Low-scaling RPA: enable MEMORY_CUT 5 / batched contraction by default

- fix bugs for memory cut larger than number of blocks - explicitly use MEMORY_CUT 1 for most regtests due to overhead for small systems
cp2k · Feb 24, 2020 · 7d7496d · 7d7496d
1 parent 8784d52
commit 7d7496d
Show file tree

Hide file tree

Showing 26 changed files with 56 additions and 25 deletions.
diff --git a/src/input_cp2k_mp2.F b/src/input_cp2k_mp2.F
@@ -1166,11 +1166,12 @@ SUBROUTINE create_low_scaling(section)
       CALL keyword_release(keyword)
 
       CALL keyword_create(keyword, __LOCATION__, name="MEMORY_CUT", &
-                          description="Cuts the memory of the M_munu_P^occ/virt matrix by the factor given. "// &
-                          "The calculation is slowed down when using a high value of MEMORY_CUT. A high value for "// &
-                          "MEMORY_CUT is recommended for large systems in order not to run out of memory.", &
+                          description="Reduces memory for sparse tensor contractions by this factor. "// &
+                          "A high value leads to some loss of performance. "// &
+                          "This memory reduction factor applies to storage of the tensors 'M occ' / 'M virt' "// &
+                          "but does not reduce storage of '3c ints'.", &
                           usage="MEMORY_CUT 16", &
-                          default_i_val=1)
+                          default_i_val=5)
       CALL section_add_keyword(section, keyword)
       CALL keyword_release(keyword)
 

diff --git a/src/mp2_integrals.F b/src/mp2_integrals.F
@@ -725,8 +725,6 @@ SUBROUTINE mp2_ri_gpw_compute_in(BIb_C, BIb_C_gw, BIb_C_bse_ij, BIb_C_bse_ab, gd
 
       ELSE
 
-         cut_memory = qs_env%mp2_env%ri_rpa_im_time%cut_memory
-
          impose_split = .NOT. qs_env%mp2_env%ri_rpa_im_time%group_size_internal
 
          IF (impose_split) THEN
@@ -759,6 +757,7 @@ SUBROUTINE mp2_ri_gpw_compute_in(BIb_C, BIb_C_gw, BIb_C_bse_ij, BIb_C_bse_ab, gd
          CALL basis_set_list_setup(basis_set_ao, "ORB", qs_kind_set)
          CALL get_particle_set(particle_set, qs_kind_set, nsgf=sizes_AO, basis=basis_set_ao)
 
+         cut_memory = qs_env%mp2_env%ri_rpa_im_time%cut_memory
          CALL create_contraction_batches(sizes_AO, cut_memory, starts_array_mc_int, ends_array_mc_int, &
                                          starts_array_mc_block_int, ends_array_mc_block_int)
          DEALLOCATE (starts_array_mc_int, ends_array_mc_int)
@@ -816,8 +815,11 @@ SUBROUTINE mp2_ri_gpw_compute_in(BIb_C, BIb_C_gw, BIb_C_bse_ij, BIb_C_bse_ab, gd
             pgrid_t3c_M = get_pgrid_from_ngroup(para_env, ngroup_im_time_P, map1_2d=[1], map2_2d=[2, 3])
          ENDIF
 
-         CALL create_contraction_batches(sizes_AO_split, cut_memory, starts_array_mc, ends_array_mc, &
-                                         starts_array_mc_block, ends_array_mc_block)
+         ASSOCIATE (cut_memory=>qs_env%mp2_env%ri_rpa_im_time%cut_memory)
+            CALL create_contraction_batches(sizes_AO_split, cut_memory, starts_array_mc, ends_array_mc, &
+                                            starts_array_mc_block, ends_array_mc_block)
+         END ASSOCIATE
+         cut_memory = qs_env%mp2_env%ri_rpa_im_time%cut_memory
 
          CALL create_3c_tensor(t_3c_M, dist_RI, dist_AO_1, dist_AO_2, pgrid_t3c_M, &
                                sizes_RI_split, sizes_AO_split, sizes_AO_split, &
@@ -947,20 +949,18 @@ FUNCTION get_pgrid_from_ngroup(para_env, ngroup, map1_2d, map2_2d) RESULT(pgrid)
    SUBROUTINE create_contraction_batches(sizes, nbatches, starts_array, ends_array, &
                                          starts_array_block, ends_array_block)
       INTEGER, DIMENSION(:), INTENT(IN)                  :: sizes
-      INTEGER, INTENT(IN)                                :: nbatches
+      INTEGER, INTENT(INOUT)                             :: nbatches
       INTEGER, ALLOCATABLE, DIMENSION(:), INTENT(OUT)    :: starts_array, ends_array, &
                                                             starts_array_block, ends_array_block
 
