def main():
d = Domain(10)
print(d, d.bounds)
t = 0
for x in d:
print(x)
t += int(x)
assert t == 45
d2 = Domain([3, 3], [1, 1])
print(d2, d2.bounds)
t2 = 0
for x in d2:
print(x)
t2 += x[0] * x[1]
assert t2 == 36
d3 = Domain([4, 5, 6], [-1, 0, 1])
print(d3, d3.bounds)
t3 = 0
for x in d3:
t3 += x[0] * x[1] * x[2]
assert t3 == 420
def main():
futures = []
for i in IndexLaunch(10):
futures.append(hi(i))
for i, future in enumerate(futures):
print("got %s" % future.get())
assert int(future.get()) == i
# Same in 2 dimensions.
futures = []
for point in IndexLaunch([3, 3]):
futures.append(hi(point))
for i, point in enumerate(Domain([3, 3])):
assert futures[i].get() == point
R = Region([4, 4], {'x': legion.float64})
P = Partition.equal(R, [2, 2])
legion.fill(R, 'x', 0)
for i in IndexLaunch([2, 2]):
hello(R, i)
for i in IndexLaunch([2, 2]):
hello(P[i], i)
# Again, with a more explicit syntax.
# ID is the name of the (implicit) loop variable.
futures = index_launch([3, 3], hi, ID)
for point in Domain([3, 3]):
assert futures[point].get() == point
index_launch([2, 2], hello, R, ID)
index_launch([2, 2], hello, P[ID], ID)
def execute_timestep(graph, num_fields, timestep, result, primary, secondary,
scratch, p_scratch, dset_max_args):
fin = timestep % num_fields
offset = c.task_graph_offset_at_timestep(graph, timestep)
width = c.task_graph_width_at_timestep(graph, timestep)
dset = c.task_graph_dependence_set_at_timestep(graph, timestep)
colors = Domain([width], [offset])
num_args = dset_max_args[dset]
if scratch is None:
point_task = execute_point_tasks[fin][num_args]
else:
point_task = execute_point_tasks_scratch[fin][num_args]
point_task_args = [
graph if use_native else encode_task_graph(graph),
num_fields,
timestep,
0 if use_native else ID,
primary[ID],
]
if scratch is not None:
point_task_args.append(p_scratch[ID])
elif not use_native:
point_task_args.append(None)
point_task_args.extend(secondary[dset][arg][ID] for arg in range(num_args))
index_launch(colors, point_task, *point_task_args)
def main():
R = Region([4, 4], {'x': legion.float64})
legion.fill(R, 'x', 0)
# Create a partition of R.
colors = [2, 2]
transform = [[2, 0], [0, 2]]
extent = [2, 2]
P = Partition.restrict(R, colors, transform, extent)
# Again, with different parameters.
colors2 = [3]
transform2 = [[1], [2]]
extent2 = Domain([2, 2], [-1, -1])
P2 = Partition.restrict(R, colors2, transform2, extent2)
assert P.color_space.volume == 4
assert P2.color_space.volume == 3
# Grab a subregion of P.
R00 = P[0, 0]
print('Parent region has volume %s' % R.ispace.volume)
assert R.ispace.volume == 16
assert check_subregion(R00).get() == 4
for Rij in P:
assert check_subregion(Rij).get() == 4
assert check_subregion(P2[0]).get() == 1
assert check_subregion(P2[1]).get() == 4
assert check_subregion(P2[2]).get() == 2
def main():
conf = parse_args(legion.input_args(True))
assert conf.num_pieces % conf.pieces_per_superpiece == 0, "pieces should be evenly distributed to superpieces"
conf.shared_nodes_per_piece = int(
math.ceil(conf.nodes_per_piece * conf.pct_shared_nodes / 100.0))
print(
"circuit settings: loops=%d prune=%d pieces=%d (pieces/superpiece=%d) nodes/piece=%d (nodes/piece=%d) wires/piece=%d pct_in_piece=%d seed=%d"
% (conf.num_loops, conf.prune, conf.num_pieces,
conf.pieces_per_superpiece, conf.nodes_per_piece,
conf.shared_nodes_per_piece, conf.wires_per_piece,
conf.pct_wire_in_piece, conf.random_seed))
num_pieces = conf.num_pieces
num_superpieces = conf.num_pieces // conf.pieces_per_superpiece
num_circuit_nodes = num_pieces * conf.nodes_per_piece
num_circuit_wires = num_pieces * conf.wires_per_piece
node = Fspace.create(
OrderedDict([
('node_cap', legion.float32),
('leakage', legion.float32),
('charge', legion.float32),
('node_voltage', legion.float32),
]))
wire = Fspace.create(
OrderedDict([
('in_ptr', legion.int64),
('in_ptr_r', legion.uint8),
('out_ptr', legion.int64),
('out_ptr_r', legion.uint8),
('inductance', legion.float32),
('resistance', legion.float32),
('wire_cap', legion.float32),
] + [('current_%d' % i, legion.float32)
for i in range(10)] + [('voltage_%d' % i, legion.float32)
for i in range(9)]))
all_nodes = Region.create([num_circuit_nodes], node)
