Cleanup

src/elpa1_compute_private.F90

@@ -160,6 +160,11 @@ module ELPA1_COMPUTE
 #define BYTESIZE 8
 #define REALCASE 1
 #define DOUBLE_PRECISION 1
+#include "precision_macros.h"
+#undef PRECISION
+#define PRECISION double
+#undef PRECISION_SUFFIX
+#define PRECISION_SUFFIX "_double"
 #include "elpa_transpose_vectors.X90"
 #include "elpa_reduce_add_vectors.X90"
 #undef DOUBLE_PRECISION_REAL
@@ -167,7 +172,8 @@ module ELPA1_COMPUTE
 #undef BYTESIZE
 #undef REALCASE
 #undef DOUBLE_PRECISION
-
+#undef PRECISION
+#undef PRECISION_SUFFIX
 ! single precision
 #ifdef WANT_SINGLE_PRECISION_REAL
 
@@ -176,12 +182,19 @@ module ELPA1_COMPUTE
 #define BYTESIZE 4
 #define REALCASE 1
 #define SINGLE_PRECISION 1
+#include "precision_macros.h"
+#undef PRECISION
+#define PRECISION single
+#undef PRECISION_SUFFIX
+#define PRECISION_SUFFIX "_single"
 #include "elpa_transpose_vectors.X90"
 #include "elpa_reduce_add_vectors.X90"
 #undef DATATYPE
 #undef BYTESIZE
 #undef REALCASE
 #undef SINGLE_PRECISION
+#undef PRECISION
+#undef PRECISION_SUFFIX
 
 #endif
 
@@ -192,6 +205,11 @@ module ELPA1_COMPUTE
 #define BYTESIZE 16
 #define COMPLEXCASE 1
 #define DOUBLE_PRECISION 1
+#include "precision_macros.h"
+#undef PRECISION
+#define PRECISION double
+#undef PRECISION_SUFFIX
+#define PRECISION_SUFFIX "_double"
 #include "elpa_transpose_vectors.X90"
 #include "elpa_reduce_add_vectors.X90"
 #undef DATATYPE
@@ -199,6 +217,8 @@ module ELPA1_COMPUTE
 #undef COMPLEXCASE
 #undef DOUBLE_PRECISION
 #undef DOUBLE_PRECISION_COMPLEX
+#undef PRECISION
+#undef PRECISION_SUFFIX
 
 #ifdef WANT_SINGLE_PRECISION_COMPLEX
 
@@ -207,6 +227,11 @@ module ELPA1_COMPUTE
 #define DATATYPE COMPLEX(kind=ck4)
 #define COMPLEXCASE 1
 #define SINGLE_PRECISION 1
+#include "precision_macros.h"
+#undef PRECISION
+#define PRECISION single
+#undef PRECISION_SUFFIX
+#define PRECISION_SUFFIX "_single"
 #include "elpa_transpose_vectors.X90"
 #include "elpa_reduce_add_vectors.X90"
 #undef DATATYPE

src/elpa1_merge_systems_real_template.X90

@@ -120,8 +120,7 @@
       endif
 #endif
 
-      call timer%start("merge_systems" // &
-      &PRECISION_SUFFIX)
+      call timer%start("merge_systems" // PRECISION_SUFFIX)
       success = .true.
       call timer%start("mpi_communication")
       call mpi_comm_rank(mpi_comm_rows,my_prow,mpierr)

src/elpa1_tools_template.X90

@@ -244,7 +244,7 @@
 #if COMPLEXCASE == 1
     subroutine hh_transform_complex_&
 #endif
-&PRECISION &
+    &PRECISION &
                    (alpha, xnorm_sq, xf, tau)
 #if REALCASE  == 1
       ! Similar to LAPACK routine DLARFP, but uses ||x||**2 instead of x(:)
@@ -282,8 +282,7 @@
      call timer%start("hh_tranform_&
                       &MATH_DATATYPE&
 		      &" // &
-		      &PRECISION_SUFFIX &
-		      )
+		      &PRECISION_SUFFIX )
 
 #if COMPLEXCASE == 1
       ALPHR = real( ALPHA, kind=REAL_DATATYPE )
@@ -347,8 +346,7 @@
       call timer%stop("hh_transform_&
       &MATH_DATATYPE&
       &" // &
-      &PRECISION_SUFFIX &
-      )
+      &PRECISION_SUFFIX )
 
