def build_logsumexp_wl(self, params, placedb, data_collections,
pin_pos_op):
"""
@brief build the op to compute log-sum-exp wirelength
@param params parameters
@param placedb placement database
@param data_collections a collection of data and variables required for constructing ops
@param pin_pos_op the op to compute pin locations according to cell locations
"""
gamma = 10 * self.base_gamma(params, placedb)
print("[I] gamma = %g" % (gamma))
wirelength_for_pin_op = logsumexp_wirelength.LogSumExpWirelength(
flat_netpin=data_collections.flat_net2pin_map,
netpin_start=data_collections.flat_net2pin_start_map,
pin2net_map=data_collections.pin2net_map,
net_mask=data_collections.net_mask_ignore_large_degrees,
gamma=torch.tensor(gamma,
dtype=data_collections.pos[0].dtype,
device=data_collections.pos[0].device),
algorithm='atomic')
# wirelength for position
def build_wirelength_op(pos):
pin_pos = pin_pos_op(pos)
return wirelength_for_pin_op(pin_pos)
# update gamma
base_gamma = self.base_gamma(params, placedb)
def build_update_gamma_op(iteration, overflow):
self.update_gamma(iteration, overflow, base_gamma)
#print("[I] update gamma to %g" % (wirelength_for_pin_op.gamma.data))
return build_wirelength_op, build_update_gamma_op
def build_logsumexp_wl(self, params, placedb, data_collections,
pin_pos_op):
"""
@brief build the op to compute log-sum-exp wirelength
@param params parameters
@param placedb placement database
@param data_collections a collection of data and variables required for constructing ops
@param pin_pos_op the op to compute pin locations according to cell locations
"""
wirelength_for_pin_op = logsumexp_wirelength.LogSumExpWirelength(
flat_netpin=data_collections.flat_net2pin_map,
netpin_start=data_collections.flat_net2pin_start_map,
pin2net_map=data_collections.pin2net_map,
net_weights=data_collections.net_weights,
net_mask=data_collections.net_mask_ignore_large_degrees,
pin_mask=data_collections.pin_mask_ignore_fixed_macros,
gamma=self.gamma,
algorithm='merged',
num_threads=params.num_threads)
# wirelength for position
def build_wirelength_op(pos):
return wirelength_for_pin_op(pin_pos_op(pos))
# update gamma
base_gamma = self.base_gamma(params, placedb)
def build_update_gamma_op(iteration, overflow):
self.update_gamma(iteration, overflow, base_gamma)
#logging.debug("update gamma to %g" % (wirelength_for_pin_op.gamma.data))
return build_wirelength_op, build_update_gamma_op
def test_logsumexp_wirelength_random(self):
pin_pos = np.array(
[[0.0, 0.0], [1.0, 2.0], [1.5, 0.2], [0.5, 3.1], [0.6, 1.1]],
dtype=np.float32) * 10
net2pin_map = np.array([np.array([0, 4]), np.array([1, 2, 3])])
pin2net_map = np.zeros(len(pin_pos), dtype=np.int32)
for net_id, pins in enumerate(net2pin_map):
for pin in pins:
pin2net_map[pin] = net_id
pin_x = pin_pos[:, 0]
pin_y = pin_pos[:, 1]
gamma = torch.tensor(0.5, dtype=torch.float32)
ignore_net_degree = 4
# net mask
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
# construct flat_net2pin_map and flat_net2pin_start_map
# flat netpin map, length of #pins
flat_net2pin_map = np.zeros(len(pin_pos), dtype=np.int32)
# starting index in netpin map for each net, length of #nets+1, the last entry is #pins
flat_net2pin_start_map = np.zeros(len(net2pin_map) + 1, dtype=np.int32)
count = 0
for i in range(len(net2pin_map)):
flat_net2pin_map[count:count +
len(net2pin_map[i])] = net2pin_map[i]
flat_net2pin_start_map[i] = count
count += len(net2pin_map[i])
flat_net2pin_start_map[len(net2pin_map)] = len(pin_pos)
print("flat_net2pin_map = ", flat_net2pin_map)
print("flat_net2pin_start_map = ", flat_net2pin_start_map)
print(np.transpose(pin_pos))
pin_pos_var = Variable(torch.from_numpy(np.transpose(pin_pos)).reshape(
[-1]),
requires_grad=True)
#pin_pos_var = torch.nn.Parameter(torch.from_numpy(np.transpose(pin_pos)).reshape([-1]))
print(pin_pos_var)
golden = build_wirelength(pin_pos_var[:pin_pos_var.numel() // 2],
pin_pos_var[pin_pos_var.numel() // 2:],
pin2net_map, net2pin_map, gamma,
ignore_net_degree)
print("golden_value = ", golden.data)
golden.backward()
golden_grad = pin_pos_var.grad.clone()
print("golden_grad = ", golden_grad.data)
# test cpu
# clone is very important, because the custom op cannot deep copy the data
pin_pos_var.grad.zero_()
custom = logsumexp_wirelength.LogSumExpWirelength(
torch.from_numpy(flat_net2pin_map),
torch.from_numpy(flat_net2pin_start_map),
torch.from_numpy(pin2net_map),
torch.from_numpy(net_mask),
torch.tensor(gamma),
algorithm='sparse')
result = custom.forward(pin_pos_var)
print("custom = ", result)
result.backward()
grad = pin_pos_var.grad.clone()
print("custom_grad = ", grad)
np.testing.assert_allclose(result.data.numpy(),
golden.data.detach().numpy())
np.testing.assert_allclose(grad.data.numpy(), golden_grad.data.numpy())
# test gpu
if torch.cuda.device_count():
pin_pos_var.grad.zero_()
custom_cuda = logsumexp_wirelength.LogSumExpWirelength(
Variable(torch.from_numpy(flat_net2pin_map)).cuda(),
Variable(torch.from_numpy(flat_net2pin_start_map)).cuda(),
torch.from_numpy(pin2net_map).cuda(),
torch.from_numpy(net_mask).cuda(),
torch.tensor(gamma).cuda(),
algorithm='sparse')
result_cuda = custom_cuda.forward(pin_pos_var.cuda())
print("custom_cuda_result = ", result_cuda.data.cpu())
result_cuda.backward()
grad_cuda = pin_pos_var.grad.clone()
print("custom_grad_cuda = ", grad_cuda.data.cpu())
np.testing.assert_allclose(result_cuda.data.cpu().numpy(),
golden.data.detach().numpy())
np.testing.assert_allclose(grad_cuda.data.cpu().numpy(),
grad.data.numpy(),
rtol=1e-7,
atol=1e-12)
# test gpu atomic
if torch.cuda.device_count():
pin_pos_var.grad.zero_()
custom_cuda = logsumexp_wirelength.LogSumExpWirelength(
Variable(torch.from_numpy(flat_net2pin_map)).cuda(),
Variable(torch.from_numpy(flat_net2pin_start_map)).cuda(),
torch.from_numpy(pin2net_map).cuda(),
torch.from_numpy(net_mask).cuda(),
torch.tensor(gamma).cuda(),
algorithm='atomic')
result_cuda = custom_cuda.forward(pin_pos_var.cuda())
print("custom_cuda_result atomic = ", result_cuda.data.cpu())
result_cuda.backward()
grad_cuda = pin_pos_var.grad.clone()
print("custom_grad_cuda atomic = ", grad_cuda.data.cpu())
np.testing.assert_allclose(result_cuda.data.cpu().numpy(),
golden.data.detach().numpy())
np.testing.assert_allclose(grad_cuda.data.cpu().numpy(),
grad.data.numpy(),
rtol=1e-7,
atol=1e-15)
