def check_gather(gatype):
if 0 == me:
print '> Checking gather (might be slow)...',
g_a = create_global_array(gatype)
a = create_local_a(gatype)
if 0 == me:
ga.put(g_a, a)
ga.sync()
ijv = np.zeros((m,2), dtype=np.int64)
random.seed(ga.nodeid()*51 + 1) # different seed for each proc
for j in range(10):
itmp = None
if MIRROR:
itmp = random.randint(0,lprocs-1)
else:
itmp = random.randint(0,nproc-1)
if itmp == me:
for loop in range(m):
ijv[loop,:] = (random.randint(0,n-1),random.randint(0,n-1))
#if ijv[loop,0] > ijv[loop,1]:
# ijv[loop,:] = ijv[loop,::-1] # reverse
result = ga.gather(g_a, ijv)
for loop in range(m):
value = ga.get(g_a, ijv[loop], ijv[loop]+1).flatten()
if not result[loop] == value:
ga.error('gather failed')
if 0 == me:
print 'OK'
ga.destroy(g_a)
def check_scatter(gatype):
nptype = ga.dtype(gatype)
if 0 == me:
print '> Checking scatter (might be slow)...',
g_a = create_global_array(gatype)
a = create_local_a(gatype)
if 0 == me:
ga.put(g_a, a)
ga.sync()
ijv = np.zeros((m,2), dtype=np.int64)
v = np.zeros(m, dtype=nptype)
random.seed(ga.nodeid()*51 + 1) # different seed for each proc
for j in range(10):
check = None
if MIRROR:
check = random.randint(0,lprocs-1) == iproc
else:
check = random.randint(0,nproc-1) == me
if check:
for loop in range(m):
ijv[loop,:] = (random.randint(0,n-1),random.randint(0,n-1))
v[loop] = ijv[loop,0]+ijv[loop,1]
ga.scatter(g_a, v, ijv)
for loop in range(m):
value = ga.get(g_a, ijv[loop], ijv[loop]+1).flatten()
if not v[loop] == value:
ga.error('scatter failed')
if 0 == me:
print 'OK'
ga.destroy(g_a)
def check_dot(gatype):
if 0 == me:
print '> Checking dot ...',
np.random.seed(12345) # everyone has same seed
g_a = create_global_array(gatype)
g_b = create_global_array(gatype)
a = create_local_a(gatype)
b = np.random.random_sample((n,n))
if MIRROR:
if 0 == iproc:
ga.put(g_b, b)
ga.put(g_a, a)
else:
if 0 == me:
ga.put(g_b, b)
ga.put(g_a, a)
ga.sync()
sum1 = np.sum(a*b)
sum2 = ga.dot(g_a, g_b)
if mismatch(sum1, sum2):
ga.error('dot wrong %s != %s' % (sum1, sum2))
if 0 == me:
print 'OK'
ga.destroy(g_a)
ga.destroy(g_b)
def srumma(g_a, g_b, g_c, chunk_size, multiplier, g_counter):
task_list = get_task_list(chunk_size, multiplier)
### get first integer from g_counter and assign to 'task_id'
# the srumma algorithm, more or less
task_prev = task_list[task_id]
a_prev,a_nb_prev = ga.nbget(g_a, task_prev.alo, task_prev.ahi)
b_prev,b_nb_prev = ga.nbget(g_b, task_prev.blo, task_prev.bhi)
### get next integer from g_counter and assign to 'task_id'
while task_id < multiplier**3:
task_next = task_list[task_id]
a_next,a_nb_next = ga.nbget(g_a, task_next.alo, task_next.ahi)
b_next,b_nb_next = ga.nbget(g_b, task_next.blo, task_next.bhi)
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev,b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
task_prev = task_next
a_prev,a_nb_prev = a_next,a_nb_next
b_prev,b_nb_prev = b_next,b_nb_next
### get next integer from g_counter and assign to 'task_id'
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev,b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
ga.sync()
def srumma(g_a, g_b, g_c, chunk_size, multiplier):
# statically partition the task list among nprocs
task_list = get_task_list(chunk_size, multiplier)
ntasks = multiplier**3 // nproc
start = me * ntasks
stop = (me + 1) * ntasks
if me + 1 == nproc:
stop += multiplier**3 % nproc
# the srumma algorithm, more or less
task_prev = task_list[start]
a_prev, a_nb_prev = ga.nbget(g_a, task_prev.alo, task_prev.ahi)
b_prev, b_nb_prev = ga.nbget(g_b, task_prev.blo, task_prev.bhi)
for i in range(start + 1, stop):
task_next = task_list[i]
a_next, a_nb_next = ga.nbget(g_a, task_next.alo, task_next.ahi)
