File: test.F

package info (click to toggle)
ga 5.9.2-2
links: PTS, VCS
area: main
in suites: forky, sid
size: 18,472 kB
sloc: ansic: 192,963; fortran: 53,761; f90: 11,218; cpp: 5,784; makefile: 2,248; sh: 1,945; python: 1,734; perl: 534; csh: 134; asm: 106
file content (4264 lines) | stat: -rw-r--r-- 109,688 bytes
parent folder | download | duplicates (3)
#if HAVE_CONFIG_H 
#   include "config.fh"
#endif
c $Id: test.F,v 1.64.2.11 2007-04-06 22:37:35 d3g293 Exp $
c vector boxes lack arithmetic precision 
# define THRESH 1d-13
# define THRESHF 2e-5

#define MISMATCH(x,y) abs(x-y)/max(1d0,abs(x)).gt.THRESH
#define MISMATCHF(x,y) abs(x-y)/max(1.0,abs(x)).gt.THRESHF 

c#define NEW_API
c#define MIRROR
#define GA3
#define NGA_GATSCAT
#define BLOCK_CYCLIC
c#define USE_SCALAPACK_DISTR
c#define USE_TILED_DISTR
c#define USE_TILED_IRREG_DISTR
c#define USE_RESTRICTED
c#define USE_MEMDEV

#ifdef USE_RESTRICTED
#  define NEW_API
#endif

#define MEM_INC 1000

#ifdef USE_SCALAPACK_DISTR
#  undef USE_TILED_DISTR
#  undef USE_TILED_IRREG_DISTR
#  define BLOCK_CYCLIC
#endif

#ifdef USE_TILED_DISTR
#  undef USE_SCALAPACK_DISTR
#  undef USE_TILED_IRREG_DISTR
#  define BLOCK_CYCLIC
#endif

#ifdef USE_TILED_IRREG_DISTR
#  undef USE_SCALAPACK_DISTR
#  undef USE_TILED_DISTR
#  define BLOCK_CYCLIC
#endif

#ifdef BLOCK_CYCLIC
#  define NEW_API
#  undef MIRROR
#endif

#ifdef USE_MEMDEV
#  define NEW_API
#  define MEMDEVICE "dram"
#endif

      program main
      implicit none
#include "mafdecls.fh"
#include "global.fh"
#include "testutil.fh"
      integer heap, stack, fudge, ma_heap, me, nproc, map(4096), block
      integer g_s, ndim, dim1, i
      integer version, subversion, patch
      logical status
      parameter (heap=800*800*4, fudge=100, stack=800*800)
c     
c***  Intitialize a message passing library
c
#include "mp3.fh"
c
c***  Initialize GA
c
c     There are 2 choices: ga_initialize or ga_initialize_ltd.
c     In the first case, there is no explicit limit on memory usage.
c     In the second, user can set limit (per processor) in bytes.
c
      call ga_initialize()
      nproc = ga_nnodes()
      me = ga_nodeid()
c     we can also use GA_set_memory_limit BEFORE first ga_create call
c
      ma_heap = heap/nproc + fudge 
      ma_heap = 2*ma_heap
#ifdef USE_RESTRICTED
      ma_heap = 2*ma_heap
#endif
      print *, 'util_mdtob(', ma_heap, ')=', util_mdtob(ma_heap)
      call GA_set_memory_limit(util_mdtob(ma_heap))
c
      call ga_version(version,subversion,patch)
      if(ga_nodeid().eq.0)then
#ifdef MIRROR
         print *,' Performing tests on Mirrored Arrays'
#endif
         print *,' GA initialized '
         print *,' '
         write(6,'(a,i1,a,i1,a,i1)') ' Version ',version,'.',
     +      subversion,'.',patch
         call ffflush(6)
#ifdef BLOCK_CYCLIC
#if !defined(USE_SCALAPACK_DISTR) && !defined(USE_TILE_DISTR) && !defined(USE_TILED_IRREG_DISTR)
         print *,' '
         print *,' Testing block-cyclic arrays '
         print *,' '
#endif
#endif
#ifdef USE_SCALAPACK_DISTR
         print *,' '
         print *,' Testing ScaLAPACK distributed arrays '
         print *,' '
#endif
#ifdef USE_TILED_DISTR
         print *,' '
         print *,' Testing tiled-distributed arrays '
         print *,' '
#endif
#ifdef USE_TILED_IRREG_DISTR
         print *,' '
         print *,' Testing irregularly tiled-distributed arrays '
         print *,' '
#endif
#ifdef USE_RESTRICTED
         print *,' '
         print *,' Testing restricted arrays '
         print *,' '
#endif
      endif
#ifdef BLOCK_CYCLIC
#if !defined(USE_SCALAPACK_DISTR) && !defined(USE_TILE_DISTR) && !defined(USE_TILED_IRREG_DISTR)
      if (me.eq.0) print *,' '
      if (me.eq.0) print *,' Testing block-cyclic layout'
      if (me.eq.0) print *,' '
#endif
#endif
#ifdef USE_SCALAPACK_DISTR
      if (me.eq.0) print *,' '
      if (me.eq.0) print *,' Testing ScaLAPACK block-cyclic layout'
      if (me.eq.0) print *,' '
#endif
#ifdef USE_TILED_DISTR
      if (me.eq.0) print *,' '
      if (me.eq.0) print *,' Testing tiled layout'
      if (me.eq.0) print *,' '
#endif
#ifdef USE_TILED_IRREG_DISTR
      if (me.eq.0) print *,' '
      if (me.eq.0) print *,' Testing irregular tiled layout'
      if (me.eq.0) print *,' '
#endif
c
c***  Initialize the MA package
c     MA must be initialized before any global array is allocated
c
      status = ma_init(MT_DCPL, stack, ma_heap)
      if (.not. status) call ga_error('ma_init failed',-1) 
c
c     Uncomment the below line to register external memory allocator
c     for dynamic arrays inside GA routines.
c      call register_ext_memory()
c
      if(me.eq.(nproc-1))then
        print *, 'using ', nproc,' process(es) ', ga_cluster_nnodes(),
     $           ' cluster nodes'
        print *,'process ', me, ' is on node ',ga_cluster_nodeid(),
     $          ' with ',  ga_cluster_nprocs(-1), ' processes' 
        call ffflush(6)
      endif
c
c   create array to force staggering of memory and uneven distribution
c   of pointers
c
      dim1 = MEM_INC
      map(1) = 1
      do i = 2, nproc
        map(i) = MEM_INC*(i-1)+1
        dim1 = dim1 + MEM_INC*i
      end do
      g_s = ga_create_handle()
      ndim = 1
      call ga_set_data(g_s,ndim,dim1,MT_INT)
      call ga_set_array_name(g_s,'s')
      call ga_set_irreg_distr(g_s,map,nproc)

#ifdef MIRROR
      if (me.eq.0) then
         write(6,*)
         write(6,*) '  TESTING MIRRORED ARRAYS  '
         write(6,*)
         call ffflush(6)
      endif
#endif
      call ga_sync()
c
c***  Check support for double precision arrays
c
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' CHECKING DOUBLES  '
         write(6,*)
         call ffflush(6)
      endif

      call check_dbl()
c
c***  Check support for single precision complex arrays
c
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' CHECKING SINGLE COMPLEX  '
         write(6,*)
         call ffflush(6)
      endif

      call check_complex_float()
c
c***  Check support for double precision complex arrays
c
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' CHECKING DOUBLE COMPLEX  '
         write(6,*)
         call ffflush(6)
      endif

      call check_complex()
c
c***  Check support for integer arrays
c
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' CHECKING INTEGERS  '
         write(6,*)
         call ffflush(6)
      endif

      call check_int()
c
c
c***  Check support for single precision
c
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' CHECKING SINGLE PRECISION  '
         write(6,*)
         call ffflush(6)
      endif
 
      call check_flt() 
c     
      if (me.eq.0) then
         write(6,*)
         write(6,*)'CHECKING Wrappers to Message Passing Collective ops'
         write(6,*)
         call ffflush(6)
      endif

      call check_wrappers
c
c***  Check if memory limits are enforced 
c
      call check_mem(util_mdtob(ma_heap*ga_nnodes()))
c
      if(me.eq.0) call ga_print_stats()
      if(me.eq.0) print *,' ' 
      if(me.eq.0) print *,'All tests successful ' 
      status = ga_destroy(g_s)
c
c***  Tidy up the GA package
c
      call ga_terminate()
c
c***  Tidy up after message-passing library
c
      call MP_FINALIZE()
c
      end


      subroutine check_dbl()
      implicit none
#include "mafdecls.fh"
#include "global.fh"
#include "testutil.fh"
c     
      integer n,m
      parameter (n = 256)
      parameter (m = 2*n)
      double precision a(n,n), b(n,n), v(m),w(m)
#ifdef MIRROR
      logical lsame
      integer g_c, p_mirror
      integer ndim, dims(2), chunk(2)
      integer alo(2), blo(2), ahi(2), bhi(2)
#else
#  ifdef NEW_API
      integer ndim, dims(2), chunk(2)
#  endif
#endif
      integer iv(m), jv(m)
#ifdef NGA_GATSCAT
      integer ijv(2,m)
#endif
      logical status
      integer g_a, g_b
      integer iran, i,j, loop,nloop,maxloop, ilo, ihi, jlo, jhi, itmp
      integer nproc, me, int, ij, inc, ii, jj
      parameter (maxloop = 100)
      integer maxproc
      parameter (maxproc = 4096)
      double precision crap, sum1, sum2, x
      double precision nwords
      logical found, nfound
      integer lprocs,inode,iproc,nnodes,lproc
#ifdef USE_RESTRICTED
      integer num_rstrctd
      integer rstrctd_list(maxproc/2)
#endif
#ifdef BLOCK_CYCLIC
      integer block_size(2), proc_grid(2)
#endif
#ifdef USE_TILED_IRREG_DISTR
      integer mapc(4096), nblocks(2)
#endif
      intrinsic int
      iran(i) = int(drand(0)*real(i)) + 1
c     
      nproc = ga_nnodes()
      me    = ga_nodeid()
      inode = ga_cluster_nodeid()
      lprocs = ga_cluster_nprocs(inode)
      nnodes = ga_cluster_nnodes()
      iproc = mod(me,lprocs)
      nloop = Min(maxloop,n)
#ifdef USE_RESTRICTED
      num_rstrctd = nproc/2
      if (num_rstrctd.eq.0) num_rstrctd = 1
      do i = 1, num_rstrctd
        rstrctd_list(i) = (num_rstrctd/2) + i-1
      end do
#endif
#ifdef BLOCK_CYCLIC
      block_size(1) = 32
      block_size(2) = 32
#if defined(USE_SCALAPACK_DISTR) || defined(USE_TILED_DISTR)
      if (mod(nproc,2).ne.0)
     +  call ga_error("Available procs must be divisible by 2",-1)
      proc_grid(1) = 2
      proc_grid(2) = nproc/2
#endif
#endif
c     
c     a() is a local copy of what the global array should start as
c
      do j = 1, n
         do i = 1, n
#ifndef MIRROR
            a(i,j) = i-1 + (j-1)*n
#else
            a(i,j) = inode + i-1 + (j-1)*n
#endif
            b(i,j) =-1. 
         enddo
      enddo
*      write(6,*) ' correct '
*      call output(a, 1, n, 1, n, n, n, 1)
*      call ffflush(6)
c
c     Create a global array
c
*     print *,ga_nodeid(), ' creating array'
*     call ffflush(6)
c     call setdbg(1)
#ifdef NEW_API
      g_a = ga_create_handle()
      ndim = 2
      dims(1) = n
      dims(2) = n
      call ga_set_data(g_a,ndim,dims,MT_DBL)
      call ga_set_array_name(g_a,'a')
#ifdef USE_RESTRICTED
      call ga_set_restricted(g_a, rstrctd_list, num_rstrctd)
#endif

#ifdef BLOCK_CYCLIC
#if defined(USE_SCALAPACK_DISTR) || defined(USE_TILED_DISTR) || defined(USE_TILED_IRREG_DISTR)
#ifdef USE_SCALAPACK_DISTR
      call ga_set_block_cyclic_proc_grid(g_a,block_size,proc_grid)
#else
#ifdef USE_TILED_DISTR
      call ga_set_tiled_proc_grid(g_a,block_size,proc_grid)
#else
      call util_create_map(mapc,block_size(1),dims,nblocks,proc_grid)
      call ga_set_tiled_irreg_proc_grid(g_a,mapc,nblocks,proc_grid)
#endif
#endif
#else
      call ga_set_block_cyclic(g_a,block_size)
#endif
#endif

