def forward(self, pos):
if self.initial_density_map is None:
self.compute_initial_density_map(pos)
# plot(0, self.initial_density_map.clone().div(self.bin_size_x*self.bin_size_y).cpu().numpy(), self.padding, 'summary/initial_potential_map')
logger.info("fixed density map: average %g, max %g, bin area %g" %
(self.initial_density_map.mean(),
self.initial_density_map.max(),
self.bin_size_x * self.bin_size_y))
# expk
M = self.num_bins_x
N = self.num_bins_y
self.exact_expkM = precompute_expk(M,
dtype=pos.dtype,
device=pos.device)
self.exact_expkN = precompute_expk(N,
dtype=pos.dtype,
device=pos.device)
# init dct2, idct2, idct_idxst, idxst_idct with expkM and expkN
self.dct2 = dct.DCT2(self.exact_expkM, self.exact_expkN)
if not self.fast_mode:
self.idct2 = dct.IDCT2(self.exact_expkM, self.exact_expkN)
self.idct_idxst = dct.IDCT_IDXST(self.exact_expkM,
self.exact_expkN)
self.idxst_idct = dct.IDXST_IDCT(self.exact_expkM,
self.exact_expkN)
# wu and wv
wu = torch.arange(M, dtype=pos.dtype, device=pos.device).mul(
2 * np.pi / M).view([M, 1])
# scale wv because the aspect ratio of a bin may not be 1
wv = torch.arange(N, dtype=pos.dtype,
device=pos.device).mul(2 * np.pi / N).view(
[1,
N]).mul_(self.bin_size_x / self.bin_size_y)
wu2_plus_wv2 = wu.pow(2) + wv.pow(2)
wu2_plus_wv2[0,
0] = 1.0 # avoid zero-division, it will be zeroed out
self.inv_wu2_plus_wv2 = 1.0 / wu2_plus_wv2
self.inv_wu2_plus_wv2[0, 0] = 0.0
self.wu_by_wu2_plus_wv2_half = wu.mul(self.inv_wu2_plus_wv2).mul_(
1. / 2)
self.wv_by_wu2_plus_wv2_half = wv.mul(self.inv_wu2_plus_wv2).mul_(
1. / 2)
return ElectricPotentialFunction.apply(
pos, self.node_size_x_clamped, self.node_size_y_clamped,
self.offset_x, self.offset_y, self.ratio, self.bin_center_x,
self.bin_center_y, self.initial_density_map, self.buf,
self.target_density, self.xl, self.yl, self.xh, self.yh,
self.bin_size_x, self.bin_size_y, self.num_movable_nodes,
self.num_filler_nodes, self.padding, self.padding_mask,
self.num_bins_x, self.num_bins_y, self.num_movable_impacted_bins_x,
self.num_movable_impacted_bins_y, self.num_filler_impacted_bins_x,
self.num_filler_impacted_bins_y, self.deterministic_flag,
self.sorted_node_map, self.exact_expkM, self.exact_expkN,
self.inv_wu2_plus_wv2, self.wu_by_wu2_plus_wv2_half,
self.wv_by_wu2_plus_wv2_half, self.dct2, self.idct2,
self.idct_idxst, self.idxst_idct, self.fast_mode, self.num_threads)
def forward(self, x):
M = x.size(-2)
N = x.size(-1)
if self.expkM is None or self.expkM.size(
-2) != M or self.expkM.dtype != x.dtype:
self.expkM = precompute_expk(M, dtype=x.dtype, device=x.device)
if self.expkN is None or self.expkN.size(
-2) != N or self.expkN.dtype != x.dtype:
self.expkN = precompute_expk(N, dtype=x.dtype, device=x.device)
if self.out is None:
self.out = torch.empty(M, N, dtype=x.dtype, device=x.device)
self.buf = torch.empty(M,
N // 2 + 1,
2,
dtype=x.dtype,
device=x.device)
return IDCT2Function.apply(x, self.expkM, self.expkN, self.out,
self.buf)
def forward(self, pos, mode="density"):
assert mode in {"density", "overflow"
}, "Only support density mode or overflow mode"
if (self.region_id is not None):
### reconstruct pos, only extract cells in this electric field
pos = pos[self.pos_mask]
if self.initial_density_map is None:
num_nodes = pos.size(0) // 2
if (self.fence_regions is not None):
if (self.placedb.num_terminals > 0):
### merge fence region density and macro density together as initial density map
### pay attention to the number of nodes, must use data from self
### here pos is reconstructed pos !
self.initial_density_map = self.compute_fence_region_map(
self.fence_regions,
pos[self.num_movable_nodes:self.num_movable_nodes +
self.num_terminals],
pos[num_nodes + self.num_movable_nodes:num_nodes +
self.num_movable_nodes + self.num_terminals],
self.node_size_x[self.num_movable_nodes:self.