b_next, b_nb_next = ga.nbget(g_b, task_next.blo, task_next.bhi)
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev, b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
task_prev = task_next
a_prev, a_nb_prev = a_next, a_nb_next
b_prev, b_nb_prev = b_next, b_nb_next
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev, b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
ga.sync()
def check_get(gatype):
"""check nloop random gets from each node"""
if 0 == me:
print '> Checking random get (%d calls)...' % nloop
g_a = create_global_array(gatype)
a = create_local_a(gatype)
if 0 == me:
ga.put(g_a, a)
ga.sync()
nwords = 0
random.seed(ga.nodeid()*51+1) # different seed for each proc
for loop in range(nloop):
ilo,ihi = random.randint(0, nloop-1),random.randint(0, nloop-1)
if ihi < ilo: ilo,ihi = ihi,ilo
jlo,jhi = random.randint(0, nloop-1),random.randint(0, nloop-1)
if jhi < jlo: jlo,jhi = jhi,jlo
nwords += (ihi-ilo+1)*(jhi-jlo+1)
ihi += 1
jhi += 1
result = ga.get(g_a, (ilo,jlo), (ihi,jhi))
if not np.all(result == a[ilo:ihi,jlo:jhi]):
ga.error('random get failed')
if 0 == me and loop % max(1,nloop/20) == 0:
print ' call %d node %d checking get((%d,%d),(%d,%d)) total %f' % (
loop, me, ilo, ihi, jlo, jhi, nwords)
if 0 == me:
print 'OK'
ga.destroy(g_a)
def srumma(g_a, g_b, g_c, chunk_size, multiplier, g_counter):
# statically partition the task list among nprocs
task_list = get_task_list(chunk_size, multiplier)
task_id = ga.read_inc(g_counter, 0)
# the srumma algorithm, more or less
task_prev = task_list[task_id]
a_prev, a_nb_prev = ga.nbget(g_a, task_prev.alo, task_prev.ahi)
b_prev, b_nb_prev = ga.nbget(g_b, task_prev.blo, task_prev.bhi)
task_id = ga.read_inc(g_counter, 0)
while task_id < multiplier**3:
task_next = task_list[task_id]
a_next, a_nb_next = ga.nbget(g_a, task_next.alo, task_next.ahi)
b_next, b_nb_next = ga.nbget(g_b, task_next.blo, task_next.bhi)
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev, b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
task_prev = task_next
a_prev, a_nb_prev = a_next, a_nb_next
b_prev, b_nb_prev = b_next, b_nb_next
task_id = ga.read_inc(g_counter, 0)
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev, b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
ga.sync()
def check_put_disjoint(gatype):
"""each node fills in disjoint sections of the array"""
if 0 == me:
print '> Checking disjoint put ...',
g_a = create_global_array(gatype)
a = create_local_a(gatype)
inc = (n-1)/20 + 1
ij = 0
for i in range(0,n,inc):
for j in range(0,n,inc):
check = False
if MIRROR:
check = ij % lprocs == iproc
else:
check = ij % nproc == me
if check:
lo = [i,j]
hi = [min(i+inc,n), min(j+inc,n)]
piece = a[ga.zip(lo,hi)]
ga.put(g_a, piece, lo, hi)
# the following check is not part of the original test.F
result = ga.get(g_a, lo, hi)
if not np.all(result == piece):
ga.error("put followed by get failed", 1)
ga.sync()
ij += 1
ga.sync()
# all nodes check all of a
b = ga.get(g_a)
if not np.all(a == b):
ga.error('put failed, exiting')
if 0 == me:
print 'OK'
ga.destroy(g_a)
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 srumma(g_a, g_b, g_c, chunk_size, multiplier, g_counter):
# statically partition the task list among nprocs
task_list = get_task_list(chunk_size, multiplier)
task_id = ga.read_inc(g_counter, 0)
# the srumma algorithm, more or less
task_prev = task_list[task_id]
a_prev,a_nb_prev = ga.nbget(g_a, task_prev.alo, task_prev.ahi)
b_prev,b_nb_prev = ga.nbget(g_b, task_prev.blo, task_prev.bhi)
task_id = ga.read_inc(g_counter, 0)
while task_id < multiplier**3:
task_next = task_list[task_id]
a_next,a_nb_next = ga.nbget(g_a, task_next.alo, task_next.ahi)
b_next,b_nb_next = ga.nbget(g_b, task_next.blo, task_next.bhi)
