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 time_put(g_a, lo, hi, buf, chunk, jump, local):
count = 0
rows = hi[0] - lo[0]
cols = hi[1] - lo[1]
shifti = [rows, 0, rows]
shiftj = [0, cols, cols]
seconds = time.time()
# distance between consecutive patches increased by jump
# to destroy locality of reference
for ilo in range(lo[0], hi[0] - chunk - jump + 1, chunk + jump):
ihi = ilo + chunk
for jlo in range(lo[1], hi[1] - chunk - jump + 1, chunk + jump):
jhi = jlo + chunk
count += 1
if local:
llo = [ilo, jlo]
lhi = [ihi, jhi]
ga.put(g_a, buf[ga.zip(llo, lhi)], llo, lhi)
else:
index = count % 3
llo = [ilo + shifti[index], jlo + shiftj[index]]
lhi = [ihi + shifti[index], jhi + shiftj[index]]
ga.put(g_a, buf[ilo:ihi, jlo:jhi], llo, lhi)
seconds = time.time() - seconds
return seconds / count
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_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 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 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 time_put(g_a, lo, hi, buf, chunk, jump, local):
count = 0
rows = hi[0]-lo[0]
cols = hi[1]-lo[1]
shifti = [rows, 0, rows]
shiftj = [0, cols, cols]
seconds = time.time()
# distance between consecutive patches increased by jump
# to destroy locality of reference
for ilo in range(lo[0], hi[0]-chunk-jump+1, chunk+jump):
ihi = ilo + chunk
for jlo in range(lo[1], hi[1]-chunk-jump+1, chunk+jump):
jhi = jlo + chunk
count += 1
if local:
llo = [ilo,jlo]
lhi = [ihi,jhi]
ga.put(g_a, buf[ga.zip(llo,lhi)], llo, lhi)
else:
index = count%3
llo = [ilo+shifti[index],jlo+shiftj[index]]
lhi = [ihi+shifti[index],jhi+shiftj[index]]
ga.put(g_a, buf[ilo:ihi,jlo:jhi], llo, lhi)
seconds = time.time() - seconds
return seconds/count
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 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 check_copy(gatype):
if 0 == me:
print '> Checking copy ...',
g_a = create_global_array(gatype)
g_b = create_global_array(gatype)
a = create_local_a(gatype)
if 0 == me:
ga.put(g_a, a)
ga.copy(g_a, g_b)
if not np.all(a == ga.get(g_b)):
ga.error('copy failed')
if 0 == me:
print 'OK'
ga.destroy(g_a)
ga.destroy(g_b)
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 time_put1(g_a, lo, hi, buf, chunk, jump, local):
count = 0
rows = hi[0]-lo[0]
shift = [rows, 2*rows, 3*rows]
seconds = time.time()
# distance between consecutive patches increased by jump
# to destroy locality of reference
for ilo in range(lo[0], hi[0]-chunk-jump+1, chunk+jump):
ihi = ilo+chunk
count += 1
if local:
ga.put(g_a, buf[ilo:ihi], [ilo], [ihi])
else:
index = count%3
llo = ilo+shift[index]
lhi = ihi+shift[index]
ga.put(g_a, buf[ilo:ihi], llo, lhi)
seconds = time.time() - seconds
return seconds/count
def time_put1(g_a, lo, hi, buf, chunk, jump, local):
count = 0
rows = hi[0] - lo[0]
shift = [rows, 2 * rows, 3 * rows]
seconds = time.time()
# distance between consecutive patches increased by jump
# to destroy locality of reference
for ilo in range(lo[0], hi[0] - chunk - jump + 1, chunk + jump):
ihi = ilo + chunk
count += 1
if local:
ga.put(g_a, buf[ilo:ihi], [ilo], [ihi])
else:
index = count % 3
llo = ilo + shift[index]
lhi = ihi + shift[index]
ga.put(g_a, buf[ilo:ihi], llo, lhi)
seconds = time.time() - seconds
return seconds / count
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)
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 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)
print "\nSuccessfully created Global Array"
# Initialize data in GA. Find data owned by neighboring processor
nghbr = (me+1)%nprocs
lo,hi = ga.distribution(g_a, nghbr)
# Create data in local buffer, assign unique value for each data element
patch_shape = hi-lo
a_buf = np.fromfunction(lambda i,j: j*NSIZE + i,
patch_shape, dtype=ga.dtype(ga.C_INT))
a_buf += lo[1,np.newaxis]
a_buf += lo[np.newaxis,0]*dims[0]
# Copy local data to GA
ga.put(g_a, a_buf, lo, hi)
ga.sync()
if me == 0:
print "\nCopied values into Global Array from local buffer\n"
# Check data in GA to see if it is correct. Find data owned by this
# processor and then copy it to local buffer
lo,hi = ga.distribution(g_a, me)
b_buf = ga.get(g_a, lo, hi)
if me == 0:
print "\nCopied values from Global Array to local buffer\n"
# Verify that data is correct
patch_shape = hi-lo
c_buf = np.fromfunction(lambda i,j: j*NSIZE + i,
patch_shape, dtype=ga.dtype(ga.C_INT))
"""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!