#  ifdef MIRROR
      p_mirror = ga_pgroup_get_mirror()
      call ga_set_pgroup(g_a,p_mirror)
#  endif
#  ifdef USE_MEMDEV
      call ga_set_memory_dev(g_a,MEMDEVICE)
#  endif
      status = ga_allocate(g_a)
#else
#  ifndef MIRROR
      status = ga_create(MT_DBL, n, n, 'a', 0, 0, g_a)
#  else
      ndim = 2
      dims(1) = n
      dims(2) = n
      chunk(1) = 0
      chunk(2) = 0
      p_mirror = ga_pgroup_get_mirror()
      status = nga_create_config(MT_DBL, ndim,  dims, 'a', chunk,
     +                  p_mirror, g_a)
#  endif
#endif
      if (.not. status) then
         write(6,*) ' ga_create failed'
         call ffflush(6)
         call ga_error('... exiting ',-1)
      endif
c
c     check if handle is valid. be quiet unless error
C
      if(.not.ga_valid_handle(g_a)) call ga_error("invalid handle",g_a)
c
#ifndef MIRROR
      call ga_distribution(g_a,me,ilo, ihi, jlo, jhi)
#else
      lproc = me - ga_cluster_procid(inode,0)
      call ga_distribution(g_a, lproc, ilo, ihi, jlo, jhi)
#endif
      call ga_sync()
c     
c     Zero the array
c
      if (me .eq. 0) then 
         write(6,21) 
 21      format('> Checking zero ... ')
         call ffflush(6)
      endif
      call ga_zero(g_a)
c
c     Check that it is indeed zero
c
      call ga_get(g_a, 1, n, 1, n, b, n)
      call ga_sync()
      do i = 1, n
         do j = 1, n
            if (b(i,j) .ne. 0.0d0) then
               write(6,*) me,' zero ', i, j, b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_zero is OK'
         write(6,*)
      endif
      call ga_sync()
c
c     Each node fills in disjoint sections of the array
c
      if (me .eq. 0) then
         write(6,2) 
 2       format('> Checking disjoint put ... ')
         call ffflush(6)
      endif
      call ga_sync()
c
      inc = (n-1)/20 + 1
      ij = 0
      do j = 1, n, inc
         do i = 1, n, inc
#ifndef MIRROR
            if (mod(ij,nproc) .eq. me) then
#else
            if (mod(ij,lprocs) .eq. iproc) then
#endif
               ilo = i
               ihi = min(i+inc, n)
               jlo = j
               jhi = min(j+inc, n)
*               write(6,4) me, ilo, ihi, jlo, jhi
* 4             format(' node ',i2,' checking put ',4i4)
*               call ffflush(6)
               call ga_put(g_a, ilo, ihi, jlo, jhi, a(ilo, jlo), n)
            endif
            ij = ij + 1
         enddo
      enddo
      call ga_sync()
c
c     All nodes check all of a
c
      call util_dfill(n*n, 0.0d0, b, 1)
*     call ga_print(g_a,1)
      call ga_get(g_a, 1, n, 1, n, b, n)
*      write(6,*) ' after get'
*      call output(b, 1, n, 1, n, n, n, 1)
c
      do i = 1, n
         do j = 1, n
            if (b(i,j) .ne. a(i,j)) then
               write(6,*) ' put ', me, i, j, a(i,j),b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_put is OK'
         write(6,*)
      endif
      call ga_sync()
c
c     Now check nloop random gets from each node
c
      if (me .eq. 0) then
         write(6,5) nloop
 5       format('> Checking random get (',i5,' calls)...')
         call ffflush(6)
      endif
      call ga_sync()
c
      nwords = 0
c
      crap = drand(ga_nodeid()*51 +1) !Different seed for each proc
      do loop = 1, nloop
         ilo = iran(loop)
         ihi = iran(loop)
         if (ihi.lt. ilo) then
            itmp = ihi
            ihi = ilo
            ilo = itmp
         endif
         jlo = iran(loop)
         jhi = iran(loop)
         if (jhi.lt. jlo) then
            itmp = jhi
            jhi = jlo
            jlo = itmp
         endif
c
         nwords = nwords + (ihi-ilo+1)*(jhi-jlo+1)
c     
         call util_dfill(n*n, 0.0d0, b, 1)
         call ga_get(g_a, ilo, ihi, jlo, jhi, b(ilo, jlo), n)
         if (me .eq. 0 .and. mod(loop-1, max(1,nloop/20)).eq.0) then
            write(6,1) loop, me, ilo, ihi, jlo, jhi, nwords
 1          format(' call ',i5, ' node ',i2,' checking get ',4i4,
     $           ' total ',d9.2)
            call ffflush(6)
         endif
         sum1 = 0.0d0
         do j = jlo, jhi
            do i = ilo, ihi
               sum1 = sum1 + b(i,j)
               if (b(i,j) .ne. a(i,j)) then
                  write(6,*) i, j, b(i,j), a(i,j)
                  call ffflush(6)
                  call ga_error('... exiting ',-1)
               endif
            enddo
         enddo
c
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_get is OK'
         write(6,*)
         call ffflush(6)
      endif
      call ga_sync()
c
c     Each node accumulates into disjoint sections of the array
c
#if 1
      if (me .eq. 0) then
         write(6,9) 
 9       format('> Checking accumulate ... ')
         call ffflush(6)
      endif
      call ga_sync()
c     
      crap = drand(12345)       ! Same seed for each process
      do j = 1, n
         do i = 1, n
c           b(i,j) = drand(0)
            b(i,j) = i+j 
         enddo
      enddo
c
      inc = (n-1)/20 + 1
      ij = 0
      do j = 1, n, inc
         do i = 1, n, inc
c           x = drand(0)
            x = 10.
            ilo = i
            ihi = min(i+inc-1, n)
            if(ihi.eq.n-1)ihi=n
c           ihi = min(i+inc, n)
            jlo = j
            jhi = min(j+inc-1, n)
            if(jhi.eq.n-1)jhi=n
c           jhi = min(j+inc-1, n)
*               call ffflush(6)
#ifndef MIRROR
            if (mod(ij,nproc) .eq. me) then
#else
            if (mod(ij,lprocs) .eq. iproc) then
#endif
c               print *, me, 'checking accumulate ',ilo,ihi,jlo,jhi,x
* 11            format(' node ',i2,' checking accumulate ',4i4)
*               call ffflush(6)
               call ga_acc(g_a, ilo, ihi, jlo, jhi, b(ilo, jlo), n, x)
            endif
            ij = ij + 1
c     
c           Each process applies all updates to its local copy
c
            do jj = jlo, jhi
               do ii = ilo, ihi
                  a(ii,jj) = a(ii,jj) + x * b(ii,jj)
               enddo
            enddo
         enddo
      enddo
      call ga_sync()
c
c     All nodes check all of a
c
      call ga_get(g_a, 1, n, 1, n, b, n)
      do j = 1, n
         do i = 1, n
            if(MISMATCH(b(i,j),a(i,j)))then
               write(6,*) ' acc ', me, i, j, a(i,j), b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' disjoint ga_acc is OK'
         write(6,*)
      endif
c
c     overlapping accumulate
      call ga_sync()
#ifdef NEW_API
      g_b = ga_create_handle()
      call ga_set_data(g_b,ndim,dims,MT_DBL)
      call ga_set_array_name(g_b,'b')

#ifdef BLOCK_CYCLIC
#if defined(USE_SCALAPACK_DISTR) || defined(USE_TILED_DISTR) || defined(USE_TILED_IRREG_DISTR)
#ifdef USE_SCALAPACK_DISTR
      call ga_set_block_cyclic_proc_grid(g_b,block_size,proc_grid)
#else
#ifdef USE_TILED_DISTR
      call ga_set_tiled_proc_grid(g_b,block_size,proc_grid)
#else
      call ga_set_tiled_irreg_proc_grid(g_b,mapc,nblocks,proc_grid)
#endif
#endif
#else
      call ga_set_block_cyclic(g_b,block_size)
#endif
#endif

#  ifdef MIRROR
      call ga_set_pgroup(g_b,p_mirror)
#  endif
#  ifdef USE_MEMDEV
      call ga_set_memory_dev(g_b,MEMDEVICE)
#  endif
      if (.not.ga_allocate(g_b)) then
#else
#  ifndef MIRROR
      if (.not. ga_create(MT_DBL, n, n, 'b', 0, 0, g_b)) then
#  else
      if (.not. nga_create_config(MT_DBL, ndim, dims, 'b', chunk,
     +                    p_mirror, g_b)) then
#  endif
#endif
         call ga_error('ga_create failed for second array ',-1)
      endif
c
      call ga_zero(g_b)
      call ga_acc(g_b, n/2, n/2, n/2, n/2, 1d0, 1, 1d0)
      call ga_sync()
#ifndef MIRROR
      if (me.eq.0) then
         call ga_get(g_b, n/2, n/2, n/2, n/2, b(1,1), 1)
         x = abs(b(1,1) -1d0*nproc)
#else
      if (iproc.eq.0) then
         call ga_get(g_b, n/2, n/2, n/2, n/2, b(1,1), 1)
         x = abs(b(1,1) -1d0*lprocs)
#endif
         if(x.gt. 1d-10)then
#ifndef MIRROR
            write(6,*)'val=',b(1,1),' expected=',nproc, x
#else
            write(6,*)'val=',b(1,1),' expected=',lprocs, x
#endif
            call ga_error('overlapping accumulate failed',-1)
         endif
         if (me.eq.0) then
           write(6,*)
           write(6,*) ' overlapping ga_acc is OK'
           write(6,*)
         endif
      endif
#endif
c
c     Check the ga_add function
c
      if (me .eq. 0) then
         write(6,91)
 91      format('> Checking add ...')
         call ffflush(6)
      endif
c
c
      crap = drand(12345)       ! Everyone has same seed
      do j = 1, n
         do i = 1, n
            b(i,j) = drand(0)
            a(i,j) = 0.1d0*a(i,j) + 0.9d0*b(i,j)
         enddo
      enddo
      
#ifndef MIRROR
      if (me.eq.0) call ga_put(g_b, 1, n, 1, n, b, n)
#else
      if (iproc.eq.0) call ga_put(g_b, 1, n, 1, n, b, n)
#endif
      call ga_add(0.1d0, g_a, 0.9d0, g_b, g_b)
      call ga_get(g_b, 1, n, 1, n, b, n)
      do j = 1, n
         do i = 1, n
            if(MISMATCH(b(i,j), a(i,j)))then
               write(6,*) ' add ', me, i, j, a(i,j), b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1) 
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' add is OK '
         write(6,*)
      endif
      call ga_sync()
c
c     Check the ddot function
c
      if (me .eq. 0) then
         write(6,19) 
 19      format('> Checking ddot ...')
         call ffflush(6)
      endif
      crap = drand(12345)       ! Everyone has same seed
      sum1 = 0.0d0
      do j = 1, n
         do i = 1, n
            b(i,j) = drand(0)
            sum1 = sum1 + a(i,j)*b(i,j)
         enddo
      enddo
#ifndef MIRROR
      if (me.eq.0) then 
#else
      if (iproc.eq.0) then 
#endif
         call ga_put(g_b, 1, n, 1, n, b, n)
         call ga_put(g_a, 1, n, 1, n, a, n)
      endif
      call ga_sync()
      sum2 = ga_ddot(g_a,g_b)
      if(MISMATCH(sum1, sum2))then
         write(6,*) ' ddot wrong ', sum1, sum2
         call ffflush(6)
         call ga_error('... exiting ',-1)
      else if (me.eq.0) then
         write(6,*)
         write(6,*) ' ddot is OK '
         write(6,*)
      endif
c
c     Check the ga_scale function
c
      if (me .eq. 0) then
         write(6,92)
 92      format('> Checking scale ...')
         call ffflush(6)
      endif
      call ga_scale(g_a, 0.123d0)
      call ga_get(g_a, 1, n, 1, n, b, n)
      do j = 1, n
         do i = 1, n
            a(i,j) = a(i,j)*0.123d0
            if (MISMATCH(b(i,j), a(i,j)))then
               write(6,*) ' dscal ', me, i, j, a(i,j), b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' scale is OK '
         write(6,*)
      endif
c
c     Check the ga_copy function
c
      if (me .eq. 0) then
         write(6,*)
         write(6,*)'> Checking copy'
         write(6,*)
         call ffflush(6)
      endif
      if(me.eq.0) call ga_put(g_a, 1, n, 1, n, a, n)
      call ga_copy(g_a, g_b)
      call ga_get(g_b, 1, n, 1, n, b, n)
      do j = 1, n
         do i = 1, n
            if (b(i,j) .ne. a(i,j)) then
               write(6,*) ' copy ', me, i, j, a(i,j), b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' copy is OK '
         write(6,*)
      endif
c
      call ga_sync()
      if (me .eq. 0) then
         write(6,*) '> Checking scatter/gather (might be slow)... '
         call ffflush(6)
      endif
      call ga_sync()
c
      crap = drand(ga_nodeid()*51 +1) !Different seed for each proc
      do j = 1, 10 
       call ga_sync()
#ifndef MIRROR
       itmp = iran(nproc)-1
       if(me.eq.itmp) then
#else
       itmp = iran(lprocs)-1
       if(me.eq.itmp) then
#endif
#ifndef NGA_GATSCAT
         do loop = 1,m 
           ilo = iran(n) 
           jlo = iran(n)
           iv(loop) = ilo
           jv(loop) = jlo
         enddo 
         call ga_gather(g_a, v, iv, jv, m)
         do loop = 1,m
           ilo= iv(loop)
           jlo= jv(loop)
           call ga_get(g_a,ilo,ilo,jlo,jlo,v(loop),1)
           if(v(loop)  .ne. a(ilo,jlo))then
             write(6,*)me,' gather ', ilo,',',jlo,',', a(ilo,jlo)
     &             ,' ',v(loop)
             call ffflush(6)
             call ga_error('... exiting ',-1)
           endif
         enddo
#else
         do loop = 1,m 
           ilo = iran(n) 
           jlo = iran(n)
           ijv(1,loop) = ilo
           ijv(2,loop) = jlo
         enddo 
         call nga_gather(g_a, v, ijv, m)
         do loop = 1,m
           ilo= ijv(1,loop)
           jlo= ijv(2,loop)
           call ga_get(g_a,ilo,ilo,jlo,jlo,v(loop),1)
           if(v(loop)  .ne. a(ilo,jlo))then
             write(6,*)me,' gather ', ilo,',',jlo,',', a(ilo,jlo)
     &             ,' ',v(loop)
             call ffflush(6)
             call ga_error('... exiting ',-1)
           endif
         enddo
#endif
       endif
      enddo
c
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' gather is  OK'
         write(6,*)
         call ffflush(6)
      endif
c
      do j = 1,10 
       call ga_sync()
#ifndef MIRROR
       if(me.eq.iran(ga_nnodes())-1) then
#else
       if(iproc.eq.iran(lprocs)-1) then
#endif
#ifndef NGA_GATSCAT
         do loop = 1,m
           ilo = iran(n)
           jlo = iran(n)
           iv(loop) = ilo
           jv(loop) = jlo
c          v(loop) = DSIN(a(ilo,jlo)+b(ilo,jlo)) 
           v(loop) = 1d0 *(ilo+jlo)
         enddo 
         call ga_scatter(g_a, v, iv, jv, m)
         do loop = 1,m
           ilo= iv(loop)
           jlo= jv(loop) 
           call ga_get(g_a,ilo,ilo,jlo,jlo,w(loop),1)
c          if(v(loop)  .ne. w(loop))then 
           if(w(loop)  .ne. 1d0 *(ilo+jlo) )then 
             write(6,*)me,' scatter ', ilo,',',jlo,',',w(loop) 
     &             ,' ', 1d0 *(ilo+jlo) 
             call ffflush(6)
             call ga_error('... exiting ',-1)
           endif
         enddo 
#else
         do loop = 1,m
           ilo = iran(n)
           jlo = iran(n)
           ijv(1,loop) = ilo
           ijv(2,loop) = jlo
c          v(loop) = DSIN(a(ilo,jlo)+b(ilo,jlo)) 
           v(loop) = 1d0 *(ilo+jlo)
         enddo 
         call nga_scatter(g_a, v, ijv, m)
         do loop = 1,m
           ilo= ijv(1,loop)
           jlo= ijv(2,loop) 
           call ga_get(g_a,ilo,ilo,jlo,jlo,w(loop),1)
c          if(v(loop)  .ne. w(loop))then 
           if(w(loop)  .ne. 1d0 *(ilo+jlo) )then 
             write(6,927)me,ilo,jlo,w(loop),1d0 *(ilo+jlo),loop
  927        format(i3,' scatter ',i5,',',i5,',',f7.1,',',f7.1,',',i5)
             call ffflush(6)
             call ga_error('... exiting ',-1)
           endif
         enddo 
#endif
       endif
       call ga_sync()
      enddo
c
      call ga_sync()
c
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' scatter is  OK'
         write(6,*)
      endif
c
      call ga_sync()
c
c     scatter-acc available in GA ver. 3.0
#ifdef GA3
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' checking scatter-accumulate'
         write(6,*)
      endif
c
      crap = drand(1234)
      call ga_zero(g_a)
c
      do j = 1, n
         do i = 1, n
            b(i,j) =0.
         enddo
      enddo
c
      x = .1d0
      ii =n
      do jj = 1,1 
        call ga_sync()
        do loop = 1, ii 

c         generate unique i,j pairs
10        continue
          i = iran(n)
          j=iran(n)
#ifndef NGA_GATSCAT
          if (found(i,j, iv, jv, loop-1) ) goto 10
          iv(loop) = i
          jv(loop) = j
#else
          if (nfound(i,j, ijv, loop-1) ) goto 10
          ijv(1,loop) = i
          ijv(2,loop) = j
#endif
          v(loop) = 1d0 *(i+j)
#ifndef MIRROR
          b(i,j) = b(i,j) + nproc*x*v(loop) ! update local ref. array
#else
          b(i,j) = b(i,j) + lprocs*x*v(loop) ! update local ref. array
#endif
        enddo