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev,b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
task_prev = task_next
a_prev,a_nb_prev = a_next,a_nb_next
b_prev,b_nb_prev = b_next,b_nb_next
task_id = ga.read_inc(g_counter, 0)
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev,b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
ga.sync()
def check_fence_and_lock(gatype):
if 0 == me:
print '> Checking ga.fence and ga.lock',
g_a = create_global_array(gatype)
ga.zero(g_a)
if not ga.create_mutexes(1):
ga.error('ga.create_mutexes failed')
if n < 2:
ga.error('insufficient n to test ga.fence', n)
ga.lock(0)
a = ga.get(g_a) # get original values
a[:,0] += 1 # add my contribution
# need to use fence to assure that coms complete before leaving
# critical section
ga.init_fence()
ga.put(g_a, a)
ga.fence()
ga.unlock(0)
if not ga.destroy_mutexes():
ga.error('mutex not destroyed')
ga.sync()
if 0 == me:
a = ga.get(g_a)
if not np.all(a[:,0] == nproc):
ga.error('fence failed')
if 0 == me:
print 'OK'
def srumma(g_a, g_b, g_c, chunk_size, multiplier):
# statically partition the task list among nprocs
task_list = get_task_list(chunk_size, multiplier)
ntasks = multiplier**3 // nproc
start = me*ntasks
stop = (me+1)*ntasks
if me+1 == nproc:
stop += multiplier**3 % nproc
# the srumma algorithm, more or less
task_prev = task_list[start]
a_prev,a_nb_prev = ga.nbget(g_a, task_prev.alo, task_prev.ahi)
b_prev,b_nb_prev = ga.nbget(g_b, task_prev.blo, task_prev.bhi)
for i in range(start+1,stop):
task_next = task_list[i]
a_next,a_nb_next = ga.nbget(g_a, task_next.alo, task_next.ahi)
b_next,b_nb_next = ga.nbget(g_b, task_next.blo, task_next.bhi)
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev,b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
task_prev = task_next
a_prev,a_nb_prev = a_next,a_nb_next
b_prev,b_nb_prev = b_next,b_nb_next
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev,b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
ga.sync()
def set_boundary_conditions_put(g_a):
# process 0 initializes global array
# this would only set the initial conditions since we are putting an entire
# zeros array with the outer elements changed
if rank == 0:
a = np.zeros((dim,dim), dtype=np.float32)
a[0,:] = 100 #top row
a[:,0] = 75 #left column
a[:,a.shape[0] - 1] = 50 #right column
ga.put(g_a, a)
ga.sync()
def matrix_multiply():
# Configure array dimensions. Force an unequal data distribution.
dims = [TOTALELEMS]*NDIM
chunk = [TOTALELEMS/nprocs-1]*NDIM
# Create a global array g_a and duplicate it to get g_b and g_c.
g_a = ga.create(ga.C_DBL, dims, "array A", chunk)
if not g_a: ga.error("create failed: A")
if not me: print "Created Array A"
g_b = ga.duplicate(g_a, "array B")
g_c = ga.duplicate(g_a, "array C")
if not g_b or not g_c: ga.eror("duplicate failed")
if not me: print "Created Arrays B and C"
# Initialize data in matrices a and b.
if not me: print "Initializing matrix A and B"
a = np.random.rand(*dims)*29
b = np.random.rand(*dims)*37
# Copy data to global arrays g_a and g_b.
if not me:
ga.put(g_a, a)
ga.put(g_b, b)
# Synchronize all processors to make sure everyone has data.
ga.sync()
# Determine which block of data is locally owned. Note that
# the same block is locally owned for all GAs.
lo,hi = ga.distribution(g_c)
# Get the blocks from g_a and g_b needed to compute this block in
# g_c and copy them into the local buffers a and b.
a = ga.get(g_a, (lo[0],0), (hi[0],dims[0]))
b = ga.get(g_b, (0,lo[1]), (dims[1],hi[1]))
# Do local matrix multiplication and store the result in local
# buffer c. Start by evaluating the transpose of b.
btrns = b.transpose()
# Multiply a and b to get c.
c = np.dot(a,b)
