Line data Source code
1 : !--------------------------------------------------------------------------------------------------!
2 : ! CP2K: A general program to perform molecular dynamics simulations !
3 : ! Copyright 2000-2026 CP2K developers group <https://cp2k.org> !
4 : ! !
5 : ! SPDX-License-Identifier: GPL-2.0-or-later !
6 : !--------------------------------------------------------------------------------------------------!
7 :
8 : ! **************************************************************************************************
9 : !> \brief RT-BSE RI-RS kernels: SEX and Hartree evaluated by collocation on grid points r_l.
10 : !> Once-built grid objects:
11 : !> φ_µ(r_l) [grid×AO], Z_lP [grid×RI],
12 : !> V^aux_PQ = [M^-1 V^tr M^-1]_PQ [RI×RI], W^0_ll' = sum_PQ Z_lP (V + W^c(ω=0))_PQ Z_l'Q.
13 : !> Per-call kernels, collocation X = φ_µ(r_l) (AO domain, written φ_lµ below):
14 : !> SEX: ρ^grid_ll' = sum_µν φ_lµ Δρ_µν φ_l'ν ;
15 : !> Σ_µν = pref * sum_ll' φ_lµ [ρ^grid ∘ W^0]_ll' φ_l'ν
16 : !> Hartree: n_l = sum_µν φ_lµ Δρ_µν φ_lν ;
17 : !> v_l = sum_PQl' Z_lP V^aux_PQ Z_l'Q n_l' (applied factorized, stage 2) ;
18 : !> V^H_µν = sum_l φ_lµ v_l φ_lν (diagonal-only, no grid×grid)
19 : !> Independent of which GW variant produced bs_env%fm_W_MIC_freq_zero.
20 : !> \author Maximilian Graml (05.26)
21 : ! **************************************************************************************************
22 : MODULE rt_bse_ri_rs
23 : USE cp_cfm_basic_linalg, ONLY: cp_cfm_scale_and_add
24 : USE cp_cfm_types, ONLY: cp_cfm_create,&
25 : cp_cfm_release,&
26 : cp_cfm_to_fm,&
27 : cp_cfm_type,&
28 : cp_fm_to_cfm
29 : USE cp_dbcsr_api, ONLY: &
30 : dbcsr_add, dbcsr_copy, dbcsr_create, dbcsr_distribution_type, dbcsr_filter, &
31 : dbcsr_get_info, dbcsr_get_occupation, dbcsr_iterator_blocks_left, &
32 : dbcsr_iterator_next_block, dbcsr_iterator_start, dbcsr_iterator_stop, dbcsr_iterator_type, &
33 : dbcsr_multiply, dbcsr_release, dbcsr_scale, dbcsr_set, dbcsr_type, dbcsr_type_no_symmetry
34 : USE cp_dbcsr_contrib, ONLY: dbcsr_frobenius_norm,&
35 : dbcsr_get_diag
36 : USE cp_dbcsr_operations, ONLY: copy_dbcsr_to_fm,&
37 : copy_fm_to_dbcsr
38 : USE cp_fm_types, ONLY: cp_fm_create,&
39 : cp_fm_release,&
40 : cp_fm_type
41 : USE gw_utils_dbcsr, ONLY: dbcsr_contract_ABA,&
42 : hadamard_product_inplace
43 : USE gw_utils_fm, ONLY: fm_contract_ABA
44 : USE kinds, ONLY: dp
45 : USE machine, ONLY: m_walltime
46 : USE message_passing, ONLY: mp_para_env_type
47 : USE mp2_ri_2c, ONLY: RI_2c_integral_mat
48 : USE physcon, ONLY: angstrom
49 : USE post_scf_bandstructure_types, ONLY: post_scf_bandstructure_type
50 : USE qs_environment_types, ONLY: qs_environment_type
51 : USE ri_rs_matrices, ONLY: release_dbcsr_topology_and_matrices,&
52 : reserve_blocks_within_radius,&
53 : setup_square_topology
54 : USE ri_rs_setup, ONLY: ensure_ri_rs_representation
55 : #include "../base/base_uses.f90"
56 :
57 : IMPLICIT NONE
58 :
59 : PRIVATE
60 :
61 : CHARACTER(len=*), PARAMETER, PRIVATE :: moduleN = 'rt_bse_ri_rs'
62 :
63 : PUBLIC :: rt_bse_ri_rs_ensure_V_grid, &
64 : rt_bse_ri_rs_ensure_W0_grid, &
65 : compute_sigma_ri_rs, &
66 : compute_sigma_ri_rs_complex, &
67 : compute_hartree_ri_rs, &
68 : compute_hartree_ri_rs_complex, &
69 : compute_hartree_ri_rs_from_diag, &
70 : hartree_potential_from_diag_ri_rs
71 :
72 : CONTAINS
73 :
74 : ! **************************************************************************************************
75 : !> \brief Build V^aux_PQ = [M^-1 V^tr M^-1]_PQ (truncated Coulomb in the RI basis, M^-1-sandwiched
76 : !> to match the W^MIC convention). The grid Coulomb V_ll' = sum_PQ Z_lP V^aux_PQ Z_l'Q is
77 : !> never materialized -- the Hartree stage applies it factorized:
78 : !> v_l = sum_PQl' Z_lP V^aux_PQ Z_l'Q n_l'
79 : !> Needed by RT-BSE RI-RS Hartree.
80 : !> \param bs_env ...
81 : !> \param qs_env ...
82 : ! **************************************************************************************************
83 22 : SUBROUTINE rt_bse_ri_rs_ensure_V_grid(bs_env, qs_env)
84 :
85 : TYPE(post_scf_bandstructure_type), POINTER :: bs_env
86 : TYPE(qs_environment_type), POINTER :: qs_env
87 :
88 : CHARACTER(LEN=*), PARAMETER :: routineN = 'rt_bse_ri_rs_ensure_V_grid'
89 :
90 : INTEGER :: handle
91 22 : INTEGER, CONTIGUOUS, DIMENSION(:), POINTER :: blk_aux, dist_row_aux
92 : REAL(KIND=dp) :: t1
93 22 : TYPE(cp_fm_type), ALLOCATABLE, DIMENSION(:, :) :: fm_Vtr_Gamma
94 : TYPE(dbcsr_distribution_type) :: dist_aux_aux
95 :
96 22 : CALL timeset(routineN, handle)
97 :
98 22 : IF (bs_env%ri_rs%V_grid_built) THEN
99 0 : CALL timestop(handle)
100 0 : RETURN
101 : END IF
102 :
103 22 : CALL ensure_ri_rs_representation(qs_env, bs_env)
104 22 : t1 = m_walltime()
105 :
106 22 : CALL setup_square_topology(bs_env%ri_rs%mat_Z_lP, 'COL', dist_aux_aux, blk_aux, dist_row_aux)
107 :
108 : CALL RI_2c_integral_mat(qs_env, fm_Vtr_Gamma, bs_env%fm_RI_RI%matrix_struct, bs_env%n_RI, &
109 22 : bs_env%trunc_coulomb)
110 : ! Apply M^-1 sandwich to match the W^MIC convention; same scale as W^c when both used.