#ifndef NGA_GATSCAT
        call ga_scatter_acc(g_a,v,iv,jv, ii,x)
#else
        call nga_scatter_acc(g_a,v,ijv,ii,x)
#endif

c
        call ga_sync()
c     
c       check the result
c
        call ga_get(g_a, 1, n, 1,n, a, n)

        do loop = 1,ii
#ifndef NGA_GATSCAT
          i = iv(loop)
          j = jv(loop)
#else
          i = ijv(1,loop)
          j = ijv(2,loop)
#endif
          if(MISMATCH(a(i,j),b(i,j)))then
             print *,'Error',i,j,loop,a(i,j),'expected=',b(i,j)  
*            if(me.eq.0)then
*              do ii=1,loop
*                   print *,'element',ii, iv(ii),jv(ii)
*              enddo
*            endif
             call ga_error('scatter-acc error in trial ',jj) 
          endif
        enddo
        call ga_sync()

      enddo

      call ga_sync()
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' scatter-accumulate is  OK'
         write(6,*)
      endif
#endif

#ifdef MIRROR
c
c     Check the merge function
c
      do j = 1, n
         do i = 1, n
            a(i,j) = inode + i-1 + (j-1)*n
         enddo
      enddo
      if (me .eq. 0) then
         write(6,23) 
 23      format('> Checking merge ...')
         call ffflush(6)
      endif
      if (iproc.eq.0) then 
         call ga_put(g_a, 1, n, 1, n, a, n)
      endif
      call ga_merge_mirrored(g_a)
      call ga_get(g_a, 1, n, 1, n, b, n)
      lsame = .true.
      ilo = 0
      ihi = nnodes*(nnodes-1)/2
      do i = 1, n
        do j = 1, n
          if ((dble(nnodes)*(a(j,i)-dble(inode))
     +        + dble(ihi)).ne.b(j,i)) then
            lsame = .false.
            ilo = ilo + 1
          endif
        end do
      end do
      if(lsame.and.me.eq.0)then
         write(6,*)
         write(6,*) ' merge is OK '
         write(6,*)
      else if (.not.lsame) then
         write(6,*)
         write(6,*) ' merge is wrong ',ilo
         write(6,*)
         call ga_error('... exiting ',-1)
      endif
c
c    Checking copy between different array configurations
c
      if (me .eq. 0) then
         write(6,20) 
 20      format('> Checking mixed copy ...')
         call ffflush(6)
      endif
#ifdef NEW_API
      g_c = ga_create_handle()
      call ga_set_data(g_c,ndim,dims,MT_DBL)
      call ga_set_array_name(g_c,'c')

#ifdef BLOCK_CYCLIC
#if defined(USE_SCALAPACK_DISTR) || defined(USE_TILED_DISTR) || defined(USE_TILED_IRREG_DISTR)
#ifdef USE_SCALAPACK_DISTR
      call ga_set_block_cyclic_proc_grid(g_c,block_size,proc_grid)
#else
#ifdef USE_TILE_DISTR
      call ga_set_tiled_proc_grid(g_c,block_size,proc_grid)
#else
      call ga_set_tiled_irreg_proc_grid(g_c,mapc,nblocks,proc_grid)
#endif
#endif
#else
      call ga_set_block_cyclic(g_c,block_size)
#endif
#endif

#  ifdef USE_MEMDEV
      call ga_set_memory_dev(g_c,MEMDEVICE)
#  endif
      if (.not.ga_allocate(g_c)) then
#else
      if (.not. nga_create(MT_DBL, ndim, dims, 'c', chunk, g_c)) then
#endif
         call ga_error('ga_create failed for third array ',-1)
      endif
      call ga_zero(g_c)
      call ga_copy(g_a,g_c)
      call ga_get(g_a, 1, n, 1, n, a, n)
      call ga_get(g_c, 1, n, 1, n, b, n)
      do i = 1, n
        do j = 1, n
          if (a(j,i).ne.b(j,i)) then
            lsame = .false.
          endif
        end do
      end do
      if(lsame.and.me.eq.0)then
         write(6,*)
         write(6,*) ' Mixed copy (mirrored to normal) is OK '
         write(6,*)
      elseif (me.eq.0) then
         write(6,*)
         write(6,*) ' Mixed copy (mirrored to normal) is wrong ',ilo
         write(6,*)
         call ga_error('... exiting ',-1)
      endif
      call ga_zero(g_a)
      call ga_copy(g_c,g_a)
      call ga_get(g_a, 1, n, 1, n, a, n)
      do i = 1, n
        do j = 1, n
          if (a(j,i).ne.b(j,i)) then
            lsame = .false.
          endif
        end do
      end do
      if(lsame.and.me.eq.0)then
         write(6,*)
         write(6,*) ' Mixed copy (normal to mirrored) is OK '
         write(6,*)
      elseif (me.eq.0) then
         write(6,*)
         write(6,*) ' Mixed copy (normal to mirrored) is wrong ',ilo
         write(6,*)
         call ga_error('... exiting ',-1)
      endif
c
c   Check the merge to distributed patch function
c
      do j = 1, n
         do i = 1, n
            a(i,j) = inode + i-1 + (j-1)*n
         enddo
      enddo
      if (me .eq. 0) then
         write(6,24) 
 24      format('> Checking merge to distributed patch ...')
         call ffflush(6)
      endif
c
c    get low and high indices of patch
c
      ilo = n/4
      ihi = ilo + n/2
      do i = 1, 2
       alo(i) = ilo
       blo(i) = ilo + 1
       ahi(i) = ihi
       bhi(i) = ihi + 1
      end do
      if (iproc.eq.0) then 
         call ga_put(g_a, 1, n, 1, n, a, n)
      endif
      call nga_merge_distr_patch(g_a, alo, ahi, g_c, blo, bhi)
      call ga_get(g_c, 1, n, 1, n, b, n)
      lsame = .true.
      ilo = 0
      ihi = nnodes*(nnodes-1)/2
      do i = alo(1), ahi(1)
        ii = i-alo(1) + blo(1)
        do j = alo(2), ahi(2)
          jj = j-alo(2) + blo(2)
          if ((dble(nnodes)*(a(j,i)-dble(inode))
     +        + dble(ihi)).ne.b(jj,ii)) then
c            if (me.eq.0) then
c              write(6,*) i,j,ii,jj, nnodes*a(j,i)+ihi, b(jj,ii)
c            endif
            lsame = .false.
            ilo = ilo + 1
          endif
        end do
      end do
      if(lsame.and.me.eq.0)then
         write(6,*)
         write(6,*) ' merge to distributed patch is OK '
         write(6,*)
      elseif (.not.lsame) then
         write(6,*)
         write(6,*) ' merge to distributed patch is wrong ',ilo
         write(6,*)
         call ga_error('... exiting ',-1)
      endif
#endif
c
c     Delete the global arrays
c
      status = ga_destroy(g_b)
      status = ga_destroy(g_a)
c     
      end

c-----------------------------------------------------------------

      subroutine check_complex_float()
      implicit none
#include "mafdecls.fh"
#include "global.fh"
#include "testutil.fh"
c
      integer n,m
      parameter (n = 60)
      parameter (m = 2*n)
      complex a(n,n), b(n,n), v(m),w(m)
#ifdef MIRROR
      integer ndim, dims(2), chunk(2), p_mirror
#else
#  ifdef NEW_API
      integer ndim, dims(2), chunk(2), p_mirror
#  endif
#endif
      integer iv(m), jv(m)
      logical status
      integer g_a, g_b
      integer iran, i,j, loop,nloop,maxloop, ilo, ihi, jlo, jhi, itmp
      integer nproc, me, int, ij, inc, ii, jj, nnodes
      parameter (maxloop = 100)
      integer maxproc
      parameter (maxproc = 4096)
      double precision crap, real
      double precision nwords
      complex   x, sum1, sum2, factor
      integer lprocs, inode, iproc, lproc
#ifdef USE_RESTRICTED
      integer num_rstrctd
      integer rstrctd_list(maxproc/2)
#endif
      intrinsic int
      iran(i) = int(drand(0)*real(i)) + 1
c
      nproc = ga_nnodes()
      me    = ga_nodeid()
      inode = ga_cluster_nodeid()
      lprocs = ga_cluster_nprocs(inode)
      nnodes = ga_cluster_nnodes()
      iproc = mod(me,lprocs)
      nloop = Min(maxloop,n)
#ifdef USE_RESTRICTED
      num_rstrctd = nproc/2
      if (num_rstrctd.eq.0) num_rstrctd = 1
      do i = 1, num_rstrctd
        rstrctd_list(i) = (num_rstrctd/2) + i-1
      end do
#endif
c
c     a() is a local copy of what the global array should start as
c
      do j = 1, n
         do i = 1, n
#ifndef MIRROR
            a(i,j) = cmplx(real(i-1), real((j-1)*n))
#else
            a(i,j) = cmplx(real(inode),0.0d00)
     +             + cmplx(real(i-1), real((j-1)*n))
#endif
            b(i,j) = cmplx(-1d0,1d0)
         enddo
      enddo
c
c     Create a global array
c
c     print *,ga_nodeid(), ' creating array'
      call ffflush(6)
c     call setdbg(1)
#ifdef NEW_API
      ndim = 2
      dims(1) = n
      dims(2) = n
      g_a = ga_create_handle()
      call ga_set_data(g_a,ndim,dims,MT_SCPL)
      call ga_set_array_name(g_a,'a')
#ifdef USE_RESTRICTED
      call ga_set_restricted(g_a, rstrctd_list, num_rstrctd)
#endif
#  ifdef MIRROR
      p_mirror = ga_pgroup_get_mirror()
      call ga_set_pgroup(g_a,p_mirror)
#  endif
#  ifdef USE_MEMDEV
      call ga_set_memory_dev(g_a,MEMDEVICE)
#  endif
      status = ga_allocate(g_a)
#else
#  ifndef MIRROR
      status = ga_create(MT_SCPL, n, n, 'a', 0, 0, g_a)
#  else
      ndim = 2
      dims(1) = n
      dims(2) = n
      chunk(1) = 0
      chunk(2) = 0
      p_mirror = ga_pgroup_get_mirror()
      status = nga_create_config(MT_SCPL, ndim, dims, 'a', chunk,
     +                           p_mirror, g_a)
#  endif
#endif
      if (.not. status) then
         write(6,*) ' ga_create failed'
         call ga_error('... exiting ',-1)
      endif
#ifdef NEW_API
      g_b = ga_create_handle()
      call ga_set_data(g_b,ndim,dims,MT_SCPL)
      call ga_set_array_name(g_b,'b')
#  ifdef MIRROR
      call ga_set_pgroup(g_b,p_mirror)
#  endif
#  ifdef USE_MEMDEV
      call ga_set_memory_dev(g_b,MEMDEVICE)
#  endif
      if (.not.ga_allocate(g_b)) then
#else
#  ifndef MIRROR
      if (.not. ga_create(MT_SCPL, n, n, 'b', 0, 0, g_b)) then
#  else
      if (.not. nga_create_config(MT_SCPL, ndim, dims, 'b', chunk,
     _                            p_mirror, g_b)) then
#  endif
#endif
         call ga_error('ga_create failed for second array ',-1)
      endif

