add timing-driven wirelength model; add flute multi thread; setup iccad2015 dataset; refactor flow

cpp_to_py/CMakeLists.txt

@@ -9,3 +9,5 @@ add_subdirectory(hpwl_cuda)
 add_subdirectory(io_parser)
 add_subdirectory(routedp)
 add_subdirectory(wa_wirelength_hpwl_cuda)
+add_subdirectory(gputimer)
+add_subdirectory(wirelength_timing_cuda)

cpp_to_py/__init__.py

@@ -10,6 +10,8 @@ __all__ = [
     "gpugr",
     "gpudp",
     "routedp",
+    "gputimer",
+    "wirelength_timing_cuda"
 ]
 from .cpybin import (
     dct_cuda,
@@ -22,5 +24,7 @@ from .cpybin import (
     gpugr,
     gpudp,
     routedp,
+    gputimer,
+    wirelength_timing_cuda
 )
 

cpp_to_py/wirelength_timing_cuda/CMakeLists.txt

@@ -0,0 +1,9 @@
+set(TARGET_NAME wirelength_timing_cuda)
+
+add_pytorch_extension(${TARGET_NAME}
+  ${TARGET_NAME}.cpp
+  ${TARGET_NAME}_kernel.cu)
+
+install(TARGETS 
+  ${TARGET_NAME}
+  DESTINATION ${XPLACE_LIB_DIR})

cpp_to_py/wirelength_timing_cuda/wirelength_timing_cuda.cpp

@@ -0,0 +1,52 @@
+#include <torch/extension.h>
+
+std::vector<torch::Tensor> wa_wirelength_timing_weight_cuda(torch::Tensor node_pos,
+                                                            torch::Tensor timing_pin_grad,
+                                                            torch::Tensor pin_id2node_id,
+                                                            torch::Tensor pin_rel_cpos,
+                                                            torch::Tensor node2pin_list,
+                                                            torch::Tensor node2pin_list_end,
+                                                            torch::Tensor hyperedge_list,
+                                                            torch::Tensor hyperedge_list_end,
+                                                            torch::Tensor net_mask,
+                                                            torch::Tensor net_weight,
+                                                            torch::Tensor hpwl_scale,
+                                                            float gamma,
+                                                            bool deterministic);
+
+#define CHECK_CUDA(x) TORCH_CHECK(x.device().is_cuda(), #x " must be a CUDA tensor")
+#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
+#define CHECK_INPUT(x) \
+    CHECK_CUDA(x);     \
+    CHECK_CONTIGUOUS(x)
+
+std::vector<torch::Tensor> wa_wirelength_timing_weight(torch::Tensor node_pos,
+                                                       torch::Tensor timing_pin_grad,
+                                                       torch::Tensor pin_id2node_id,
+                                                       torch::Tensor pin_rel_cpos,
+                                                       torch::Tensor node2pin_list,
+                                                       torch::Tensor node2pin_list_end,
+                                                       torch::Tensor hyperedge_list,
+                                                       torch::Tensor hyperedge_list_end,
+                                                       torch::Tensor net_mask,
+                                                       torch::Tensor net_weight,
+                                                       torch::Tensor hpwl_scale,
+                                                       float gamma,
+                                                       bool deterministic) {
+    CHECK_INPUT(node_pos);
+    CHECK_INPUT(timing_pin_grad);
+    CHECK_INPUT(pin_id2node_id);
+    CHECK_INPUT(pin_rel_cpos);
+    CHECK_INPUT(node2pin_list);
+    CHECK_INPUT(node2pin_list_end);
+    CHECK_INPUT(hyperedge_list);
+    CHECK_INPUT(hyperedge_list_end);
+    CHECK_INPUT(net_mask);
+    CHECK_INPUT(net_weight);
+    CHECK_INPUT(hpwl_scale);
+
+    return wa_wirelength_timing_weight_cuda(
+        node_pos, timing_pin_grad, pin_id2node_id, pin_rel_cpos, node2pin_list, node2pin_list_end, hyperedge_list, hyperedge_list_end, net_mask, net_weight, hpwl_scale, gamma, deterministic);
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { m.def("merged_wl_loss_grad_timing", &wa_wirelength_timing_weight, "calculate timing-driven WA wirelength pin grad"); }

