diff --git a/code/m_calculate.f90 b/code/m_calculate.f90
index d62ce53..53e7327 100644
--- a/code/m_calculate.f90
+++ b/code/m_calculate.f90
@@ -31,15 +31,12 @@ contains
 
      call ludcmp(nxp,nyp,nzp,1,0,0)      ! 1,1,0
 
-     allocate(xsrc(nyp, nxp), stat=ierr)
-     allocate(xdst(nyp, nxp), stat=ierr)
+     allocate(xsrc(nb, nxp), stat=ierr)
+     allocate(xdst(nb, nxp), stat=ierr)
 
      allocate(rsrc(nxp, nzp), stat=ierr)
      allocate(rdst(nxp, nzp), stat=ierr)
 
-     allocate(zsrc(nyp, nzp, nxp), stat=ierr)
-     allocate(zdst(nyp, nzp, nxp), stat=ierr)
-
      end subroutine m_calculate_init
 
 
@@ -53,9 +50,6 @@ contains
      deallocate(rsrc)
      deallocate(rdst)
 
-     deallocate(zsrc)
-     deallocate(zdst)
-
      end subroutine m_calculate_finalize
 
 
@@ -65,16 +59,21 @@ contains
         real*8, dimension(nxp,nyp,nzp), intent(out) :: dst
 
         integer :: i, j ,k
+        integer :: ju
 
 
         do k = 1,nzp
+        do j = 1,nyp,nb
 
-        call tp2(xsrc, src(:,:,k), nyp, nxp)
+        ju = min(j+nb-1,nyp)
 
-        call dfnonp(nxp, hxp, xsrc, xdst, nyp, 1)
+        call tp2(xsrc, src(:,j:ju,k), nb, nxp)
 
-        call tp2(dst(:,:,k), xdst, nxp, nyp)
+        call dfnonp(nxp, hxp, xsrc, xdst, nb, 1)
 
+        call tp2(dst(:,j:ju,k), xdst, nxp, nb)
+
+        end do
         end do
 
      end subroutine ddx
@@ -104,27 +103,6 @@ contains
 
         integer :: i, j ,k
 
-
-        call tp2(zsrc, src, nyp*nzp, nxp)
-
-        do i = 1,nxp
-
-        call dfp(nzp, hzp, zsrc(:,:,i), zdst(:,:,i), nyp, 3)
-
-        end do
-
-        call tp2(dst, zdst, nxp, nyp*nzp)
-
-     end subroutine ddz
-
-
-     subroutine ddz_ref(dst, src)
-
-        real*8, dimension(nxp,nyp,nzp), intent(in) :: src
-        real*8, dimension(nxp,nyp,nzp), intent(out) :: dst
-
-        integer :: i, j ,k
-
         do j = 1,nyp
 
         do k = 1,nzp
@@ -139,7 +117,7 @@ contains
 
         end do
 
-     end subroutine ddz_ref
+     end subroutine ddz
 
 
      subroutine d2dx(dst, src)
@@ -148,16 +126,21 @@ contains
         real*8, dimension(nxp,nyp,nzp), intent(out) :: dst
 
         integer :: i, j ,k
+        integer :: ju
 
 
         do k = 1,nzp
+        do j = 1,nyp,nb
 
-        call tp2(xsrc, src(:,:,k), nyp, nxp)
+        ju = min(j+nb-1,nyp)
 
-        call d2fnonp(nxp, hxp, xsrc, xdst, nyp, 1)
+        call tp2(xsrc, src(:,j:ju,k), nb, nxp)
 
-        call tp2(dst(:,:,k), xdst, nxp, nyp)
+        call d2fnonp(nxp, hxp, xsrc, xdst, nb, 1)
 
+        call tp2(dst(:,j:ju,k), xdst, nxp, nb)
+
+        end do
         end do
 
      end subroutine d2dx
@@ -187,27 +170,6 @@ contains
 
         integer :: i, j ,k
 
-
-        call tp2(zsrc, src, nyp*nzp, nxp)
-
-        do i = 1,nxp
-
-        call d2fp(nzp, hzp, zsrc(:,:,i), zdst(:,:,i), nyp, 3)
-
-        end do
-
-        call tp2(dst, zdst, nxp, nyp*nzp)
-
-     end subroutine d2dz
-
-
-     subroutine d2dz_ref(dst, src)
-
-        real*8, dimension(nxp,nyp,nzp), intent(in) :: src
-        real*8, dimension(nxp,nyp,nzp), intent(out) :: dst
-
-        integer :: i, j ,k
-
         do j = 1,nyp
 
         do k = 1,nzp
@@ -222,7 +184,7 @@ contains
 
         end do
 
-     end subroutine d2dz_ref
+     end subroutine d2dz
 
 
      subroutine tp(a, b, nx)
diff --git a/code/makefile b/code/makefile
index 93aa037..927d663 100644
--- a/code/makefile
+++ b/code/makefile
@@ -4,7 +4,7 @@
 #Cluster_2(16.161) flags 
 
 
-BLOCKSIZE?=16
+BLOCKSIZE?=32
 
 flags = -cpp -DBLOCKSIZE=$(BLOCKSIZE) -Wall -O3 -ffree-line-length-0 -fdefault-integer-8 -fdefault-double-8 -fdefault-real-8 -march=native