merged

TACC · Dec 10, 2014 · a18ec9f · a18ec9f
2 parents 07547e9 + 28eb1c9
commit a18ec9f
Show file tree

Hide file tree

Showing 4 changed files with 70 additions and 35 deletions.
diff --git a/build.rtm b/build.rtm
@@ -50,6 +50,7 @@ CXX=mpicxx
 FC=mpif90
 F77=mpif90
 
+EXTRA="--with-phdf5=$PHDF5_DIR"
 
 case $SYSHOST in
   ls4)
@@ -61,10 +62,9 @@ case $SYSHOST in
      PHDF5_DIR=$TACC_HDF5_DIR
     ;;
 
-  kraken)
+  xc30)
      PREFIX="$SCRATCH/t3pio"
-     EXTRA="--enable-static --disable-shared"
-     PHDF5_DIR=$HDF5_DIR
+     EXTRA="--enable-static --disable-shared LDFLAGS=-static"
      CC=cc
      CXX=CC
      FC=ftn
@@ -79,7 +79,7 @@ esac
 
 
 cmdA=("-make distclean"
-      "FCFLAGS=\"$FLAGS\" FFLAGS=\"$FLAGS\" CFLAGS=\"$FLAGS\" CXXFLAGS=\"$FLAGS\" FC=$FC F77=$F77 CC=$CC CXX=$CXX ./configure --prefix=$PREFIX --with-phdf5=$PHDF5_DIR $EXTRA"
+      "FCFLAGS=\"$FLAGS\" FFLAGS=\"$FLAGS\" CFLAGS=\"$FLAGS\" CXXFLAGS=\"$FLAGS\" FC=$FC F77=$F77 CC=$CC CXX=$CXX ./configure --prefix=$PREFIX $EXTRA"
       "make"
       "make install"
       )

diff --git a/cubeTest/cmdline.F90 b/cubeTest/cmdline.F90
@@ -12,6 +12,7 @@ module cmdline
    integer :: GblFileSz     ! Global File Size in GB.
    integer :: Numvar        ! number of variables
    integer :: MaxWriters    ! the max number of writers.
+   logical :: PartYZ        ! partition processor grid in y-z only.
    logical :: UseT3PIO      ! if true then use T3PIO (on by default).
    logical :: VersionFlag   ! if true then report version and quit.
    logical :: HelpFlag      ! if true then print usage and quit
@@ -52,6 +53,7 @@ subroutine parse()
       HDF5Flag      = .false.
       LuaOutput     = .false.
       TableOutput   = .true.
+      PartYZ        = .false.
 
 #ifdef USE_HDF5
       ROMIO         = .false.
@@ -82,6 +84,8 @@ subroutine parse()
             i = i + 1
             call getarg(i,optarg)
             read(optarg,*, err=11) nDim
+         elseif (arg == "--yz") then
+            PartYZ      = .true.
          elseif (arg == "--lua") then
             LuaOutput   = .true.
             TableOutput = .false.

diff --git a/cubeTest/master.F90 b/cubeTest/master.F90
@@ -25,7 +25,7 @@ program main
       call exit()
    end if
 
-   call partitionProc(nDim)
+   call partitionProc(nDim, PartYZ)
 
    ! partition grid to local locations.
    call partitionGrid(global, local)

diff --git a/cubeTest/parallel.F90 b/cubeTest/parallel.F90
@@ -29,48 +29,79 @@ subroutine msg_init(comm)
 
    end subroutine msg_init
 
-   subroutine partitionProc(ndim)
+   subroutine partitionProc(ndim, yzPart)
       implicit none
 
-      integer :: ndim
-      real(8) :: xp
-      integer :: exponent, npx, npy, npz, npxy, rem
+      integer :: ndim, npA(2)
+      logical :: yzPart
+      integer :: rem
 
+      if (yzPart) then
+         call build2d_partition(npA)
+         nProcA(1) = 1
+         nProcA(2) = npA(1)
+         nProcA(3) = npA(2)
+         iProcA(3) = p % myProc/(nProcA(1) * nProcA(2))
+         rem       = mod(p % myProc,(nProcA(1) * nProcA(2)))
+         iProcA(2) = rem / nProcA(1)
+         iProcA(1) = mod(rem,nProcA(1))
+         return
+      end if
+
       if (ndim == 2) then
-         xp       = sqrt(real(p % nProcs))
-         exponent = int(log(xp)/log(2.0)) + 1
-         npx      = int(2.0**exponent + 1.5)
-         do while ( npx * (p % nProcs/npx) /= p % nProcs)
-            npx = npx - 1
-         end do
-
-         nProcA(1) = npx
-         nProcA(2) = p % nProcs / npx
-
+         call build2d_partition(nProcA)
          iProcA(1) = mod(p % myProc, nProcA(1))
          iProcA(2) = p % myProc/nProcA(1)
 
       else if (ndim == 3) then
-         xp  = p % nProcs
-         npz = xp**(1.0/3.0) + 1.5
-         do while ( npz * (p % nProcs/npz) /= p % nProcs )
-            npz = npz - 1
-         end do
-         npxy = p % nProcs/npz
-         npx  = sqrt(real(npxy)) + 1.5
-         do while ( npx *(npxy/npx) /= npxy)
-            npx = npx - 1
-         end do
-         npy = npxy/npx
-
-         nProcA(1) = npx
-         nProcA(2) = npy
-         nProcA(3) = npz
-
+         call build3d_partition(nProcA)
          iProcA(3) = p % myProc/(nProcA(1) * nProcA(2))
          rem       = mod(p % myProc,(nProcA(1) * nProcA(2)))
          iProcA(2) = rem / nProcA(1)
          iProcA(1) = mod(rem,nProcA(1))
       end if
    end subroutine partitionProc
+
+   subroutine build2d_partition(np)
+      integer np(2)
+      real(8) :: xp
+      integer :: exponent, npx
+
+      xp       = sqrt(real(p % nProcs))
+      exponent = int(log(xp)/log(2.0)) + 1
+      npx      = int(2.0**exponent + 1.5)
+      do while ( npx * (p % nProcs/npx) /= p % nProcs)
+         npx = npx - 1
+      end do
+
+      np(1) = npx
+      np(2) = p % nProcs / npx
+   end subroutine build2d_partition
+
+   subroutine build3d_partition(np)
+      integer np(3)
+      real(8) :: xp
+      integer :: exponent, npx, npy, npz, npxy
+
+      xp  = p % nProcs
+      npz = xp**(1.0/3.0) + 1.5
+      do while ( npz * (p % nProcs/npz) /= p % nProcs )
+         npz = npz - 1
+      end do
+      npxy = p % nProcs/npz
+      npx  = sqrt(real(npxy)) + 1.5
+      do while ( npx *(npxy/npx) /= npxy)
+         npx = npx - 1
+      end do
+      npy = npxy/npx
+
+      np(1) = npx
+      np(2) = npy
+      np(3) = npz
+
+   end subroutine build3d_partition
+
+
+
+
 end module parallel