cpp_to_py/wirelength_timing_cuda/wirelength_timing_cuda_kernel.cu

@@ -0,0 +1,210 @@
+#include <ATen/cuda/CUDAContext.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+#include <torch/extension.h>
+
+#include <vector>
+
+__global__ void node_pos_to_pin_pos_cuda_kernel(
+    const torch::PackedTensorAccessor32<float, 2, torch::RestrictPtrTraits> node_pos,
+    const torch::PackedTensorAccessor32<int64_t, 1, torch::RestrictPtrTraits> pin_id2node_id,
+    torch::PackedTensorAccessor32<float, 2, torch::RestrictPtrTraits> pin_pos,
+    int num_pins) {
+    const int index = blockIdx.x * blockDim.x + threadIdx.x;
+    const int i = index >> 1;  // pin index
+    if (i < num_pins) {
+        const int c = index & 1;  // channel index
+        int64_t node_id = pin_id2node_id[i];
+        pin_pos[i][c] += node_pos[node_id][c];
+    }
+}
+
+__global__ void calc_node_grad_deterministic_cuda_kernel(
+    torch::PackedTensorAccessor32<float, 2, torch::RestrictPtrTraits> node_grad,
+    const torch::PackedTensorAccessor32<float, 2, torch::RestrictPtrTraits> pin_grad,
+    const torch::PackedTensorAccessor32<int64_t, 1, torch::RestrictPtrTraits> node2pin_list,
+    const torch::PackedTensorAccessor32<int64_t, 1, torch::RestrictPtrTraits> node2pin_list_end,
+    int num_nodes) {
+    const int index = blockIdx.x * blockDim.x + threadIdx.x;
+    const int i = index >> 1;  // node index
+    if (i < num_nodes) {
+        const int c = index & 1;  // channel index
+        int64_t start_idx = 0;
+        if (i != 0) {
+            start_idx = node2pin_list_end[i - 1];
+        }
+        int64_t end_idx = node2pin_list_end[i];
+        if (end_idx != start_idx) {
+            node_grad[i][c] += pin_grad[node2pin_list[start_idx]][c];
+            for (int64_t idx = start_idx + 1; idx < end_idx; idx++) {
+                node_grad[i][c] += pin_grad[node2pin_list[idx]][c];
+            }
+        }
+    }
+}
+
+__global__ void wa_wirelength_pin_root_timing_kernel(
+    const torch::PackedTensorAccessor32<float, 2, torch::RestrictPtrTraits> pin_pos,
+    const torch::PackedTensorAccessor32<float, 1, torch::RestrictPtrTraits> timing_pin_weight,
+    const torch::PackedTensorAccessor32<int64_t, 1, torch::RestrictPtrTraits> hyperedge_list,
+    const torch::PackedTensorAccessor32<int64_t, 1, torch::RestrictPtrTraits> hyperedge_list_end,
+    const torch::PackedTensorAccessor32<bool, 1, torch::RestrictPtrTraits> net_mask,
+    const torch::PackedTensorAccessor32<float, 1, torch::RestrictPtrTraits> net_weight,
+    const torch::PackedTensorAccessor32<float, 1, torch::RestrictPtrTraits> hpwl_scale,
+    torch::PackedTensorAccessor32<float, 2, torch::RestrictPtrTraits> partial_wa_wl,
+    torch::PackedTensorAccessor32<float, 2, torch::RestrictPtrTraits> partial_hpwl,
+    torch::PackedTensorAccessor32<float, 2, torch::RestrictPtrTraits> pin_grad,
+    int num_nets,
+    float inv_gamma) {
+    const int index = blockIdx.x * blockDim.x + threadIdx.x;
+    const int i = index >> 1;  // net index
+    if (i < num_nets && net_mask[i]) {
+        const int c = index & 1;  // channel index
+        int64_t start_idx = 0;
+        if (i != 0) {
+            start_idx = hyperedge_list_end[i - 1];
+        }
+        int64_t end_idx = hyperedge_list_end[i];
+        if (end_idx != start_idx) {
+            int64_t root_id = hyperedge_list[start_idx];
+            float x_min = pin_pos[root_id][c];
+            float x_max = pin_pos[root_id][c];
+            float root_x = pin_pos[root_id][c];
+            float recenter_exp_r = exp((root_x - x_max) * inv_gamma);
+            float recenter_exp_nr = exp((x_min - root_x) * inv_gamma);
+            for (int64_t idx = start_idx + 1; idx < end_idx; idx++) {
+                float cur_x = pin_pos[hyperedge_list[idx]][c];
+                x_min = min(cur_x, x_min);
+                x_max = max(cur_x, x_max);
