def test_ispd2005(design, algorithm, device_str):
with gzip.open(design, "rb") as f:
if sys.version_info[0] < 3:
data_collections = pickle.load(f)
else:
data_collections = pickle.load(f, encoding='bytes')
node_size_x = data_collections[0]
node_size_y = data_collections[1]
flat_net2pin_map = data_collections[2]
flat_net2pin_start_map = data_collections[3]
pin2net_map = data_collections[4]
flat_node2pin_map = data_collections[5]
flat_node2pin_start_map = data_collections[6]
pin2node_map = data_collections[7]
pin_offset_x = data_collections[8]
pin_offset_y = data_collections[9]
net_mask_ignore_large_degrees = data_collections[10]
xl = data_collections[11]
yl = data_collections[12]
xh = data_collections[13]
yh = data_collections[14]
site_width = data_collections[15]
row_height = data_collections[16]
num_bins_x = data_collections[17]
num_bins_y = data_collections[18]
num_movable_nodes = data_collections[19]
num_filler_nodes = data_collections[20]
pos = data_collections[21]
#net_mask = net_mask_ignore_large_degrees
net_mask = np.ones_like(net_mask_ignore_large_degrees)
for i in range(1, len(flat_net2pin_start_map)):
degree = flat_net2pin_start_map[i] - flat_net2pin_start_map[i - 1]
if degree > 100:
net_mask[i - 1] = 0
net_mask = torch.from_numpy(net_mask)
#max_node_degree = 0
#for i in range(1, len(flat_node2pin_start_map)):
# if i <= num_movable_nodes:
# max_node_degree = max(max_node_degree, flat_node2pin_start_map[i]-flat_node2pin_start_map[i-1])
#print("max node degree %d" % (max_node_degree))
device = torch.device(device_str)
#if device.type == 'cuda':
# num_bins_x = int(2**(math.ceil(np.log2(math.sqrt(num_movable_nodes / 512)))))
# num_bins_y = num_bins_x
print("bins %dx%d" % (num_bins_x, num_bins_y))
print("num_movable_nodes %d, num_nodes %d" %
(num_movable_nodes, node_size_x.numel() - num_filler_nodes))
# test cpu/cuda
custom = global_swap.GlobalSwap(
node_size_x=node_size_x.float().to(device),
node_size_y=node_size_y.float().to(device),
flat_net2pin_map=flat_net2pin_map.to(device),
flat_net2pin_start_map=flat_net2pin_start_map.to(device),
pin2net_map=pin2net_map.to(device),
flat_node2pin_map=flat_node2pin_map.to(device),
flat_node2pin_start_map=flat_node2pin_start_map.to(device),
pin2node_map=pin2node_map.to(device),
pin_offset_x=pin_offset_x.float().to(device),
pin_offset_y=pin_offset_y.float().to(device),
net_mask=net_mask.to(device),
xl=xl,
yl=yl,
xh=xh,
yh=yh,
site_width=site_width,
row_height=row_height,
num_bins_x=num_bins_x // 2,
num_bins_y=num_bins_y // 2,
num_movable_nodes=num_movable_nodes,
num_filler_nodes=num_filler_nodes,
batch_size=256,
max_iters=2,
algorithm=algorithm,
num_threads=20)
result = custom(pos.float().to(device))
def build_detailed_placement(self, params, placedb, data_collections,
device):
"""
@brief detailed placement consisting of global swap and independent set matching
@param params parameters
@param placedb placement database
@param data_collections a collection of all data and variables required for constructing the ops
@param device cpu or cuda
"""
gs = global_swap.GlobalSwap(
node_size_x=data_collections.node_size_x,
node_size_y=data_collections.node_size_y,
flat_region_boxes=data_collections.flat_region_boxes,
flat_region_boxes_start=data_collections.flat_region_boxes_start,
node2fence_region_map=data_collections.node2fence_region_map,
flat_net2pin_map=data_collections.flat_net2pin_map,