 #if REALCASE == 1
     end subroutine hh_transform_real_&
@@ -356,4 +354,4 @@
 #if COMPLEXCASE == 1
     end subroutine hh_transform_complex_&
 #endif
-&PRECISION
+    &PRECISION

src/elpa2_bandred_template.X90

@@ -258,7 +258,7 @@
          na_rows = na
 #endif
         na_cols = na
-#endif
+#endif /* WITH_MPI */
 
         ! Here we convert the regular host array into a pinned host array
         successCUDA = cuda_malloc(a_dev, lda*na_cols*  &
@@ -1326,9 +1326,9 @@
 
          if (tile_size < istep*nbw) then
            call elpa_reduce_add_vectors_&
-&MATH_DATATYPE&
-&_&
-&PRECISION &
+                &MATH_DATATYPE&
+                &_&
+                &PRECISION &
                                                (vmrCUDA(cur_l_rows * n_cols + 1),cur_l_rows,mpi_comm_rows, &
                                                umcCUDA, cur_l_cols, mpi_comm_cols, &
                                                istep*nbw, n_cols, nblk)
@@ -1401,7 +1401,7 @@
          endif
 
          call symm_matrix_allreduce_&
-&PRECISION &
+         &PRECISION &
 	                          (n_cols,vav, nbw,nbw,mpi_comm_cols)
 
          successCUDA = cuda_memcpy(vav_dev, loc(vav(1,1)), nbw*nbw*size_of_PRECISION_real,cudaMemcpyHostToDevice)
@@ -1427,9 +1427,9 @@
 
          ! Transpose umc -> umr (stored in vmr, second half)
          call elpa_transpose_vectors_&
-&MATH_DATATYPE&
-&_&
-&PRECISION &
+         &MATH_DATATYPE&
+         &_&
+         &PRECISION &
                                            (umcCUDA, cur_l_cols, mpi_comm_cols, &
                                             vmrCUDA(cur_l_rows * n_cols + 1), cur_l_rows, mpi_comm_rows, &
                                             1, istep*nbw, n_cols, nblk)
@@ -1466,9 +1466,9 @@
          ! Or if we used the Algorithm 4
          if (tile_size < istep*nbw .or. n_way > 1) then
            call elpa_reduce_add_vectors_&
-&MATH_DATATYPE&
-&_&
-&PRECISION &
+           &MATH_DATATYPE&
+           &_&
+           &PRECISION &
                                             (vmrCPU(1,n_cols+1),ubound(vmrCPU,dim=1),mpi_comm_rows, &
                                              umcCPU, ubound(umcCPU,dim=1), mpi_comm_cols, &
                                              istep*nbw, n_cols, nblk)
@@ -1512,7 +1512,7 @@
                     ubound(tmat,dim=1), vav, ubound(vav,dim=1))
 	 call timer%stop("blas")
          call symm_matrix_allreduce_&
-&PRECISION &
+         &PRECISION &
 	                            (n_cols,vav, nbw, nbw ,mpi_comm_cols)
 
          ! U = U - 0.5 * V * VAV
@@ -1522,9 +1522,9 @@
 	 call timer%stop("blas")
          ! Transpose umc -> umr (stored in vmr, second half)
          call elpa_transpose_vectors_&
-&MATH_DATATYPE&
-&_&
-&PRECISION &
+         &MATH_DATATYPE&
+         &_&
+         &PRECISION &
 	                                (umcCPU, ubound(umcCPU,dim=1), mpi_comm_cols, &
                                          vmrCPU(1,n_cols+1), ubound(vmrCPU,dim=1), mpi_comm_rows, &
                                          1, istep*nbw, n_cols, nblk)
@@ -1588,9 +1588,9 @@
 #if COMPLEXCASE == 1
         if (tile_size < istep*nbw) then
           call elpa_reduce_add_vectors_&
-&MATH_DATATYPE&
-&_&
-&PRECISION &
+          &MATH_DATATYPE&
+          &_&
+          &PRECISION &
                                           (vmrCPU(1,n_cols+1),ubound(vmrCPU,dim=1),mpi_comm_rows, &
                                           umcCPU, ubound(umcCPU,dim=1), mpi_comm_cols, &
                                           istep*nbw, n_cols, nblk)
@@ -1683,7 +1683,7 @@
           endif
 
           call herm_matrix_allreduce_&
-&PRECISION &
+          &PRECISION &
                                     (n_cols,vav, nbw, nbw,mpi_comm_cols)
 
           successCUDA = cuda_memcpy(vav_dev,loc(vav(1,1)),nbw*nbw*size_of_PRECISION_complex,cudaMemcpyHostToDevice)
@@ -1699,7 +1699,7 @@
                      ubound(tmat,dim=1), vav, ubound(vav,dim=1))
           call timer%stop("blas")
           call herm_matrix_allreduce_&
-&PRECISION &
+          &PRECISION &
                                     (n_cols,vav,nbw,nbw,mpi_comm_cols)
         endif
 