+            }
+            partial_hpwl[i][c] = round((x_max - x_min) * hpwl_scale[c]);
+
+            float sum_x_exp_x = root_x * recenter_exp_r;
+            float sum_x_exp_nx = root_x * recenter_exp_nr;
+            float sum_exp_x = recenter_exp_r;
+            float sum_exp_nx = recenter_exp_nr;
+            float wl_sum_x_exp_x = sum_x_exp_x;
+            float wl_sum_x_exp_nx = sum_x_exp_nx;
+            float wl_sum_exp_x = sum_exp_x;
+            float wl_sum_exp_nx = sum_exp_nx;
+            // pin-root gradient
+            for (int64_t idx = start_idx + 1; idx < end_idx; idx++) {
+                int64_t pin_id = hyperedge_list[idx];
+                float cur_x = pin_pos[pin_id][c];
+                float recenter_exp_x = exp((cur_x - x_max) * inv_gamma);
+                float recenter_exp_nx = exp((x_min - cur_x) * inv_gamma);
+
+                float sum_x_exp_x = cur_x * recenter_exp_x + root_x * recenter_exp_r;
+                float sum_x_exp_nx = cur_x * recenter_exp_nx + root_x * recenter_exp_nr;
+                float sum_exp_x = recenter_exp_x + recenter_exp_r;
+                float sum_exp_nx = recenter_exp_nx + recenter_exp_nr;
+                wl_sum_x_exp_x += cur_x * recenter_exp_x;
+                wl_sum_x_exp_nx += cur_x * recenter_exp_nx;
+                wl_sum_exp_x += recenter_exp_x;
+                wl_sum_exp_nx += recenter_exp_nx;
+
+                float inv_sum_exp_x = 1 / sum_exp_x;
+                float inv_sum_exp_nx = 1 / sum_exp_nx;
+                float s_x = sum_x_exp_x * inv_sum_exp_x;
+                float ns_nx = sum_x_exp_nx * inv_sum_exp_nx;
+                partial_wa_wl[i][c] += s_x - ns_nx;
+                float x_coeff = inv_gamma * inv_sum_exp_x;
+                float nx_coeff = -inv_gamma * inv_sum_exp_nx;
+                float grad_const = (1 - inv_gamma * s_x) * inv_sum_exp_x;
+                float grad_nconst = (1 + inv_gamma * ns_nx) * inv_sum_exp_nx;
+
+                float x_grad = (grad_const + x_coeff * cur_x) * recenter_exp_x -
+                               (grad_nconst + nx_coeff * cur_x) * recenter_exp_nx;
+                float root_grad = (grad_const + x_coeff * root_x) * recenter_exp_r -
+                                  (grad_nconst + nx_coeff * root_x) * recenter_exp_nr;
+
+                float delta_x = timing_pin_weight[pin_id];
+                pin_grad[pin_id][c] = x_grad * delta_x;
+                pin_grad[root_id][c] += root_grad * delta_x;
+            }
+        }
+    }
+}
+
+void calc_node_grad_cuda(torch::Tensor node_grad,
+                         torch::Tensor pin_id2node_id,
+                         torch::Tensor pin_grad,
+                         torch::Tensor node2pin_list,
+                         torch::Tensor node2pin_list_end,
+                         int num_nodes,
+                         bool deterministic) {
+    if (deterministic) {
+        auto stream = at::cuda::getCurrentCUDAStream();
+        const int threads = 128;
+        const int blocks = (num_nodes * 2 + threads - 1) / threads;
+        calc_node_grad_deterministic_cuda_kernel<<<blocks, threads, 0, stream>>>(
+            node_grad.packed_accessor32<float, 2, torch::RestrictPtrTraits>(),
+            pin_grad.packed_accessor32<float, 2, torch::RestrictPtrTraits>(),
+            node2pin_list.packed_accessor32<int64_t, 1, torch::RestrictPtrTraits>(),
+            node2pin_list_end.packed_accessor32<int64_t, 1, torch::RestrictPtrTraits>(),
+            num_nodes);
+    } else {
+        const auto pin_id2node_id_view = pin_id2node_id.unsqueeze(1).expand({-1, 2});
+        node_grad.scatter_add_(0, pin_id2node_id_view, pin_grad);
+    }
+}
+
+std::vector<torch::Tensor> wa_wirelength_timing_weight_cuda(torch::Tensor node_pos,
+                                                            torch::Tensor timing_pin_weight,
