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 time_acc(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.acc(g_a, buf[ga.zip(llo, lhi)], llo, lhi, 1)
else:
index = count % 3
llo = [ilo + shifti[index], jlo + shiftj[index]]
lhi = [ihi + shifti[index], jhi + shiftj[index]]
ga.acc(g_a, buf[ilo:ihi, jlo:jhi], llo, lhi, 1)
seconds = time.time() - seconds
return seconds / count
def time_acc(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.acc(g_a, buf[ga.zip(llo,lhi)], llo, lhi, 1)
else:
index = count%3
llo = [ilo+shifti[index],jlo+shiftj[index]]
lhi = [ihi+shifti[index],jhi+shiftj[index]]
ga.acc(g_a, buf[ilo:ihi,jlo:jhi], llo, lhi, 1)
seconds = time.time() - seconds
return seconds/count
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 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 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 time_acc1(g_a, lo, hi, buf, chunk, jump, local):
# Note: differs from test.F because the passed buffer must be the same
# size/shape as the patch. The slicing should be fast as the buffer is 1D
# and contiguous (and so is the slice).
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.acc(g_a, buf[ilo:ihi], [ilo], [ihi], 1.0)
else:
index = count%3
ga.acc(g_a, buf[ilo:ihi], ilo+shift[index], ihi+shift[index], 1.0)
seconds = time.time() - seconds
return seconds/count
def time_acc1(g_a, lo, hi, buf, chunk, jump, local):
# Note: differs from test.F because the passed buffer must be the same
# size/shape as the patch. The slicing should be fast as the buffer is 1D
# and contiguous (and so is the slice).
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.acc(g_a, buf[ilo:ihi], [ilo], [ihi], 1.0)
else:
index = count % 3
ga.acc(g_a, buf[ilo:ihi], ilo + shift[index], ihi + shift[index],
1.0)
seconds = time.time() - seconds
return seconds / count
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_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)
nprocs = ga.nnodes()
myrank = ga.nodeid()
g_pi = ga.create(ga.C_DBL, [1])
one_time = False
if len(sys.argv) == 2:
n = int(sys.argv[1])
one_time = True
while True:
if not one_time:
if myrank == 0:
n = get_n()
n = ga.brdcst(n)
else:
n = ga.brdcst(0)
if n == 0:
break
ga.zero(g_pi)
mypi = comp_pi(n, myrank, nprocs)
ga.acc(g_pi, mypi)
ga.sync()
if myrank == 0:
pi = ga.get(g_pi)[0]
prn_pi(pi, PI)
if one_time:
break
ga.destroy(g_pi)
nprocs = ga.nnodes()
myrank = ga.nodeid()
g_pi = ga.create(ga.C_DBL, [1])
one_time = False
if len(sys.argv) == 2:
n = int(sys.argv[1])
one_time = True
while True:
if not one_time:
if myrank == 0:
n = get_n()
n = ga.brdcst(n)
else:
n = ga.brdcst(0)
if n == 0:
break
ga.zero(g_pi)
mypi = comp_pi(n, myrank, nprocs)
ga.acc(g_pi, mypi)
ga.sync()
if myrank == 0:
pi = ga.get(g_pi)[0]
prn_pi(pi, PI)
if one_time:
break
ga.destroy(g_pi)