# Copy c back to g_c.
ga.put(g_c, c, lo, hi)
verify(g_a, g_b, g_c)
# Deallocate arrays.
ga.destroy(g_a)
ga.destroy(g_b)
ga.destroy(g_c)
def set_boundary_conditions_access(g_a):
# this will reset the outer (ghost) elements back to the boundary cond.
if rlo == 0 or clo == 0 or chi == dim:
a = ga.access_ghosts(g_a)
if rlo == 0:
a[0,:] = 100 # I own a top row
if clo == 0:
a[:,0] = 75 # I own a left column
if chi == dim:
a[:,-1] = 50 # I own a right column
ga.release_update_ghosts(g_a)
ga.sync()
def check_scale(gatype):
if 0 == me:
print '> Checking scale ...',
g_a = create_global_array(gatype)
a = create_local_a(gatype)
if 0 == me:
ga.put(g_a, a)
ga.sync()
ga.scale(g_a, 0.123)
a *= 0.123
if np.any(np.vectorize(mismatch)(a,ga.get(g_a))):
ga.error('add failed')
if 0 == me:
print 'OK'
ga.destroy(g_a)
def TRANSPOSE1D():
# Configure array dimensions. Force an unequal data distribution.
dims = [nprocs*TOTALELEMS + nprocs/2]
chunk = [TOTALELEMS] # minimum data on each process
# create a global array g_a and duplicate it to get g_b
g_a = ga.create(ga.C_INT, dims, "array A", chunk)
if not g_a: ga.error("create failed: A")
if not me: print "Created Array A"
g_b = ga.duplicate(g_a, "array B")
if not g_b: ga.error("duplicate failed")
if not me: print "Created Array B"
# initialize data in g_a
if not me:
print "Initializing matrix A"
ga.put(g_a, np.arange(dims[0], dtype=np.int32))
# Synchronize all processors to guarantee that everyone has data
# before proceeding to the next step.
ga.sync()
# Start initial phase of inversion by inverting the data held locally on
# each processor. Start by finding out which data each processor owns.
lo,hi = ga.distribution(g_a)
# Get locally held data and copy it into local buffer a
a = ga.get(g_a, lo, hi)
# Invert data locally
b = a[::-1]
# Invert data globally by copying locally inverted blocks into
# their inverted positions in the GA
ga.put(g_b, b, dims[0]-hi[0], dims[0]-lo[0])
# Synchronize all processors to make sure inversion is complete
ga.sync()
# Check to see if inversion is correct
if not me: verify(g_a, g_b)
# Deallocate arrays
ga.destroy(g_a)
ga.destroy(g_b)
def TRANSPOSE1D():
# Configure array dimensions. Force an unequal data distribution.
dims = [nprocs * TOTALELEMS + nprocs / 2]
chunk = [TOTALELEMS] # minimum data on each process
# create a global array g_a and duplicate it to get g_b
g_a = ga.create(ga.C_INT, dims, "array A", chunk)
if not g_a: ga.error("create failed: A")
if not me: print "Created Array A"
g_b = ga.duplicate(g_a, "array B")
if not g_b: ga.error("duplicate failed")
if not me: print "Created Array B"
# initialize data in g_a
if not me:
print "Initializing matrix A"
ga.put(g_a, np.arange(dims[0], dtype=np.int32))
# Synchronize all processors to guarantee that everyone has data
# before proceeding to the next step.
ga.sync()