num_movable_nodes +
self.num_terminals],
self.node_size_y[self.num_movable_nodes:self.
num_movable_nodes +
self.num_terminals])
else:
self.initial_density_map = self.compute_fence_region_map(
self.fence_regions)
else:
self.compute_initial_density_map(pos)
## sync the initial density map with
# self.compute_initial_density_map(pos)
# plot(0, self.initial_density_map.clone().div(self.bin_size_x*self.bin_size_y).cpu().numpy(), self.padding, 'summary/initial_potential_map')
logger.info("fixed density map: average %g, max %g, bin area %g" %
(self.initial_density_map.mean(),
self.initial_density_map.max(),
self.bin_size_x * self.bin_size_y))
# expk
M = self.num_bins_x
N = self.num_bins_y
self.exact_expkM = precompute_expk(M,
dtype=pos.dtype,
device=pos.device)
self.exact_expkN = precompute_expk(N,
dtype=pos.dtype,
device=pos.device)
# init dct2, idct2, idct_idxst, idxst_idct with expkM and expkN
self.dct2 = dct.DCT2(self.exact_expkM, self.exact_expkN)
if not self.fast_mode:
self.idct2 = dct.IDCT2(self.exact_expkM, self.exact_expkN)
self.idct_idxst = dct.IDCT_IDXST(self.exact_expkM,
self.exact_expkN)
self.idxst_idct = dct.IDXST_IDCT(self.exact_expkM,
self.exact_expkN)
# wu and wv
wu = torch.arange(M, dtype=pos.dtype, device=pos.device).mul(
2 * np.pi / M).view([M, 1])
# scale wv because the aspect ratio of a bin may not be 1
wv = torch.arange(N, dtype=pos.dtype,
device=pos.device).mul(2 * np.pi / N).view(
[1,
N]).mul_(self.bin_size_x / self.bin_size_y)
wu2_plus_wv2 = wu.pow(2) + wv.pow(2)
wu2_plus_wv2[0,
0] = 1.0 # avoid zero-division, it will be zeroed out
self.inv_wu2_plus_wv2 = 1.0 / wu2_plus_wv2
self.inv_wu2_plus_wv2[0, 0] = 0.0
self.wu_by_wu2_plus_wv2_half = wu.mul(self.inv_wu2_plus_wv2).mul_(
1. / 2)
self.wv_by_wu2_plus_wv2_half = wv.mul(self.inv_wu2_plus_wv2).mul_(
1. / 2)
if (mode == "density"):
return ElectricPotentialFunction.apply(
pos, self.node_size_x_clamped, self.node_size_y_clamped,
self.offset_x, self.offset_y, self.ratio, self.bin_center_x,
self.bin_center_y, self.initial_density_map,
self.target_density, self.xl, self.yl, self.xh, self.yh,
self.bin_size_x, self.bin_size_y, self.num_movable_nodes,
self.num_filler_nodes, self.padding, self.padding_mask,
self.num_bins_x, self.num_bins_y,
self.num_movable_impacted_bins_x,
self.num_movable_impacted_bins_y,
self.num_filler_impacted_bins_x,
self.num_filler_impacted_bins_y, self.deterministic_flag,
self.sorted_node_map, self.exact_expkM, self.exact_expkN,
self.inv_wu2_plus_wv2, self.wu_by_wu2_plus_wv2_half,
self.wv_by_wu2_plus_wv2_half, self.dct2, self.idct2,
self.idct_idxst, self.idxst_idct, self.fast_mode)
elif (mode == "overflow"):
### num_filler_nodes is set 0
density_map = ElectricDensityMapFunction.forward(
pos, self.node_size_x_clamped, self.node_size_y_clamped,
self.offset_x, self.offset_y, self.ratio, self.bin_center_x,
self.bin_center_y, self.initial_density_map,
self.target_density, self.xl, self.yl, self.xh, self.yh,
self.bin_size_x, self.bin_size_y, self.num_movable_nodes, 0,
self.padding, self.padding_mask, self.num_bins_x,
self.num_bins_y, self.num_movable_impacted_bins_x,
self.num_movable_impacted_bins_y,
self.num_filler_impacted_bins_x,
self.num_filler_impacted_bins_y, self.deterministic_flag,
self.sorted_node_map)
bin_area = self.bin_size_x * self.bin_size_y
density_cost = (density_map -
self.target_density * bin_area).clamp_(
min=0.0).sum()
return density_cost, density_map.max() / bin_area
def forward(self, pos):