flat_net2pin_start_map=data_collections.flat_net2pin_start_map,
pin2net_map=data_collections.pin2net_map,
flat_node2pin_map=data_collections.flat_node2pin_map,
flat_node2pin_start_map=data_collections.flat_node2pin_start_map,
pin2node_map=data_collections.pin2node_map,
pin_offset_x=data_collections.pin_offset_x,
pin_offset_y=data_collections.pin_offset_y,
net_mask=data_collections.net_mask_ignore_large_degrees,
xl=placedb.xl,
yl=placedb.yl,
xh=placedb.xh,
yh=placedb.yh,
site_width=placedb.site_width,
row_height=placedb.row_height,
#num_bins_x=placedb.num_bins_x//16, num_bins_y=placedb.num_bins_y//16,
num_bins_x=placedb.num_bins_x // 2,
num_bins_y=placedb.num_bins_y // 2,
num_movable_nodes=placedb.num_movable_nodes,
num_terminal_NIs=placedb.num_terminal_NIs,
num_filler_nodes=placedb.num_filler_nodes,
batch_size=256,
max_iters=2,
algorithm='concurrent')
kr = k_reorder.KReorder(
node_size_x=data_collections.node_size_x,
node_size_y=data_collections.node_size_y,
flat_region_boxes=data_collections.flat_region_boxes,
flat_region_boxes_start=data_collections.flat_region_boxes_start,
node2fence_region_map=data_collections.node2fence_region_map,
flat_net2pin_map=data_collections.flat_net2pin_map,
flat_net2pin_start_map=data_collections.flat_net2pin_start_map,
pin2net_map=data_collections.pin2net_map,
flat_node2pin_map=data_collections.flat_node2pin_map,
flat_node2pin_start_map=data_collections.flat_node2pin_start_map,
pin2node_map=data_collections.pin2node_map,
pin_offset_x=data_collections.pin_offset_x,
pin_offset_y=data_collections.pin_offset_y,
net_mask=data_collections.net_mask_ignore_large_degrees,
xl=placedb.xl,
yl=placedb.yl,
xh=placedb.xh,
yh=placedb.yh,
site_width=placedb.site_width,
row_height=placedb.row_height,
num_bins_x=placedb.num_bins_x,
num_bins_y=placedb.num_bins_y,
num_movable_nodes=placedb.num_movable_nodes,
num_terminal_NIs=placedb.num_terminal_NIs,
num_filler_nodes=placedb.num_filler_nodes,
K=4,
max_iters=2)
ism = independent_set_matching.IndependentSetMatching(
node_size_x=data_collections.node_size_x,
node_size_y=data_collections.node_size_y,
flat_region_boxes=data_collections.flat_region_boxes,
flat_region_boxes_start=data_collections.flat_region_boxes_start,
node2fence_region_map=data_collections.node2fence_region_map,
flat_net2pin_map=data_collections.flat_net2pin_map,
flat_net2pin_start_map=data_collections.flat_net2pin_start_map,
pin2net_map=data_collections.pin2net_map,
flat_node2pin_map=data_collections.flat_node2pin_map,
flat_node2pin_start_map=data_collections.flat_node2pin_start_map,
pin2node_map=data_collections.pin2node_map,
pin_offset_x=data_collections.pin_offset_x,
pin_offset_y=data_collections.pin_offset_y,
net_mask=data_collections.net_mask_ignore_large_degrees,
xl=placedb.xl,
yl=placedb.yl,
xh=placedb.xh,
yh=placedb.yh,
site_width=placedb.site_width,
row_height=placedb.row_height,
num_bins_x=placedb.num_bins_x,
num_bins_y=placedb.num_bins_y,
num_movable_nodes=placedb.num_movable_nodes,
num_terminal_NIs=placedb.num_terminal_NIs,
num_filler_nodes=placedb.num_filler_nodes,
batch_size=2048,
set_size=128,
max_iters=50,
algorithm='concurrent')
# wirelength for position
def build_detailed_placement_op(pos):
logging.info("Start ABCDPlace for refinement")
pos1 = pos
legal = self.op_collections.legality_check_op(pos1)
logging.info("ABCDPlace input legal flag = %d" % (legal))