# Start initial phase of inversion by inverting the data held locally on
# each processor. Start by finding out which data each processor owns.
lo, hi = ga.distribution(g_a)
# Get locally held data and copy it into local buffer a
a = ga.get(g_a, lo, hi)
# Invert data locally
b = a[::-1]
# Invert data globally by copying locally inverted blocks into
# their inverted positions in the GA
ga.put(g_b, b, dims[0] - hi[0], dims[0] - lo[0])
# Synchronize all processors to make sure inversion is complete
ga.sync()
# Check to see if inversion is correct
if not me: verify(g_a, g_b)
# Deallocate arrays
ga.destroy(g_a)
ga.destroy(g_b)
def TestPutGetAcc(g_a, n, chunk, buf, lo, hi, local):
if 0 == me:
print ''
if local:
print 'Local 2-D Array Section'
else:
print 'Remote 2-D Array Section'
print '%15s %19s %19s %19s' % ('section', 'get', 'put', 'accumulate')
print '%7s %7s %9s %9s %9s %9s %9s %9s' % (
'bytes','dim','usec','MB/s','usec','MB/s','usec','MB/s')
ga.sync()
bytes = 0
jump = 0
num_chunks = len(chunk)
for loop in range(num_chunks):
tg = 0.0
tp = 0.0
ta = 0.0
bytes = 8*chunk[loop]*chunk[loop] # how much data is accessed
jump = n/(60*(loop+1)) # jump between consecutive patches
if loop+1 == num_chunks:
jump = 0
# everybody touches own data
ga.fill(g_a, me*(loop+1), [0,0], [n,n])
if 0 == me:
tg = time_get(g_a, lo, hi, buf, chunk[loop], jump, local)
else:
time.sleep(1)
# everybody touches own data
ga.fill(g_a, me*(loop+1), [0,0], [n,n])
if 0 == me:
tp = time_put(g_a, lo, hi, buf, chunk[loop], jump, local)
else:
time.sleep(1)
# everybody touches own data
ga.fill(g_a, me*(loop+1), [0,0], [n,n])
if 0 == me:
ta = time_acc(g_a, lo, hi, buf, chunk[loop], jump, local)
else:
time.sleep(1)
if 0 == me:
print '%7d %7d %8.3e %8.3e %8.3e %8.3e %8.3e %8.3e' % (
bytes, chunk[loop],
tg/1e-6, 1e-6*bytes/tg,
tp/1e-6, 1e-6*bytes/tp,
ta/1e-6, 1e-6*bytes/ta)
def TestPutGetAcc(g_a, n, chunk, buf, lo, hi, local):
if 0 == me:
print ''
if local:
print 'Local 2-D Array Section'
else:
print 'Remote 2-D Array Section'
print '%15s %19s %19s %19s' % ('section', 'get', 'put', 'accumulate')
print '%7s %7s %9s %9s %9s %9s %9s %9s' % (
'bytes', 'dim', 'usec', 'MB/s', 'usec', 'MB/s', 'usec', 'MB/s')
ga.sync()
bytes = 0
jump = 0
num_chunks = len(chunk)
for loop in range(num_chunks):
tg = 0.0
tp = 0.0
ta = 0.0
bytes = 8 * chunk[loop] * chunk[loop] # how much data is accessed
jump = n / (60 * (loop + 1)) # jump between consecutive patches
if loop + 1 == num_chunks:
jump = 0
# everybody touches own data
ga.fill(g_a, me * (loop + 1), [0, 0], [n, n])
if 0 == me:
tg = time_get(g_a, lo, hi, buf, chunk[loop], jump, local)
else:
time.sleep(1)
# everybody touches own data
ga.fill(g_a, me * (loop + 1), [0, 0], [n, n])
if 0 == me:
tp = time_put(g_a, lo, hi, buf, chunk[loop], jump, local)
else:
time.sleep(1)
# everybody touches own data
ga.fill(g_a, me * (loop + 1), [0, 0], [n, n])
if 0 == me:
ta = time_acc(g_a, lo, hi, buf, chunk[loop], jump, local)
else:
time.sleep(1)
if 0 == me:
print '%7d %7d %8.3e %8.3e %8.3e %8.3e %8.3e %8.3e' % (
bytes, chunk[loop], tg / 1e-6, 1e-6 * bytes / tg, tp / 1e-6,
1e-6 * bytes / tp, ta / 1e-6, 1e-6 * bytes / ta)
def verify(g_a, g_b, g_c):
g_chk = ga.duplicate(g_a, "array check")
if not g_chk: ga.error("duplicate failed")
ga.sync()
ga.gemm(False, False, TOTALELEMS, TOTALELEMS, TOTALELEMS, 1.0, g_a, g_b,
0.0, g_chk);
ga.sync()
ga.add(g_c, g_chk, g_chk, 1.0, -1.0)
rchk = ga.dot(g_chk, g_chk)
if not me:
print "Normed difference in matrices: %12.4f" % rchk
if not (-TOLERANCE < rchk < TOLERANCE):
ga.error("Matrix multiply verify failed")