111 : ! M^-1(k=0) V^tr M^-1(k=0) -> fm_Vtr_Gamma
112 22 : CALL fm_contract_ABA(bs_env%fm_Minv_Gamma, fm_Vtr_Gamma(:, 1))
113 :
114 : ! Store the M^-1-sandwiched RI-basis Coulomb; the grid kernel Z V Z^T is applied factorized.
115 : CALL dbcsr_create(bs_env%ri_rs%mat_V_aux_rtbse, "V_aux_rtbse", dist_aux_aux, &
116 22 : dbcsr_type_no_symmetry, blk_aux, blk_aux)
117 : CALL copy_fm_to_dbcsr(fm_Vtr_Gamma(1, 1), bs_env%ri_rs%mat_V_aux_rtbse, &
118 22 : keep_sparsity=.FALSE.)
119 :
120 22 : bs_env%ri_rs%V_grid_built = .TRUE.
121 :
122 22 : IF (bs_env%unit_nr > 0) THEN
123 : WRITE (bs_env%unit_nr, '(T2,A,T57,A,F7.1,A)') &
124 11 : 'Precomputed RT-BSE RI-RS V_aux kernel,', ' Execution time', &
125 22 : m_walltime() - t1, ' s'
126 11 : WRITE (bs_env%unit_nr, '(A)') ' '
127 : END IF
128 :
129 22 : CALL release_dbcsr_topology_and_matrices(dist=dist_aux_aux, mapped_dist=dist_row_aux)
130 22 : CALL cp_fm_release(fm_Vtr_Gamma)
131 :
132 22 : CALL timestop(handle)
133 :
134 44 : END SUBROUTINE rt_bse_ri_rs_ensure_V_grid
135 :
136 : ! **************************************************************************************************
137 : !> \brief Build W^0_ll' = sum_PQ Z_lP (V + W^c(ω=0))_PQ Z_l'Q (statically screened W on the grid).
138 : !> Needed by RT-BSE RI-RS SEX/COH. Reuses bs_env%fm_W_MIC_freq_zero which must already
139 : !> contain M^-1 W^c(ω=0) M^-1 (built by either GW path under the BSE rtp_method gate).
140 : !> W^0 enters only through the Hadamard ρ^grid ∘ W^0 -- its elements are needed, so it is
141 : !> the one persistent grid×grid object of the kernel layer.
142 : !> RTBSE%CUTOFF_RADIUS_W0 > 0 keeps only block pairs whose grid centroids lie within that
143 : !> radius; the per-step ρ^grid inherits the pattern, so both grid×grid objects shrink.
144 : !> \param bs_env ...
145 : !> \param qs_env ...
146 : ! **************************************************************************************************
147 22 : SUBROUTINE rt_bse_ri_rs_ensure_W0_grid(bs_env, qs_env)
148 :
149 : TYPE(post_scf_bandstructure_type), POINTER :: bs_env
150 : TYPE(qs_environment_type), POINTER :: qs_env
151 :
152 : CHARACTER(LEN=*), PARAMETER :: routineN = 'rt_bse_ri_rs_ensure_W0_grid'
153 :
154 : INTEGER :: handle
155 22 : INTEGER, CONTIGUOUS, DIMENSION(:), POINTER :: blk_aux, blk_grid, dist_col_grid, &
156 22 : dist_row_aux
157 : LOGICAL :: use_cutoff_w0
158 : REAL(KIND=dp) :: fnorm_w0, occ_w0, t1
159 22 : TYPE(cp_fm_type), ALLOCATABLE, DIMENSION(:, :) :: fm_Vtr_Gamma
160 : TYPE(dbcsr_distribution_type) :: dist_aux_aux, dist_grid_grid
161 : TYPE(dbcsr_type) :: matrix_V_aux, matrix_W_aux
162 :
163 22 : CALL timeset(routineN, handle)
164 :
165 22 : IF (bs_env%ri_rs%W0_grid_built) THEN
166 0 : CALL timestop(handle)
167 0 : RETURN
168 : END IF
169 :
170 : ! W(w=0) is built by the GW step only under its RTBSE rtp_method gate; reaching this consumer
171 : ! without it means gate and consumer disagree. Abort rather than read a never-created cp_fm.
172 22 : IF (.NOT. ASSOCIATED(bs_env%fm_W_MIC_freq_zero%matrix_struct)) THEN
173 : CALL cp_abort(__LOCATION__, &
174 : "RT-BSE RI-RS kernel needs the screened interaction W(w=0), which the GW "// &
175 : "step did not build. Select the RT-BSE propagator with '&RTBSE' or "// &
176 0 : "'&RTBSE RTBSE', not '&RTBSE TDDFT'.")
177 : END IF
178 :
179 22 : CALL ensure_ri_rs_representation(qs_env, bs_env)
180 22 : t1 = m_walltime()
181 :
182 : CALL setup_square_topology(bs_env%ri_rs%mat_phi_mu_l, 'ROW', dist_grid_grid, blk_grid, &
183 22 : dist_col_grid)
184 22 : CALL setup_square_topology(bs_env%ri_rs%mat_Z_lP, 'COL', dist_aux_aux, blk_aux, dist_row_aux)
185 :
186 : CALL RI_2c_integral_mat(qs_env, fm_Vtr_Gamma, bs_env%fm_RI_RI%matrix_struct, bs_env%n_RI, &
187 22 : bs_env%trunc_coulomb)
188 : ! M^-1(k=0) V^tr M^-1(k=0) -> fm_Vtr_Gamma
189 22 : CALL fm_contract_ABA(bs_env%fm_Minv_Gamma, fm_Vtr_Gamma(:, 1))
190 :
191 : CALL dbcsr_create(matrix_V_aux, "V_aux_rtbse_W0", dist_aux_aux, dbcsr_type_no_symmetry, &
192 22 : blk_aux, blk_aux)
193 22 : CALL copy_fm_to_dbcsr(fm_Vtr_Gamma(1, 1), matrix_V_aux, keep_sparsity=.FALSE.)