if not legal:
return pos1
# integer factorization to prime numbers
def prime_factorization(num):
lt = []
while num != 1:
for i in range(2, int(num + 1)):
if num % i == 0: # i is a prime factor
lt.append(i)
num = num / i # get the quotient for further factorization
break
return lt
# compute the scale factor for detailed placement
# as the algorithms prefer integer coordinate systems
scale_factor = params.scale_factor
if params.scale_factor != 1.0:
inv_scale_factor = int(round(1.0 / params.scale_factor))
prime_factors = prime_factorization(inv_scale_factor)
target_inv_scale_factor = 1
for factor in prime_factors:
if factor != 2 and factor != 5:
target_inv_scale_factor = inv_scale_factor
break
scale_factor = 1.0 / target_inv_scale_factor
logging.info("Deriving from system scale factor %g (1/%d)" %
(params.scale_factor, inv_scale_factor))
logging.info(
"Use scale factor %g (1/%d) for detailed placement" %
(scale_factor, target_inv_scale_factor))
for i in range(1):
pos1 = kr(pos1, scale_factor)
legal = self.op_collections.legality_check_op(pos1)
logging.info("K-Reorder legal flag = %d" % (legal))
if not legal:
return pos1
pos1 = ism(pos1, scale_factor)
legal = self.op_collections.legality_check_op(pos1)
logging.info("Independent set matching legal flag = %d" %
(legal))
if not legal:
return pos1
pos1 = gs(pos1, scale_factor)
legal = self.op_collections.legality_check_op(pos1)
logging.info("Global swap legal flag = %d" % (legal))
if not legal:
return pos1
pos1 = kr(pos1, scale_factor)
legal = self.op_collections.legality_check_op(pos1)
logging.info("K-Reorder legal flag = %d" % (legal))
if not legal:
return pos1
return pos1
return build_detailed_placement_op
def test_random(self):
dtype = np.float64
xx = np.array([1.0, 1.5, 3.0]).astype(dtype)
yy = np.array([1.0, 3.0, 1.0]).astype(dtype)
node_size_x = np.array([1.0, 2.0, 1.0]).astype(dtype)
node_size_y = np.array([2.0, 2.0, 2.0]).astype(dtype)
num_nodes = len(xx)
net2pin_map = np.array([np.array([0, 4]), np.array([1, 2, 3])])
pin2net_map, flat_net2pin_map, flat_net2pin_start_map = flatten_2D_map(
net2pin_map)
node2pin_map = np.array(
[np.array([0, 1]),
np.array([2]), np.array([3, 4])])
pin2node_map, flat_node2pin_map, flat_node2pin_start_map = flatten_2D_map(
node2pin_map)
# net mask
ignore_net_degree = 4
net_mask = np.ones(len(net2pin_map), dtype=np.uint8)
for i in range(len(net2pin_map)):
if len(net2pin_map[i]) >= ignore_net_degree:
net_mask[i] = 0
# pin offset
pin_offset_x = np.zeros(len(pin2net_map))
pin_offset_y = np.zeros(len(pin2net_map))
xl = 1.0
yl = 1.0
xh = 5.0
yh = 5.0
num_terminals = 0
num_terminal_NIs = 0
num_filler_nodes = 0
num_movable_nodes = len(
xx) - num_terminals - num_terminal_NIs - num_filler_nodes
site_width = 1
row_height = 2
num_bins_x = 2
num_bins_y = 2
batch_size = 32
print("flat_net2pin_map = ", flat_net2pin_map)
print("flat_net2pin_start_map = ", flat_net2pin_start_map)
print("flat_node2pin_map = ", flat_node2pin_map)
print("flat_node2pin_start_map = ", flat_node2pin_start_map)
plot(
"initial.png", xx, yy, node_size_x, node_size_y, xl, yl, xh, yh,
num_bins_x, num_bins_y, num_movable_nodes + num_terminals +
num_terminal_NIs + num_filler_nodes, num_movable_nodes,
num_movable_nodes + num_terminals + num_terminal_NIs,
num_filler_nodes)
# test cpu
custom = global_swap.GlobalSwap(
torch.from_numpy(node_size_x),
torch.from_numpy(node_size_y),