@@ -1723,9 +1723,9 @@
             stop
           endif
           call elpa_transpose_vectors_&
-&MATH_DATATYPE&
-&_&
-&PRECISION &
+          &MATH_DATATYPE&
+          &_&
+          &PRECISION &
 	                                       (umcCPU, ubound(umcCPU,dim=1), mpi_comm_cols, &
                                                 vmrCPU(1,n_cols+1), ubound(vmrCPU,dim=1), mpi_comm_rows, &
                                                 1, istep*nbw, n_cols, nblk)
@@ -1755,9 +1755,9 @@
           call timer%stop("blas")
           ! Transpose umc -> umr (stored in vmr, second half)
           call elpa_transpose_vectors_&
-&MATH_DATATYPE&
-&_&
-&PRECISION &
+          &MATH_DATATYPE&
+          &_&
+          &PRECISION &
 	                                       (umcCPU, ubound(umcCPU,dim=1), mpi_comm_cols, &
                                                 vmrCPU(1,n_cols+1), ubound(vmrCPU,dim=1), mpi_comm_rows, &
                                                 1, istep*nbw, n_cols, nblk)
@@ -1826,12 +1826,22 @@
 !       endif
 !#endif
        ! this is not needed since a_dev is passed along from one subroutine to the other
+
+       successCUDA = cuda_memcpy ( &
 #if REALCASE == 1
-       successCUDA = cuda_memcpy ( loc (a), a_dev, lda*na_cols*size_of_PRECISION_real,cudaMemcpyDeviceToHost)
+                                   loc(a),      &
 #endif
 #if COMPLEXCASE == 1
-       successCUDA = cuda_memcpy ( loc(a(1,1)), a_dev, lda*na_cols*size_of_PRECISION_complex,cudaMemcpyDeviceToHost)
+                                   loc(a(1,1)), &
+#endif
+                                   a_dev, lda*na_cols*  &
+#if REALCASE == 1
+                                   size_of_PRECISION_real,   &
+#endif
+#if COMPLEXCASE ==1
+                                   size_of_PRECISION_complex,&
 #endif
+                                   cudaMemcpyDeviceToHost)
        if (.not.(successCUDA)) then
          print *,"bandred_&
 	 &MATH_DATATYPE&

src/elpa2_compute.F90

@@ -116,25 +116,43 @@ module ELPA2_compute
 #define REAL_DATATYPE rk8
 #define BYTESIZE 8
 #define REALCASE 1
+#define DOUBLE_PRECISION 1
+#include "precision_macros.h"
+#undef PRECISION_SUFFIX
+#define PRECISION_SUFFIX "_double"
+#define PRECISION double
 #include "redist_band.X90"
+
 #undef DOUBLE_PRECISION_REAL
 #undef REAL_DATATYPE
 #undef BYTESIZE
 #undef REALCASE
+#undef DOUBLE_PRECISION
+#undef PRECISION_SUFFIX
+#undef PRECISION
 
 ! single precision
 #ifdef WANT_SINGLE_PRECISION_REAL
 
 #undef DOUBLE_PRECISION_REAL
+#undef DOUBLE_PRECISION
 #define REAL_DATATYPE rk4
 #define BYTESIZE 4
 #define REALCASE 1
+#include "precision_macros.h"
+#undef PRECISION_SUFFIX
+#define PRECISION_SUFFIX "_single"
+#undef PRECISION
+#define PRECISION single
 #include "redist_band.X90"
+
 #undef REAL_DATATYPE
 #undef BYTESIZE
 #undef REALCASE
+#undef PRECISION_SUFFIX
+#undef PRECISION
 
-#endif
+#endif /* WANT_SINGLE_PRECISION_REAL */
 
 ! double precision
 #define DOUBLE_PRECISION_COMPLEX 1
@@ -142,22 +160,40 @@ module ELPA2_compute
 #define COMPLEX_DATATYPE ck8
 #define BYTESIZE 16
 #define COMPLEXCASE 1
+#define DOUBLE_PRECISION
+#include "precision_macros.h"
+#undef PRECISION_SUFFIX
+#define PRECISION_SUFFIX "_double"
+#undef PRECISION
+#define PRECISION double
 #include "redist_band.X90"
+
 #undef COMPLEX_DATATYPE
 #undef BYTESIZE
 #undef COMPLEXCASE
+#undef DOUBLE_PRECISION
 #undef DOUBLE_PRECISION_COMPLEX
+#undef PRECISION_SUFFIX
+#undef PRECISION
 