194 :
195 : CALL dbcsr_create(matrix_W_aux, "W_aux_rtbse", dist_aux_aux, dbcsr_type_no_symmetry, &
196 22 : blk_aux, blk_aux)
197 22 : CALL copy_fm_to_dbcsr(bs_env%fm_W_MIC_freq_zero, matrix_W_aux, keep_sparsity=.FALSE.)
198 22 : CALL dbcsr_add(matrix_W_aux, matrix_V_aux, 1.0_dp, 1.0_dp)
199 :
200 : CALL dbcsr_create(bs_env%ri_rs%mat_W0_grid_rtbse, "W0_grid_rtbse", dist_grid_grid, &
201 22 : dbcsr_type_no_symmetry, blk_grid, blk_grid)
202 :
203 : ! CUTOFF_RADIUS_W0: pre-seed only the in-radius block pairs, then let retain_sparsity confine
204 : ! the contraction to them. Gate on the value, never on ALLOCATED alone -- .AND. does not
205 : ! short-circuit in Fortran.
206 22 : use_cutoff_w0 = bs_env%ri_rs%cutoff_radius_w0 > 0.0_dp
207 :
208 : ! chunk_centroids is built only by the non-periodic RI-RS GW driver, so on a periodic cell
209 : ! the cut would silently fall through to the exact path while the log reports it as active.
210 22 : IF (use_cutoff_w0 .AND. .NOT. ALLOCATED(bs_env%ri_rs%chunk_centroids)) THEN
211 : CALL cp_abort(__LOCATION__, &
212 : "RTBSE%CUTOFF_RADIUS_W0 needs the per-grid-block centroids, which only the "// &
213 : "non-periodic RI-RS GW driver builds; this cell is periodic. Use '&CELL "// &
214 0 : "PERIODIC NONE', or drop CUTOFF_RADIUS_W0 to propagate with the exact W0.")
215 : END IF
216 :
217 : IF (use_cutoff_w0) THEN
218 : CALL reserve_blocks_within_radius(bs_env%ri_rs%mat_W0_grid_rtbse, &
219 : bs_env%ri_rs%chunk_centroids, &
220 6 : bs_env%ri_rs%cutoff_radius_w0)
221 : ! TODO: eps_filter is INERT here -- DBCSR skips multrec_filtering whenever
222 : ! keep_sparsity is set (dbcsr_mm_multrec.F: "IF (use_eps .AND. .NOT. keep_sparsity)").
223 : ! So a radius wider than the cell equals the uncut build only while eps_filter prunes
224 : ! nothing on W0, which holds for every system measured so far but is not guaranteed.
225 : CALL dbcsr_contract_ABA("N", "T", bs_env%ri_rs%mat_Z_lP, matrix_W_aux, &
226 : bs_env%ri_rs%mat_W0_grid_rtbse, bs_env%eps_filter, &
227 6 : retain_sparsity=.TRUE.)
228 : ELSE
229 : CALL dbcsr_contract_ABA("N", "T", bs_env%ri_rs%mat_Z_lP, matrix_W_aux, &
230 16 : bs_env%ri_rs%mat_W0_grid_rtbse, bs_env%eps_filter)
231 : END IF
232 :
233 22 : bs_env%ri_rs%W0_grid_built = .TRUE.
234 :
235 : ! Occupancy is a storage statistic with no magnitude, so pair it with the retained Frobenius
236 : ! weight -- the two together are the x-axis of a CUTOFF_RADIUS_W0 convergence scan. Both
237 : ! reduce over all ranks; occupancy saturates at 1.000 and says nothing about magnitude.
238 22 : occ_w0 = dbcsr_get_occupation(bs_env%ri_rs%mat_W0_grid_rtbse)
239 22 : fnorm_w0 = dbcsr_frobenius_norm(bs_env%ri_rs%mat_W0_grid_rtbse)
240 :
241 22 : IF (bs_env%unit_nr > 0) THEN
242 : WRITE (bs_env%unit_nr, '(T2,A,T67,F14.4)') &
243 11 : 'W0CUT| Cutoff radius, negative = off [angstrom]', bs_env%ri_rs%cutoff_radius_w0*angstrom
244 11 : WRITE (bs_env%unit_nr, '(T2,A,T67,F14.6)') 'W0CUT| Retained block fraction', occ_w0
245 11 : WRITE (bs_env%unit_nr, '(T2,A,T59,ES22.8)') 'W0CUT| Retained Frobenius norm', fnorm_w0
246 : WRITE (bs_env%unit_nr, '(T2,A,T57,A,F7.1,A)') &
247 11 : 'Precomputed RT-BSE RI-RS W0_grid kernel,', ' Execution time', &
248 22 : m_walltime() - t1, ' s'
249 11 : WRITE (bs_env%unit_nr, '(A)') ' '
250 : END IF
251 :
252 22 : CALL release_dbcsr_topology_and_matrices(dist=dist_grid_grid, mapped_dist=dist_col_grid)
253 : CALL release_dbcsr_topology_and_matrices(dist=dist_aux_aux, mapped_dist=dist_row_aux, &
254 22 : m1=matrix_V_aux, m2=matrix_W_aux)
255 22 : CALL cp_fm_release(fm_Vtr_Gamma)
256 :
257 22 : CALL timestop(handle)
258 :
259 44 : END SUBROUTINE rt_bse_ri_rs_ensure_W0_grid
260 :
261 : ! **************************************************************************************************
262 : !> \brief AO-domain SEX: Σ_µν = pref * sum_ll' φ_lµ [ρ^grid ∘ W^0]_ll' φ_l'ν,
263 : !> ρ^grid_ll' = sum_µν φ_lµ Δρ_µν φ_l'ν.
264 : !> The grid×grid ρ^grid is intrinsic to SEX -- the Hadamard needs W^0's elements, so no
265 : !> factorized application exists (unlike the Hartree V_ll'). Real input, real output;
266 : !> used for COH (input S^-1) and the init reference (ρ^0); dynamic Δρ goes through the
267 : !> complex variant. Mirrors the AO-RI get_sigma(rtbse_env, sigma_fm, prefactor, rho_fm) API.
268 : !> \param bs_env ...
269 : !> \param sigma_AO_fm result, AO x AO
270 : !> \param prefactor scaling applied to the final result
271 : !> \param rho_AO_fm input density-like matrix, AO x AO
272 : !> \param grid_diag_accum ...