torch.from_numpy(flat_net2pin_map),
torch.from_numpy(flat_net2pin_start_map),
torch.from_numpy(pin2net_map),
torch.from_numpy(flat_node2pin_map),
torch.from_numpy(flat_node2pin_start_map),
torch.from_numpy(pin2node_map),
torch.from_numpy(pin_offset_x),
torch.from_numpy(pin_offset_y),
torch.from_numpy(net_mask),
xl=xl,
yl=yl,
xh=xh,
yh=yh,
site_width=site_width,
row_height=row_height,
num_bins_x=num_bins_x,
num_bins_y=num_bins_y,
num_movable_nodes=num_movable_nodes,
num_terminal_NIs=num_terminal_NIs,
num_filler_nodes=num_filler_nodes,
batch_size=batch_size)
pos = Variable(torch.from_numpy(np.concatenate([xx, yy])))
result = custom(pos)
print("initial result = ", np.concatenate([xx, yy]))
print("custom_result = ", result)
print("average displacement = %g" %
(np.sum(np.absolute(result.numpy() - np.concatenate([xx, yy]))) /
num_movable_nodes))
plot(
"final.png",
result.numpy()[0:len(xx)],
result.numpy()[len(xx):], node_size_x, node_size_y, xl, yl, xh, yh,
num_bins_x, num_bins_y, num_movable_nodes + num_terminals +
num_terminal_NIs + num_filler_nodes, num_movable_nodes,
num_movable_nodes + num_terminals + num_terminal_NIs,
num_filler_nodes)
# test cuda
custom_cuda = global_swap.GlobalSwap(
torch.from_numpy(node_size_x).cuda(),
torch.from_numpy(node_size_y).cuda(),
torch.from_numpy(flat_net2pin_map).cuda(),
torch.from_numpy(flat_net2pin_start_map).cuda(),
torch.from_numpy(pin2net_map).cuda(),
torch.from_numpy(flat_node2pin_map).cuda(),
torch.from_numpy(flat_node2pin_start_map).cuda(),
torch.from_numpy(pin2node_map).cuda(),
torch.from_numpy(pin_offset_x).cuda(),
torch.from_numpy(pin_offset_y).cuda(),
torch.from_numpy(net_mask).cuda(),
xl=xl,
yl=yl,
xh=xh,
yh=yh,
site_width=site_width,
row_height=row_height,
num_bins_x=num_bins_x,
num_bins_y=num_bins_y,
num_movable_nodes=num_movable_nodes,
num_terminal_NIs=num_terminal_NIs,
num_filler_nodes=num_filler_nodes,
batch_size=batch_size)
pos = Variable(torch.from_numpy(np.concatenate([xx, yy]))).cuda()
result_cuda = custom_cuda(pos)
print("custom_result = ", result_cuda.data.cpu())
def build_detailed_placement(self, params, placedb, data_collections,
device):
"""
@brief detailed placement consisting of global swap and independent set matching
@param params parameters
@param placedb placement database
@param data_collections a collection of all data and variables required for constructing the ops
@param device cpu or cuda
"""
gs = global_swap.GlobalSwap(
node_size_x=data_collections.node_size_x,
node_size_y=data_collections.node_size_y,
flat_region_boxes=data_collections.flat_region_boxes,
flat_region_boxes_start=data_collections.flat_region_boxes_start,
node2fence_region_map=data_collections.node2fence_region_map,
flat_net2pin_map=data_collections.flat_net2pin_map,
flat_net2pin_start_map=data_collections.flat_net2pin_start_map,
pin2net_map=data_collections.pin2net_map,
flat_node2pin_map=data_collections.flat_node2pin_map,
flat_node2pin_start_map=data_collections.flat_node2pin_start_map,
pin2node_map=data_collections.pin2node_map,
pin_offset_x=data_collections.pin_offset_x,
pin_offset_y=data_collections.pin_offset_y,
net_mask=data_collections.net_mask_ignore_large_degrees,
xl=placedb.xl,
yl=placedb.yl,
xh=placedb.xh,
yh=placedb.yh,
site_width=placedb.site_width,
row_height=placedb.row_height,
#num_bins_x=placedb.num_bins_x//16, num_bins_y=placedb.num_bins_y//16,
num_bins_x=placedb.num_bins_x // 2,