       INTEGER                                            :: bsum, imem, nblocks
 
       nblocks = SIZE(sizes)
 
+      CALL contiguous_tensor_dist(nblocks, nbatches, sizes, limits_start=starts_array_block, limits_end=ends_array_block)
+
       ALLOCATE (starts_array(nbatches))
       ALLOCATE (ends_array(nbatches))
-      ALLOCATE (starts_array_block(nbatches))
-      ALLOCATE (ends_array_block(nbatches))
-
-      CALL contiguous_tensor_dist(nblocks, nbatches, sizes, limits_start=starts_array_block, limits_end=ends_array_block)
 
       bsum = 0
       DO imem = 1, nbatches

diff --git a/src/qs_tensors.F b/src/qs_tensors.F
@@ -1266,14 +1266,20 @@ SUBROUTINE build_3c_integrals(t3c, filter_eps, qs_env, &
 !> \param dist ...
 ! **************************************************************************************************
    SUBROUTINE contiguous_tensor_dist(nel, nbin, weights, limits_start, limits_end, dist)
-      INTEGER, INTENT(IN)                                :: nel, nbin
+      INTEGER, INTENT(IN)                                :: nel
+      INTEGER, INTENT(INOUT)                             :: nbin
       INTEGER, DIMENSION(nel), INTENT(IN)                :: weights
-      INTEGER, DIMENSION(nbin), INTENT(OUT), OPTIONAL    :: limits_start, limits_end
+      INTEGER, ALLOCATABLE, DIMENSION(:), INTENT(OUT), &
+         OPTIONAL                                        :: limits_start, limits_end
       INTEGER, DIMENSION(nel), INTENT(OUT), OPTIONAL     :: dist
 
       INTEGER                                            :: el_end, el_start, end_weight, ibin, &
                                                             nel_div, nel_rem, nel_split, nel_w, &
                                                             w_partialsum
+      INTEGER, ALLOCATABLE, DIMENSION(:)                 :: lim_e, lim_s
+
+      ALLOCATE (lim_s(nbin), lim_e(nbin))
+      lim_s = 0; lim_e = 0
 
       nel_w = SUM(weights)
       nel_div = nel_w/nbin
@@ -1295,12 +1301,23 @@ SUBROUTINE contiguous_tensor_dist(nel, nbin, weights, limits_start, limits_end,
             !IF (ABS(w_partialsum + weights(el_end) - end_weight) > ABS(w_partialsum - end_weight)) EXIT
             el_end = el_end + 1
             w_partialsum = w_partialsum + weights(el_end)
+            IF (el_end == nel) EXIT
          ENDDO
+
          IF (PRESENT(dist)) dist(el_start:el_end) = ibin - 1
-         IF (PRESENT(limits_start)) limits_start(ibin) = el_start
-         IF (PRESENT(limits_end)) limits_end(ibin) = el_end
+         lim_s(ibin) = el_start
+         lim_e(ibin) = el_end
+
+         IF (el_end == nel) EXIT
       ENDDO
 
+      IF (PRESENT(limits_start) .AND. PRESENT(limits_end)) THEN
+         ALLOCATE (limits_start(ibin)); limits_start(:ibin) = lim_s(:ibin)
+         ALLOCATE (limits_end(ibin)); limits_end(:ibin) = lim_e(:ibin)
+      ENDIF
+
+      nbin = ibin
+
    END SUBROUTINE contiguous_tensor_dist
 
 ! **************************************************************************************************

diff --git a/src/rpa_im_time.F b/src/rpa_im_time.F
@@ -355,7 +355,7 @@ SUBROUTINE compute_mat_P_omega(mat_P_omega, fm_scaled_dm_occ_tau, &
 
                   ibounds_2(:, 1) = [starts_array_mc(i_mem), ends_array_mc(i_mem)]
 
-                  IF (unit_nr_dbcsr > 0) WRITE (UNIT=unit_nr_dbcsr, FMT="(T3,A,I2,1X,I2)") &
+                  IF (unit_nr_dbcsr > 0) WRITE (UNIT=unit_nr_dbcsr, FMT="(T3,A,I3,1X,I3)") &
                      "RPA_LOW_SCALING_INFO| Memory Cut iteration", i_mem, j_mem
 