273 : ! **************************************************************************************************
274 4600 : SUBROUTINE compute_sigma_ri_rs(bs_env, sigma_AO_fm, prefactor, rho_AO_fm, &
275 4600 : grid_diag_accum)
276 :
277 : TYPE(post_scf_bandstructure_type), POINTER :: bs_env
278 : TYPE(cp_fm_type), INTENT(INOUT) :: sigma_AO_fm
279 : REAL(KIND=dp), INTENT(IN) :: prefactor
280 : TYPE(cp_fm_type), INTENT(IN) :: rho_AO_fm
281 : REAL(KIND=dp), INTENT(INOUT), OPTIONAL :: grid_diag_accum(:)
282 :
283 : CHARACTER(LEN=*), PARAMETER :: routineN = 'compute_sigma_ri_rs'
284 :
285 : INTEGER :: handle, n_grid
286 4600 : INTEGER, CONTIGUOUS, DIMENSION(:), POINTER :: blk_ao, dist_row_ao
287 4600 : REAL(KIND=dp), ALLOCATABLE :: diag_local(:)
288 : TYPE(dbcsr_distribution_type) :: dist_ao_ao
289 : TYPE(dbcsr_type) :: matrix_rho_AO, matrix_rho_grid, &
290 : matrix_Sigma_AO
291 :
292 4600 : CALL timeset(routineN, handle)
293 :
294 4600 : CPASSERT(bs_env%ri_rs%W0_grid_built)
295 :
296 : ! Only the AO×AO topology is built here; ρ^grid takes its grid×grid one from W^0's template.
297 4600 : CALL setup_square_topology(bs_env%ri_rs%mat_phi_mu_l, 'COL', dist_ao_ao, blk_ao, dist_row_ao)
298 :
299 : CALL dbcsr_create(matrix_rho_AO, "rho_AO_ri_rs", dist_ao_ao, dbcsr_type_no_symmetry, &
300 4600 : blk_ao, blk_ao)
301 4600 : CALL copy_fm_to_dbcsr(rho_AO_fm, matrix_rho_AO, keep_sparsity=.FALSE.)
302 :
303 : ! The projection below is shared with the Hartree, so this cuts SEX and Hartree alike.
304 4600 : IF (bs_env%ri_rs%eps_filter_rho >= 0.0_dp) THEN
305 0 : CALL dbcsr_filter(matrix_rho_AO, bs_env%ri_rs%eps_filter_rho)
306 : END IF
307 :
308 : ! ρ^grid_ll' = sum_µν φ_lµ Δρ_µν φ_l'ν (the SEX-intrinsic grid×grid transient),
309 : ! built directly into W^0's block pattern. Exact: the Hadamard below zeroes every block
310 : ! without a W^0 partner anyway, so restricting the multiply drops only computed zeros.
311 : ! No eps_filter on the SEX projections either -- a Δρ-dependent block drop breaks kernel
312 : ! self-adjointness (L non-Hermitian); W^0's pattern is symmetric and Δρ-independent.
313 : ! TODO: seeds the pattern by copying W0's values and discarding them every RK4 substep --
314 : ! unquantified overhead, paid even when no cut is set. Cache the zeroed pattern instead.
315 4600 : CALL dbcsr_create(matrix_rho_grid, template=bs_env%ri_rs%mat_W0_grid_rtbse)
316 4600 : CALL dbcsr_copy(matrix_rho_grid, bs_env%ri_rs%mat_W0_grid_rtbse)
317 4600 : CALL dbcsr_set(matrix_rho_grid, 0.0_dp)
318 : CALL dbcsr_contract_ABA("N", "T", bs_env%ri_rs%mat_phi_mu_l, matrix_rho_AO, &
319 4600 : matrix_rho_grid, 0.0_dp, retain_sparsity=.TRUE.)
320 :
321 : ! Harvest n_l = diag(ρ^grid) BEFORE the in-place Hadamard destroys it; the Hartree reuses it
322 : ! (compute_hartree_ri_rs_from_diag) instead of rebuilding φρφ^T. Bare accumulate (caller
323 : ! pre-zeroes): the cross-spin Hartree density is the spin SUM of these, and spin_degeneracy is
324 : ! applied to the V_H OUTPUT (post-filter, bit-identical) -- never to the diagonal, which would
325 : ! shift the stage-3 eps_filter cut (coarse-filter sensitivity).
326 4600 : IF (PRESENT(grid_diag_accum)) THEN
327 3040 : n_grid = SIZE(grid_diag_accum)
328 9120 : ALLOCATE (diag_local(n_grid))
329 3040 : diag_local = 0.0_dp
330 3040 : CALL dbcsr_get_diag(matrix_rho_grid, diag_local)
331 3040 : CALL bs_env%para_env%sum(diag_local)
332 889440 : grid_diag_accum(:) = grid_diag_accum(:) + diag_local(:)
333 3040 : DEALLOCATE (diag_local)
334 : END IF
335 :
336 : ! ρ^grid_ll' <- ρ^grid_ll' * W^0_ll' (in place; blocks without a W^0 partner zeroed)
337 : CALL hadamard_product_inplace(matrix_rho_grid, bs_env%ri_rs%mat_W0_grid_rtbse, &
338 4600 : 1.0_dp)
339 :
340 : ! Σ_µν = sum_ll' φ_lµ [ρ^grid ∘ W^0]_ll' φ_l'ν
341 4600 : CALL dbcsr_create(matrix_Sigma_AO, template=matrix_rho_AO)
342 : ! Unfiltered for self-adjointness (see the forward projection above; PERF note there).
343 : CALL dbcsr_contract_ABA("T", "N", bs_env%ri_rs%mat_phi_mu_l, matrix_rho_grid, &
344 4600 : matrix_Sigma_AO, 0.0_dp)
345 :
346 4600 : CALL dbcsr_scale(matrix_Sigma_AO, prefactor)
347 4600 : CALL copy_dbcsr_to_fm(matrix_Sigma_AO, sigma_AO_fm)
348 :
349 : CALL release_dbcsr_topology_and_matrices(dist=dist_ao_ao, mapped_dist=dist_row_ao, &
350 4600 : m1=matrix_rho_AO, m2=matrix_Sigma_AO)
351 4600 : CALL release_dbcsr_topology_and_matrices(m1=matrix_rho_grid)
352 :
353 4600 : CALL timestop(handle)
354 :
355 9200 : END SUBROUTINE compute_sigma_ri_rs
356 :
357 : ! **************************************************************************************************
358 : !> \brief Complex-input AO SEX via Re/Im split: the kernel is real, so complex linearity holds as
359 : !> Σ[Δρ] = Σ[Re Δρ] + i Σ[Im Δρ]. Required for non-Hermitian Δρ inputs
360 : !> (TDA OV-only / ABBA OV+VO).
