RI HFX: minor tweaks for block sizes and complete output of parameters

cp2k · Mar 7, 2020 · e32a109 · e32a109
1 parent 3568f8f
commit e32a109
Show file tree

Hide file tree

Showing 3 changed files with 45 additions and 27 deletions.
diff --git a/src/hfx_ri.F b/src/hfx_ri.F
@@ -710,8 +710,7 @@ SUBROUTINE hfx_ri_update_ks_mo(qs_env, ri_data, ks_matrix, mo_coeff, &
       TYPE(cp_para_env_type), POINTER                    :: para_env
       TYPE(dbcsr_distribution_type)                      :: ks_dist
       TYPE(dbcsr_t_pgrid_type)                           :: pgrid_1, pgrid_2, pgrid_2d
-      TYPE(dbcsr_t_type)                                 :: ks_t, ks_t_split, mo_coeff_t, &
-                                                            mo_coeff_t_split
+      TYPE(dbcsr_t_type)                                 :: ks_t, mo_coeff_t, mo_coeff_t_split
       TYPE(dbcsr_t_type), DIMENSION(1, 1)                :: t_3c_ctr_KS, t_3c_ctr_KS_copy, &
                                                             t_3c_ctr_RI, t_3c_int_mo_1, &
                                                             t_3c_int_mo_2, t_3c_int_mo_3
@@ -749,24 +748,24 @@ SUBROUTINE hfx_ri_update_ks_mo(qs_env, ri_data, ks_matrix, mo_coeff, &
          CALL dbcsr_t_pgrid_create(para_env%group, pdims, pgrid_2)
 
          CALL create_3c_tensor(t_3c_int_mo_1(1, 1), dist1, dist2, dist3, ri_data%pgrid_1, &
-                               ri_data%bsizes_AO_split, ri_data%bsizes_RI_split, mo_bsizes, &
+                               ri_data%bsizes_AO, ri_data%bsizes_RI_split, mo_bsizes, &
                                [1, 2], [3], name="(AO RI | MO)")
 
          DEALLOCATE (dist1, dist2, dist3)
 
          CALL create_3c_tensor(t_3c_int_mo_2(1, 1), dist1, dist2, dist3, ri_data%pgrid_2, &
-                               mo_bsizes, ri_data%bsizes_RI_split, ri_data%bsizes_AO_split, &
+                               mo_bsizes, ri_data%bsizes_RI_split, ri_data%bsizes_AO, &
                                [1], [2, 3], "(MO | RI AO)")
 
          DEALLOCATE (dist1, dist2, dist3)
 
          CALL create_3c_tensor(t_3c_int_mo_3(1, 1), dist1, dist2, dist3, pgrid_2, &
-                               ri_data%bsizes_RI_split, mo_bsizes, ri_data%bsizes_AO_split, &
+                               ri_data%bsizes_RI_split, mo_bsizes, ri_data%bsizes_AO, &
                                [1], [2, 3], name="(RI | MO AO)")
 
          DEALLOCATE (dist1, dist2, dist3)
          CALL create_3c_tensor(t_3c_ctr_KS(1, 1), dist1, dist2, dist3, pgrid_2, &
-                               ri_data%bsizes_RI_split, mo_bsizes, ri_data%bsizes_AO_split, &
+                               ri_data%bsizes_RI_split, mo_bsizes, ri_data%bsizes_AO, &
                                [1, 2], [3], name="(RI MO | AO)")
          DEALLOCATE (dist1, dist2, dist3)
          CALL dbcsr_t_pgrid_destroy(pgrid_2)
@@ -781,11 +780,7 @@ SUBROUTINE hfx_ri_update_ks_mo(qs_env, ri_data, ks_matrix, mo_coeff, &
          CALL dbcsr_distribution_get(ks_dist, group=comm_2d, nprows=pdims_2d(1), npcols=pdims_2d(2))
          pgrid_2d = dbcsr_t_nd_mp_comm(comm_2d, [1], [2], pdims_2d=pdims_2d)
 
-         CALL create_2c_tensor(ks_t_split, dist1, dist2, pgrid_2d, ri_data%bsizes_AO_split, ri_data%bsizes_AO_split, &
-                               name="(AO | AO)")
-         DEALLOCATE (dist1, dist2)
-
-         CALL create_2c_tensor(mo_coeff_t_split, dist1, dist2, pgrid_2d, ri_data%bsizes_AO_split, mo_bsizes, &
+         CALL create_2c_tensor(mo_coeff_t_split, dist1, dist2, pgrid_2d, ri_data%bsizes_AO, mo_bsizes, &
                                name="(AO | MO)")
 
          DEALLOCATE (dist1, dist2)
@@ -867,7 +862,7 @@ SUBROUTINE hfx_ri_update_ks_mo(qs_env, ri_data, ks_matrix, mo_coeff, &
 