                   IF (do_Gamma_RPA) THEN

diff --git a/tests/QS/regtest-gw-cubic/G0W0_H2O_PBE0.inp b/tests/QS/regtest-gw-cubic/G0W0_H2O_PBE0.inp
@@ -55,6 +55,7 @@
         &END
         &LOW_SCALING
           EPS_FILTER  1.0E-7
+          MEMORY_CUT 1
         &END
         &RI_RPA
           &HF

diff --git a/tests/QS/regtest-gw-cubic/G0W0_H2O_PBE0_trunc_Clenshaw.inp b/tests/QS/regtest-gw-cubic/G0W0_H2O_PBE0_trunc_Clenshaw.inp
@@ -56,6 +56,7 @@
         &END
         &LOW_SCALING
           EPS_FILTER  1.0E-7
+          MEMORY_CUT 1
         &END
         &RI_RPA
           &HF

diff --git a/tests/QS/regtest-gw-cubic/G0W0_H2O_PBE_periodic.inp b/tests/QS/regtest-gw-cubic/G0W0_H2O_PBE_periodic.inp
@@ -48,6 +48,7 @@
         &END
         &LOW_SCALING
           EPS_FILTER  1.0E-7
+          MEMORY_CUT 1
         &END
         &RI_RPA
           &HF

diff --git a/tests/QS/regtest-gw-cubic/G0W0_OH_PBE.inp b/tests/QS/regtest-gw-cubic/G0W0_OH_PBE.inp
@@ -49,6 +49,7 @@
           POTENTIAL_TYPE IDENTITY
         &END
         &LOW_SCALING
+          MEMORY_CUT 1
         &END LOW_SCALING
         &RI_RPA
           &HF

diff --git a/tests/QS/regtest-gw-cubic/G0W0_OH_PBE_svd.inp b/tests/QS/regtest-gw-cubic/G0W0_OH_PBE_svd.inp
@@ -50,6 +50,7 @@
           POTENTIAL_TYPE IDENTITY
         &END
         &LOW_SCALING
+          MEMORY_CUT 1
         &END LOW_SCALING
         &RI_RPA
           &HF

diff --git a/tests/QS/regtest-gw-cubic/evGW_H2O_PBE0_trunc_Clenshaw.inp b/tests/QS/regtest-gw-cubic/evGW_H2O_PBE0_trunc_Clenshaw.inp
@@ -55,6 +55,7 @@
           CUTOFF_RADIUS 3.0
         &END
         &LOW_SCALING
+          MEMORY_CUT 1
           EPS_FILTER  1.0E-7
         &END
         &RI_RPA

diff --git a/tests/QS/regtest-gw-cubic/scGW0_H2O_PBE0_trunc_Clenshaw.inp b/tests/QS/regtest-gw-cubic/scGW0_H2O_PBE0_trunc_Clenshaw.inp
@@ -55,6 +55,7 @@
         &END
         &LOW_SCALING
           EPS_FILTER  1.0E-7
+          MEMORY_CUT 1
         &END
         &RI_RPA
           &HF

diff --git a/tests/QS/regtest-gw-cubic/scGW0_H2O_PBE0_trunc_minimax.inp b/tests/QS/regtest-gw-cubic/scGW0_H2O_PBE0_trunc_minimax.inp
@@ -54,6 +54,7 @@
           CUTOFF_RADIUS 3.0
         &END
         &LOW_SCALING
+          MEMORY_CUT 1
           EPS_FILTER  1.0E-7
         &END
         &RI_RPA

diff --git a/tests/QS/regtest-gw-ic-model/IC_no_wf_update.inp b/tests/QS/regtest-gw-ic-model/IC_no_wf_update.inp
@@ -57,6 +57,7 @@
           POTENTIAL_TYPE IDENTITY
         &END
         &LOW_SCALING
+          MEMORY_CUT 1
         &END
         &RI_RPA
           RPA_NUM_QUAD_POINTS  5

diff --git a/tests/QS/regtest-gw/evGW_H2O_PBE_Fermi_level_offset.inp b/tests/QS/regtest-gw/evGW_H2O_PBE_Fermi_level_offset.inp
@@ -58,6 +58,7 @@
           EPS_FILTER  1.0E-12
         &END
         &LOW_SCALING
+            MEMORY_CUT 1
         &END
         &RI_RPA
           &HF

diff --git a/tests/QS/regtest-ri-laplace-mp2-cubic/RI_laplace_cubic_MP2_CH3.inp b/tests/QS/regtest-ri-laplace-mp2-cubic/RI_laplace_cubic_MP2_CH3.inp
@@ -46,6 +46,7 @@
           QUADRATURE_POINTS  5
         &END
         &LOW_SCALING
+          MEMORY_CUT 1
         &END
         &WFC_GPW
           CUTOFF  100

diff --git a/tests/QS/regtest-ri-laplace-mp2-cubic/RI_laplace_cubic_MP2_CH3_svd.inp b/tests/QS/regtest-ri-laplace-mp2-cubic/RI_laplace_cubic_MP2_CH3_svd.inp
@@ -46,6 +46,7 @@
           QUADRATURE_POINTS  5
         &END
         &LOW_SCALING
+          MEMORY_CUT 1
         &END
         &WFC_GPW
           CUTOFF  100