361 : !> \param bs_env ...
362 : !> \param sigma_AO_cfm result, AO x AO (complex)
363 : !> \param prefactor scaling applied to the final result
364 : !> \param rho_AO_cfm input AO x AO complex matrix
365 : !> \param grid_diag_re_accum optional: accumulate diag(φ.Re(ρ).φ^T) (bare; for the Hartree reuse)
366 : !> \param grid_diag_im_accum optional: accumulate diag(φ.Im(ρ).φ^T) (bare; for the Hartree reuse)
367 : ! **************************************************************************************************
368 11500 : SUBROUTINE compute_sigma_ri_rs_complex(bs_env, sigma_AO_cfm, prefactor, rho_AO_cfm, &
369 2300 : grid_diag_re_accum, grid_diag_im_accum)
370 :
371 : TYPE(post_scf_bandstructure_type), POINTER :: bs_env
372 : TYPE(cp_cfm_type), INTENT(INOUT) :: sigma_AO_cfm
373 : REAL(KIND=dp), INTENT(IN) :: prefactor
374 : TYPE(cp_cfm_type), INTENT(IN) :: rho_AO_cfm
375 : REAL(KIND=dp), INTENT(INOUT), OPTIONAL :: grid_diag_re_accum(:), &
376 : grid_diag_im_accum(:)
377 :
378 : CHARACTER(LEN=*), PARAMETER :: routineN = 'compute_sigma_ri_rs_complex'
379 :
380 : INTEGER :: handle
381 : TYPE(cp_cfm_type) :: cfm_real_part
382 : TYPE(cp_fm_type) :: fm_rho, fm_sigma
383 :
384 2300 : CALL timeset(routineN, handle)
385 :
386 2300 : CALL cp_fm_create(fm_rho, bs_env%fm_s_Gamma%matrix_struct)
387 2300 : CALL cp_fm_create(fm_sigma, bs_env%fm_s_Gamma%matrix_struct)
388 2300 : CALL cp_cfm_create(cfm_real_part, bs_env%fm_s_Gamma%matrix_struct)
389 :
390 : ! Re/Im each harvest into their own accumulator; absent optionals propagate as absent.
391 2300 : CALL cp_cfm_to_fm(msource=rho_AO_cfm, mtargetr=fm_rho)
392 : CALL compute_sigma_ri_rs(bs_env, fm_sigma, prefactor, fm_rho, &
393 2680 : grid_diag_accum=grid_diag_re_accum)
394 2300 : CALL cp_fm_to_cfm(msourcer=fm_sigma, mtarget=cfm_real_part)
395 :
396 2300 : CALL cp_cfm_to_fm(msource=rho_AO_cfm, mtargeti=fm_rho)
397 : CALL compute_sigma_ri_rs(bs_env, fm_sigma, prefactor, fm_rho, &
398 3480 : grid_diag_accum=grid_diag_im_accum)
399 2300 : CALL cp_fm_to_cfm(msourcei=fm_sigma, mtarget=sigma_AO_cfm)
400 :
401 : CALL cp_cfm_scale_and_add(CMPLX(1.0_dp, 0.0_dp, kind=dp), sigma_AO_cfm, &
402 2300 : CMPLX(1.0_dp, 0.0_dp, kind=dp), cfm_real_part)
403 :
404 2300 : CALL cp_fm_release(fm_rho)
405 2300 : CALL cp_fm_release(fm_sigma)
406 2300 : CALL cp_cfm_release(cfm_real_part)
407 :
408 2300 : CALL timestop(handle)
409 :
410 2300 : END SUBROUTINE compute_sigma_ri_rs_complex
411 :
412 : ! **************************************************************************************************
413 : !> \brief AO-domain Hartree via RI-RS:
414 : !> n_l = sum_µν φ_lµ Δρ_µν φ_lν = (φ ρ φ^T)_ll (diagonal of materialized grid×grid) ;
415 : !> v_l = sum_PQl' Z_lP V^aux_PQ Z_l'Q n_l' (factorized, stage 2) ;
416 : !> V^H_µν = sum_l φ_lµ v_l φ_lν (diagonal-only row-scale, stage 3)
417 : !> Real input, real output; complex inputs go through compute_hartree_ri_rs_complex.
418 : !> \param bs_env ...
419 : !> \param rho_AO_fm input AO x AO density matrix
420 : !> \param V_H_AO_fm output AO x AO Hartree potential
421 : ! **************************************************************************************************
422 578 : SUBROUTINE compute_hartree_ri_rs(bs_env, rho_AO_fm, V_H_AO_fm)
423 :
424 : TYPE(post_scf_bandstructure_type), POINTER :: bs_env
425 : TYPE(cp_fm_type), INTENT(IN) :: rho_AO_fm
426 : TYPE(cp_fm_type), INTENT(INOUT) :: V_H_AO_fm
427 :
428 : CHARACTER(LEN=*), PARAMETER :: routineN = 'compute_hartree_ri_rs'
429 :
430 : INTEGER :: handle, n_grid
431 578 : INTEGER, CONTIGUOUS, DIMENSION(:), POINTER :: blk_ao, blk_grid, dist_col_grid, &
432 578 : dist_row_ao
433 : REAL(KIND=dp), ALLOCATABLE :: n_vec(:)
434 : TYPE(dbcsr_distribution_type) :: dist_ao_ao, dist_grid_grid
435 : TYPE(dbcsr_type) :: matrix_rho_AO, matrix_rho_grid
436 :
437 578 : CALL timeset(routineN, handle)
438 :
439 578 : CPASSERT(bs_env%ri_rs%V_grid_built)
440 :
441 578 : CALL setup_square_topology(bs_env%ri_rs%mat_phi_mu_l, 'COL', dist_ao_ao, blk_ao, dist_row_ao)
442 : CALL setup_square_topology(bs_env%ri_rs%mat_phi_mu_l, 'ROW', dist_grid_grid, blk_grid, &
443 578 : dist_col_grid)
444 :
445 : CALL dbcsr_create(matrix_rho_AO, "rho_AO_hartree", dist_ao_ao, dbcsr_type_no_symmetry, &
446 578 : blk_ao, blk_ao)
447 578 : CALL copy_fm_to_dbcsr(rho_AO_fm, matrix_rho_AO, keep_sparsity=.FALSE.)
448 :
449 6710 : n_grid = SUM(blk_grid)
450 1734 : ALLOCATE (n_vec(n_grid))
451 :
452 : ! stage 1: n_l = (φ ρ φ^T)_ll. A grid×AO row-dot (iterate φρ, look up φ) silently drops
453 : ! off-rank pairs at ≥2 ranks — φ is sparse and the product is not co-located with φ — so
454 : ! materialize φρφ^T and take its diagonal, co-location-safe like the SEX grid kernel.