#ifndef MIRROR
      call ga_distribution(g_a,me,ilo, ihi, jlo, jhi)
#else
      lproc = me - ga_cluster_procid(inode,0)
      call ga_distribution(g_a, lproc, ilo, ihi, jlo, jhi)
#endif
      call ga_sync()
c
c     Zero the array
c
      if (me .eq. 0) then
         write(6,21)
 21      format(/'> Checking zero ... ')
         call ffflush(6)
      endif
      call ga_zero(g_a)
c
c     Check that it is indeed zero
c
      call ga_get(g_a, 1, n, 1, n, b, n)
      call ga_sync()
      do i = 1, n
         do j = 1, n
            if(b(i,j).ne.(0d0,0d0)) then
               write(6,*) me,' zero ', i, j, b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_zero is OK'
         write(6,*)
      endif
      call ga_sync()
c
c     Each node fills in disjoint sections of the array
c
      if (me .eq. 0) then
         write(6,2)
 2       format(/'> Checking disjoint put ... ')
         call ffflush(6)
      endif
      call ga_sync()
c
      inc = (n-1)/4 + 1
      ij = 0
      do j = 1, n, inc
         do i = 1, n, inc
#ifndef MIRROR
            if (mod(ij,nproc) .eq. me) then
#else
            if (mod(ij,lprocs) .eq. iproc) then
#endif
               ilo = i
               ihi = min(i+inc, n)
               jlo = j
               jhi = min(j+inc, n)
               call ga_put(g_a, ilo, ihi, jlo, jhi, a(ilo, jlo), n)
            endif
            ij = ij + 1
         enddo
      enddo
      call ga_sync()
c
c     All nodes check all of a
c
      call ga_get(g_a, 1, n, 1, n, b, n)
c
      do i = 1, n
         do j = 1, n
            if (b(i,j) .ne. a(i,j)) then
               write(6,*) ' put ', me, i, j, a(i,j),b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_put is OK'
         write(6,*)
      endif
      call ga_sync()
c
c     Now check nloop random gets from each node
c
      if (me .eq. 0) then
         write(6,5) nloop
 5       format(/'> Checking random get (',i5,' calls)...')
         call ffflush(6)
      endif
      call ga_sync()
c
      nwords = 0
c
      crap = drand(ga_nodeid()*51 +1) !Different seed for each proc
      do loop = 1, nloop
         ilo = iran(loop)
         ihi = iran(loop)
         if (ihi.lt. ilo) then
            itmp = ihi
            ihi = ilo
            ilo = itmp
         endif
         jlo = iran(loop)
         jhi = iran(loop)
         if (jhi.lt. jlo) then
            itmp = jhi
            jhi = jlo
            jlo = itmp
         endif
c
         nwords = nwords + (ihi-ilo+1)*(jhi-jlo+1)
c
         call ga_get(g_a, ilo, ihi, jlo, jhi, b(ilo, jlo), n)
         if (me .eq. 0 .and. mod(loop-1, max(1,nloop/20)).eq.0) then
            write(6,1) loop, me, ilo, ihi, jlo, jhi, nwords
 1          format(' call ',i5, ' node ',i2,' checking get ',4i4,
     $           ' total ',d9.2)
            call ffflush(6)
         endif
         do j = jlo, jhi
            do i = ilo, ihi
               if (b(i,j) .ne. a(i,j)) then
                  write(6,*)'error:', i, j, b(i,j), a(i,j)
                  call ga_error('... exiting ',-1)
               endif
            enddo
         enddo
c
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_get is OK'
         write(6,*)
         call ffflush(6)
      endif
      call ga_sync()
c
c     Each node accumulates into disjoint sections of the array
c
      if (me .eq. 0) then
         write(6,9)
 9       format(/'> Checking accumulate ... ')
         call ffflush(6)
      endif
      call ga_sync()
c
      crap = drand(12345)       ! Same seed for each process
      do j = 1, n
         do i = 1, n
            b(i,j) = cmplx(drand(0),drand(1))
            b(i,j) = cmplx(real(i),real(j))
         enddo
      enddo
c
      inc = (n-1)/20 + 1
      ij = 0
      do j = 1, n, inc
         do i = 1, n, inc
c           x = cmplx(drand(0),0.333d0)
c           x = cmplx(0.333d0,0)
*           x = cmplx(0d0,0d0)
            x = 0 
            ilo = i
            ihi = min(i+inc-1, n)
            if(ihi.eq.n-1)ihi=n
            jlo = j
            jhi = min(j+inc-1, n)
            if(jhi.eq.n-1)jhi=n
#ifndef MIRROR
            if (mod(ij,nproc) .eq. me) then
#else
            if (mod(ij,lprocs) .eq. iproc) then
#endif
               call ga_acc(g_a, ilo, ihi, jlo, jhi, b(ilo, jlo), n, x)
            endif
            ij = ij + 1
c
c           Each process applies all updates to its local copy
c
            do jj = jlo, jhi
               do ii = ilo, ihi
                  a(ii,jj) = a(ii,jj) + x * b(ii,jj)
               enddo
            enddo
         enddo
      enddo
      call ga_sync()
c
c     All nodes check all of a
c
      call ga_get(g_a, 1, n, 1, n, b, n)
      do j = 1, n
         do i = 1, n
            if (MISMATCHF(b(i,j), a(i,j)))then
               write(6,*) ' acc ', me, i, j, a(i,j), b(i,j)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' disjoint ga_acc is OK'
         write(6,*)
      endif
c
c     overlapping accumulate
c
      call ga_zero(g_b)
      call ga_acc(g_b, n/2, n/2, n/2, n/2, (1,-1), 1, (1,0))
      call ga_sync()
#ifndef MIRROR
      if (me.eq.0) then
         call ga_get(g_b, n/2, n/2, n/2, n/2, x, 1)
         if (MISMATCHF(x, ((1d0,-1d0)*nproc)))then
#else
      if (iproc.eq.0) then
         call ga_get(g_b, n/2, n/2, n/2, n/2, x, 1)
         if (MISMATCHF(x, ((1d0,-1d0)*lprocs)))then
#endif
c        if(error.gt. (1d-8))then
#ifndef MIRROR
            write(6,*)'val=',x,' expected=(',nproc,',',-nproc,')'
#else
            write(6,*)'val=',x,' expected=(',lprocs,',',-lprocs,')'
#endif
            call ga_error('overlapping accumulate failed',-1)
         endif
         write(6,*)
         write(6,*) ' overlapping ga_acc is OK'
         write(6,*)
      endif
c
c     Check the ga_copy function
c
      if (me .eq. 0) then
         write(6,*)
         write(6,*)'> Checking copy'
         write(6,*)
         call ffflush(6)
      endif
      call ga_sync()
#ifndef MIRROR
      if(me.eq.0) call ga_put(g_a, 1, n, 1, n, a, n)
#else
      if(iproc.eq.0) call ga_put(g_a, 1, n, 1, n, a, n)
#endif
      call ga_copy(g_a, g_b)
      call ga_get(g_b, 1, n, 1, n, b, n)
      do j = 1, n
         do i = 1, n
            if (b(i,j) .ne. a(i,j)) then
               write(6,*) ' copy ', me, i, j, a(i,j), b(i,j)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' copy is OK '
         write(6,*)
      endif
c
c
c     Check the ga_scale function
c
      if (me .eq. 0) then
         write(6,92)
 92      format(/'> Checking scale ...')
         call ffflush(6)
      endif
      factor = (1d0,-1d0)
      call ga_scale(g_a, factor)
      call ga_get(g_a, 1, n, 1, n, b, n)
      do j = 1, n
         do i = 1, n
            a(i,j) = a(i,j)*factor
            if (MISMATCHF(b(i,j), a(i,j)))then
               write(6,*) ' fscal ', me, i, j, a(i,j), b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' scale is OK '
         write(6,*)
      endif
c
c     Check scatter&gather
c
      call ga_sync()
      if (me .eq. 0) then
         write(6,*) '> Checking scatter/gather (might be slow)... '
         call ffflush(6)
         if(me.eq.0) call ga_put(g_a, 1, n, 1, n, a, n)
      endif
      call ga_sync()
c
      crap = drand(ga_nodeid()*51 +1) !Different seed for each proc
      do j = 1, 10
       call ga_sync()
#ifndef MIRROR
       itmp = iran(nproc)-1
       if(me.eq.itmp) then
#else
       itmp = iran(lprocs)-1
       if(iproc.eq.itmp) then
#endif
         do loop = 1,m
           ilo = iran(n)
           jlo = iran(n)
           iv(loop) = ilo
           jv(loop) = jlo
         enddo
         call ga_gather(g_a, v, iv, jv, m)
         do loop = 1,m
           ilo= iv(loop)
           jlo= jv(loop)
           call ga_get(g_a,ilo,ilo,jlo,jlo,v(loop),1)
           if(v(loop)  .ne. a(ilo,jlo))then
             write(6,*)me,' gather ', ilo,',',jlo,',', a(ilo,jlo)
     &             ,' ',v(loop)
             call ffflush(6)
             call ga_error('... exiting ',-1)
           endif
         enddo
       endif
      enddo
c
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' gather is  OK'
         write(6,*)
         call ffflush(6)
      endif
c
      do j = 1,10
       call ga_sync()
#ifndef MIRROR
       if(me.eq.iran(ga_nnodes())-1) then
#else
       if(me.eq.iran(lprocs)-1) then
#endif
         do loop = 1,m
           ilo = iran(n)
           jlo = iran(n)
           iv(loop) = ilo
           jv(loop) = jlo
           v(loop) = (1d0,-1d0) *(ilo+jlo)
         enddo
         call ga_scatter(g_a, v, iv, jv, m)
         do loop = 1,m
           ilo= iv(loop)
           jlo= jv(loop)
           call ga_get(g_a,ilo,ilo,jlo,jlo,w(loop),1)
           if(w(loop)  .ne. (1d0,-1d0) *(ilo+jlo) )then
             write(6,*)me,' scatter ', ilo,',',jlo,',',w(loop)
     &             ,' ', (1d0,-1d0) *(ilo+jlo)
             call ffflush(6)
           endif
         enddo
       endif
       call ga_sync()
      enddo
c
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' scatter is  OK'
         write(6,*)
      endif
c
c     Check ga_add
c
      if (me .eq. 0) then
         write(6,91)
 91      format(/'> Checking add ...')
         call ffflush(6)
      endif
      call ga_get(g_a, 1, n, 1, n, a, n)
      crap = drand(12345)       ! Everyone has same seed
      do j = 1, n
         do i = 1, n
            b(i,j) = cmplx(drand(0), drand(1))
            a(i,j) = (0.1d0,-.1d0)*a(i,j) + (.9d0,-.9d0)*b(i,j)
         enddo
      enddo

#ifndef MIRROR
      if (me.eq.0) call ga_put(g_b, 1, n, 1, n, b, n)
#else
      if (iproc.eq.0) call ga_put(g_b, 1, n, 1, n, b, n)
#endif
      call ga_add((0.1,-.1), g_a, (0.9,-.9), g_b, g_b)
      call ga_get(g_b, 1, n, 1, n, b, n)
      do j = 1, n
         do i = 1, n
            if (MISMATCHF(b(i,j), a(i,j)))then
               write(6,*) ' add ', me, i, j, a(i,j), b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' add is OK '
         write(6,*)
      endif
      call ga_sync()
c
c     Check the cdot function
c
      if (me .eq. 0) then
         write(6,19)
 19      format(/'> Checking cdot ...')
         call ffflush(6)
      endif
      crap = drand(12345)       ! Everyone has same seed
      sum1 = (0.0d0,0.d0)
      do j = 1, n
         do i = 1, n
            b(i,j) = cmplx(drand(0), drand(1))
            sum1 = sum1 + a(i,j)*b(i,j)
         enddo
      enddo
#ifndef MIRROR
      if (me.eq.0) then
#else
      if (iproc.eq.0) then
#endif
         call ga_put(g_b, 1, n, 1, n, b, n)
         call ga_put(g_a, 1, n, 1, n, a, n)
      endif
      call ga_sync()
      sum2 =  ga_cdot(g_a,g_b)
      if (MISMATCHF(sum1, sum2))then
         write(6,*) ' cdot wrong ', sum1, sum2
         call ffflush(6)
         call ga_error('... exiting ',-1)
      else if (me.eq.0) then
         write(6,*)
         write(6,*) ' cdot is OK '
         write(6,*)
      endif
c
c     Delete the global arrays
c

      status = ga_destroy(g_b)
      status = ga_destroy(g_a)

c
      end
c-----------------------------------------------------------------

      subroutine check_complex()
      implicit none
#include "mafdecls.fh"
#include "global.fh"
#include "testutil.fh"
c
      integer n,m
      parameter (n = 60)
      parameter (m = 2*n)
      double complex a(n,n), b(n,n), v(m),w(m)
#ifdef MIRROR
      integer ndim, dims(2), chunk(2), p_mirror
#else
#  ifdef NEW_API
      integer ndim, dims(2), chunk(2), p_mirror
#  endif
#endif
      integer iv(m), jv(m)
      logical status
      integer g_a, g_b
      integer iran, i,j, loop,nloop,maxloop, ilo, ihi, jlo, jhi, itmp
      integer nproc, me, int, ij, inc, ii, jj, nnodes
      parameter (maxloop = 100)
      integer maxproc
      parameter (maxproc = 4096)
      double precision crap, real
      double precision nwords
      double complex   x, sum1, sum2, factor
      integer lprocs, inode, iproc, lproc
#ifdef USE_RESTRICTED
      integer num_rstrctd
      integer rstrctd_list(maxproc/2)
#endif
#ifdef BLOCK_CYCLIC
      integer block_size(2), proc_grid(2)
#endif
#ifdef USE_TILED_IRREG_DISTR
      integer mapc(4096), nblocks(2)
#endif
      intrinsic int
      iran(i) = int(drand(0)*real(i)) + 1
c
      nproc = ga_nnodes()
      me    = ga_nodeid()
      inode = ga_cluster_nodeid()
      lprocs = ga_cluster_nprocs(inode)
      nnodes = ga_cluster_nnodes()
      iproc = mod(me,lprocs)
      nloop = Min(maxloop,n)
#ifdef USE_RESTRICTED
      num_rstrctd = nproc/2
      if (num_rstrctd.eq.0) num_rstrctd = 1
      do i = 1, num_rstrctd
        rstrctd_list(i) = (num_rstrctd/2) + i-1
      end do
#endif
#ifdef BLOCK_CYCLIC
      block_size(1) = 32
      block_size(2) = 32
#if defined(USE_SCALAPACK_DISTR) || defined(USE_TILED_DISTR)
      if (mod(nproc,2).ne.0)
     +  call ga_error("Available procs must be divisible by 2",-1)
      proc_grid(1) = 2
      proc_grid(2) = nproc/2
#endif
#endif
c
c     a() is a local copy of what the global array should start as
c
      do j = 1, n
         do i = 1, n
#ifndef MIRROR
            a(i,j) = cmplx(dble(i-1), dble((j-1)*n))
#else
            a(i,j) = cmplx(dble(inode),0.0d00)
     +             + cmplx(dble(i-1), dble((j-1)*n))
#endif
            b(i,j) = cmplx(-1d0,1d0)
         enddo
      enddo
c
c     Create a global array
c
c     print *,ga_nodeid(), ' creating array'
      call ffflush(6)
c     call setdbg(1)
#ifdef NEW_API
      ndim = 2
      dims(1) = n
      dims(2) = n
      g_a = ga_create_handle()
      call ga_set_data(g_a,ndim,dims,MT_DCPL)
      call ga_set_array_name(g_a,'a')
#ifdef USE_RESTRICTED
      call ga_set_restricted(g_a, rstrctd_list, num_rstrctd)
#endif
#ifdef BLOCK_CYCLIC
#if defined(USE_SCALAPACK_DISTR) || defined(USE_TILED_DISTR) || defined(USE_TILED_IRREG_DISTR)
#ifdef USE_SCALAPACK_DISTR
      call ga_set_block_cyclic_proc_grid(g_a,block_size,proc_grid)
#else
#ifdef USE_TILED_DISTR
      call ga_set_tiled_proc_grid(g_a,block_size,proc_grid)
#else
      call util_create_map(mapc,block_size(1),dims,nblocks,proc_grid)
      call ga_set_tiled_irreg_proc_grid(g_a,mapc,nblocks,proc_grid)
#endif
#endif
#else
      call ga_set_block_cyclic(g_a,block_size)
#endif
#endif
#  ifdef MIRROR
      p_mirror = ga_pgroup_get_mirror()
      call ga_set_pgroup(g_a,p_mirror)
#  endif
#  ifdef USE_MEMDEV
      call ga_set_memory_dev(g_a,MEMDEVICE)
#  endif
      status = ga_allocate(g_a)
#else
#  ifndef MIRROR
      status = ga_create(MT_DCPL, n, n, 'a', 0, 0, g_a)
#  else
      ndim = 2
      dims(1) = n
      dims(2) = n
      chunk(1) = 0
      chunk(2) = 0
      p_mirror = ga_pgroup_get_mirror()
      status = nga_create_config(MT_DCPL, ndim, dims, 'a', chunk,
     +                           p_mirror, g_a)
#  endif
#endif
      if (.not. status) then
         write(6,*) ' ga_create failed'
         call ga_error('... exiting ',-1)
      endif
#ifdef NEW_API
      g_b = ga_create_handle()
      call ga_set_data(g_b,ndim,dims,MT_DCPL)
      call ga_set_array_name(g_b,'b')
#ifdef BLOCK_CYCLIC
#if defined(USE_SCALAPACK_DISTR) || defined(USE_TILED_DISTR) || defined(USE_TILED_IRREG_DISTR)
#ifdef USE_SCALAPACK_DISTR
      call ga_set_block_cyclic_proc_grid(g_b,block_size,proc_grid)
#else
#ifdef USE_TILED_DISTR
      call ga_set_tiled_proc_grid(g_b,block_size,proc_grid)
#else
      call ga_set_tiled_irreg_proc_grid(g_b,mapc,nblocks,proc_grid)
#endif
#endif
#else
      call ga_set_block_cyclic(g_b,block_size)
#endif
#endif
#  ifdef MIRROR
      call ga_set_pgroup(g_b,p_mirror)
#  endif
#  ifdef USE_MEMDEV
      call ga_set_memory_dev(g_b,MEMDEVICE)
#  endif
      if (.not.ga_allocate(g_b)) then
#else
#  ifndef MIRROR
      if (.not. ga_create(MT_DCPL, n, n, 'b', 0, 0, g_b)) then
#  else
      if (.not. nga_create_config(MT_DCPL, ndim, dims, 'b', chunk,
     _                            p_mirror, g_b)) then
#  endif
#endif
         call ga_error('ga_create failed for second array ',-1)
      endif