all_wires = Region.create([num_circuit_wires], wire)
node_size = np.dtype(list(
map(lambda x: (x[0], x[1].numpy_type), node.field_types.items())),
align=True).itemsize
wire_size = np.dtype(list(
map(lambda x: (x[0], x[1].numpy_type), wire.field_types.items())),
align=True).itemsize
print("Circuit memory usage:")
print(" Nodes : %10d * %4d bytes = %12d bytes" %
(num_circuit_nodes, node_size, num_circuit_nodes * node_size))
print(" Wires : %10d * %4d bytes = %12d bytes" %
(num_circuit_wires, wire_size, num_circuit_wires * wire_size))
total = ((num_circuit_nodes * node_size) + (num_circuit_wires * wire_size))
print(" Total %12d bytes" % total)
snpp = conf.shared_nodes_per_piece
pnpp = conf.nodes_per_piece - conf.shared_nodes_per_piece
pps = conf.pieces_per_superpiece
num_shared_nodes = num_pieces * snpp
privacy_coloring = Region.create([2], {'rect': legion.rect1d})
np.copyto(privacy_coloring.rect,
np.array([(num_shared_nodes, num_circuit_nodes - 1),
(0, num_shared_nodes - 1)],
dtype=privacy_coloring.rect.dtype),
casting='no')
privacy_part = Partition.create_by_restriction(privacy_coloring, [2],
np.eye(1), [1],
disjoint_complete)
all_nodes_part = Partition.create_by_image(all_nodes, privacy_part, 'rect',
[2], disjoint_complete)
all_private = all_nodes_part[0]
all_shared = all_nodes_part[1]
launch_domain = Ispace.create([num_superpieces])
private_part = Partition.create_by_restriction(
all_private, launch_domain,
np.eye(1) * pnpp * pps,
Domain.create([pnpp * pps], [num_shared_nodes]), disjoint_complete)
shared_part = Partition.create_by_restriction(all_shared, launch_domain,
np.eye(1) * snpp * pps,
[snpp * pps],
disjoint_complete)
wires_part = Partition.create_equal(all_wires, launch_domain)
ghost_ranges = Region.create([num_superpieces],
OrderedDict([('rect', legion.rect1d)]))
ghost_ranges_part = Partition.create_equal(ghost_ranges, launch_domain)
for i in IndexLaunch(launch_domain):
init_piece(int(i), conf[0], ghost_ranges_part[i], private_part[i],
shared_part[i], all_shared, wires_part[i])
ghost_part = Partition.create_by_image(all_shared, ghost_ranges_part,
'rect', launch_domain)
for i in IndexLaunch(launch_domain):
init_pointers(private_part[i], shared_part[i], ghost_part[i],
wires_part[i])
steps = conf.steps
prune = conf.prune
num_loops = conf.num_loops + 2 * prune
for j in range(num_loops):
for i in IndexLaunch(launch_domain):
calculate_new_currents(j == prune, steps, private_part[i],
shared_part[i], ghost_part[i],
wires_part[i])
for i in IndexLaunch(launch_domain):
distribute_charge(private_part[i], shared_part[i], ghost_part[i],
wires_part[i])
for i in IndexLaunch(launch_domain):
update_voltages(j == num_loops - prune - 1, private_part[i],
shared_part[i])
def main():
print_once('Running circuit_sparse.py')
conf = parse_args(legion.input_args(True))
assert conf.num_pieces % conf.pieces_per_superpiece == 0, "pieces should be evenly distributed to superpieces"
conf.shared_nodes_per_piece = int(
math.ceil(conf.nodes_per_piece * conf.pct_shared_nodes / 100.0))
print_once(
"circuit settings: loops=%d prune=%d pieces=%d (pieces/superpiece=%d) nodes/piece=%d (nodes/piece=%d) wires/piece=%d pct_in_piece=%d seed=%d"
% (conf.num_loops, conf.prune, conf.num_pieces,
conf.pieces_per_superpiece, conf.nodes_per_piece,
conf.shared_nodes_per_piece, conf.wires_per_piece,
conf.pct_wire_in_piece, conf.random_seed))
num_pieces = conf.num_pieces
num_superpieces = conf.num_pieces // conf.pieces_per_superpiece
num_circuit_nodes = num_pieces * conf.nodes_per_piece
num_circuit_wires = num_pieces * conf.wires_per_piece
node = Fspace(
OrderedDict([
('node_cap', legion.float32),
('leakage', legion.float32),
('charge', legion.float32),
('node_voltage', legion.float32),
]))
wire = Fspace(
OrderedDict([
('in_ptr', legion.int64),
('in_ptr_r', legion.uint8),
('out_ptr', legion.int64),
('out_ptr_r', legion.uint8),
('inductance', legion.float32),
('resistance', legion.float32),
('wire_cap', legion.float32),
] + [('current_%d' % i, legion.float32)
for i in range(WIRE_SEGMENTS)] +
[('voltage_%d' % i, legion.float32)
for i in range(WIRE_SEGMENTS - 1)]))
all_nodes = Region([num_circuit_nodes], node)
all_wires = Region([num_circuit_wires], wire)