"""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!
frames_seg = np.zeros([size, 2], dtype=int)
for iblock in range(size):
frames_seg[iblock, :] = iblock * bsize, (iblock + 1) * bsize
d = dict([key, frames_seg[key]] for key in range(size))
start, stop = d[rank][0], d[rank][1]
# Block-RMSD in Parallel
start3 = time.time()
out = block_rmsd(index, topology, trajectory, xref0)
# Communication
start4 = time.time()
print(np.shape(out[0]), start, stop)
ga.put(g_a, out[0], (start, 0), (stop, 2))
start5 = time.time()
if rank == 0:
buf = ga.get(g_a, lo=None, hi=None)
start6 = time.time()
if rank == 0:
data = np.zeros([size, 5], dtype=float)
else:
data = None
comm.Gather(np.array(out[1:], dtype=float), data, root=0)
start7 = time.time()
v = np.dot(a,b)
val = int(np.abs(np.sum(c-v))>0.0001)
val = ga.gop_add(val)
return val == 0
if __name__ == '__main__':
if nproc > MULTIPLIER**3:
if 0 == me:
print "You must use less than %s processors" % (MULTIPLIER**3+1)
else:
g_a = ga.create(ga.C_DBL, [N,N])
g_b = ga.create(ga.C_DBL, [N,N])
g_c = ga.create(ga.C_DBL, [N,N])
# put some fake data into input arrays A and B
if me == 0:
ga.put(g_a, np.random.random(N*N))
ga.put(g_b, np.random.random(N*N))
ga.sync()
if me == 0:
print "srumma...",
srumma(g_a, g_b, g_c, CHUNK_SIZE, MULTIPLIER)
if me == 0:
print "done"
if me == 0:
print "verifying using ga.gemm...",
ok = verify_using_ga(g_a, g_b, g_c)
if me == 0:
if ok:
print "OKAY"
else:
print "FAILED"
def load_ga(handle, h0, num_m, num_k):
if True:
ga.put(handle, h0[:num_m*num_k])
else:
a = np.arange(num_m*num_k, dtype=np.float64)
ga.put(handle, a)
return value < EPSILON
# create GA, distribute entire rows
g_a = ga.create(ga.C_FLOAT, (dim,dim), chunk=(0,dim))
# create a duplicate GA for the convergence test
g_b = ga.duplicate(g_a)
# process 0 initializes global array
# Note: alternatively, each process could initialize its local data using
# ga.access() and ga.distribution()
a = np.zeros((dim,dim), dtype=np.float32)
if rank == 0:
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()
# which piece of array do I own?
# note that rhi and chi follow python range conventions i.e. [lo,hi)
(rlo,clo),(rhi,chi) = ga.distribution(g_a)
iteration = 0
start = ga.wtime()
while True:
iteration += 1
if iteration % HOW_MANY_STEPS_BEFORE_CONVERGENCE_TEST == 0:
# check for convergence will occur, so make a copy of the GA
ga.sync()
ga.copy(g_a, g_b)
# the iteration
return val == 0
if __name__ == '__main__':
if nproc > MULTIPLIER**3:
if 0 == me:
print "You must use less than %s processors" % (MULTIPLIER**3 + 1)
else:
g_a = ga.create(ga.C_DBL, [N, N])
g_b = ga.create(ga.C_DBL, [N, N])
g_c = ga.create(ga.C_DBL, [N, N])
g_counter = ga.create(ga.C_INT, [1])
ga.zero(g_counter)
# put some fake data into input arrays A and B
if me == 0:
ga.put(g_a, np.random.random(N * N))
ga.put(g_b, np.random.random(N * N))
ga.sync()
if me == 0:
print "srumma...",
srumma(g_a, g_b, g_c, CHUNK_SIZE, MULTIPLIER, g_counter)
if me == 0:
print "done"
if me == 0:
print "verifying using ga.gemm...",
ok = verify_using_ga(g_a, g_b, g_c)
if me == 0:
if ok:
print "OKAY"
else:
print "FAILED"