else:
print "Matrix Multiply OK"
ga.destroy(g_chk)
def check_accumulate_overlap(gatype):
if 0 == me:
print '> Checking overlapping accumulate ...',
g_a = create_global_array(gatype)
ga.zero(g_a)
ga.acc(g_a, [1], (n/2,n/2), (n/2+1,n/2+1), 1)
ga.sync()
if MIRROR:
if 0 == iproc:
x = abs(ga.get(g_a, (n/2,n/2), (n/2+1,n/2+1))[0,0] - lprocs)
if not 0 == x:
ga.error('overlapping accumulate failed -- expected %s got %s'%(
x, lprocs))
else:
if 0 == me:
x = abs(ga.get(g_a, (n/2,n/2), (n/2+1,n/2+1))[0,0] - nproc)
if not 0 == x:
ga.error('overlapping accumulate failed -- expected %s got %s'%(
x, nproc))
if 0 == me:
print 'OK'
ga.destroy(g_a)
def check_add(gatype):
if 0 == me:
print '> Checking add ...',
g_a = create_global_array(gatype)
g_b = create_global_array(gatype)
a = create_local_a(gatype)
b = create_local_b(gatype)
alpha = None
beta = None
if 0 == me:
ga.put(g_a, a)
ga.sync();
np.random.seed(12345) # everyone has same seed
if gatype in [ga.C_SCPL,ga.C_DCPL]:
b_real = np.random.random_sample((n,n))
b_imag = np.random.random_sample((n,n))
b[:] = np.vectorize(complex)(b_real,b_imag)
alpha = complex(0.1,-0.1)
beta = complex(0.9,-0.9)
else:
b[:] = np.random.random_sample((n,n))
alpha = 0.1
beta = 0.9
a = alpha*a + beta*b
if MIRROR:
if 0 == iproc:
ga.put(g_b, b)
else:
if 0 == me:
ga.put(g_b, b)
ga.sync()
ga.add(g_a, g_b, g_b, alpha, beta)
b = ga.get(g_b, buffer=b)
if np.any(np.vectorize(mismatch)(b,a)):
ga.error('add failed')
if 0 == me:
print 'OK'
ga.destroy(g_a)
ga.destroy(g_b)
def check_accumulate_disjoint(gatype):
"""Each node accumulates into disjoint sections of the array."""
if 0 == me:
print '> Checking disjoint accumulate ...',
g_a = create_global_array(gatype)
a = create_local_a(gatype)
b = np.fromfunction(lambda i,j: i+j+2, (n,n), dtype=ga.dtype(gatype))
if 0 == me:
ga.put(g_a, a)
ga.sync()
inc = (n-1)/20 + 1
ij = 0
for i in range(0,n,inc):
for j in range(0,n,inc):
x = 10.0
lo = [i,j]
hi = [min(i+inc,n), min(j+inc,n)]
piece = b[ga.zip(lo,hi)]
check = False
if MIRROR:
check = ij % lprocs == iproc
else:
check = ij % nproc == me
if check:
ga.acc(g_a, piece, lo, hi, x)
ga.sync()
ij += 1
# each process applies all updates to its local copy
a[ga.zip(lo,hi)] += x * piece
ga.sync()
# all nodes check all of a
if not np.all(ga.get(g_a) == a):
ga.error('acc failed')
if 0 == me:
print 'OK'
ga.destroy(g_a)
me = ga.nodeid()
nproc = ga.nnodes()
def parallel_task():
me = ga.pgroup_nodeid()
nproc = ga.pgroup_nnodes()
if not me:
print "This is process 0 on group %s" % ga.pgroup_get_default()
g_a = ga.create(ga.C_DBL, (3,4,5))
ga.randomize(g_a)
if me == 0:
print np.sum(ga.access(g_a))
midproc = nproc//2
proclist_first = range(0,midproc)
proclist_last = range(midproc,nproc)
group_id_first = ga.pgroup_create(proclist_first)
group_id_last = ga.pgroup_create(proclist_last)
if me in proclist_first:
ga.pgroup_set_default(group_id_first)
parallel_task()
ga.pgroup_set_default(ga.pgroup_get_world())
ga.sync()
if me in proclist_last:
ga.pgroup_set_default(group_id_last)
parallel_task()
ga.pgroup_set_default(ga.pgroup_get_world())
ga.sync()
if not me:
print "All done with groups"
"""Use ga.access() to sum locally per SMP node."""
import mpi4py.MPI
from ga4py import ga
import numpy as np
# Okay, we create the global array
g_a = ga.create(ga.C_DBL, (3, 4, 5, 6))
if world_id == 0:
ga.put(g_a, np.arange(3 * 4 * 5 * 6))
ga.sync()
# You're on your own!
def print_sync(obj):
for proc in range(size):
if rank == proc:
print "%d: %s" % (proc, obj)
ga.sync()
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.'
