suppressed ipass, PERFORM_CUTHILL_MCKEE, ACTUALLY_IMPLEMENT_PERM* and…

… FURTHER_REDUCE_CACHE_MISSES, which were making the code too complex with no significant performance improvement. Also suppressed GENERATE_PARAVER_TRACES, which is not used any more.

setup/constants.h.in

@@ -33,33 +33,12 @@
 ! maximum size of the X and Z directions of a JPEG image generated by the display routine
   integer, parameter :: NX_NZ_IMAGE_MAX = 10000
 
-! further reduce cache misses inner/outer in two passes in the case of an MPI simulation
-! this flag is ignored in the case of a serial simulation
-  logical, parameter :: FURTHER_REDUCE_CACHE_MISSES = .true.
-
-! for inverse Cuthill-McKee (1969) permutation
-  logical, parameter :: INVERSE = .true.
-  logical, parameter :: FACE = .false.
-  integer, parameter :: NGNOD_QUADRANGLE = 4
-! perform classical or multi-level Cuthill-McKee ordering
-  logical, parameter :: CMcK_MULTI = .false.
-! maximum size if multi-level Cuthill-McKee ordering
-  integer, parameter :: LIMIT_MULTI_CUTHILL = 50
-
-! implement Cuthill-McKee or replace with identity permutation
-  logical, parameter :: ACTUALLY_IMPLEMENT_PERM_OUT = .false.
-  logical, parameter :: ACTUALLY_IMPLEMENT_PERM_INN = .false.
-  logical, parameter :: ACTUALLY_IMPLEMENT_PERM_WHOLE = .true.
-
 ! create file DATA/model_velocity.dat_output or not (can be huge and slow, thus off by default)
   logical, parameter :: OUTPUT_MODEL_VELOCITY_FILE = .false.
 
 ! in PostScript files about mesh quality, show all elements above this threshold
   double precision, parameter :: THRESHOLD_POSTSCRIPT = 95.d0 / 100.d0
 
-! add MPI barriers and suppress seismograms if we generate traces of the run for analysis with "ParaVer"
-  logical, parameter :: GENERATE_PARAVER_TRACES = .false.
-
 ! option to display only part of the mesh and not the whole mesh,
 ! for instance to analyze Cuthill-McKee mesh partitioning etc.
 ! Possible values are:

src/meshfem2D/read_parameter_file.F90

@@ -137,9 +137,6 @@ module parameter_file
 ! non linear display to enhance small amplitudes in color images
   double precision :: POWER_DISPLAY_COLOR
 
-! perform inverse Cuthill-McKee (1969) permutation for mesh numbering
-  logical :: PERFORM_CUTHILL_MCKEE
-
 ! output seismograms in Seismic Unix format (adjoint traces will be read in the same format)
   logical :: SU_FORMAT
 
@@ -201,9 +198,6 @@ subroutine read_parameter_file()
   call read_value_integer_p(partitioning_method, 'mesher.partitioning_method')
   if(err_occurred() /= 0) stop 'error reading parameter 5a in Par_file'
 
-  call read_value_logical_p(PERFORM_CUTHILL_MCKEE, 'mesher.PERFORM_CUTHILL_MCKEE')
-  if(err_occurred() /= 0) stop 'error reading parameter 5b in Par_file'
-
   call read_value_integer_p(ngnod, 'mesher.ngnod')
   if(err_occurred() /= 0) stop 'error reading parameter 6 in Par_file'
 

src/meshfem2D/save_databases.f90

@@ -197,9 +197,6 @@ subroutine save_databases(nspec,num_material,region_pml_external_mesh, &
     write(15,*) 'POWER_DISPLAY_COLOR'
     write(15,*) POWER_DISPLAY_COLOR
 
-    write(15,*) 'PERFORM_CUTHILL_MCKEE'
-    write(15,*) PERFORM_CUTHILL_MCKEE
-
     write(15,*) 'SU_FORMAT'
     write(15,*) SU_FORMAT
 

src/specfem2D/Makefile.in

@@ -128,7 +128,6 @@ OBJS_SPECFEM2D = \
 	$O/force_ftz.o \
 	$O/get_global.o \
 	$O/get_MPI.o \
-	$O/get_perm_cuthill_mckee.o \
 	$O/get_poroelastic_velocities.o \
 	$O/gll_library.o \
 	$O/gmat01.o \
@@ -309,9 +308,6 @@ $O/get_global.o: ${S}/get_global.f90 ${SETUP}/constants.h
 $O/get_MPI.o: ${S}/get_MPI.F90 ${SETUP}/constants.h
 	${F90} $(FLAGS_CHECK) -c -o $O/get_MPI.o ${S}/get_MPI.F90
 