diff --git a/tests/QS/regtest-ri-laplace-mp2-cubic/RI_laplace_cubic_MP2_H2O.inp b/tests/QS/regtest-ri-laplace-mp2-cubic/RI_laplace_cubic_MP2_H2O.inp
@@ -47,6 +47,7 @@
           QUADRATURE_POINTS  5
         &END
         &LOW_SCALING
+          MEMORY_CUT 1
         &END
         &WFC_GPW
           CUTOFF  100

diff --git a/tests/QS/regtest-ri-rpa-rse/Cubic_RPA_RSE_H2.inp b/tests/QS/regtest-ri-rpa-rse/Cubic_RPA_RSE_H2.inp
@@ -51,6 +51,7 @@
 !        ERI_METHOD MME
         ERI_METHOD OS
         &LOW_SCALING
+          MEMORY_CUT 1
         &END
         &RI_RPA
           RPA_NUM_QUAD_POINTS  6

diff --git a/tests/QS/regtest-rpa-cubic-scaling/Cubic_RPA_CH3.inp b/tests/QS/regtest-rpa-cubic-scaling/Cubic_RPA_CH3.inp
@@ -46,6 +46,7 @@
         &END
         &LOW_SCALING
           EPS_FILTER  1.0E-7
+          MEMORY_CUT 1
         &END
         &RI_RPA
           RPA_NUM_QUAD_POINTS  6

diff --git a/tests/QS/regtest-rpa-cubic-scaling/Cubic_RPA_H2O_check_group_sizes.inp b/tests/QS/regtest-rpa-cubic-scaling/Cubic_RPA_H2O_check_group_sizes.inp
@@ -45,12 +45,6 @@
         ERI_METHOD MME
         &LOW_SCALING
           EPS_FILTER  1.0E-7
-          !MEMORY_INFO
-          ! normally, increase MEMORY_CUT for large systems
-          ! not to run out of memory
-          MEMORY_CUT 2
-          ! should be increased with N^2 (N: group size) not
-          ! to run OOM
           &TEST
             GROUP_SIZE_3c 2
             ! normally, GROUP_SIZE_P 1 is fine

diff --git a/tests/QS/regtest-rpa-cubic-scaling/Cubic_RPA_H2O_standard_svd.inp b/tests/QS/regtest-rpa-cubic-scaling/Cubic_RPA_H2O_standard_svd.inp
@@ -50,6 +50,7 @@
 !        ERI_METHOD MME
         ERI_METHOD OS
         &LOW_SCALING
+          MEMORY_CUT 1
         &END
         &RI_RPA
           RPA_NUM_QUAD_POINTS  6

diff --git a/tests/QS/regtest-rpa-cubic-scaling/Cubic_RPA_H2O_trunc.inp b/tests/QS/regtest-rpa-cubic-scaling/Cubic_RPA_H2O_trunc.inp
@@ -45,6 +45,7 @@
         &END
         &LOW_SCALING
           EPS_FILTER  1.0E-6
+          MEMORY_CUT 1
         &END
         &RI_RPA
           RPA_NUM_QUAD_POINTS  6

diff --git a/tests/QS/regtest-rpa-cubic-scaling/RPA_kpoints_H2O_batched.inp b/tests/QS/regtest-rpa-cubic-scaling/RPA_kpoints_H2O_batched.inp
@@ -57,7 +57,6 @@
         NUMBER_PROC  1
         &LOW_SCALING
           EPS_FILTER 1.0E-7
-          MEMORY_CUT 2
           !MEMORY_INFO
           DO_KPOINTS
         &END LOW_SCALING

diff --git a/tests/QS/regtest-rpa-lr/H2O-rpa-cubic-lr-mme.inp b/tests/QS/regtest-rpa-lr/H2O-rpa-cubic-lr-mme.inp
@@ -53,6 +53,7 @@
         &END
         ERI_METHOD MME
         &LOW_SCALING
+          MEMORY_CUT 1
         &END
         &RI_RPA
           RPA_NUM_QUAD_POINTS  6

diff --git a/tests/QS/regtest-rpa-lr/H2O-rpa-cubic-lr.inp b/tests/QS/regtest-rpa-lr/H2O-rpa-cubic-lr.inp
@@ -55,6 +55,7 @@
         &END
         ERI_METHOD OS
         &LOW_SCALING
+          MEMORY_CUT 1
         &END
         &RI_RPA
           RPA_NUM_QUAD_POINTS  6

diff --git a/tests/QS/regtest-sos-mp2-lr/H2O-sos-mp2-cubic-lr.inp b/tests/QS/regtest-sos-mp2-lr/H2O-sos-mp2-cubic-lr.inp
@@ -51,6 +51,7 @@
           QUADRATURE_POINTS  5
         &END
         &LOW_SCALING
+          MEMORY_CUT 1
         &END
         &WFC_GPW
           CUTOFF  100