#ifndef MIRROR
      call ga_distribution(g_a,me,ilo, ihi, jlo, jhi)
#else
      lproc = me - ga_cluster_procid(inode,0)
      call ga_distribution(g_a, lproc, ilo, ihi, jlo, jhi)
#endif
      call ga_sync()
c
c     Zero the array
c
      if (me .eq. 0) then
         write(6,21)
 21      format('> Checking zero ... ')
         call ffflush(6)
      endif
      call ga_zero(g_a)
c
c     Check that it is indeed zero
c
      call ga_get(g_a, 1, n, 1, n, b, n)
      call ga_sync()
      do i = 1, n
         do j = 1, n
            if(b(i,j).ne.(0d0,0d0)) then
               write(6,*) me,' zero ', i, j, b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_zero is OK'
         write(6,*)
      endif
      call ga_sync()
c
c     Each node fills in disjoint sections of the array
c
      if (me .eq. 0) then
         write(6,2)
 2       format('> Checking disjoint put ... ')
         call ffflush(6)
      endif
      call ga_sync()
c
      inc = (n-1)/20 + 1
      ij = 0
      do j = 1, n, inc
         do i = 1, n, inc
#ifndef MIRROR
            if (mod(ij,nproc) .eq. me) then
#else
            if (mod(ij,lprocs) .eq. iproc) then
#endif
               ilo = i
               ihi = min(i+inc, n)
               jlo = j
               jhi = min(j+inc, n)
               call ga_put(g_a, ilo, ihi, jlo, jhi, a(ilo, jlo), n)
            endif
            ij = ij + 1
         enddo
      enddo
      call ga_sync()
c
c     All nodes check all of a
c
      call util_qfill(n*n, (0d0,0d0), b, 1)
      call ga_get(g_a, 1, n, 1, n, b, n)
c
      do i = 1, n
         do j = 1, n
            if (b(i,j) .ne. a(i,j)) then
               write(6,*) ' put ', me, i, j, a(i,j),b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_put is OK'
         write(6,*)
      endif
      call ga_sync()
c
c     Now check nloop random gets from each node
c
      if (me .eq. 0) then
         write(6,5) nloop
 5       format('> Checking random get (',i5,' calls)...')
         call ffflush(6)
      endif
      call ga_sync()
c
      nwords = 0
c
      crap = drand(ga_nodeid()*51 +1) !Different seed for each proc
      do loop = 1, nloop
         ilo = iran(loop)
         ihi = iran(loop)
         if (ihi.lt. ilo) then
            itmp = ihi
            ihi = ilo
            ilo = itmp
         endif
         jlo = iran(loop)
         jhi = iran(loop)
         if (jhi.lt. jlo) then
            itmp = jhi
            jhi = jlo
            jlo = itmp
         endif
c
         nwords = nwords + (ihi-ilo+1)*(jhi-jlo+1)
c
         call util_qfill(n*n, (0.0d0,0d0), b, 1)
         call ga_get(g_a, ilo, ihi, jlo, jhi, b(ilo, jlo), n)
         if (me .eq. 0 .and. mod(loop-1, max(1,nloop/20)).eq.0) then
            write(6,1) loop, me, ilo, ihi, jlo, jhi, nwords
 1          format(' call ',i5, ' node ',i2,' checking get ',4i4,
     $           ' total ',d9.2)
            call ffflush(6)
         endif
         do j = jlo, jhi
            do i = ilo, ihi
               if (b(i,j) .ne. a(i,j)) then
                  write(6,*)'error:', i, j, b(i,j), a(i,j)
                  call ga_error('... exiting ',-1)
               endif
            enddo
         enddo
c
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_get is OK'
         write(6,*)
         call ffflush(6)
      endif
      call ga_sync()
c
c     Each node accumulates into disjoint sections of the array
c
      if (me .eq. 0) then
         write(6,9)
 9       format('> Checking accumulate ... ')
         call ffflush(6)
      endif
      call ga_sync()
c
      crap = drand(12345)       ! Same seed for each process
      do j = 1, n
         do i = 1, n
            b(i,j) = cmplx(drand(0),drand(1))
            b(i,j) = cmplx(dble(i),dble(j))
         enddo
      enddo
c
      inc = (n-1)/20 + 1
      ij = 0
      do j = 1, n, inc
         do i = 1, n, inc
c           x = cmplx(drand(0),0.333d0)
c           x = cmplx(0.333d0,0)
*           x = cmplx(0d0,0d0)
            x = 0 
            ilo = i
            ihi = min(i+inc-1, n)
            if(ihi.eq.n-1)ihi=n
            jlo = j
            jhi = min(j+inc-1, n)
            if(jhi.eq.n-1)jhi=n
#ifndef MIRROR
            if (mod(ij,nproc) .eq. me) then
#else
            if (mod(ij,lprocs) .eq. iproc) then
#endif
               call ga_acc(g_a, ilo, ihi, jlo, jhi, b(ilo, jlo), n, x)
            endif
            ij = ij + 1
c
c           Each process applies all updates to its local copy
c
            do jj = jlo, jhi
               do ii = ilo, ihi
                  a(ii,jj) = a(ii,jj) + x * b(ii,jj)
               enddo
            enddo
         enddo
      enddo
      call ga_sync()
c
c     All nodes check all of a
c
      call ga_get(g_a, 1, n, 1, n, b, n)
      do j = 1, n
         do i = 1, n
            if (MISMATCH(b(i,j), a(i,j)))then
               write(6,*) ' acc ', me, i, j, a(i,j), b(i,j)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' disjoint ga_acc is OK'
         write(6,*)
      endif
c
c     overlapping accumulate
c
      call ga_zero(g_b)
      call ga_acc(g_b, n/2, n/2, n/2, n/2, (1d0,-1d0), 1, (1d0,0d0))
      call ga_sync()
#ifndef MIRROR
      if (me.eq.0) then
         call ga_get(g_b, n/2, n/2, n/2, n/2, x, 1)
         if (MISMATCH(x, ((1d0,-1d0)*nproc)))then
#else
      if (iproc.eq.0) then
         call ga_get(g_b, n/2, n/2, n/2, n/2, x, 1)
         if (MISMATCH(x, ((1d0,-1d0)*lprocs)))then
#endif
c        if(error.gt. (1d-8))then
#ifndef MIRROR
            write(6,*)'val=',x,' expected=(',nproc,',',-nproc,')'
#else
            write(6,*)'val=',x,' expected=(',lprocs,',',-lprocs,')'
#endif
            call ga_error('overlapping accumulate failed',-1)
         endif
         write(6,*)
         write(6,*) ' overlapping ga_acc is OK'
         write(6,*)
      endif
c
c     Check the ga_copy function
c
      if (me .eq. 0) then
         write(6,*)
         write(6,*)'> Checking copy'
         write(6,*)
         call ffflush(6)
      endif
      call ga_sync()
#ifndef MIRROR
      if(me.eq.0) call ga_put(g_a, 1, n, 1, n, a, n)
#else
      if(iproc.eq.0) call ga_put(g_a, 1, n, 1, n, a, n)
#endif
      call ga_copy(g_a, g_b)
      call ga_get(g_b, 1, n, 1, n, b, n)
      do j = 1, n
         do i = 1, n
            if (b(i,j) .ne. a(i,j)) then
               write(6,*) ' copy ', me, i, j, a(i,j), b(i,j)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' copy is OK '
         write(6,*)
      endif
c
c
c     Check the ga_scale function
c
      if (me .eq. 0) then
         write(6,92)
 92      format('> Checking scale ...')
         call ffflush(6)
      endif
      factor = (1d0,-1d0)
      call ga_scale(g_a, factor)
      call ga_get(g_a, 1, n, 1, n, b, n)
      do j = 1, n
         do i = 1, n
            a(i,j) = a(i,j)*factor
            if (MISMATCH(b(i,j), a(i,j)))then
               write(6,*) ' dscal ', me, i, j, a(i,j), b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' scale is OK '
         write(6,*)
      endif
c
c     Check scatter&gather
c
      call ga_sync()
      if (me .eq. 0) then
         write(6,*) '> Checking scatter/gather (might be slow)... '
         call ffflush(6)
         if(me.eq.0) call ga_put(g_a, 1, n, 1, n, a, n)
      endif
      call ga_sync()
c
      crap = drand(ga_nodeid()*51 +1) !Different seed for each proc
      do j = 1, 10
       call ga_sync()
#ifndef MIRROR
       itmp = iran(nproc)-1
       if(me.eq.itmp) then
#else
       itmp = iran(lprocs)-1
       if(iproc.eq.itmp) then
#endif
         do loop = 1,m
           ilo = iran(n)
           jlo = iran(n)
           iv(loop) = ilo
           jv(loop) = jlo
         enddo
         call ga_gather(g_a, v, iv, jv, m)
         do loop = 1,m
           ilo= iv(loop)
           jlo= jv(loop)
           call ga_get(g_a,ilo,ilo,jlo,jlo,v(loop),1)
           if(v(loop)  .ne. a(ilo,jlo))then
             write(6,*)me,' gather ', ilo,',',jlo,',', a(ilo,jlo)
     &             ,' ',v(loop)
             call ffflush(6)
             call ga_error('... exiting ',-1)
           endif
         enddo
       endif
      enddo
c
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' gather is  OK'
         write(6,*)
         call ffflush(6)
      endif
c
      do j = 1,10
       call ga_sync()
#ifndef MIRROR
       if(me.eq.iran(ga_nnodes())-1) then
#else
       if(me.eq.iran(lprocs)-1) then
#endif
         do loop = 1,m
           ilo = iran(n)
           jlo = iran(n)
           iv(loop) = ilo
           jv(loop) = jlo
           v(loop) = (1d0,-1d0) *(ilo+jlo)
         enddo
         call ga_scatter(g_a, v, iv, jv, m)
         do loop = 1,m
           ilo= iv(loop)
           jlo= jv(loop)
           call ga_get(g_a,ilo,ilo,jlo,jlo,w(loop),1)
           if(w(loop)  .ne. (1d0,-1d0) *(ilo+jlo) )then
             write(6,*)me,' scatter ', ilo,',',jlo,',',w(loop)
     &             ,' ', (1d0,-1d0) *(ilo+jlo)
             call ffflush(6)
           endif
         enddo
       endif
       call ga_sync()
      enddo
c
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' scatter is  OK'
         write(6,*)
      endif
c
c     Check ga_add
c
      if (me .eq. 0) then
         write(6,91)
 91      format('> Checking add ...')
         call ffflush(6)
      endif
      call ga_get(g_a, 1, n, 1, n, a, n)
      crap = drand(12345)       ! Everyone has same seed
      do j = 1, n
         do i = 1, n
            b(i,j) = cmplx(drand(0), drand(1))
            a(i,j) = (0.1d0,-.1d0)*a(i,j) + (.9d0,-.9d0)*b(i,j)
         enddo
      enddo

#ifndef MIRROR
      if (me.eq.0) call ga_put(g_b, 1, n, 1, n, b, n)
#else
      if (iproc.eq.0) call ga_put(g_b, 1, n, 1, n, b, n)
#endif
      call ga_add((0.1d0,-.1d0), g_a, (0.9d0,-.9d0), g_b, g_b)
      call ga_get(g_b, 1, n, 1, n, b, n)
      do j = 1, n
         do i = 1, n
            if (MISMATCH(b(i,j), a(i,j)))then
               write(6,*) ' add ', me, i, j, a(i,j), b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' add is OK '
         write(6,*)
      endif
      call ga_sync()
c
c     Check the zdot function
c
      if (me .eq. 0) then
         write(6,19)
 19      format('> Checking zdot ...')
         call ffflush(6)
      endif
      crap = drand(12345)       ! Everyone has same seed
      sum1 = (0.0d0,0.d0)
      do j = 1, n
         do i = 1, n
            b(i,j) = cmplx(drand(0), drand(1))
            sum1 = sum1 + a(i,j)*b(i,j)
         enddo
      enddo
#ifndef MIRROR
      if (me.eq.0) then
#else
      if (iproc.eq.0) then
#endif
         call ga_put(g_b, 1, n, 1, n, b, n)
         call ga_put(g_a, 1, n, 1, n, a, n)
      endif
      call ga_sync()
      sum2 =  ga_zdot(g_a,g_b)
      if (MISMATCH(sum1, sum2))then
         write(6,*) ' zdot wrong ', sum1, sum2
         call ffflush(6)
         call ga_error('... exiting ',-1)
      else if (me.eq.0) then
         write(6,*)
         write(6,*) ' zdot is OK '
         write(6,*)
      endif
c
c     Delete the global arrays
c
      status = ga_destroy(g_b)
      status = ga_destroy(g_a)
c
      end
c-----------------------------------------------------------------