455 : CALL dbcsr_create(matrix_rho_grid, "rho_grid_hartree", dist_grid_grid, &
456 578 : dbcsr_type_no_symmetry, blk_grid, blk_grid)
457 : ! No eps_filter: a Δρ-dependent diagonal-block drop zeroes n_l inconsistently between OV
458 : ! pairs and breaks Hartree self-adjointness (L non-Hermitian).
459 : ! PERF(rirs-selfadjoint): this full grid×grid is built only for its diagonal n_l; a
460 : ! diagonal-only build would avoid it. NOT YET APPLIED, needs a perf test.
461 : CALL dbcsr_contract_ABA("N", "T", bs_env%ri_rs%mat_phi_mu_l, matrix_rho_AO, &
462 578 : matrix_rho_grid, 0.0_dp)
463 578 : n_vec = 0.0_dp
464 578 : CALL dbcsr_get_diag(matrix_rho_grid, n_vec)
465 578 : CALL bs_env%para_env%sum(n_vec)
466 : CALL release_dbcsr_topology_and_matrices(dist=dist_grid_grid, mapped_dist=dist_col_grid, &
467 578 : m1=matrix_rho_grid)
468 : CALL release_dbcsr_topology_and_matrices(dist=dist_ao_ao, mapped_dist=dist_row_ao, &
469 578 : m1=matrix_rho_AO)
470 :
471 : ! stages 2-3: factorized Coulomb v = Z V^aux Z^T n, then V^H = φ^T diag(v) φ.
472 578 : CALL hartree_potential_from_diag_ri_rs(bs_env, n_vec, V_H_AO_fm)
473 :
474 578 : DEALLOCATE (n_vec)
475 :
476 578 : CALL timestop(handle)
477 :
478 1156 : END SUBROUTINE compute_hartree_ri_rs
479 :
480 : ! **************************************************************************************************
481 : !> \brief Hartree stages 2-3 from a precomputed grid density n_l (skips the stage-1 φρφ^T build):
482 : !> v_l = sum_PQl' Z_lP V^aux_PQ Z_l'Q n_l' (factorized Coulomb) ;
483 : !> V^H_µν = sum_l φ_lµ v_l φ_lν (Φ_lν = v_l φ_lν rowscale, then V^H = φ^T Φ).
484 : !> n_l is harvested as diag(φρφ^T) inside compute_sigma_ri_rs (the SEX grid kernel), so the
485 : !> Hartree never rebuilds the grid×grid product. Real in/out.
486 : !> \param bs_env ...
487 : !> \param n_vec grid density n_l (length n_grid, replicated)
488 : !> \param V_H_AO_fm output AO x AO Hartree potential
489 : ! **************************************************************************************************
490 3138 : SUBROUTINE hartree_potential_from_diag_ri_rs(bs_env, n_vec, V_H_AO_fm)
491 :
492 : TYPE(post_scf_bandstructure_type), POINTER :: bs_env
493 : REAL(KIND=dp), INTENT(IN) :: n_vec(:)
494 : TYPE(cp_fm_type), INTENT(INOUT) :: V_H_AO_fm
495 :
496 : CHARACTER(LEN=*), PARAMETER :: routineN = 'hartree_potential_from_diag_ri_rs'
497 :
498 : INTEGER :: handle
499 3138 : INTEGER, CONTIGUOUS, DIMENSION(:), POINTER :: blk_ao, dist_row_ao
500 : REAL(KIND=dp), ALLOCATABLE :: u_ri(:), v_vec(:), w_ri(:)
501 : TYPE(dbcsr_distribution_type) :: dist_ao_ao
502 : TYPE(dbcsr_type) :: matrix_Phi, matrix_V_H_AO
503 :
504 3138 : CALL timeset(routineN, handle)
505 :
506 3138 : CPASSERT(bs_env%ri_rs%V_grid_built)
507 :
508 3138 : CALL setup_square_topology(bs_env%ri_rs%mat_phi_mu_l, 'COL', dist_ao_ao, blk_ao, dist_row_ao)
509 9414 : ALLOCATE (v_vec(SIZE(n_vec)))
510 :
511 : ! stage 2: v_l = sum_PQl' Z_lP V^aux_PQ Z_l'Q n_l' (factorized; helper zeroes each output)
512 12552 : ALLOCATE (w_ri(bs_env%n_RI), u_ri(bs_env%n_RI))
513 : ! w_Q = sum_l Z_lQ n_l
514 : CALL dbcsr_matvec_replicated(bs_env%ri_rs%mat_Z_lP, n_vec, w_ri, bs_env%para_env, &
515 3138 : transposed=.TRUE.)
516 : ! u_P = sum_Q V^aux_PQ w_Q
517 3138 : CALL dbcsr_matvec_replicated(bs_env%ri_rs%mat_V_aux_rtbse, w_ri, u_ri, bs_env%para_env)
518 : ! v_l = sum_P Z_lP u_P
519 3138 : CALL dbcsr_matvec_replicated(bs_env%ri_rs%mat_Z_lP, u_ri, v_vec, bs_env%para_env)
520 3138 : DEALLOCATE (w_ri, u_ri)
521 :
522 : ! stage 3: V^H_µν = sum_l φ_lµ v_l φ_lν (Φ_lν = v_l φ_lν rowscale, then V^H = φ^T Φ)
523 3138 : CALL dbcsr_create(matrix_Phi, template=bs_env%ri_rs%mat_phi_mu_l)
524 3138 : CALL dbcsr_copy(matrix_Phi, bs_env%ri_rs%mat_phi_mu_l)
525 3138 : CALL dbcsr_scale_rows_replicated(matrix_Phi, v_vec)
526 :
527 : CALL dbcsr_create(matrix_V_H_AO, "V_H_AO_hartree", dist_ao_ao, dbcsr_type_no_symmetry, &
528 3138 : blk_ao, blk_ao)
529 : ! No eps_filter: a Δρ-dependent block drop breaks Hartree kernel self-adjointness.