 #ifdef WANT_SINGLE_PRECISION_COMPLEX
 
 #undef DOUBLE_PRECISION_COMPLEX
 #undef DOUBLE_PRECISION_REAL
+#undef DOUBLE_PRECISION
 #define COMPLEX_DATATYPE ck4
 #define COMPLEXCASE 1
+#include "precision_macros.h"
+#undef PRECISION_SUFFIX
+#define PRECISION_SUFFIX "_single"
+#undef PRECISION
+#define PRECISION single
 #include "redist_band.X90"
+
 #undef COMPLEX_DATATYPE
 #undef BYTESIZE
 #undef COMPLEXCASE
+#undef PRECISION_SUFFIX
 
 #endif /* WANT_SINGLE_PRECISION_COMPLEX */
 

src/elpa2_compute_real_template.X90

@@ -141,22 +141,6 @@
       character(200)                           :: errorMessage
 
       call timer%start("band_band_real" // PRECISION_SUFFIX)
-!      if (na .lt. 2*nb) then
-!        print *,"na lt 2*nb ",na,2*nb
-!        stop
-!      endif
-!      if (na .lt. 2*nb2) then
-!        print *,"na lt 2*nb2 ",na,2*nb2
-!        stop
-!      endif
-!      if (na .lt. nbCol) then
-!        print *,"na lt nbCol ",na,nbCol
-!        stop
-!      endif
-!      if (na .lt. nb2Col) then
-!        print *,"na lt nb2Col ",na,nb2Col
-!        stop
-!      endif
 
       call timer%start("mpi_communication")
       call mpi_comm_rank(mpi_comm,my_pe,mpierr)

src/elpa2_template.X90

@@ -69,7 +69,7 @@
 
     call timer%start("solve_evp_&
     &MATH_DATATYPE&
-    &_2stage_" // &
+    &_2stage" // &
     &PRECISION_SUFFIX &
     )
 
@@ -483,7 +483,7 @@
 
      call timer%stop("solve_evp_&
      &MATH_DATATYPE&
-     &_2stage_" // &
+     &_2stage" // &
      &PRECISION_SUFFIX &
      )
 1    format(a,f10.3)

src/elpa2_trans_ev_tridi_to_band_template.X90

@@ -238,7 +238,7 @@
 
       call timer%start("trans_ev_tridi_to_band_&
       &MATH_DATATYPE&
-      &_" // &
+      &" // &
       &PRECISION_SUFFIX &
       )
 
@@ -275,7 +275,7 @@
         na_cols = na
 #endif
       endif
-#endif /* COMPLEXCASEc*/
+#endif /* COMPLEXCASE */
 
       if (mod(nbw,nblk)/=0) then
         if (my_prow==0 .and. my_pcol==0) then
@@ -370,7 +370,9 @@
 #endif
 
         endif ! useGPU
+
 #else /* WITH_OPENMP */
+
         ! Suggested stripe width is 48 since 48*64 real*8 numbers should fit into
         ! every primary cache
 	! Suggested stripe width is 48 - should this be reduced for the complex case ???
@@ -553,7 +555,7 @@
 #endif
           print *,"trans_ev_tridi_to_band_&
           &MATH_DATATYPE&
-          &^: error when allocating aIntern"//errorMessage
+          &: error when allocating aIntern"//errorMessage
           stop
         endif
 
@@ -721,9 +723,9 @@
 #endif /* WITH_MPI */
 
                 call unpack_row_&
-&MATH_DATATYPE&
-&_cpu_&
-&PRECISION &
+                &MATH_DATATYPE&
+                &_cpu_&
+                &PRECISION &
                                 (aIntern, row,i-limits(ip), stripe_count, stripe_width, last_stripe_width)
               endif ! useGPU
 #endif /* WITH_OPENMP */
@@ -776,9 +778,9 @@
 !$omp parallel do private(my_thread), schedule(static, 1)
               do my_thread = 1, max_threads
 	        call unpack_row_&
-&MATH_DATATYPE&
-&_cpu_openmp_&
-&PRECISION &
+                &MATH_DATATYPE&
+                &_cpu_openmp_&