-$O/get_perm_cuthill_mckee.o: ${S}/get_perm_cuthill_mckee.f90 ${SETUP}/constants.h
-	${F90} $(FLAGS_CHECK) -c -o $O/get_perm_cuthill_mckee.o ${S}/get_perm_cuthill_mckee.f90
-
 $O/get_poroelastic_velocities.o: ${S}/get_poroelastic_velocities.f90 ${SETUP}/constants.h
 	${F90} $(FLAGS_CHECK) -c -o $O/get_poroelastic_velocities.o ${S}/get_poroelastic_velocities.f90
 

src/specfem2D/createnum_fast.f90

@@ -42,15 +42,15 @@
 !
 !========================================================================
 
-  subroutine createnum_fast(knods,ibool,shape,coorg,nglob,npgeo,nspec,ngnod,myrank,ipass)
+  subroutine createnum_fast(knods,ibool,shape,coorg,nglob,npgeo,nspec,ngnod,myrank)
 
 ! same as subroutine "createnum_slow" but with a faster algorithm
 
   implicit none
 
   include "constants.h"
 
-  integer nglob,npgeo,nspec,ngnod,myrank,ipass
+  integer nglob,npgeo,nspec,ngnod,myrank
   integer knods(ngnod,nspec),ibool(NGLLX,NGLLZ,nspec)
   double precision shape(ngnod,NGLLX,NGLLX)
   double precision coorg(NDIM,npgeo)
@@ -70,7 +70,7 @@ subroutine createnum_fast(knods,ibool,shape,coorg,nglob,npgeo,nspec,ngnod,myrank
 
 
 !----  create global mesh numbering
-  if(myrank == 0 .and. ipass == 1) then
+  if(myrank == 0) then
     write(IOUT,*)
     write(IOUT,*)
     write(IOUT,*) 'Generating global mesh numbering (fast version)...'
@@ -231,12 +231,6 @@ subroutine createnum_fast(knods,ibool,shape,coorg,nglob,npgeo,nspec,ngnod,myrank
 ! check the numbering obtained
   if(minval(ibool) /= 1 .or. maxval(ibool) /= nglob) call exit_MPI('Error while generating global numbering')
 
-! if(myrank == 0 .and. ipass == 1) then
-!   write(IOUT,*)
-!   write(IOUT,*) 'Total number of points of the global mesh on slice 0: ',nglob
-!   write(IOUT,*)
-! endif
-
   end subroutine createnum_fast
 
 

src/specfem2D/createnum_slow.f90

@@ -42,15 +42,15 @@
 !
 !========================================================================
 
-  subroutine createnum_slow(knods,ibool,nglob,nspec,ngnod,myrank,ipass)
+  subroutine createnum_slow(knods,ibool,nglob,nspec,ngnod,myrank)
 
 ! generate the global numbering
 
   implicit none
 
   include "constants.h"
 
-  integer nglob,nspec,ngnod,myrank,ipass
+  integer nglob,nspec,ngnod,myrank
 
   integer knods(ngnod,nspec),ibool(NGLLX,NGLLZ,nspec)
 
@@ -63,7 +63,7 @@ subroutine createnum_slow(knods,ibool,nglob,nspec,ngnod,myrank,ipass)
 
 
 !----  create global mesh numbering
-  if(myrank == 0 .and. ipass == 1) then
+  if(myrank == 0) then
     write(IOUT,*)
     write(IOUT,*) 'Generating global mesh numbering (slow version)...'
     write(IOUT,*)
@@ -311,14 +311,5 @@ subroutine createnum_slow(knods,ibool,nglob,nspec,ngnod,myrank,ipass)
 ! verification de la coherence de la numerotation generee
   if(minval(ibool) /= 1 .or. maxval(ibool) /= nglob) call exit_MPI('Error while generating global numbering')
 
-! if(myrank == 0 .and. ipass == 1) then
-!   write(IOUT,*) 'Total number of points of the global mesh on slice 0: ',nglob,' distributed as follows:'
-!   write(IOUT,*)
-!   write(IOUT,*) 'Number of interior points: ',nglob-npedge-npcorn
-!   write(IOUT,*) 'Number of edge points (without corners): ',npedge
-!   write(IOUT,*) 'Number of corner points: ',npcorn
-!   write(IOUT,*)
-! endif
-
   end subroutine createnum_slow
 

src/specfem2D/get_MPI.F90

@@ -55,7 +55,7 @@ subroutine get_MPI(nspec,ibool,knods,ngnod,nglob,elastic,poroelastic, &
                     inum_interfaces_poroelastic, &
                     ninterface_acoustic, ninterface_elastic, ninterface_poroelastic, &
                     mask_ispec_inner_outer, &
-                    myrank,ipass,coord)
+                    myrank,coord)
 