530 : CALL dbcsr_multiply("T", "N", 1.0_dp, bs_env%ri_rs%mat_phi_mu_l, matrix_Phi, &
531 3138 : 0.0_dp, matrix_V_H_AO, filter_eps=0.0_dp)
532 3138 : CALL dbcsr_release(matrix_Phi)
533 :
534 3138 : CALL copy_dbcsr_to_fm(matrix_V_H_AO, V_H_AO_fm)
535 :
536 3138 : DEALLOCATE (v_vec)
537 :
538 : CALL release_dbcsr_topology_and_matrices(dist=dist_ao_ao, mapped_dist=dist_row_ao, &
539 3138 : m1=matrix_V_H_AO)
540 :
541 3138 : CALL timestop(handle)
542 :
543 6276 : END SUBROUTINE hartree_potential_from_diag_ri_rs
544 :
545 : ! **************************************************************************************************
546 : !> \brief Complex Hartree from precomputed grid diagonals: V^H = V^H[n_re] + i V^H[n_im], each via
547 : !> hartree_potential_from_diag_ri_rs (stages 2-3 only). n_re/n_im are the spin-summed grid
548 : !> densities harvested in the SEX kernel; this is the cross-spin / TDA complex consumer that
549 : !> replaces compute_hartree_ri_rs_complex when SEX already built the grid. n_im optional: when
550 : !> absent the result is purely real (matches the Re-only real-input Hartree).
551 : !> \param bs_env ...
552 : !> \param n_re grid density Re part (length n_grid, replicated)
553 : !> \param V_H_AO_cfm output AO x AO complex Hartree potential
554 : !> \param n_im optional grid density Im part (length n_grid, replicated)
555 : ! **************************************************************************************************
556 1600 : SUBROUTINE compute_hartree_ri_rs_from_diag(bs_env, n_re, V_H_AO_cfm, n_im)
557 :
558 : TYPE(post_scf_bandstructure_type), POINTER :: bs_env
559 : REAL(KIND=dp), INTENT(IN) :: n_re(:)
560 : TYPE(cp_cfm_type), INTENT(INOUT) :: V_H_AO_cfm
561 : REAL(KIND=dp), INTENT(IN), OPTIONAL :: n_im(:)
562 :
563 : CHARACTER(LEN=*), PARAMETER :: routineN = 'compute_hartree_ri_rs_from_diag'
564 :
565 : INTEGER :: handle
566 : TYPE(cp_cfm_type) :: cfm_real_part
567 : TYPE(cp_fm_type) :: fm_v
568 :
569 1600 : CALL timeset(routineN, handle)
570 :
571 1600 : CALL cp_fm_create(fm_v, bs_env%fm_s_Gamma%matrix_struct)
572 :
573 1600 : CALL hartree_potential_from_diag_ri_rs(bs_env, n_re, fm_v)
574 1600 : IF (PRESENT(n_im)) THEN
575 960 : CALL cp_cfm_create(cfm_real_part, bs_env%fm_s_Gamma%matrix_struct)
576 960 : CALL cp_fm_to_cfm(msourcer=fm_v, mtarget=cfm_real_part)
577 960 : CALL hartree_potential_from_diag_ri_rs(bs_env, n_im, fm_v)
578 960 : CALL cp_fm_to_cfm(msourcei=fm_v, mtarget=V_H_AO_cfm)
579 : CALL cp_cfm_scale_and_add(CMPLX(1.0_dp, 0.0_dp, kind=dp), V_H_AO_cfm, &
580 960 : CMPLX(1.0_dp, 0.0_dp, kind=dp), cfm_real_part)
581 960 : CALL cp_cfm_release(cfm_real_part)
582 : ELSE
583 640 : CALL cp_fm_to_cfm(msourcer=fm_v, mtarget=V_H_AO_cfm)
584 : END IF
585 :
586 1600 : CALL cp_fm_release(fm_v)
587 :
588 1600 : CALL timestop(handle)
589 :
590 1600 : END SUBROUTINE compute_hartree_ri_rs_from_diag
591 :
592 : ! **************************************************************************************************
593 : !> \brief Complex-input Hartree potential via RI-RS. Re/Im split: feed each part to the real
594 : !> compute_hartree_ri_rs and reassemble. Real-input Hartree on a non-Hermitian input
595 : !> would silently drop Im and break Hermitian conjugacy of OV+VO contributions in TDA.
596 : !> \param bs_env ...
597 : !> \param rho_AO_cfm input AO x AO complex density-like matrix
598 : !> \param V_H_AO_cfm output AO x AO complex Hartree potential
599 : ! **************************************************************************************************
600 1144 : SUBROUTINE compute_hartree_ri_rs_complex(bs_env, rho_AO_cfm, V_H_AO_cfm)
601 :
602 : TYPE(post_scf_bandstructure_type), POINTER :: bs_env
603 : TYPE(cp_cfm_type), INTENT(IN) :: rho_AO_cfm
604 : TYPE(cp_cfm_type), INTENT(INOUT) :: V_H_AO_cfm
605 :
606 : CHARACTER(LEN=*), PARAMETER :: routineN = 'compute_hartree_ri_rs_complex'
607 :
608 : INTEGER :: handle
609 : TYPE(cp_cfm_type) :: cfm_real_part
610 : TYPE(cp_fm_type) :: fm_rho, fm_v
611 :
612 286 : CALL timeset(routineN, handle)
613 :
614 286 : CALL cp_fm_create(fm_rho, bs_env%fm_s_Gamma%matrix_struct)
615 286 : CALL cp_fm_create(fm_v, bs_env%fm_s_Gamma%matrix_struct)
616 286 : CALL cp_cfm_create(cfm_real_part, bs_env%fm_s_Gamma%matrix_struct)
617 :
618 286 : CALL cp_cfm_to_fm(msource=rho_AO_cfm, mtargetr=fm_rho)
619 286 : CALL compute_hartree_ri_rs(bs_env, fm_rho, fm_v)
620 286 : CALL cp_fm_to_cfm(msourcer=fm_v, mtarget=cfm_real_part)
621 :
622 286 : CALL cp_cfm_to_fm(msource=rho_AO_cfm, mtargeti=fm_rho)
623 286 : CALL compute_hartree_ri_rs(bs_env, fm_rho, fm_v)
624 286 : CALL cp_fm_to_cfm(msourcei=fm_v, mtarget=V_H_AO_cfm)
625 :
626 : CALL cp_cfm_scale_and_add(CMPLX(1.0_dp, 0.0_dp, kind=dp), V_H_AO_cfm, &
627 286 : CMPLX(1.0_dp, 0.0_dp, kind=dp), cfm_real_part)
628 :
629 286 : CALL cp_fm_release(fm_rho)
630 286 : CALL cp_fm_release(fm_v)
631 286 : CALL cp_cfm_release(cfm_real_part)
632 :
633 286 : CALL timestop(handle)
634 :
635 286 : END SUBROUTINE compute_hartree_ri_rs_complex
636 :
637 : ! **************************************************************************************************
638 : !> \brief Scale each row of a dbcsr matrix by a replicated full-length vector:
639 : !> block(ir,ic) <- vec(global_row(ir)) * block(ir,ic). Value mutation only.