          CALL timeset(routineN//"_3Cx3C", handle2)
          CALL dbcsr_t_contract(dbcsr_scalar(1.0_dp), t_3c_ctr_KS(1, 1), t_3c_ctr_KS_copy(1, 1), &
-                               dbcsr_scalar(0.0_dp), ks_t_split, &
+                               dbcsr_scalar(0.0_dp), ks_t, &
                                contract_1=[1, 2], notcontract_1=[3], &
                                contract_2=[1, 2], notcontract_2=[3], &
                                map_1=[1], map_2=[2], filter_eps=ri_data%filter_eps, &
@@ -877,8 +872,6 @@ SUBROUTINE hfx_ri_update_ks_mo(qs_env, ri_data, ks_matrix, mo_coeff, &
          CALL dbcsr_t_destroy(t_3c_ctr_KS(1, 1))
          CALL dbcsr_t_destroy(t_3c_ctr_KS_copy(1, 1))
 
-         CALL dbcsr_t_copy(ks_t_split, ks_t, move_data=.TRUE.)
-         CALL dbcsr_t_destroy(ks_t_split)
          CALL dbcsr_t_filter(ks_t, ri_data%filter_eps)
          CALL dbcsr_t_copy_tensor_to_matrix(ks_t, ks_matrix(ispin, 1)%matrix, summation=.TRUE.)
          CALL dbcsr_t_destroy(ks_t)

diff --git a/src/hfx_types.F b/src/hfx_types.F
@@ -369,7 +369,7 @@ MODULE hfx_types
       REAL(KIND=dp) :: filter_eps, filter_eps_2c, filter_eps_3c, filter_eps_mo, &
                        eps_lanczos, eps_pgf_orb
       INTEGER :: t2c_sqrt_order, max_iter_lanczos, flavor, unit_nr_dbcsr, unit_nr, &
-                 max_bsize, max_bsize_mo, t2c_method, nelectron_total
+                 max_bsize_RI, max_bsize_mo, max_bsize_AO, t2c_method, nelectron_total
       LOGICAL :: check_2c_inv
 
       TYPE(libint_potential_type) :: ri_metric
@@ -1130,7 +1130,8 @@ SUBROUTINE hfx_ri_init_read_input(ri_data, ri_section, qs_kind_set, &
       CALL section_vals_val_get(ri_section, "MAX_ITER_LANCZOS", i_val=ri_data%max_iter_lanczos)
       CALL section_vals_val_get(ri_section, "RI_FLAVOR", i_val=ri_data%flavor)
       CALL section_vals_val_get(ri_section, "EPS_PGF_ORB", r_val=ri_data%eps_pgf_orb)
-      CALL section_vals_val_get(ri_section, "MAX_BLOCK_SIZE", i_val=ri_data%max_bsize)
+      CALL section_vals_val_get(ri_section, "MAX_BLOCK_SIZE_RI", i_val=ri_data%max_bsize_RI)
+      CALL section_vals_val_get(ri_section, "MAX_BLOCK_SIZE_AO", i_val=ri_data%max_bsize_AO)
       CALL section_vals_val_get(ri_section, "MAX_BLOCK_SIZE_MO", i_val=ri_data%max_bsize_mo)
 
       ri_data%orb_basis_type = orb_basis_type
@@ -1266,8 +1267,8 @@ SUBROUTINE hfx_ri_init(ri_data, qs_kind_set, particle_set, para_env, basis_cntrl
 
       ! initialize tensors expressed in basis representation
 
-      CALL split_block_sizes(ri_data%bsizes_AO, ri_data%bsizes_AO_split, ri_data%max_bsize)
-      CALL split_block_sizes(ri_data%bsizes_RI, ri_data%bsizes_RI_split, ri_data%max_bsize)
+      CALL split_block_sizes(ri_data%bsizes_AO, ri_data%bsizes_AO_split, ri_data%max_bsize_AO)
+      CALL split_block_sizes(ri_data%bsizes_RI, ri_data%bsizes_RI_split, ri_data%max_bsize_RI)
 
       IF (ri_data%flavor == ri_pmat) THEN
          ALLOCATE (ri_data%t_3c_int_ctr_1(1, 1))
@@ -1329,13 +1330,13 @@ SUBROUTINE hfx_ri_init(ri_data, qs_kind_set, particle_set, para_env, basis_cntrl
          CALL dbcsr_t_pgrid_create(para_env%group, pdims, ri_data%pgrid_2)
 
          CALL create_3c_tensor(ri_data%t_3c_int_ctr_1(1, 1), ri_data%dist1_ao_1, ri_data%dist1_ri, ri_data%dist1_ao_2, &
-                               ri_data%pgrid_1, ri_data%bsizes_AO_split, ri_data%bsizes_RI_split, ri_data%bsizes_AO_split, &
+                               ri_data%pgrid_1, ri_data%bsizes_AO, ri_data%bsizes_RI_split, ri_data%bsizes_AO, &
                                [1, 2], [3], name="(AO RI | AO)")
 