      subroutine check_int()
      implicit none
#include "mafdecls.fh"
#include "global.fh"
#include "testutil.fh"
c     
      integer n
      parameter (n = 128)
      integer a(n,n), b(n,n)
      logical status
      integer g_a
      integer iran, i, j, loop, nloop, ilo, ihi, jlo, jhi, itmp
      integer nproc, me, int, ij, inc, dimi, dimj, ii, jj
      integer ndim, dims(2), chunk(2), p_mirror, nnodes
      integer lprocs, inode, iproc,lproc
      double precision nwords
      parameter (nloop = 100)
      integer maxproc
      parameter (maxproc = 4096)
      double precision crap, sum1, real
      integer buf
#ifdef USE_RESTRICTED
      integer num_rstrctd
      integer rstrctd_list(maxproc/2)
#endif
#ifdef BLOCK_CYCLIC
      integer block_size(2), proc_grid(2)
      integer map(4,maxproc), block_list(maxproc), found, np,k
      integer blo(2),bhi(2)
#else
      integer map(5,maxproc), found, np,k
#endif
#ifdef USE_TILED_IRREG_DISTR
      integer mapc(4096), nblocks(2)
#endif
      intrinsic int
      iran(i) = int(drand(0)*real(i)) + 1
c     
      nproc = ga_nnodes()
      me    = ga_nodeid()
      inode = ga_cluster_nodeid()
      lprocs = ga_cluster_nprocs(inode)
      nnodes = ga_cluster_nnodes()
      iproc = mod(me,lprocs)
#ifdef USE_RESTRICTED
      num_rstrctd = nproc/2
      if (num_rstrctd.eq.0) num_rstrctd = 1
      do i = 1, num_rstrctd
        rstrctd_list(i) = (num_rstrctd/2) + i-1
      end do
#endif
#ifdef BLOCK_CYCLIC
      block_size(1) = 32
      block_size(2) = 32
#if defined(USE_SCALAPACK_DISTR) || defined(USE_TILED_DISTR)
      if (mod(nproc,2).ne.0)
     +  call ga_error("Available procs must be divisible by 2",-1)
      proc_grid(1) = 2
      proc_grid(2) = nproc/2
#endif
#endif
c     
c     a() is a local copy of what the global array should start as
c
      do j = 1, n
         do i = 1, n
#ifndef MIRROR
            a(i,j) = i-1 + (j-1)*1000
#else
            a(i,j) = inode + i-1 + (j-1)*1000
#endif
         enddo
      enddo
c     
c     Create a global array
c
      ndim = 2
      dims(1) = n
      dims(2) = n
      chunk(1) = 0
      chunk(2) = 0
#ifdef NEW_API
      g_a = ga_create_handle()
      call ga_set_data(g_a,ndim,dims,MT_INT)
      call ga_set_array_name(g_a,'a')
#ifdef USE_RESTRICTED
      call ga_set_restricted(g_a, rstrctd_list, num_rstrctd)
#endif
#ifdef BLOCK_CYCLIC
#if defined(USE_SCALAPACK_DISTR) || defined(USE_TILED_DISTR) || defined(USE_TILED_IRREG_DISTR)
#ifdef USE_SCALAPACK_DISTR
      call ga_set_block_cyclic_proc_grid(g_a,block_size,proc_grid)
#else
#ifdef USE_TILED_DISTR
      call ga_set_tiled_proc_grid(g_a,block_size,proc_grid)
#else
      call util_create_map(mapc,block_size(1),dims,nblocks,proc_grid)
      call ga_set_tiled_irreg_proc_grid(g_a,mapc,nblocks,proc_grid)
#endif
#endif
#else
      call ga_set_block_cyclic(g_a,block_size)
#endif
#endif
#  ifdef MIRROR
      p_mirror = ga_pgroup_get_mirror()
      call ga_set_pgroup(g_a,p_mirror)
#  endif
#  ifdef USE_MEMDEV
      call ga_set_memory_dev(g_a,MEMDEVICE)
#  endif
      if (.not.ga_allocate(g_a)) then
#else
#  ifndef MIRROR
      if (.not. ga_create(MT_INT, n, n, 'a', 0, 0, g_a)) then
#  else
      p_mirror = ga_pgroup_get_mirror()
      if (.not. nga_create_config(MT_INT, ndim, dims, 'a', chunk,
     +                            p_mirror, g_a)) then
#  endif
#endif
         write(6,*) ' ga_create failed'
         call ffflush(6)
         call ga_error('... exiting ',-1)
      endif
c     
c     Zero the array
c
      if (me .eq. 0) then
         write(6,21) 
 21      format('> Checking zero ... ')
         call ffflush(6)
      endif
      call ga_zero(g_a)
c
c     Check that it is indeed zero
c
      call ga_get(g_a, 1, n, 1, n, b, n)
      do i = 1, n
         do j = 1, n
            if (b(i,j) .ne. 0) then
               write(6,*) ' zero ', me, i, j, b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_zero is OK'
         write(6,*)
      endif
      call ga_sync()
c
c     Each node fills in disjoint sections of the array
c
      if (me .eq. 0) then
         write(6,2) 
 2       format('> Checking disjoint put ... ')
         call ffflush(6)
      endif
      call ga_sync()
c
      inc = (n-1)/20 + 1
      ij = 0
      do j = 1, n, inc
         do i = 1, n, inc
#ifndef MIRROR
            if (mod(ij,nproc) .eq. me) then
#else
            if (mod(ij,lprocs) .eq. iproc) then
#endif
               ilo = i
               ihi = min(i+inc, n)
               jlo = j
               jhi = min(j+inc, n)
c              write(6,4) me, ilo, ihi, jlo, jhi
c4             format(' node ',i2,' checking put ',4i4)
c              call ffflush(6)
               call ga_put(g_a, ilo, ihi, jlo, jhi, a(ilo, jlo), n)
            endif
            ij = ij + 1
         enddo
      enddo
      call ga_sync()
c
c     All nodes check all of a
c
#ifndef MIRROR
      if(me.eq.0)then
#else
      if(iproc.eq.0)then
#endif
        call ga_get(g_a, 1, n, 1, n, b, n)
        do i = 1, n
           do j = 1, n
              if (b(i,j) .ne. a(i,j)) then
                 write(6,*) ' put ', me, i, j, a(i,j),b(i,j)
                 call ffflush(6)
                 call ga_error('... exiting ',-1)
              endif
           enddo
        enddo
      endif
      call ga_sync()
c
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_put is OK'
         write(6,*)
      endif
c
c     Now check nloop random gets from each node
c
      if (me .eq. 0) then
         write(6,5) nloop
 5       format('> Checking random get (',i5,' calls)...')
         call ffflush(6)
      endif
      call ga_sync()
c
      nwords = 0
c
      crap = drand(ga_nodeid()*51 +1) !Different seed for each proc
      do loop = 1, nloop
         ilo = iran(loop)
         ihi = iran(loop)
         if (ihi.lt. ilo) then
            itmp = ihi
            ihi = ilo
            ilo = itmp
         endif
         jlo = iran(loop)
         jhi = iran(loop)
         if (jhi.lt. jlo) then
            itmp = jhi
            jhi = jlo
            jlo = itmp
         endif
c     
         nwords = nwords + (ihi-ilo+1)*(jhi-jlo+1)
c
         call util_ifill(n*n, 0.0d0, b, 1)
         call ga_get(g_a, ilo, ihi, jlo, jhi, b(ilo, jlo), n)
c
         if (me .eq. 0 .and. mod(loop-1, max(1,nloop/20)).eq.0) then
            write(6,1) loop, me, ilo, ihi, jlo, jhi, nwords
 1          format(' call ',i5, ' node ',i2,' checking get ',4i4,
     $           ' total ',d9.2)
            call ffflush(6)
         endif
c
         sum1 = 0.0d0
         do j = jlo, jhi
            do i = ilo, ihi
               sum1 = sum1 + b(i,j)
               if (b(i,j) .ne. a(i,j)) then
                  write(6,*) 'error ', i, j, b(i,j), a(i,j)
                  call ffflush(6)
                  call ga_error('... exiting ',-1)
               endif
            enddo
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_get is OK'
         write(6,*)
      endif
c
      call ga_sync()
      if (me .eq. 0 .and. n.gt.7) then
         write(6,*)
         write(6,*) '> Checking ga_print_patch --- should see '
         write(6,*)' [2002 3002 4002 5002 6002]'  
         write(6,*)' [2003 3003 4003 5003 6003]'  
         write(6,*)' [2004 3004 4004 5004 6004]'  
         write(6,*)
         call ffflush(6)
      endif
      if(n.gt.5) call ga_print_patch(g_a,3,5,3,7,0)
c
      call ga_sync()
c
      call ga_sync()
      if (me .eq. 0) then
         write(6,*)
         write(6,*) '> Checking read_inc ... '
         write(6,*)
         call ffflush(6)
      endif
      call ga_sync()
c
      crap = drand(ga_nodeid()*51 +1) !Different seed for each proc
      inc =5 
c     every processor will be operating on somebody elses data
c
#ifndef MIRROR
      lproc = nproc - me - 1
#else
      lproc = nnodes-iproc-1
      lproc = inode*lprocs + lproc
      lproc = lprocs - me-ga_cluster_procid(inode,0)
#endif
c
#ifndef MIRROR
      call ga_distribution(g_a,lproc,ilo,ihi,jlo,jhi)
#else
      lproc = me - ga_cluster_procid(inode,0)
      call ga_distribution(g_a, lproc, ilo, ihi, jlo, jhi)
#endif
c
      dimi = ihi-ilo
      dimj = jhi-jlo
c     write(6,111) me,ilo,ihi,jlo,jhi,dimi,dimj
  111 format(i2,'..',4i8,'.',2i8)
      call ga_sync()
      if(ilo .gt.0 .and. jhi .gt. 0)then
       do loop = 1,nloop 
         ii= IABS(iran(dimi)) - 1
         jj= IABS(iran(dimj)) - 1
         i=ilo + Mod(ii,dimi)
         j=jlo + Mod(jj,dimj)
c
c        write(6,*) me,'..',ilo,ihi,jlo,jhi,'.',dimi,dimj,'..',i,j
c        call ffflush(6)
         buf = ga_read_inc(g_a,i,j,inc)
         if(a(i,j).ne. buf)then   
            write(6,114) me,ilo,ihi,jlo,jhi
            write(6,112) me,i,j,a(i,j),buf,loop
c            write(6,*)me,'READ_inc ', i,',',j,',', a(i,j),' ',buf,me 
            call ffflush(6)
         endif
  112    format(i3,' READ_inc ',i3,',',i3,',',i8,' ',i8,' loop ',i8)
         call ga_get(g_a, i,i,j,j, buf,1) 
         a(i,j) = a(i,j)+inc 
         if(a(i,j).ne.  buf)then   
            write(6,114) me,ilo,ihi,jlo,jhi
            write(6,113) me,i,j,a(i,j),buf,loop
c            write(6,*)me,'read_INC ', i,',',j,',', a(i,j),' ',buf,me
            call ffflush(6)
         endif
  113    format(i3,' read_INC ',i3,',',i3,',',i8,' ',i8,' loop ',i8)
  114    format(i3,' ilo ',i3,' ihi ',i3,' jlo ',i3,' jhi ',i3)
       enddo
      endif
      call ga_sync()
c
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' read_inc is OK'
         write(6,*)
      endif
c
c
c
#ifndef MIRROR
      if (me.eq.0) then
         write(6,*)
         write(6,*) '> checking ga_fence and ga_lock'
         write(6,*)
         call ffflush(6)
      endif
c
      call ga_zero(g_a)
c
c***  use ga_read_inc and elements g_a(1:2,1) to implement a lock
c***  compute g_a(:,n) = sum (1(:)) for P processors
c     
      status = ga_create_mutexes(1)
      if (.not. status) then
         call ga_error('ga_create_mutexes failed ',-1)
      endif

      if(n.lt.2)call ga_error('insufficient n to test ga_fence',n)

      call ga_lock(0)
c     call my_lock(g_a)

c          get original values  g_a(:,n)
           call ga_get(g_a, 1,n, n,n, b,n)
c          add my contribution
           do i =1,n
              b(i,1)= b(i,1)+1
           enddo
c
c          need to use fence to assure that coms complete before leaving
c          Critical Section
c
           call ga_init_fence()
                call ga_put(g_a, 1,n, n,n, b,n)
           call ga_fence()
      call ga_unlock(0)
c     call my_unlock(g_a)
c
333   if(.not.ga_destroy_mutexes())
     $   call ga_error('mutex not destroyed',-1)

      call ga_sync()
      if (me.eq.0) then
         call ga_get(g_a, 1,n, n,n, b,n)
         do i =1,n
              if(b(i,1).ne. nproc)then
                  print *, 'mismatch',b(i,1),nproc
                  call ga_error('fence failed',i)
              endif
         enddo
         write(6,*)
         write(6,*) ' ga_fence and ga_lock are OK'
         write(6,*)
      endif 
#endif
c
c
      if (me.eq.0) then
         write(6,*)
         write(6,*) '> checking ga_locate_region'
         write(6,*)
         call ffflush(6)
      endif

#ifndef BLOCK_CYCLIC
      status =  ga_locate_region(g_a, 1, n, 1, n, map, np)
      found = 0
      do j=1,n
         do i=1,n
            b(i,j)=-1
         enddo
      enddo
#ifndef MIRROR
      if(me.eq.0)call ga_put(g_a,1,n,1,n,b,n)
#else
      if(iproc.eq.0)call ga_put(g_a,1,n,1,n,b,n)
#endif
      call ga_sync()
      do k = 1, np
#ifndef MIRROR
         if(map(5,k).eq.me)then
#else
         if(map(5,k).eq.iproc)then
#endif
                if(found.eq.1) then
                    write(6,*)'double entry in map for proc ',me
                    call ffflush(6)
                endif
                do j= map(3,k), map(4,k)
                        do i= map(1,k), map(2,k)
#ifndef MIRROR
                        b(i,j)=1*me
#else
                        b(i,j)=1*iproc
#endif
                        enddo
                enddo
                call ga_put(g_a, map(1,k),map(2,k),map(3,k),map(4,k),
     &                 b(map(1,k),map(3,k)),n)
                found = 1
         endif
      enddo
      call ga_sync()
c
      do k = 1, np
#ifndef MIRROR
         if(map(5,k).eq.me)then
#else
         if(map(5,k).eq.iproc)then
#endif
                call ga_get(g_a, map(1,k),map(2,k),map(3,k),map(4,k),
     &                 a(map(1,k),map(3,k)),n)
                do j= map(3,k), map(4,k)
                      do i= map(1,k), map(2,k)
                        if(b(i,j).ne.a(i,j)) then
                           write(6,*)
     &                     'proc ',me, 'overlap with ',a(i,j)
                           call ffflush(6)
                        endif
                      enddo
                enddo
         endif
      enddo
      call ga_sync()
c
#ifndef MIRROR
      if(me.eq.0)then
#else
      if(iproc.eq.0)then
#endif
        call ga_get(g_a,1,n,1,n,a,n)
        do j=1,n
         do i=1,n
            if(a(i,j).eq.-1)then
               write(6,*)'i=',i,' j=',j, ' not assigned '
               call ga_error('... exiting ',-1)
            endif
         enddo
        enddo
      endif
#else
c
c   block cyclic distributions
c
      blo(1) = 1
      bhi(1) = n
      blo(2) = 1
      bhi(2) = n
      status = nga_locate_region(g_a, blo, bhi, map, block_list, np)
      found = 0
      do j=1,n
         do i=1,n
            b(i,j)=-1
         enddo
      enddo
      if(me.eq.0)call ga_put(g_a,1,n,1,n,b,n)
      call ga_sync()
#if !defined(USE_SCALAPACK_DISTR) && !defined(USE_TILED_DISTR) && !defined(USE_TILED_IRREG_DISTR)
      do k = 1, np
        if(mod(block_list(k),nproc).eq.me)then
          do j= map(2,k), map(4,k)
            do i= map(1,k), map(3,k)
              b(i,j)=1*me
            enddo
          enddo
          call ga_put(g_a, map(1,k),map(3,k),map(2,k),map(4,k),
     &                b(map(1,k),map(2,k)),n)
          found = 1
        endif
      enddo
      call ga_sync()
c
      do k = 1, np
         if(mod(block_list(k),nproc).eq.me)then
                call ga_get(g_a, map(1,k),map(3,k),map(2,k),map(4,k),
     &                 a(map(1,k),map(2,k)),n)
                do j= map(2,k), map(4,k)
                      do i= map(1,k), map(3,k)
                        if(b(i,j).ne.a(i,j)) then
                           write(6,*)
     &                     'proc ',me, 'overlap with ',a(i,j)
                           call ffflush(6)
                        endif
                      enddo
                enddo
         endif
      enddo
      call ga_sync()
#else
c#if defined(USE_SCALAPACK_DISTR) || defined(USE_TILED_DISTR)
      do k = 1, np
        i = (dims(1)-1)/block_size(1);
        j = (dims(2)-1)/block_size(2);
        ii = mod(block_list(k),i)
        jj = (block_list(k)-ii)/i
        if(mod((ii+i*jj),nproc).eq.me)then
          do j= map(2,k), map(4,k)
            do i= map(1,k), map(3,k)
              b(i,j)=1*me
            enddo
          enddo
          call ga_put(g_a, map(1,k),map(3,k),map(2,k),map(4,k),
     &                b(map(1,k),map(2,k)),n)
          found = 1
        endif
      enddo
      call ga_sync()
c
      do k = 1, np
        i = (dims(1)-1)/block_size(1);
        j = (dims(2)-1)/block_size(2);
        ii = mod(block_list(k),i)
        jj = (block_list(k)-ii)/i
        if(mod((ii+i*jj),nproc).eq.me)then
          call ga_get(g_a, map(1,k),map(3,k),map(2,k),map(4,k),
     &                a(map(1,k),map(2,k)),n)
          do j= map(2,k), map(4,k)
            do i= map(1,k), map(3,k)
              if(b(i,j).ne.a(i,j)) then
                write(6,*)
     &               'proc ',me, 'overlap with ',a(i,j)
                call ffflush(6)
              endif
            enddo
          enddo
        endif
      enddo
      call ga_sync()
c#endif
c#ifdef USE_TILED_IRREG_DISTR
c#endif
#endif
#endif

c
      if(me.eq.0)then
        call ga_get(g_a,1,n,1,n,a,n)
        do j=1,n
         do i=1,n
            if(a(i,j).eq.-1)then
               write(6,*)'i=',i,' j=',j, ' not assigned '
               call ga_error('... exiting ',-1)
            endif
         enddo
        enddo
      endif

