Only print Kernel FLOP/s with detailed timings

src/elpa2_compute_complex_template.X90

@@ -3334,9 +3334,9 @@
 #endif
 
       ! Just for measuring the kernel performance
-      real(kind=c_double)                         :: kernel_time  ! MPI_WTIME always needs double
+      real(kind=c_double)                         :: kernel_time, kernel_time_recv  ! MPI_WTIME always needs double
       ! long integer
-      integer(kind=lik)                           :: kernel_flops
+      integer(kind=lik)                           :: kernel_flops, kernel_flops_recv
 
       logical, intent(in)                         :: wantDebug
       integer(kind=ik)                            :: istat
@@ -5949,10 +5949,23 @@
 #ifdef WITH_MPI
         if (ANY(result_send_request /= MPI_REQUEST_NULL)) write(error_unit,*) '*** ERROR result_send_request ***',my_prow,my_pcol
         if (ANY(result_recv_request /= MPI_REQUEST_NULL)) write(error_unit,*) '*** ERROR result_recv_request ***',my_prow,my_pcol
+
+#ifdef HAVE_DETAILED_TIMINGS
+       call MPI_ALLREDUCE(kernel_flops, kernel_flops_recv, 1, MPI_INTEGER8, MPI_SUM, MPI_COMM_ROWS, mpierr)
+       kernel_flops = kernel_flops_recv
+       call MPI_ALLREDUCE(kernel_flops, kernel_flops_recv, 1, MPI_INTEGER8, MPI_SUM, MPI_COMM_COLS, mpierr)
+       kernel_flops = kernel_flops_recv
+
+       call MPI_ALLREDUCE(kernel_time, kernel_time_recv, 1, MPI_REAL8, MPI_MAX, MPI_COMM_ROWS, mpierr)
+       kernel_time_recv = kernel_time
+       call MPI_ALLREDUCE(kernel_time, kernel_time_recv, 1, MPI_REAL8, MPI_MAX, MPI_COMM_COLS, mpierr)
+       kernel_time_recv = kernel_time
 #endif
+
+#else /* WITH_MPI */
         if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
           write(error_unit,'(" Kernel time:",f10.3," MFlops: ",es12.5)') kernel_time, kernel_flops/kernel_time*1.d-6
-
+#endif /* WITH_MPI */
         ! deallocate all working space
 
         if (.not.(useGPU)) then

src/elpa2_compute_real_template.X90

@@ -3851,9 +3851,9 @@
       integer(kind=ik), parameter              :: result_recv_tag = 333
 
       ! Just for measuring the kernel performance
-      real(kind=c_double)                      :: kernel_time ! MPI_WTIME always needs double
+      real(kind=c_double)                      :: kernel_time, kernel_time_recv ! MPI_WTIME always needs double
       ! long integer
-      integer(kind=lik)                        :: kernel_flops
+      integer(kind=lik)                        :: kernel_flops, kernel_flops_recv
 
 #ifdef WITH_OPENMP
       integer(kind=ik)                         :: max_threads, my_thread
@@ -6292,11 +6292,25 @@
 #ifdef WITH_MPI
      if (ANY(result_send_request /= MPI_REQUEST_NULL)) write(error_unit,*) '*** ERROR result_send_request ***',my_prow,my_pcol
      if (ANY(result_recv_request /= MPI_REQUEST_NULL)) write(error_unit,*) '*** ERROR result_recv_request ***',my_prow,my_pcol
+
+#ifdef HAVE_DETAILED_TIMINGS
+       call MPI_ALLREDUCE(kernel_flops, kernel_flops_recv, 1, MPI_INTEGER8, MPI_SUM, MPI_COMM_ROWS, mpierr)
+       kernel_flops = kernel_flops_recv
+       call MPI_ALLREDUCE(kernel_flops, kernel_flops_recv, 1, MPI_INTEGER8, MPI_SUM, MPI_COMM_COLS, mpierr)
+       kernel_flops = kernel_flops_recv
+
+       call MPI_ALLREDUCE(kernel_time, kernel_time_recv, 1, MPI_REAL8, MPI_MAX, MPI_COMM_ROWS, mpierr)
+       kernel_time_recv = kernel_time
+       call MPI_ALLREDUCE(kernel_time, kernel_time_recv, 1, MPI_REAL8, MPI_MAX, MPI_COMM_COLS, mpierr)
+       kernel_time_recv = kernel_time
 #endif
 
+#else /* WITH_MPI */
+
      if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
          write(error_unit,'(" Kernel time:",f10.3," MFlops: ",es12.5)')  kernel_time, kernel_flops/kernel_time*1.d-6
 
+#endif /* WITH_MPI */
 
      ! copy q to q_dev needed in trans_ev_band_to_full
 #ifdef DOUBLE_PRECISION_REAL