640 : !> \param matrix ...
641 : !> \param vec full-length replicated row-scaling vector
642 : ! **************************************************************************************************
643 3138 : SUBROUTINE dbcsr_scale_rows_replicated(matrix, vec)
644 :
645 : TYPE(dbcsr_type), INTENT(INOUT) :: matrix
646 : REAL(KIND=dp), INTENT(IN) :: vec(:)
647 :
648 : INTEGER :: col_blk, ic, ir, nblkrows_total, &
649 : row_blk, row_off
650 3138 : INTEGER, ALLOCATABLE :: row_offset(:)
651 3138 : INTEGER, DIMENSION(:), POINTER :: row_blk_size
652 3138 : REAL(KIND=dp), DIMENSION(:, :), POINTER :: blk
653 : TYPE(dbcsr_iterator_type) :: iter
654 :
655 3138 : CALL dbcsr_get_info(matrix, nblkrows_total=nblkrows_total, row_blk_size=row_blk_size)
656 9414 : ALLOCATE (row_offset(nblkrows_total + 1))
657 3138 : row_offset(1) = 0
658 35830 : DO ir = 1, nblkrows_total
659 35830 : row_offset(ir + 1) = row_offset(ir) + row_blk_size(ir)
660 : END DO
661 :
662 3138 : CALL dbcsr_iterator_start(iter, matrix)
663 35830 : DO WHILE (dbcsr_iterator_blocks_left(iter))
664 32692 : CALL dbcsr_iterator_next_block(iter, row_blk, col_blk, blk)
665 32692 : row_off = row_offset(row_blk)
666 231982 : DO ic = 1, SIZE(blk, 2)
667 5594092 : DO ir = 1, SIZE(blk, 1)
668 5561400 : blk(ir, ic) = vec(row_off + ir)*blk(ir, ic)
669 : END DO
670 : END DO
671 : END DO
672 3138 : CALL dbcsr_iterator_stop(iter)
673 :
674 3138 : DEALLOCATE (row_offset)
675 :
676 6276 : END SUBROUTINE dbcsr_scale_rows_replicated
677 :
678 : ! **************************************************************************************************
679 : !> \brief Replicated matvec: vec_out = matrix * vec_in (or matrix^T * vec_in if transposed) for a
680 : !> possibly rectangular distributed dbcsr matrix and replicated full-length vectors.
681 : !> Iterates over local blocks and reduces.
682 : !> \param matrix distributed dbcsr matrix (may be rectangular)
683 : !> \param vec_in full input vector, replicated on all ranks (column length, or row length if transposed)
684 : !> \param vec_out full output vector, replicated on all ranks (zeroed on entry; sum-reduced on exit)
685 : !> \param para_env ...
686 : !> \param transposed if .TRUE. compute vec_out = matrix^T * vec_in
687 : ! **************************************************************************************************
688 9414 : SUBROUTINE dbcsr_matvec_replicated(matrix, vec_in, vec_out, para_env, transposed)
689 :
690 : TYPE(dbcsr_type), INTENT(INOUT) :: matrix
691 : REAL(KIND=dp), INTENT(IN) :: vec_in(:)
692 : REAL(KIND=dp), INTENT(INOUT) :: vec_out(:)
693 : TYPE(mp_para_env_type), POINTER :: para_env
694 : LOGICAL, INTENT(IN), OPTIONAL :: transposed
695 :
696 : INTEGER :: col_blk, col_off, ic, ir, &
697 : nblkcols_total, nblkrows_total, &
698 : row_blk, row_off
699 9414 : INTEGER, ALLOCATABLE :: col_offset(:), row_offset(:)
700 9414 : INTEGER, DIMENSION(:), POINTER :: col_blk_size, row_blk_size
701 : LOGICAL :: my_trans
702 9414 : REAL(KIND=dp), DIMENSION(:, :), POINTER :: block
703 : TYPE(dbcsr_iterator_type) :: iter
704 :
705 9414 : my_trans = .FALSE.
706 3138 : IF (PRESENT(transposed)) my_trans = transposed
707 :
708 : CALL dbcsr_get_info(matrix, nblkrows_total=nblkrows_total, &
709 : nblkcols_total=nblkcols_total, &
710 9414 : row_blk_size=row_blk_size, col_blk_size=col_blk_size)
711 :
712 47070 : ALLOCATE (row_offset(nblkrows_total + 1), col_offset(nblkcols_total + 1))
713 9414 : row_offset(1) = 0
714 81074 : DO ir = 1, nblkrows_total
715 81074 : row_offset(ir + 1) = row_offset(ir) + row_blk_size(ir)
716 : END DO
717 9414 : col_offset(1) = 0
718 28242 : DO ic = 1, nblkcols_total
719 28242 : col_offset(ic + 1) = col_offset(ic) + col_blk_size(ic)
720 : END DO
721 :
722 1192318 : vec_out(:) = 0.0_dp
723 :
724 9414 : CALL dbcsr_iterator_start(iter, matrix)
725 81074 : DO WHILE (dbcsr_iterator_blocks_left(iter))
726 71660 : CALL dbcsr_iterator_next_block(iter, row_blk, col_blk, block)
727 71660 : row_off = row_offset(row_blk)
728 71660 : col_off = col_offset(col_blk)
729 81074 : IF (my_trans) THEN
730 784608 : DO ic = 1, SIZE(block, 2)
731 21351392 : DO ir = 1, SIZE(block, 1)
732 : vec_out(col_off + ic) = vec_out(col_off + ic) + &
733 21318700 : block(ir, ic)*vec_in(row_off + ir)
734 : END DO
735 : END DO
736 : ELSE
737 935232 : DO ic = 1, SIZE(block, 2)
738 24822020 : DO ir = 1, SIZE(block, 1)
739 : vec_out(row_off + ir) = vec_out(row_off + ir) + &
740 24783052 : block(ir, ic)*vec_in(col_off + ic)
741 : END DO
742 : END DO
743 : END IF
744 : END DO
745 9414 : CALL dbcsr_iterator_stop(iter)
746 :
747 9414 : DEALLOCATE (row_offset, col_offset)
748 :
749 2375222 : CALL para_env%sum(vec_out)
750 :
751 21966 : END SUBROUTINE dbcsr_matvec_replicated
752 :
753 : END MODULE rt_bse_ri_rs
|