      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_locate_region is OK'
         write(6,*)
         call ffflush(6)
      endif
c     
c     Delete the global array
c
      status = ga_destroy(g_a)
c     
      end

c---------------------------------------------------------------------

      subroutine check_flt()
      implicit none
#include "mafdecls.fh"
#include "global.fh"
#include "testutil.fh"
      integer n, m
      parameter (n =100) 
      parameter (m=2*n)
      real a(n,n), b(n,n), v(m), w(m)
      integer iv(m), jv(m)
      logical status
      integer g_a, g_b
      integer iran, i, j, loop, nloop, maxloop, ilo, ihi, jlo, jhi, itmp
      integer nproc, me, int, ij, inc, ii, jj, nnodes
      double precision nwords
      parameter (maxloop = 100)
      integer maxproc
      integer ndim, dims(2), chunk(2), p_mirror
      parameter (maxproc = 4096)
      double precision crap
      real alpha, beta, x, sum1, sum2
      logical found
      integer lprocs, inode, iproc, lproc
#ifdef USE_RESTRICTED
      integer num_rstrctd
      integer rstrctd_list(maxproc/2)
#endif
#ifdef BLOCK_CYCLIC
      integer block_size(2), proc_grid(2)
#endif
      intrinsic int
      iran(i) = int(drand(0)*real(i)) + 1

      nproc = ga_nnodes()
      me    = ga_nodeid()
      inode = ga_cluster_nodeid()
      lprocs = ga_cluster_nprocs(inode)
      nnodes = ga_cluster_nnodes()
      iproc = mod(me,lprocs)
      nloop = Min(maxloop,n)
#ifdef USE_RESTRICTED
      num_rstrctd = nproc/2
      if (num_rstrctd.eq.0) num_rstrctd = 1
      do i = 1, num_rstrctd
        rstrctd_list(i) = (num_rstrctd/2) + i-1
      end do
#endif
#ifdef BLOCK_CYCLIC
      block_size(1) = 32
      block_size(2) = 32
#if defined(USE_SCALAPACK_DISTR) || defined(USE_TILED_DISTR)
      if (mod(nproc,2).ne.0)
     +  call ga_error("Available procs must be divisible by 2",-1)
      proc_grid(1) = 2
      proc_grid(2) = nproc/2
#endif
#endif
c     
c     a() is a local copy of what the global array should start as
c
      do j = 1, n
         do i = 1, n
#ifndef MIRROR
            a(i,j) = i-1 + (j-1)*n
#else
            a(i,j) = inode + i-1 + (j-1)*n
#endif
            b(i,j) = -1.
         enddo
      enddo

c     
c     Create a global array
c
      ndim = 2
      dims(1) = n
      dims(2) = n
      chunk(1) = 0
      chunk(2) = 0
#ifdef NEW_API
      g_a = ga_create_handle()
      call ga_set_data(g_a,ndim,dims,MT_REAL)
      call ga_set_array_name(g_a,'a')
#ifdef USE_RESTRICTED
      call ga_set_restricted(g_a, rstrctd_list, num_rstrctd)
#endif
#ifdef BLOCK_CYCLIC
#if defined(USE_SCALAPACK_DISTR) || defined(USE_TILED_DISTR)
#ifdef USE_SCALAPACK_DISTR
      call ga_set_block_cyclic_proc_grid(g_a,block_size,proc_grid)
#else
      call ga_set_tiled_proc_grid(g_a,block_size,proc_grid)
#endif
#else
      call ga_set_block_cyclic(g_a,block_size)
#endif
#endif
#  ifdef MIRROR
      write(6,*) 'Getting p_mirror'
      p_mirror = ga_pgroup_get_mirror()
      write(6,*) 'Setting proc config'
      call ga_set_pgroup(g_a,p_mirror)
#  endif
#  ifdef USE_MEMDEV
      call ga_set_memory_dev(g_a,MEMDEVICE)
#  endif
      if (.not.ga_allocate(g_a)) then
#else
#  ifndef MIRROR
      if (.not. ga_create(MT_REAL, n, n, 'a', 0, 0, g_a)) then
#  else
cBJP
      p_mirror = ga_pgroup_get_mirror()
      if (.not. nga_create_config(MT_REAL, ndim, dims, 'a', chunk,
     +                            p_mirror, g_a)) then
cBJP
      write(6,*) 'Created g_a'
#  endif
#endif
         write(6,*) ' ga_create failed'
         call ffflush(6)
         call ga_error('... exiting ',-1)
      endif
c
c     check if handle is valid. be quiet unless error
C
      if(.not.ga_valid_handle(g_a)) call ga_error("invalid handle",g_a)
c
#ifndef MIRROR
      call ga_distribution(g_a,me,ilo, ihi, jlo, jhi)
#else
      lproc = me - ga_cluster_procid(inode,0)
      call ga_distribution(g_a, lproc, ilo, ihi, jlo, jhi)
#endif
      call ga_sync()      
c     
c     Zero the array
c
      if (me .eq. 0) then
         write(6,21) 
 21      format('> Checking zero ... ')          
         call ffflush(6)
      endif
      call ga_zero(g_a)
c
c     Check that it is indeed zero
c
      call ga_get(g_a, 1, n, 1, n, b, n)
      call ga_sync()
      do i = 1, n
         do j = 1, n
            if (b(i,j) .ne. 0.0) then
               write(6,*) ' zero ', me, i, j, b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_zero is OK'
         write(6,*)
      endif
      call ga_sync()
c
c     Each node fills in disjoint sections of the array
c
      if (me .eq. 0) then
         write(6,2) 
 2       format('> Checking disjoint put ... ')   
         call ffflush(6)
      endif
      call ga_sync()
c
      inc = (n-1)/20 + 1
      ij = 0
      do j = 1, n, inc
         do i = 1, n, inc
#ifndef MIRROR
            if (mod(ij,nproc) .eq. me) then
#else
            if (mod(ij,lprocs) .eq. iproc) then
#endif
               ilo = i
               ihi = min(i+inc, n)
               jlo = j
               jhi = min(j+inc, n)
c               write(6,4) me, ilo, ihi, jlo, jhi
c 4             format(' node ',i2,' checking put ',4i4)
c               call ffflush(6)
               call ga_put(g_a, ilo, ihi, jlo, jhi, a(ilo, jlo), n)
            endif
            ij = ij + 1
         enddo
      enddo
      call ga_sync()
c
c     All nodes check all of a
c
      call ga_get(g_a, 1, n, 1, n, b, n)
      do i = 1, n
         do j = 1, n
            if (b(i,j) .ne. a(i,j)) then
               write(6,*) ' put ', me, i, j, a(i,j),b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      call ga_sync()
c
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_put is OK'
         write(6,*)
      endif
c
c     Now check nloop random gets from each node
c
      if (me .eq. 0) then
         write(6,5) nloop
 5       format('> Checking random get (',i5,' calls)...')
         call ffflush(6)
      endif
      call ga_sync()
c
      nwords = 0
c
      crap = drand(ga_nodeid()*51 +1) !Different seed for each proc
      do loop = 1, nloop
         ilo = iran(loop)
         ihi = iran(loop)
         if (ihi.lt. ilo) then
            itmp = ihi
            ihi = ilo
            ilo = itmp
         endif
         jlo = iran(loop)
         jhi = iran(loop)
         if (jhi.lt. jlo) then
            itmp = jhi
            jhi = jlo
            jlo = itmp
         endif
c     
         nwords = nwords + (ihi-ilo+1)*(jhi-jlo+1)
c
         call util_rfill(n*n, 0.0, b, 1)
         call ga_get(g_a, ilo, ihi, jlo, jhi, b(ilo, jlo), n)
c
         if (me .eq. 0 .and. mod(loop-1, max(1,nloop/20)).eq.0) then
            write(6,1) loop, me, ilo, ihi, jlo, jhi, nwords
 1          format(' call ',i5, ' node ',i2,' checking get ',4i4,
     $           ' total ',d9.2)
            call ffflush(6)
         endif
c
         sum1 = 0.0d0
         do j = jlo, jhi
            do i = ilo, ihi
               sum1 = sum1 + b(i,j)
               if (b(i,j) .ne. a(i,j)) then
                  write(6,*) 'error ', i, j, b(i,j), a(i,j)
                  call ffflush(6)
                  call ga_error('... exiting ',-1)
               endif
            enddo
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_get is OK'
         write(6,*)
      endif
      call ga_sync()
c
c     Each node accumulates into disjoint sections of the array
c
      if (me .eq. 0) then
         write(6,9)
 9       format('> Checking accumulate ... ')
         call ffflush(6)
      endif
      call ga_sync()

c
      crap = drand(12345)       ! Same seed for each process
      do j = 1, n
         do i = 1, n
c           b(i,j) = real(drand(0))
            b(i,j) = i+j
         enddo
      enddo
c
      inc = (n-1)/20 + 1
      ij = 0
      do j = 1, n, inc
         do i = 1, n, inc
c           x = real(drand(0))
            x = 10.
            ilo = i
            ihi = min(i+inc-1, n)
            if(ihi.eq.n-1)ihi=n
c           ihi = min(i+inc, n)
            jlo = j
            jhi = min(j+inc-1, n)
            if(jhi.eq.n-1)jhi=n
c           jhi = min(j+inc-1, n)
*               call ffflush(6)
#ifndef MIRROR
            if (mod(ij,nproc) .eq. me) then
#else
            if (mod(ij,lprocs) .eq. iproc) then
#endif
c               print *, me, 'checking accumulate ',ilo,ihi,jlo,jhi,x
* 11            format(' node ',i2,' checking accumulate ',4i4)
*               call ffflush(6)
               call ga_acc(g_a, ilo, ihi, jlo, jhi, b(ilo, jlo), n, x)
            endif
            ij = ij + 1
c
c           Each process applies all updates to its local copy
c
            do jj = jlo, jhi
               do ii = ilo, ihi
                  a(ii,jj) = a(ii,jj) + x * b(ii,jj)
               enddo
            enddo
         enddo
      enddo
      call ga_sync()                       
c
c     All nodes check all of a
      call ga_get(g_a, 1, n, 1, n, b, n)
c
      do j = 1, n
         do i = 1, n
            if(MISMATCHF(b(i,j),a(i,j)))then
               write(6,*) ' acc ', me, i, j, a(i,j), b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' disjoint ga_acc is OK'
         write(6,*)
      endif
c
c     overlapping accumulate
      call ga_sync()
#ifndef MIRROR
      if (.not. ga_create(MT_REAL, n, n, 'b', 0, 0, g_b)) then
#else
      if (.not. nga_create_config(MT_REAL, ndim, dims, 'b', chunk,
     +                            p_mirror, g_b)) then
#endif
         call ga_error('ga_create failed for second array ',-1)
      endif
c
      call ga_zero(g_b)
      call ga_acc(g_b, n/2, n/2, n/2, n/2, 1.0, 1, 1.0)
      call ga_sync()
      if (me.eq.0) then
         call ga_get(g_b, n/2, n/2, n/2, n/2, b(1,1), 1)
#ifndef MIRROR
         x = abs(b(1,1) -1*nproc)
#else
         x = abs(b(1,1) -1*lprocs)
#endif
         if(x.gt. 1e-10)then
#ifndef MIRROR
            write(6,*)'val=',b(1,1),' expected=',nproc, x
#else
            write(6,*)'val=',b(1,1),' expected=',lprocs, x
#endif
            call ga_error('overlapping accumulate failed',-1)
         endif
         write(6,*)
         write(6,*) ' overlapping ga_acc is OK'
         write(6,*)
      endif
c
c     Check the ga_add function
c
      if (me .eq. 0) then
         write(6,91)
 91      format('> Checking add ...')
         call ffflush(6)
      endif                           
c
      crap = drand(12345)       ! Everyone has same seed
      do j = 1, n
         do i = 1, n
            b(i,j) = real(drand(0)*real(i)) + 1
            a(i,j) = 0.1*a(i,j) + 0.9*b(i,j)
         enddo
      enddo
#ifndef MIRROR
      if (me.eq.0) call ga_put(g_b, 1, n, 1, n, b, n)
#else
      if (iproc.eq.0) call ga_put(g_b, 1, n, 1, n, b, n)
#endif
      alpha = 0.1
      beta = 0.9
      call ga_add(alpha, g_a, beta, g_b, g_b)
      call ga_get(g_b, 1, n, 1, n, b, n)
      do j = 1, n
         do i = 1, n
            if(MISMATCHF(b(i,j), a(i,j)))then
               write(6,*) ' add ', me, i, j, a(i,j), b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' add is OK '
         write(6,*)
      endif
      call ga_sync()
c
c     Check the sdot function
c
      if (me .eq. 0) then
         write(6,19)
 19      format('> Checking sdot ...')
         call ffflush(6)
      endif
      crap = drand(12345)       ! Everyone has same seed
      sum1 = 0.0
      do j = 1, n
         do i = 1, n
            b(i,j) = drand(0)
            sum1 = sum1 + a(i,j)*b(i,j)
         enddo
      enddo
#ifndef MIRROR
      if (me.eq.0) then
#else
      if (iproc.eq.0) then
#endif
         call ga_put(g_b, 1, n, 1, n, b, n)
         call ga_put(g_a, 1, n, 1, n, a, n)
      endif
      call ga_sync()
      sum2 = ga_sdot(g_a,g_b)
      if(MISMATCHF(sum1, sum2))then
         write(6,*) ' fdot wrong ', sum1, sum2
         call ffflush(6)
            call ga_error('... exiting ',-1)
      else if (me.eq.0) then
         write(6,*)
         write(6,*) ' sdot is OK '
         write(6,*)
      endif
c
c     Check the ga_scale function
c
      if (me .eq. 0) then
         write(6,92)
 92      format('> Checking scale ...')
         call ffflush(6)
      endif
      call ga_scale(g_a, 0.123)
      call ga_get(g_a, 1, n, 1, n, b, n)
      do j = 1, n
         do i = 1, n
            a(i,j) = a(i,j)*0.123
            if (MISMATCHF(b(i,j), a(i,j)))then
               write(6,*) ' dscal ', me, i, j, a(i,j), b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' scale is OK '
         write(6,*)
      endif
c                                
c     Check the ga_copy function
c
      if (me .eq. 0) then
         write(6,*)
         write(6,*)'> Checking copy'
         write(6,*)
         call ffflush(6)
      endif
      if(me.eq.0) call ga_put(g_a, 1, n, 1, n, a, n)
      call ga_copy(g_a, g_b)
      call ga_get(g_b, 1, n, 1, n, b, n)
      do j = 1, n
         do i = 1, n
            if (b(i,j) .ne. a(i,j)) then
               write(6,*) ' copy ', me, i, j, a(i,j), b(i,j)
               call ffflush(6)
               call ga_error('... exiting ',-1)
            endif
         enddo
      enddo
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' copy is OK '
         write(6,*)
      endif
c
      call ga_sync()
      if (me .eq. 0) then
         write(6,*) '> Checking scatter/gather (might be slow)... '
         call ffflush(6)
      endif
      call ga_sync()
c
      crap = drand(ga_nodeid()*51 +1) !Different seed for each proc
      do j = 1, 10
       call ga_sync()
#ifndef MIRROR
       itmp = iran(nproc)-1
       if(me.eq.itmp) then
#else
       itmp = iran(lprocs)-1
       if(iproc.eq.itmp) then
#endif
         do loop = 1,m
           ilo = iran(n)
           jlo = iran(n)
           iv(loop) = ilo
           jv(loop) = jlo
         enddo
         call ga_gather(g_a, v, iv, jv, m)
         do loop = 1,m
           ilo= iv(loop)
           jlo= jv(loop)
           call ga_get(g_a,ilo,ilo,jlo,jlo,v(loop),1)
           if(v(loop)  .ne. a(ilo,jlo))then
             write(6,*)me,' gather ', ilo,',',jlo,',', a(ilo,jlo)
     &             ,' ',v(loop)
             call ffflush(6)
             call ga_error('... exiting ',-1)
           endif                                     
         enddo
       endif
      enddo
c
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' gather is  OK'
         write(6,*)
         call ffflush(6)
      endif
c
      do j = 1,10
       call ga_sync()
#ifndef MIRROR
       if(me.eq.iran(ga_nnodes())-1) then
#else
       if(iproc.eq.iran(lprocs)-1) then
#endif
         do loop = 1,m
           ilo = iran(n)
           jlo = iran(n)
           iv(loop) = ilo
           jv(loop) = jlo
c          v(loop) = DSIN(a(ilo,jlo)+b(ilo,jlo))
           v(loop) = 1.0 *(ilo+jlo)
         enddo
         call ga_scatter(g_a, v, iv, jv, m)
         do loop = 1,m
           ilo= iv(loop)
           jlo= jv(loop)
           call ga_get(g_a,ilo,ilo,jlo,jlo,w(loop),1)
c          if(v(loop)  .ne. w(loop))then
           if(w(loop)  .ne. 1.0 *(ilo+jlo) )then
             write(6,*)me,' scatter ', ilo,',',jlo,',',w(loop)
     &             ,' ', 1.0 *(ilo+jlo)
             call ffflush(6)
             call ga_error('... exiting ',-1)
           endif
         enddo
       endif
       call ga_sync()
      enddo
c
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' scatter is  OK'
         write(6,*)
      endif
c
      call ga_sync()
c
c     scatter-acc available in GA ver. 3.0
#ifdef GA3                                         
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' checking scatter-accumulate'
         write(6,*)
      endif
c
      crap = drand(1234)
      call ga_zero(g_a)
c
      do j = 1, n
         do i = 1, n
            b(i,j) =0.
         enddo
      enddo
c
      x = .1d0
      ii =n
      do jj = 1,1
          call ga_sync()
        do loop = 1, ii
 
