))
l_string.append("")

2
src/Determinants/two_body_dm_map.irp.f
View File

@ -187,7 +187,7 @@ subroutine add_values_to_two_body_dm_map(mask_ijkl)
print*,'n_elements = ',n_elements
call insert_into_two_body_dm_ab_map(n_elements,buffer_i,buffer_value,&
real(mo_integrals_threshold,integral_kind))
call map_unique(two_body_dm_ab_map)
call map_merge(two_body_dm_ab_map)
deallocate(buffer_i,buffer_value)

10
src/Integrals_Bielec/mo_bi_integrals.irp.f
View File

@ -146,7 +146,7 @@ subroutine set_integrals_jj_into_map
enddo
call insert_into_mo_integrals_map(n_integrals,buffer_i,buffer_value,&
real(mo_integrals_threshold,integral_kind))
call map_unique(mo_integrals_map)
call map_merge(mo_integrals_map)
end
subroutine set_integrals_exchange_jj_into_map
@ -167,7 +167,7 @@ subroutine set_integrals_exchange_jj_into_map
enddo
call insert_into_mo_integrals_map(n_integrals,buffer_i,buffer_value,&
real(mo_integrals_threshold,integral_kind))
call map_unique(mo_integrals_map)
call map_merge(mo_integrals_map)
end
@ -458,7 +458,7 @@ subroutine add_integrals_to_map(mask_ijkl)
real(mo_integrals_threshold,integral_kind))
deallocate(buffer_i, buffer_value)
!$OMP END PARALLEL
call map_unique(mo_integrals_map)
call map_merge(mo_integrals_map)
call wall_time(wall_2)
call cpu_time(cpu_2)
@ -773,7 +773,7 @@ subroutine add_integrals_to_map_three_indices(mask_ijk)
real(mo_integrals_threshold,integral_kind))
deallocate(buffer_i, buffer_value)
!$OMP END PARALLEL
call map_unique(mo_integrals_map)
call map_merge(mo_integrals_map)
call wall_time(wall_2)
call cpu_time(cpu_2)
@ -1035,7 +1035,7 @@ subroutine add_integrals_to_map_no_exit_34(mask_ijkl)
! print*, 'Communicating the map'
! call communicate_mo_integrals()
!IRP_ENDIF
call map_unique(mo_integrals_map)
call map_merge(mo_integrals_map)
call wall_time(wall_2)
call cpu_time(cpu_2)

49
src/Utils/map_module.f90
View File

@ -13,7 +13,7 @@ module map_module
! cache_map using a binary search
!
! When using the map_update subroutine to build the map,
! the map_unique subroutine
! the map_merge subroutine
! should be called before getting data from the map.
use omp_lib
@ -274,7 +274,7 @@ subroutine map_sort(map)
end
subroutine cache_map_unique(map)
subroutine cache_map_merge(map)
use map_module
implicit none
type (cache_map_type), intent(inout) :: map
@ -298,6 +298,28 @@ subroutine cache_map_unique(map)
end
subroutine cache_map_unique(map)
use map_module
implicit none
type (cache_map_type), intent(inout) :: map
integer(cache_key_kind) :: prev_key
integer(cache_map_size_kind) :: i, j
call cache_map_sort(map)
prev_key = -1_8
j=0
do i=1,map%n_elements
if (map%key(i) /= prev_key) then
j = j+1
map%value(j) = map%value(i)
map%key(j) = map%key(i)
prev_key = map%key(i)
endif
enddo
map%n_elements = j
end
subroutine cache_map_shrink(map,thr)
use map_module
implicit none
@ -338,6 +360,27 @@ subroutine map_unique(map)
end
subroutine map_merge(map)
use map_module
implicit none
type (map_type), intent(inout) :: map
integer(map_size_kind) :: i
integer(map_size_kind) :: icount
icount = 0_8
!$OMP PARALLEL DO SCHEDULE(dynamic,1000) DEFAULT(SHARED) PRIVATE(i)&
!$OMP REDUCTION(+:icount)
do i=0_8,map%map_size
call omp_set_lock(map%map(i)%lock)
call cache_map_merge(map%map(i))
call omp_unset_lock(map%map(i)%lock)
icount = icount + map%map(i)%n_elements
enddo
!$OMP END PARALLEL DO
map%n_elements = icount
end
subroutine map_shrink(map,thr)
use map_module
implicit none
@ -402,7 +445,7 @@ subroutine map_update(map, key, value, sze, thr)
else
! Assert that the map has a proper size
if (local_map%n_elements == local_map%map_size) then
call cache_map_unique(local_map)
call cache_map_merge(local_map)
call cache_map_reallocate(local_map, local_map%n_elements + local_map%n_elements)
call cache_map_shrink(local_map,thr)
endif