num_bins_y=placedb.num_bins_y // 2,
num_movable_nodes=placedb.num_movable_nodes,
num_terminal_NIs=placedb.num_terminal_NIs,
num_filler_nodes=placedb.num_filler_nodes,
batch_size=256,
max_iters=2,
algorithm='concurrent',
num_threads=params.num_threads)
kr = k_reorder.KReorder(
node_size_x=data_collections.node_size_x,
node_size_y=data_collections.node_size_y,
flat_region_boxes=data_collections.flat_region_boxes,
flat_region_boxes_start=data_collections.flat_region_boxes_start,
node2fence_region_map=data_collections.node2fence_region_map,
flat_net2pin_map=data_collections.flat_net2pin_map,
flat_net2pin_start_map=data_collections.flat_net2pin_start_map,
pin2net_map=data_collections.pin2net_map,
flat_node2pin_map=data_collections.flat_node2pin_map,
flat_node2pin_start_map=data_collections.flat_node2pin_start_map,
pin2node_map=data_collections.pin2node_map,
pin_offset_x=data_collections.pin_offset_x,
pin_offset_y=data_collections.pin_offset_y,
net_mask=data_collections.net_mask_ignore_large_degrees,
xl=placedb.xl,
yl=placedb.yl,
xh=placedb.xh,
yh=placedb.yh,
site_width=placedb.site_width,
row_height=placedb.row_height,
num_bins_x=placedb.num_bins_x,
num_bins_y=placedb.num_bins_y,
num_movable_nodes=placedb.num_movable_nodes,
num_terminal_NIs=placedb.num_terminal_NIs,
num_filler_nodes=placedb.num_filler_nodes,
K=4,
max_iters=2,
num_threads=params.num_threads)
ism = independent_set_matching.IndependentSetMatching(
node_size_x=data_collections.node_size_x,
node_size_y=data_collections.node_size_y,
flat_region_boxes=data_collections.flat_region_boxes,
flat_region_boxes_start=data_collections.flat_region_boxes_start,
node2fence_region_map=data_collections.node2fence_region_map,
flat_net2pin_map=data_collections.flat_net2pin_map,
flat_net2pin_start_map=data_collections.flat_net2pin_start_map,
pin2net_map=data_collections.pin2net_map,
flat_node2pin_map=data_collections.flat_node2pin_map,
flat_node2pin_start_map=data_collections.flat_node2pin_start_map,
pin2node_map=data_collections.pin2node_map,
pin_offset_x=data_collections.pin_offset_x,
pin_offset_y=data_collections.pin_offset_y,
net_mask=data_collections.net_mask_ignore_large_degrees,
xl=placedb.xl,
yl=placedb.yl,
xh=placedb.xh,
yh=placedb.yh,
site_width=placedb.site_width,
row_height=placedb.row_height,
num_bins_x=placedb.num_bins_x,
num_bins_y=placedb.num_bins_y,
num_movable_nodes=placedb.num_movable_nodes,
num_terminal_NIs=placedb.num_terminal_NIs,
num_filler_nodes=placedb.num_filler_nodes,
batch_size=2048,
set_size=128,
max_iters=50,
algorithm='concurrent',
num_threads=params.num_threads)
# wirelength for position
def build_detailed_placement_op(pos):
logging.info("Start ABCDPlace for refinement")
pos1 = pos
for i in range(1):
pos1 = kr(pos1)
legal = self.op_collections.legality_check_op(pos1)
logging.info("K-Reorder legal flag = %d" % (legal))
if not legal:
return pos1
pos1 = ism(pos1)
legal = self.op_collections.legality_check_op(pos1)
logging.info("Independent set matching legal flag = %d" %
(legal))
if not legal:
return pos1
pos1 = gs(pos1)
legal = self.op_collections.legality_check_op(pos1)
logging.info("Global swap legal flag = %d" % (legal))
if not legal:
return pos1
pos1 = kr(pos1)
legal = self.op_collections.legality_check_op(pos1)
logging.info("K-Reorder legal flag = %d" % (legal))
if not legal:
return pos1
return pos1
return build_detailed_placement_op