def main():
conf = parse_args(legion.input_args(True))
nbloated = np.array([conf.nx, conf.ny])
nt = np.array([conf.ntx, conf.nty])
init = conf.init
n = nbloated - 2 * RADIUS
assert np.all(n >= nt), "grid too small"
grid = Ispace.create(n + nt * 2 * RADIUS)
tiles = Ispace.create(nt)
point = Fspace.create(OrderedDict([
('input', DTYPE),
('output', DTYPE),
]))
points = Region.create(grid, point)
private = make_private_partition(points, tiles, n, nt)
interior = make_interior_partition(points, tiles, n, nt)
exterior = make_exterior_partition(points, tiles, n, nt)
xm = Region.create([nt[0] * RADIUS, n[1]], point)
xp = Region.create([nt[0] * RADIUS, n[1]], point)
ym = Region.create([n[0], nt[1] * RADIUS], point)
yp = Region.create([n[0], nt[1] * RADIUS], point)
pxm_in = make_ghost_x_partition(xm, tiles, n, nt, -1)
pxp_in = make_ghost_x_partition(xp, tiles, n, nt, 1)
pym_in = make_ghost_y_partition(ym, tiles, n, nt, -1)
pyp_in = make_ghost_y_partition(yp, tiles, n, nt, 1)
pxm_out = make_ghost_x_partition(xm, tiles, n, nt, 0)
pxp_out = make_ghost_x_partition(xp, tiles, n, nt, 0)
pym_out = make_ghost_y_partition(ym, tiles, n, nt, 0)
pyp_out = make_ghost_y_partition(yp, tiles, n, nt, 0)
init = conf.init
for r in [points, xm, xp, ym, yp]:
for f in ['input', 'output']:
legion.fill(r, f, init)
tsteps = conf.tsteps
tprune = conf.tprune
for t in range(tsteps):
for i in IndexLaunch(tiles):
stencil(private[i], interior[i], pxm_in[i], pxp_in[i], pym_in[i],
pyp_in[i], t == tprune)
for i in IndexLaunch(tiles):
increment(private[i], exterior[i], pxm_out[i], pxp_out[i],
pym_out[i], pyp_out[i], t == tsteps - tprune - 1)
for i in IndexLaunch(tiles):
check(private[i], interior[i], tsteps, init)
def main():
app = app_create(legion.input_args())
graphs = app_task_graphs(app)
if once_only():
c.app_display(app)
num_fields = max_fields
result, primary, secondary, scratch, p_scratch, dset_max_args = init_partitions(
graphs, num_fields)
legion.execution_fence(block=True)
start_time = legion.c.legion_get_current_time_in_nanos()
execute_main_loop(graphs, num_fields, result, primary, secondary, scratch,
p_scratch, dset_max_args)
legion.execution_fence(block=True)
stop_time = legion.c.legion_get_current_time_in_nanos()
total_time = (stop_time - start_time) / 1e9
if once_only():
c.app_report_timing(app, total_time)
def main():
print_once('Running stencil.py')
conf = parse_args(legion.input_args(True))
nbloated = np.array([conf.nx, conf.ny])
nt = np.array([conf.ntx, conf.nty])
init = conf.init
n = nbloated - 2*RADIUS
assert np.all(n >= nt), "grid too small"
grid = Ispace(n + nt*2*RADIUS)
tiles = Ispace(nt)
point = Fspace(OrderedDict([
('input', DTYPE),
('output', DTYPE),
]))
points = Region(grid, point)
private = make_private_partition(points, tiles, n, nt)
interior = make_interior_partition(points, tiles, n, nt)
exterior = make_exterior_partition(points, tiles, n, nt)
xm = Region([nt[0]*RADIUS, n[1]], point)
xp = Region([nt[0]*RADIUS, n[1]], point)
ym = Region([n[0], nt[1]*RADIUS], point)
yp = Region([n[0], nt[1]*RADIUS], point)
pxm_in = make_ghost_x_partition(xm, tiles, n, nt, -1)
pxp_in = make_ghost_x_partition(xp, tiles, n, nt, 1)
pym_in = make_ghost_y_partition(ym, tiles, n, nt, -1)
pyp_in = make_ghost_y_partition(yp, tiles, n, nt, 1)
pxm_out = make_ghost_x_partition(xm, tiles, n, nt, 0)
pxp_out = make_ghost_x_partition(xp, tiles, n, nt, 0)
pym_out = make_ghost_y_partition(ym, tiles, n, nt, 0)
pyp_out = make_ghost_y_partition(yp, tiles, n, nt, 0)
init = conf.init
for r in [points, xm, xp, ym, yp]:
for f in ['input', 'output']:
legion.fill(r, f, init)
tsteps = conf.tsteps + 2 * conf.tprune
tprune = conf.tprune
trace = Trace()
for t in range(tsteps):
if t == tprune:
legion.execution_fence(block=True)
start_time = legion.c.legion_get_current_time_in_nanos()
with trace:
if _constant_time_launches:
index_launch(tiles, stencil, private[ID], interior[ID], pxm_in[ID], pxp_in[ID], pym_in[ID], pyp_in[ID], False)
index_launch(tiles, increment, private[ID], exterior[ID], pxm_out[ID], pxp_out[ID], pym_out[ID], pyp_out[ID], False)
else:
for i in IndexLaunch(tiles):
stencil(private[i], interior[i], pxm_in[i], pxp_in[i], pym_in[i], pyp_in[i], False)
for i in IndexLaunch(tiles):
increment(private[i], exterior[i], pxm_out[i], pxp_out[i], pym_out[i], pyp_out[i], False)
if t == tsteps - tprune - 1:
legion.execution_fence(block=True)
stop_time = legion.c.legion_get_current_time_in_nanos()
if _constant_time_launches:
index_launch(tiles, check, private[ID], interior[ID], tsteps, init)
else:
for i in IndexLaunch(tiles):
check(private[i], interior[i], tsteps, init)
print_once('ELAPSED TIME = %7.3f s' % ((stop_time - start_time)/1e9))
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)