 ! sets up the MPI interface for communication between partitions
 
@@ -85,7 +85,7 @@ subroutine get_MPI(nspec,ibool,knods,ngnod,nglob,elastic,poroelastic, &
 
   logical, dimension(nspec), intent(inout)  :: mask_ispec_inner_outer
 
-  integer :: myrank,ipass
+  integer :: myrank
   double precision, dimension(NDIM,nglob) :: coord
 
   !local parameters
@@ -231,7 +231,7 @@ subroutine get_MPI(nspec,ibool,knods,ngnod,nglob,elastic,poroelastic, &
   ! outputs total number of MPI interface points
   call MPI_REDUCE(num_points2, num_points1, 1, MPI_INTEGER, &
                     MPI_SUM, 0, MPI_COMM_WORLD, ier)
-  if( myrank == 0 .and. ipass == 1 ) then
+  if( myrank == 0 ) then
     write(IOUT,*) 'total MPI interface points: ',num_points1
   endif
 
@@ -276,7 +276,7 @@ subroutine get_MPI(nspec,ibool,knods,ngnod,nglob,elastic,poroelastic, &
   call MPI_REDUCE(inum, num_points1, 1, MPI_INTEGER, &
                     MPI_SUM, 0, MPI_COMM_WORLD, ier)
 
-  if( myrank == 0 .and. ipass == 1 ) then
+  if( myrank == 0 ) then
     write(IOUT,*) '       acoustic interface points: ',num_points1
   endif
 
@@ -310,9 +310,7 @@ subroutine get_MPI(nspec,ibool,knods,ngnod,nglob,elastic,poroelastic, &
                     MPI_SUM, 0, MPI_COMM_WORLD, ier)
 
   if( myrank == 0 ) then
-    if( ipass == 1 ) then
-      write(IOUT,*) '       assembly acoustic MPI interface points:',num_points2
-    endif
+    write(IOUT,*) '       assembly acoustic MPI interface points:',num_points2
 
     ! they don't need to fit, somehow..
     !if( num_points2 /= num_points1 ) then

src/specfem2D/get_global.f90

@@ -42,13 +42,12 @@
 !
 !========================================================================
 
-
-  subroutine get_global(nspec_outer,nspec,nglob,ibool)
+  subroutine get_global(nspec,nglob,ibool)
 
   implicit none
   include "constants.h"
 
-  integer :: nspec_outer,nspec,nglob
+  integer :: nspec,nglob
 
   integer, dimension(NGLLX,NGLLZ,nspec) :: ibool
 
@@ -69,46 +68,6 @@ subroutine get_global(nspec_outer,nspec,nglob,ibool)
 
   inumber = 0
 
-  if(.not. ACTUALLY_IMPLEMENT_PERM_WHOLE) then
-
-  ! first reduce cache misses in outer elements, since they are taken first
-  ! loop over spectral elements
-    do ispec = 1,nspec_outer
-      do j=1,NGLLZ
-        do i=1,NGLLX
-          if(mask_ibool(copy_ibool_ori(i,j,ispec)) == -1) then
-            ! create a new point
-            inumber = inumber + 1
-            ibool(i,j,ispec) = inumber
-            mask_ibool(copy_ibool_ori(i,j,ispec)) = inumber
-          else
-            ! use an existing point created previously
-            ibool(i,j,ispec) = mask_ibool(copy_ibool_ori(i,j,ispec))
-          endif
-        enddo
-      enddo
-    enddo
-
-  ! then reduce cache misses in inner elements, since they are taken second
-  ! loop over spectral elements
-    do ispec = nspec_outer+1,nspec
-      do j=1,NGLLZ
-        do i=1,NGLLX
-          if(mask_ibool(copy_ibool_ori(i,j,ispec)) == -1) then
-            ! create a new point
-            inumber = inumber + 1
-            ibool(i,j,ispec) = inumber
-            mask_ibool(copy_ibool_ori(i,j,ispec)) = inumber
-          else
-            ! use an existing point created previously
-            ibool(i,j,ispec) = mask_ibool(copy_ibool_ori(i,j,ispec))
-          endif
-        enddo
-      enddo
-    enddo
-
-  else ! if ACTUALLY_IMPLEMENT_PERM_WHOLE
-
   ! reduce cache misses in all the elements
   ! loop over spectral elements
     do ispec = 1,nspec
@@ -127,8 +86,6 @@ subroutine get_global(nspec_outer,nspec,nglob,ibool)
       enddo
     enddo
 
-  endif
-
   deallocate(mask_ibool,copy_ibool_ori)
 
   end subroutine get_global