if self.initial_density_map is None:
if self.num_terminals == 0:
num_fixed_impacted_bins_x = 0
num_fixed_impacted_bins_y = 0
else:
num_fixed_impacted_bins_x = int(
((self.node_size_x[
self.num_movable_nodes:self.num_movable_nodes +
self.num_terminals].max() + self.bin_size_x) /
self.bin_size_x).ceil().clamp(max=self.num_bins_x))
num_fixed_impacted_bins_y = int(
((self.node_size_y[
self.num_movable_nodes:self.num_movable_nodes +
self.num_terminals].max() + self.bin_size_y) /
self.bin_size_y).ceil().clamp(max=self.num_bins_y))
if pos.is_cuda:
self.initial_density_map = electric_potential_cuda.fixed_density_map(
pos.view(pos.numel()), self.node_size_x, self.node_size_y,
self.bin_center_x, self.bin_center_y, self.xl, self.yl,
self.xh, self.yh, self.bin_size_x, self.bin_size_y,
self.num_movable_nodes, self.num_terminals,
self.num_bins_x, self.num_bins_y,
num_fixed_impacted_bins_x, num_fixed_impacted_bins_y)
else:
self.initial_density_map = electric_potential_cpp.fixed_density_map(
pos.view(pos.numel()), self.node_size_x, self.node_size_y,
self.bin_center_x, self.bin_center_y, self.xl, self.yl,
self.xh, self.yh, self.bin_size_x, self.bin_size_y,
self.num_movable_nodes, self.num_terminals,
self.num_bins_x, self.num_bins_y,
num_fixed_impacted_bins_x, num_fixed_impacted_bins_y,
self.num_threads)
# plot(0, self.initial_density_map.clone().div(self.bin_size_x*self.bin_size_y).cpu().numpy(), self.padding, 'summary/initial_potential_map')
# scale density of fixed macros
self.initial_density_map.mul_(self.target_density)
# expk
M = self.num_bins_x
N = self.num_bins_y
self.exact_expkM = precompute_expk(M,
dtype=pos.dtype,
device=pos.device)
self.exact_expkN = precompute_expk(N,
dtype=pos.dtype,
device=pos.device)
# init dct2, idct2, idct_idxst, idxst_idct with expkM and expkN
self.dct2 = dct.DCT2(self.exact_expkM, self.exact_expkN)
if not self.fast_mode:
self.idct2 = dct.IDCT2(self.exact_expkM, self.exact_expkN)
self.idct_idxst = dct.IDCT_IDXST(self.exact_expkM,
self.exact_expkN)
self.idxst_idct = dct.IDXST_IDCT(self.exact_expkM,
self.exact_expkN)
# wu and wv
wu = torch.arange(M, dtype=pos.dtype, device=pos.device).mul(
2 * np.pi / M).view([M, 1])
# scale wv because the aspect ratio of a bin may not be 1
wv = torch.arange(N, dtype=pos.dtype,
device=pos.device).mul(2 * np.pi / N).view(
[1,
N]).mul_(self.bin_size_x / self.bin_size_y)
wu2_plus_wv2 = wu.pow(2) + wv.pow(2)
wu2_plus_wv2[0,
0] = 1.0 # avoid zero-division, it will be zeroed out
self.inv_wu2_plus_wv2 = 1.0 / wu2_plus_wv2
self.inv_wu2_plus_wv2[0, 0] = 0.0
self.wu_by_wu2_plus_wv2_half = wu.mul(self.inv_wu2_plus_wv2).mul_(
1. / 2)
self.wv_by_wu2_plus_wv2_half = wv.mul(self.inv_wu2_plus_wv2).mul_(
1. / 2)
return ElectricPotentialFunction.apply(
pos, self.node_size_x_clamped, self.node_size_y_clamped,
self.offset_x, self.offset_y, self.ratio, self.bin_center_x,
self.bin_center_y, self.initial_density_map, self.target_density,
self.xl, self.yl, self.xh, self.yh, self.bin_size_x,
self.bin_size_y, self.num_movable_nodes, self.num_filler_nodes,
self.padding, self.padding_mask, self.num_bins_x, self.num_bins_y,
self.num_movable_impacted_bins_x, self.num_movable_impacted_bins_y,
self.num_filler_impacted_bins_x, self.num_filler_impacted_bins_y,
self.sorted_node_map, self.exact_expkM, self.exact_expkN,
self.inv_wu2_plus_wv2, self.wu_by_wu2_plus_wv2_half,
self.wv_by_wu2_plus_wv2_half, self.dct2, self.idct2,
self.idct_idxst, self.idxst_idct, self.fast_mode, self.num_threads)