          ALLOCATE (ri_data%t_3c_int_ctr_2(1, 1))
          CALL create_3c_tensor(ri_data%t_3c_int_ctr_2(1, 1), ri_data%dist2_ao_1, ri_data%dist2_ri, ri_data%dist2_ao_2, &
-                               ri_data%pgrid_2, ri_data%bsizes_AO_split, ri_data%bsizes_RI_split, ri_data%bsizes_AO_split, &
-                               [1], [2, 3], name="(AO | RI AO)")
+                               ri_data%pgrid_2, ri_data%bsizes_AO, ri_data%bsizes_RI_split, ri_data%bsizes_AO, &
+                               [1], [2, 3], "(AO | RI AO)")
 
       ENDIF
       CALL timestop(handle)
@@ -2504,18 +2505,37 @@ SUBROUTINE hfx_print_ri_info(ri_data, hfx_section)
             WRITE (UNIT=iw, FMT="(T3, A, T78, A)") &
                "HFX_RI_INFO| RI flavor: ", "RHO"
          END SELECT
+         SELECT CASE (ri_data%do_loc)
+         CASE (.TRUE.)
+            WRITE (UNIT=iw, FMT="(T3, A, T80)") &
+               "HFX_RI_INFO| Localized MOs: ", "T"
+         CASE (.FALSE.)
+            WRITE (UNIT=iw, FMT="(T3, A, T80)") &
+               "HFX_RI_INFO| Localized MOs: ", "F"
+         END SELECT
          SELECT CASE (ri_data%t2c_method)
          CASE (hfx_ri_do_2c_iter)
             WRITE (UNIT=iw, FMT="(T3, A, T69, A)") &
                "HFX_RI_INFO| Matrix SQRT/INV", "DBCSR / iter"
          CASE (hfx_ri_do_2c_diag)
             WRITE (UNIT=iw, FMT="(T3, A, T65, A)") &
-               "HFX_RI_INFO| Matrix SQRT/INV", "Scalapack / diag"
+               "HFX_RI_INFO| Matrix SQRT/INV", "Dense / diag"
          END SELECT
          WRITE (UNIT=iw, FMT="(T3, A, T73, ES8.1)") &
             "HFX_RI_INFO| EPS_FILTER", ri_data%filter_eps
+         WRITE (UNIT=iw, FMT="(T3, A, T73, ES8.1)") &
+            "HFX_RI_INFO| EPS_FILTER 2-center", ri_data%filter_eps_2c
+         WRITE (UNIT=iw, FMT="(T3, A, T73, ES8.1)") &
+            "HFX_RI_INFO| EPS_FILTER 3-center", ri_data%filter_eps_3c
+         WRITE (UNIT=iw, FMT="(T3, A, T73, ES8.1)") &
+            "HFX_RI_INFO| EPS_FILTER MO", ri_data%filter_eps_mo
+         WRITE (UNIT=iw, FMT="(T3, A, T73, ES8.1)") &
+            "HFX_RI_INFO| EPS_PGF_ORB", ri_data%eps_pgf_orb
          WRITE (UNIT=iw, FMT="((T3, A, T73, ES8.1))") &
-            "HFX_INFO| EPS_SCHWARZ:     ", ri_data%eps_schwarz
+            "HFX_RI_INFO| EPS_SCHWARZ:     ", ri_data%eps_schwarz
+         WRITE (UNIT=iw, FMT="(T3, A, T70, I3, 1X, I3, 1X, I3)") &
+            "HFX_RI_INFO| Maximum block size RI, MO, AO", ri_data%max_bsize_RI, ri_data%max_bsize_mo, ri_data%max_bsize_AO
+
       ENDIF
 
    END SUBROUTINE

diff --git a/src/input_cp2k_hfx.F b/src/input_cp2k_hfx.F
@@ -625,15 +625,20 @@ SUBROUTINE create_hf_ri_section(section)
       CALL section_add_keyword(section, keyword)
       CALL keyword_release(keyword)
 
-      CALL keyword_create(keyword, __LOCATION__, name="MAX_BLOCK_SIZE", &
-                          description="Maximum block size in one dimension for sparse tensors. Blocking affects load balancing.", &
+      CALL keyword_create(keyword, __LOCATION__, name="MAX_BLOCK_SIZE_RI", &
+                          description="Maximum block size for RI orbitals.", &
                           default_i_val=10)
       CALL section_add_keyword(section, keyword)
       CALL keyword_release(keyword)
 
       CALL keyword_create(keyword, __LOCATION__, name="MAX_BLOCK_SIZE_MO", &
-                          description="Maximum block size for molecular orbitals. "// &
-                          "Blocking affects parallelization strategy and load balancing.", &
+                          description="Maximum block size for molecular orbitals.", &
+                          default_i_val=10)
+      CALL section_add_keyword(section, keyword)
+      CALL keyword_release(keyword)
+
+      CALL keyword_create(keyword, __LOCATION__, name="MAX_BLOCK_SIZE_AO", &
+                          description="Maximum block size for atomic orbitals.", &
                           default_i_val=10)
       CALL section_add_keyword(section, keyword)
       CALL keyword_release(keyword)