def create_global_array(gatype):
if NEW_API:
g_a = ga.create_handle()
ga.set_data(g_a, [n,n], gatype)
ga.set_array_name(g_a, 'a')
if USE_RESTRICTED:
num_restricted = nproc/2 or 1
restricted_list = np.arange(num_restricted) + num_restricted/2
ga.set_restricted(g_a, restricted_list)
if BLOCK_CYCLIC:
if USE_SCALAPACK_DISTR:
if nproc % 2 == 0:
ga.error('Available procs must be divisible by 2',nproc)
ga.set_block_cyclic_proc_grid(g_a, block_size, proc_grid)
else:
ga.set_block_cyclic(g_a, block_size)
if MIRROR:
p_mirror = ga.pgroup_get_mirror()
ga.set_pgroup(g_a, p_mirror)
ga.allocate(g_a)
else:
if MIRROR:
p_mirror = ga.pgroup_get_mirror()
ga.create_config(gatype, (n,n), 'a', None, p_mirror)
else:
g_a = ga.create(gatype, (n,n), 'a')
if 0 == g_a:
ga.error('ga.create failed')
if MIRROR:
lproc = me - ga.cluster_procid(inode, 0)
lo,hi = ga.distribution(g_a, lproc)
else:
lo,hi = ga.distribution(g_a, me)
ga.sync()
return g_a
def main():
# TODO there's got to be a loopless, more pythonic way to do this
ii = 0
for i in range(num1*num1):
ii += 1
if ii > num1:
ii = 0
h0[i] = ii
# compute times assuming 500 mflops and 5 second target time
# ntimes = max(3.0, 5.0/(4.0-9*num**3))
ntimes = 5
for ii in range(howmany):
num_m = nums_m[ii]
num_n = nums_n[ii]
num_k = nums_k[ii]
a = 0.5/(num_m*num_n)
if num_m > nummax or num_n > nummax or num_k > nummax:
ga.error('Insufficient memory: check nummax')
if BLOCK_CYCLIC:
block_size = [128,128]
g_c = ga.create_handle()
ga.set_data(g_c, (num_m,num_n), ga.C_DBL)
ga.set_array_name(g_c, 'g_c')
ga.set_block_cyclic(g_c, block_size)
if not ga.allocate(g_c):
ga.error('create failed')
block_size = [128,128]
g_b = ga.create_handle()
ga.set_data(g_b, (num_k,num_n), ga.C_DBL)
ga.set_array_name(g_b, 'g_b')
ga.set_block_cyclic(g_b, block_size)
if not ga.allocate(g_b):
ga.error('create failed')
block_size = [128,128]
g_a = ga.create_handle()
ga.set_data(g_a, (num_m,num_k), ga.C_DBL)
ga.set_array_name(g_a, 'g_a')
ga.set_block_cyclic(g_a, block_size)
if not ga.allocate(g_a):
ga.error('create failed')
else:
g_a = ga.create(ga.C_DBL, (num_m,num_k), 'g_a')
g_b = ga.create(ga.C_DBL, (num_k,num_n), 'g_b')
g_c = ga.create(ga.C_DBL, (num_m,num_n), 'g_c')
for handle in [g_a,g_b,g_c]:
if 0 == handle:
ga.error('create failed')
# initialize matrices A and B
if 0 == me:
load_ga(g_a, h0, num_m, num_k)
load_ga(g_b, h0, num_k, num_n)
ga.zero(g_c)
ga.sync()
if 0 == me:
print '\nMatrix Multiplication C = A[%d,%d] x B[%d,%d]\n' % (
num_m, num_k, num_k, num_n)
print ' %4s %12s %12s %7s %7s'%(
"Run#", "Time (seconds)", "mflops/proc",
"A trans", "B trans")
avg_t[:] = 0
avg_mf[:] = 0
for itime in range(ntimes):
for i in range(ntrans):
ga.sync()
ta = transa[i]
tb = transb[i]
t1 = time.time()
ga.gemm(ta,tb,num_m,num_n,num_k,1,g_a,g_b,0,g_c)
t1 = time.time() - t1
if 0 == me:
mf = 2*num_m*num_n*num_k/t1*10**-6/nproc
avg_t[i] += t1
avg_mf[i] += mf
print ' %4d %12.4f %12.1f %7s %7s'%(
itime+1, t1, mf, ta, tb)
if VERIFY and itime == 0:
verify_ga_gemm(ta, tb, num_m, num_n, num_k,
1.0, g_a, g_b, 0.0, g_c)
if 0 == me:
print ''
for i in range(ntrans):
print 'Average: %12.4f seconds %12.1f mflops/proc %s %s'%(
avg_t[i]/ntimes, avg_mf[i]/ntimes,
transa[i], transb[i])
if VERIFY:
print 'All ga.gemms are verified...O.K.'
def main():
if 0 == me:
if MIRROR:
print ' Performing tests on Mirrored Arrays'
print ' GA initialized'
# note that MA is not used, so no need to initialize it
# "import ga" registers malloc/free as memory allocators internally
#if nproc-1 == me:
if 0 == me:
print 'using %d process(es) %d custer nodes' % (
nproc, ga.cluster_nnodes())
print 'process %d is on node %d with %d processes' % (
me, ga.cluster_nodeid(), ga.cluster_nprocs(-1))
# create array to force staggering of memory and uneven distribution
# of pointers
dim1 = MEM_INC
mapc = [0]*nproc
for i in range(nproc):
mapc[i] = MEM_INC*i
dim1 += MEM_INC*i
g_s = ga.create_handle()
ga.set_data(g_s, [dim1], ga.C_INT)
ga.set_array_name(g_s, 's')
ga.set_irreg_distr(g_s, mapc, [nproc])
if MIRROR:
if 0 == me:
print '\nTESTING MIRRORED ARRAYS\n'
# check support for single precision arrays
if 0 == me:
print '\nCHECKING SINGLE PRECISION\n'
check_float()
# check support for double precision arrays
if 0 == me:
print '\nCHECKING DOUBLE PRECISION\n'
check_double()
# check support for single precision complex arrays
if 0 == me:
print '\nCHECKING SINGLE COMPLEX\n'
check_complex_float()
# check support for double precision complex arrays
if 0 == me:
print '\nCHECKING DOUBLE COMPLEX\n'
check_complex_double()
# check support for integer arrays
if 0 == me:
print '\nCHECKING INT\n'
check_int()
# check support for long integer arrays
if 0 == me:
print '\nCHECKING LONG INT\n'
check_long()
if 0 == me:
print '\nCHECKING Wrappers to Message Passing Collective ops\n'
check_wrappers()
# check if memory limits are enforced
#check_mem(ma_heap*ga.nnodes())
if 0 == me: ga.print_stats()
if 0 == me: print ' '
if 0 == me: print 'All tests successful'