Blocking factor in trans_ev_band_to_full tunable

The blocking factor can be set now with the new interface
set method. The default is 3

Autotuning still has to be implemented.

This closses issue #44

.gitlab-ci.yml

@@ -3346,7 +3346,7 @@ minsky:
     - miy
   script:
     - module purge
-    - module load gcc/5.4 pgi/17.5 ompi/pgi/17.5/1.10.2 essl/5.5 cuda/8.0
+    - module load gcc/5.4 pgi/17.7 ompi/pgi/17.7/1.10.2 essl/5.5 cuda/8.0
     - module list
     - export OMPI_CC=gcc
     - export OMPI_FC=pgfortran

src/elpa2/elpa2_trans_ev_band_to_full_template.F90

@@ -122,7 +122,7 @@
       integer(kind=ik)                       :: l_cols, l_rows, l_colh, n_cols
       integer(kind=ik)                       :: istep, lc, ncol, nrow, nb, ns
 
-      MATH_DATATYPE(kind=rck), allocatable  :: tmp1(:), tmp2(:), hvb(:), hvm(:,:)
+      MATH_DATATYPE(kind=rck), allocatable   :: tmp1(:), tmp2(:), hvb(:), hvm(:,:)
       ! hvm_dev is fist used and set in this routine
       ! q is changed in trans_ev_tridi on the host, copied to device and passed here. this can be adapted
       ! tmp_dev is first used in this routine
@@ -132,22 +132,27 @@
       integer(kind=ik)                       :: i
 
 #ifdef BAND_TO_FULL_BLOCKING
-      MATH_DATATYPE(kind=rck), allocatable  :: tmat_complete(:,:), t_tmp(:,:), t_tmp2(:,:)
+      MATH_DATATYPE(kind=rck), allocatable   :: tmat_complete(:,:), t_tmp(:,:), t_tmp2(:,:)
       integer(kind=ik)                       :: cwy_blocking, t_blocking, t_cols, t_rows
 #endif
 
       integer(kind=ik)                       :: istat
       character(200)                         :: errorMessage
       logical                                :: successCUDA
-      integer(kind=c_intptr_t), parameter      :: size_of_datatype = size_of_&
+      integer(kind=c_intptr_t), parameter    :: size_of_datatype = size_of_&
                                                                    &PRECISION&
                                                                    &_&
                                                                    &MATH_DATATYPE
+      integer                                :: blocking_factor
       call obj%timer%start("trans_ev_band_to_full_&
       &MATH_DATATYPE&
       &" // &
       &PRECISION_SUFFIX &
       )
+#ifdef BAND_TO_FULL_BLOCKING
+      call obj%get("blocking_in_band_to_full",blocking_factor)
+      print *,"Blocking factor: ", blocking_factor
+#endif
       call obj%timer%start("mpi_communication")
 
       call mpi_comm_rank(mpi_comm_rows,my_prow,mpierr)
@@ -432,7 +437,7 @@
 
 #ifdef BAND_TO_FULL_BLOCKING
         ! t_blocking was formerly 2; 3 is a better choice
-        t_blocking = 3 ! number of matrices T (tmat) which are aggregated into a new (larger) T matrix (tmat_complete) and applied at once
+        t_blocking = blocking_factor ! number of matrices T (tmat) which are aggregated into a new (larger) T matrix (tmat_complete) and applied at once
 
         ! we only use the t_blocking if we could call it fully, this is might be better but needs to benchmarked.
 !       if ( na >= ((t_blocking+1)*nbw) ) then

src/elpa_index.c

@@ -132,7 +132,6 @@ static const elpa_index_int_entry_t int_entries[] = {
         INT_PARAMETER_ENTRY("process_row", "Process row number in the 2D domain decomposition", NULL),
         INT_PARAMETER_ENTRY("process_col", "Process column number in the 2D domain decomposition", NULL),
         INT_PARAMETER_ENTRY("bandwidth", "If specified, a band matrix with this bandwidth is expected as input; bandwidth must be multiply of nblk", bw_is_valid),
-
         INT_ANY_ENTRY("mpi_comm_rows", "Communicator for inter-row communication"),
         INT_ANY_ENTRY("mpi_comm_cols", "Communicator for inter-column communication"),
         INT_ANY_ENTRY("mpi_comm_parent", "Parent communicator"),
@@ -144,6 +143,8 @@ static const elpa_index_int_entry_t int_entries[] = {
         INT_ENTRY("complex_kernel", "Complex kernel to use if 'solver' is set to ELPA_SOLVER_2STAGE", ELPA_2STAGE_COMPLEX_DEFAULT, \
                         number_of_complex_kernels, complex_kernel_enumerate, \
                         complex_kernel_is_valid, complex_kernel_name),
+
+	INT_ENTRY("blocking_in_band_to_full", "Loop blocking, default 3", 3, NULL, NULL, NULL, NULL),
         BOOL_ENTRY("qr", "Use QR decomposition, only used for ELPA_SOLVER_2STAGE, real case", 0),
         BOOL_ENTRY("gpu", "Use GPU acceleration", 0),
         BOOL_ENTRY("timings", "Enable time measurement", 0),