node_size = np.dtype(list(
map(lambda x: (x[0], x[1].numpy_type), node.field_types.items())),
align=True).itemsize
wire_size = np.dtype(list(
map(lambda x: (x[0], x[1].numpy_type), wire.field_types.items())),
align=True).itemsize
print_once("Circuit memory usage:")
print_once(" Nodes : %10d * %4d bytes = %12d bytes" %
(num_circuit_nodes, node_size, num_circuit_nodes * node_size))
print_once(" Wires : %10d * %4d bytes = %12d bytes" %
(num_circuit_wires, wire_size, num_circuit_wires * wire_size))
total = ((num_circuit_nodes * node_size) + (num_circuit_wires * wire_size))
print_once(" Total %12d bytes" % total)
snpp = conf.shared_nodes_per_piece
pnpp = conf.nodes_per_piece - conf.shared_nodes_per_piece
pps = conf.pieces_per_superpiece
num_shared_nodes = num_pieces * snpp
privacy_coloring = Region([2], {'rect': legion.rect1d})
np.copyto(privacy_coloring.rect,
np.array([(num_shared_nodes, num_circuit_nodes - 1),
(0, num_shared_nodes - 1)],
dtype=privacy_coloring.rect.dtype),
casting='no')
privacy_part = Partition.restrict(privacy_coloring, [2], np.eye(1), [1],
disjoint_complete)
all_nodes_part = Partition.image(all_nodes, privacy_part, 'rect', [2],
disjoint_complete)
all_private = all_nodes_part[0]
all_shared = all_nodes_part[1]
launch_domain = Ispace([num_superpieces])
private_part = Partition.restrict(all_private, launch_domain,
np.eye(1) * pnpp * pps,
Domain([pnpp * pps], [num_shared_nodes]),
disjoint_complete)
shared_part = Partition.restrict(all_shared, launch_domain,
np.eye(1) * snpp * pps, [snpp * pps],
disjoint_complete)
wires_part = Partition.equal(all_wires, launch_domain)
ghost_ranges = Region([num_superpieces],
OrderedDict([('rect', legion.rect1d)]))
ghost_ranges_part = Partition.equal(ghost_ranges, launch_domain)
if _constant_time_launches:
c = Future(conf[0], value_type=Config)
index_launch(launch_domain, init_piece, ID, c, ghost_ranges_part[ID],
private_part[ID], shared_part[ID], all_shared,
wires_part[ID])
else:
for i in IndexLaunch(launch_domain):
init_piece(i, conf[0], ghost_ranges_part[i], private_part[i],
shared_part[i], all_shared, wires_part[i])
ghost_part = Partition.image(all_shared, ghost_ranges_part, 'rect',
launch_domain)
if _constant_time_launches:
index_launch(launch_domain, init_pointers, private_part[ID],
shared_part[ID], ghost_part[ID], wires_part[ID])
else:
for i in IndexLaunch(launch_domain):
init_pointers(private_part[i], shared_part[i], ghost_part[i],
wires_part[i])
steps = conf.steps
prune = conf.prune
num_loops = conf.num_loops + 2 * prune
trace = Trace()
for j in range(num_loops):
if j == prune:
legion.execution_fence(block=True)
start_time = legion.c.legion_get_current_time_in_nanos()
with trace:
if _constant_time_launches:
index_launch(launch_domain, calculate_new_currents, False,
steps, private_part[ID], shared_part[ID],
ghost_part[ID], wires_part[ID])
index_launch(launch_domain, distribute_charge,
private_part[ID], shared_part[ID], ghost_part[ID],
wires_part[ID])
index_launch(launch_domain, update_voltages, False,
private_part[ID], shared_part[ID])
else:
for i in IndexLaunch(launch_domain):
calculate_new_currents(False, steps, private_part[i],
shared_part[i], ghost_part[i],
wires_part[i])
for i in IndexLaunch(launch_domain):
distribute_charge(private_part[i], shared_part[i],
ghost_part[i], wires_part[i])
for i in IndexLaunch(launch_domain):
update_voltages(False, private_part[i], shared_part[i])
if j == num_loops - prune - 1:
legion.execution_fence(block=True)
stop_time = legion.c.legion_get_current_time_in_nanos()
sim_time = (stop_time - start_time) / 1e9
print_once('ELAPSED TIME = %7.3f s' % sim_time)
# Compute the floating point operations per second
num_circuit_nodes = conf.num_pieces * conf.nodes_per_piece
num_circuit_wires = conf.num_pieces * conf.wires_per_piece
# calculate currents
operations = num_circuit_wires * (WIRE_SEGMENTS * 6 +
(WIRE_SEGMENTS - 1) * 4) * conf.steps
# distribute charge
operations += (num_circuit_wires * 4)
# update voltages
operations += (num_circuit_nodes * 4)
# multiply by the number of loops
operations *= conf.num_loops
# Compute the number of gflops
gflops = (1e-9 * operations) / sim_time
print_once("GFLOPS = %7.3f GFLOPS" % gflops)