Timer for MPI communication in elpa_reduce_add_vectors.X90

src/elpa_reduce_add_vectors.X90

@@ -90,6 +90,9 @@ subroutine elpa_reduce_add_vectors_complex_single(vmat_s,ld_s,comm_s,vmat_t,ld_t
 !   use ELPA1 ! for least_common_multiple
 #ifdef WITH_OPENMP
    use omp_lib
+#endif
+#ifdef HAVE_DETAILED_TIMINGS
+   use timings
 #endif
    use elpa_mpi
 #ifdef HAVE_DETAILED_TIMINGS
@@ -115,19 +118,28 @@ subroutine elpa_reduce_add_vectors_complex_single(vmat_s,ld_s,comm_s,vmat_t,ld_t
    call timer%start("elpa_reduce_add_vectors_complex_single")
 #endif
 
-#else
+#else /*COMPLEXCASE == 1 */
+
 #ifdef DOUBLE_PRECISION_REAL
    call timer%start("elpa_reduce_add_vectors_double")
 #else
    call timer%start("elpa_reduce_add_vectors_single")
 #endif
 
-#endif
+#endif /* COMPLEXCASE == 1 */
+
+#endif /* HAVE_DETAILED_TIMINGS */
+
+#ifdef HAVE_DETAILED_TIMINGS
+   call timer%start("mpi_communication")
 #endif
    call mpi_comm_rank(comm_s,myps,mpierr)
    call mpi_comm_size(comm_s,nps ,mpierr)
    call mpi_comm_rank(comm_t,mypt,mpierr)
    call mpi_comm_size(comm_t,npt ,mpierr)
+#ifdef HAVE_DETAILED_TIMINGS
+   call timer%stop("mpi_communication")
+#endif
 
    ! Look to elpa_transpose_vectors for the basic idea!
 
@@ -175,6 +187,9 @@ subroutine elpa_reduce_add_vectors_complex_single(vmat_s,ld_s,comm_s,vmat_t,ld_t
 #endif
 
 #ifdef WITH_MPI
+#ifdef HAVE_DETAILED_TIMINGS
+   call timer%start("mpi_communication")
+#endif
 
 #if REALCASE==1
 
@@ -196,6 +211,10 @@ subroutine elpa_reduce_add_vectors_complex_single(vmat_s,ld_s,comm_s,vmat_t,ld_t
 
 #endif /* COMPLEXCASE == 1 */
 
+#ifdef HAVE_DETAILED_TIMINGS
+   call timer%stop("mpi_communication")
+#endif
+
 #else /* WITH_MPI */
          if(k>0) aux2 = aux1
 #endif /* WITH_MPI */
@@ -237,7 +256,8 @@ subroutine elpa_reduce_add_vectors_complex_single(vmat_s,ld_s,comm_s,vmat_t,ld_t
    call timer%stop("elpa_reduce_add_vectors_complex_single")
 #endif
 
-#else
+#else /* COMPLEXCASE == 1 */
+
 #ifdef DOUBLE_PRECISION_REAL
    call timer%stop("elpa_reduce_add_vectors_double")
 #else
@@ -245,7 +265,7 @@ subroutine elpa_reduce_add_vectors_complex_single(vmat_s,ld_s,comm_s,vmat_t,ld_t
 #endif
 
 #endif
-#endif
+#endif /* HAVE_DETAILED_TIMINGS */
 
 end subroutine