c          generate unique i,j pairs
10         continue
              i = iran(n)
              j=iran(n)
           if (found(i,j, iv, jv, loop-1) ) goto 10

           iv(loop) = i
           jv(loop) = j
           v(loop) = 1.0 *(i+j)
#ifndef MIRROR
           b(i,j) = b(i,j) + nproc*x*v(loop) ! update local ref. array
#else
           b(i,j) = b(i,j) + lprocs*x*v(loop) ! update local ref. array
#endif
        enddo
 
        call ga_scatter_acc(g_a,v,iv,jv, ii,x)
 
c
        call ga_sync()
c
c       check the result
c
        call ga_get(g_a, 1, n, 1,n, a, n)
 
        do loop = 1,ii
          i = iv(loop)
          j = jv(loop)
          if(MISMATCHF(a(i,j),b(i,j)))then
             print *,'Error',i,j,loop,a(i,j),'expected=',b(i,j)
*            if(me.eq.0)then
*              do ii=1,loop
*                   print *,'element',ii, iv(ii),jv(ii)
*              enddo
*            endif
             call ga_error('scatter-acc error in trial ',jj)
          endif
        enddo                                 
          call ga_sync()
 
      enddo
 
      call ga_sync()
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' scatter-accumulate is  OK'
         write(6,*)
      endif
#endif                      
c
c     Delete the global array
c
      status = ga_destroy(g_a)
c     
      end
c_____________________________________________________________

      subroutine check_wrappers
      implicit none
#include "mafdecls.fh"
#include "global.fh"
#include "testutil.fh"
      integer n
      parameter (n=10000)
      double precision sum(n)
      double complex zsum(n), check(n)
      integer isum(n), ibuf(2)
      integer me, nproc, i
      real fsum(n)
c
      nproc = ga_nnodes()
      me = ga_nodeid()
c
      call ga_sync()
      if (me .eq. 0) then
         write(6,*)
         write(6,*)'> Checking ga_gop (integer)'
         write(6,*)
         call ffflush(6)
      endif
      do i = 1, n 
         isum(i) =  i + me
      enddo
      call ga_gop(MT_INT, isum, n, '+')
      do i = 1, n 
         if(isum(i).ne.(i*nproc + (nproc-1)*nproc/2))then
            call ga_error('ga_gop error -2',isum(i))
         endif
      enddo
      call ga_sync()
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_gop (integer) is  OK'
         write(6,*)
      endif
      call ga_sync()
c
      call ga_sync()
      if (me .eq. 0) then
         write(6,*)
         write(6,*)'> Checking ga_gop (double)'
         write(6,*)
         call ffflush(6)
      endif
      do i = 1, n
         sum(i) = i + 1d0 * me
      enddo
      call ga_gop(MT_DBL, sum, n, '+')
      do i = 1, n
         if(Int(sum(i)).ne.(i*nproc + (nproc-1)*nproc/2))then
            call ga_error('ga_gop error',Int(sum(i)))
         endif
      enddo
      call ga_sync()
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_gop (double) is  OK'
         write(6,*)
      endif
c
      call ga_sync()
      if (me .eq. 0) then
         write(6,*)
         write(6,*)'> Checking ga_gop (real)'
         write(6,*)
         call ffflush(6)
      endif
      do i = 1, n 
         fsum(i) = i + 1.0 * me
      enddo
      call ga_gop(MT_REAL, fsum, n, '+')
      do i = 1, n 
         if(Int(fsum(i)).ne.(i*nproc + (nproc-1)*nproc/2))then
            call ga_error('ga_gop error',Int(fsum(i)))
         endif
      enddo
      call ga_sync()
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_gop (real) is  OK'
         write(6,*)
      endif
c
      call ga_sync()
      if (me .eq. 0) then
         write(6,*)
         write(6,*)'> Checking ga_gop (double complex)'
         write(6,*)
         call ffflush(6)
      endif
      do i = 1, n
         zsum(i) = cmplx(i+1d0, i + 1d0 * me)
         check(i)= cmplx(nproc*i+nproc*1d0, nproc*i+(nproc-1)*nproc/2)
      enddo
      call ga_gop(MT_DCPL, zsum, n, '+')
      do i = 1, n 
         if(zsum(i).ne.check(i)) then
            call ga_error('ga_gop error', -1)
         endif
      enddo
      call ga_sync()
      if (me.eq.0) then
         write(6,*)
         write(6,*) ' ga_gop (double complex) is  OK'
         write(6,*)
      endif
c                                   
      call ga_sync()
      if (me .eq. 0) then
         write(6,*)
         write(6,*)'> Checking ga_brdcst'
         write(6,*)
         call ffflush(6)
      endif
      if(me.eq.nproc-1)then
        ibuf(1) = me
        ibuf(2) = nproc
      endif
      call ga_brdcst(1000,ibuf,util_mitob(2),nproc-1)
      if(ibuf(1).ne.nproc-1)call ga_error('ibuf(1) error',ibuf(1))
      if(ibuf(2).ne.nproc)call ga_error('ibuf(2) error',ibuf(2))
      call ga_sync()
      if (me .eq. 0) then
         write(6,*)
         write(6,*)'> ga_brdcst is OK '
         write(6,*)
         call ffflush(6)
      endif
      call ga_sync()
      end


      subroutine check_mem(mem_size)
      implicit none
      integer mem_size
#include "mafdecls.fh"
#include "global.fh"
#include "testutil.fh"
c
      integer n,nmax,left,need, me,procs,g_a, g_b 
      logical status
c
      if(.not. ga_memory_limited())return
      me = ga_nodeid()
      procs = ga_nnodes()
      nmax = int(dsqrt(dble(mem_size/util_mitob(1))))
      left =  mem_size/procs - ga_inquire_memory()
      n = nmax/2
      need = util_mdtob(n*n)/procs
c
      if(me.eq.0)then
        write(6,*)' '
        if(ga_uses_ma())then
           write(6,*)' CHECKING GA MEMORY RESTRICTIONS (MA used)'
        else
           write(6,*)' CHECKING GA MEMORY RESTRICTIONS (MA not used)'
        endif
        write(6,*)' '
        write(6,*)' '
        call print_mem_info(n,left, need, mem_size/procs)
      endif
c
      status = ga_create(MT_DBL, n, n, 'a', 0, 0, g_a)
c
      if(me.eq.0) then
        if(status) then
              write(6,*) '  success'
        else
              write(6,*) '  failure'
        endif
        call ffflush(6)
      endif
c
      n = nmax
      left =  mem_size/procs - ga_inquire_memory()
      need = util_mdtob(n*n)/procs
      if(me.eq.0)then
        call print_mem_info(n,left, need, mem_size/procs)
      endif
c
      status = ga_create(MT_DBL, n, n, 'b', 0, 0, g_b)
c
      if(me.eq.0) then
        if(status) then
              write(6,*) '  success'
        else
              write(6,*) '  failure'
        endif
        write(6,*)' '
        write(6,*)' '
        call ffflush(6)
      endif
      status = ga_destroy(g_a)
      end

     

      subroutine print_mem_info(n,left, need, total)
      implicit none
      integer n,left, need, total
c
      write(6,*)' '
      if(left - need .ge. 0) then
        write(6,1)n,n 
1       format('> Creating array ',i4,' by ',i4,' -- should succeed')
      else
        write(6,2)n,n 
2       format('> Creating array ',i4,' by ',i4,' -- SHOULD FAIL')
      endif
      write(6,3) need, left, total
3     format('  (need ',i7,' and  ',i7,' out of ',i7,' bytes are left)')
      write(6,*)' '
      call ffflush(6)
c
      end


      
      subroutine my_lock(g_b)
      implicit none
#include "global.fh"
      integer g_b, val, flag, i
      logical first_time
      double precision  dummy 
      common /lock/ val 
      common /dum/ dummy
      data first_time /.true./
c
c     this awkward initialization is to avoid a weird problem 
C     with block data on SUN
      if(first_time)then
        first_time = .false.
        dummy = .0
      endif
c
      val = ga_read_inc(g_b,1,1, 1)
10    call ga_get(g_b, 2,2,1,1, flag, 1)
           if(flag.eq.val) return 
c
c          to reduce memory stress,  wait a while before retrying
           do i = 1, 100
              dummy = dummy + .1 
           enddo 
      goto 10
      end


      subroutine my_unlock(g_b)
      implicit none
#include "global.fh"
      integer g_b, val
      common /lock/ val
c
      call ga_put(g_b, 2,2,1,1, val+1, 1) 
      end


      logical function found(i,j, iv, jv, n)
      integer n
      integer i,j, iv(n), jv(n)
      integer loop

      found = .false.
      do loop = 1, n
               if(i .eq. iv(loop) .and. j .eq.jv(loop))then
                 found = .true.
                 goto 99
               endif
            enddo
99    continue
      return
      end 

      logical function nfound(i,j, ijv, n)
      integer n
      integer i,j, ijv(2,n)
      integer loop

      nfound = .false.
      do loop = 1, n
               if(i .eq. ijv(1,loop) .and. j .eq.ijv(2,loop))then
                 nfound = .true.
                 goto 99
               endif
            enddo
99    continue
      return
      end 
      
            
      subroutine proc_remap()
      implicit none
#include "global.fh"
      integer proc(100),nproc,i
      nproc = ga_nnodes()
      if(nproc.gt.100)
     $   call ga_error("remap requires<=100 processes",nproc)
      do i = 1, nproc
         proc(i) = nproc-i
      enddo
c     call ga_register_proclist(proc,nproc)
      end

      
      subroutine util_rfill(n,val,a,ia)
      implicit none
      real  a(*), val
      integer n, ia, i
c
c     initialise real array to scalar value
c
      if (ia.eq.1) then
         do 10 i = 1, n
            a(i) = val
 10      continue
      else
         do 20 i = 1,(n-1)*ia+1,ia
            a(i) = val
 20      continue
      endif
c
      end

      
      subroutine util_dfill(n,val,a,ia)
      implicit none
      double precision  a(*), val
      integer n, ia, i
c
c     initialise double precision array to scalar value
c
      if (ia.eq.1) then
         do 10 i = 1, n
            a(i) = val
 10      continue
      else
         do 20 i = 1,(n-1)*ia+1,ia
            a(i) = val
 20      continue
      endif
c
      end

      subroutine util_ifill(n,val,a,ia)
      implicit none
      integer n, ia, i, a(*),val
c
c     initialise integer array to scalar value
c
      if (ia.eq.1) then
         do 10 i = 1, n
            a(i) = val
 10      continue
      else
         do 20 i = 1,(n-1)*ia+1,ia
            a(i) = val
 20      continue
      endif
c
      end

      subroutine util_qfill(n,val,a,ia)
      implicit none
      double  complex  a(*), val
      integer n, ia, i
c
c     initialise double complex array to scalar value
c
      if (ia.eq.1) then
         do 10 i = 1, n
            a(i) = val
 10      continue
      else
         do 20 i = 1,(n-1)*ia+1,ia
            a(i) = val
 20      continue
      endif
c
      end

      subroutine util_create_map(mapc, base, dims, nblocks, proc_grid)
      implicit none
#include "global.fh"
      integer mapc(4096)
      integer dims(2), nblocks(2), proc_grid(2)
      integer primes(256), nprime
      integer factors(256), nfactors
      integer base
      integer i, j, nproc, pmax, px, py, lo, hi
      logical ok
!
!  Factor nproc into 2D processor grid. Start by finding prime
!  factors
!
      nproc = ga_nnodes()
      pmax = nproc
      nprime = 1 
      primes(1) = 2
      do i = 3, pmax
        ok = .true.
        do j = 1, nprime
!
! check to see if any of the existing primes divides i
!
          if (mod(i,primes(j)).eq.0) then
             ok = .false.
          endif
        end do
!
!  none of the existing primes divides i, so add it to the list of
!  primes
!
        if (ok) then
          nprime = nprime + 1
          primes(nprime) = i
        endif
      end do
!
!  find prime factors of nproc
!
      nfactors = 0
      j = nproc
      i = 1
      do while (i.le.nprime.and.j.gt.1)
        do while (mod(j,primes(i)).eq.0.and.j.gt.1)
          nfactors = nfactors + 1
          factors(nfactors) = primes(i)
          j = j/primes(i)
        end do
        i = i + 1
      end do
      if (nfactors.eq.0) then
        nfactors = 1
        primes(1) = nproc
      endif
!
!  prime factors are available. Create px and py
!
      px = 1
      py = 1
      do i = nfactors, 1, -1
        if (px.lt.py) then
          px = px*factors(i) 
        else
          py = py*factors(i)
        endif
      end do
      proc_grid(1) = px
      proc_grid(2) = py
!
!  proc grid is complete, now create map. Start with x-axis
!
      lo = 1
      hi = 0
      i = 0
      j = 0
      do while (hi.lt.dims(1))
        i = i + 1
        hi = lo + base + mod(i+1,2) - 1
!        hi = lo + base - 1
        if (hi.gt.dims(1)) hi = dims(1)
        j = j + 1
        mapc(j) = lo
        lo = hi + 1
      end do
      nblocks(1) = i
      lo = 1
      hi = 0
      i = 0
      do while (hi.lt.dims(2))
        i = i + 1
        hi = lo + base + mod(i,2) - 1
!        hi = lo + base - 1
        if (hi.gt.dims(1)) hi = dims(1)
        j = j + 1
        mapc(j) = lo
        lo = hi + 1
      end do
      nblocks(2) = i
      end