+                                                            torch::Tensor pin_id2node_id,
+                                                            torch::Tensor pin_rel_cpos,
+                                                            torch::Tensor node2pin_list,
+                                                            torch::Tensor node2pin_list_end,
+                                                            torch::Tensor hyperedge_list,
+                                                            torch::Tensor hyperedge_list_end,
+                                                            torch::Tensor net_mask,
+                                                            torch::Tensor net_weight,
+                                                            torch::Tensor hpwl_scale,
+                                                            float gamma,
+                                                            bool deterministic) {
+    cudaSetDevice(node_pos.get_device());
+    auto stream = at::cuda::getCurrentCUDAStream();
+
+    const auto num_nodes = node_pos.size(0);
+    const auto num_pins = pin_id2node_id.size(0);
+    const auto num_nets = hyperedge_list_end.size(0);
+    const auto num_channels = 2;  // x, y
+
+    auto pin_pos = pin_rel_cpos.clone();  // pin
+    auto partial_wa_wl = torch::zeros({num_nets, num_channels}, torch::dtype(pin_pos.dtype()).device(pin_pos.device()));
+    auto partial_hpwl = torch::zeros({num_nets, num_channels}, torch::dtype(pin_pos.dtype()).device(pin_pos.device()));
+    auto pin_grad = torch::zeros({num_pins, num_channels}, torch::dtype(pin_pos.dtype()).device(pin_pos.device()));
+
+    const int threads = 128;
+    const int blocks = (num_pins * 2 + threads - 1) / threads;
+
+    node_pos_to_pin_pos_cuda_kernel<<<blocks, threads, 0, stream>>>(
+        node_pos.packed_accessor32<float, 2, torch::RestrictPtrTraits>(),
+        pin_id2node_id.packed_accessor32<int64_t, 1, torch::RestrictPtrTraits>(),
+        pin_pos.packed_accessor32<float, 2, torch::RestrictPtrTraits>(),
+        num_pins);
+
+    const int threads2 = 128;
+    const int blocks2 = (num_nets * 2 + threads2 - 1) / threads2;
+
+    float inv_gamma = 1 / gamma;
+    wa_wirelength_pin_root_timing_kernel<<<blocks2, threads2, 0, stream>>>(
+        pin_pos.packed_accessor32<float, 2, torch::RestrictPtrTraits>(),
+        timing_pin_weight.packed_accessor32<float, 1, torch::RestrictPtrTraits>(),
+        hyperedge_list.packed_accessor32<int64_t, 1, torch::RestrictPtrTraits>(),
+        hyperedge_list_end.packed_accessor32<int64_t, 1, torch::RestrictPtrTraits>(),
+        net_mask.packed_accessor32<bool, 1, torch::RestrictPtrTraits>(),
+        net_weight.packed_accessor32<float, 1, torch::RestrictPtrTraits>(),
+        hpwl_scale.packed_accessor32<float, 1, torch::RestrictPtrTraits>(),
+        partial_wa_wl.packed_accessor32<float, 2, torch::RestrictPtrTraits>(),
+        partial_hpwl.packed_accessor32<float, 2, torch::RestrictPtrTraits>(),
+        pin_grad.packed_accessor32<float, 2, torch::RestrictPtrTraits>(),
+        num_nets,
+        inv_gamma);
+
+    auto node_grad = torch::zeros({num_nodes, num_channels}, torch::dtype(pin_grad.dtype()).device(pin_grad.device()));
+    calc_node_grad_cuda(
+        node_grad, pin_id2node_id, pin_grad, node2pin_list, node2pin_list_end, num_nodes, deterministic);
+
+    return {partial_wa_wl, node_grad, partial_hpwl};
+}

main.py

@@ -49,6 +49,18 @@ def get_option():
     parser.add_argument('--pseudo_weight', type=float, default=0, help='the weight of pseudo net')
     parser.add_argument('--visualize_cgmap', type=str2bool, default=False, help='visualize congestion map')
 
+    # timing opt params
+    parser.add_argument('--timing_opt', type=str2bool, default=True, help='perform timing optimization') 
+    parser.add_argument('--timing_freq', type=int, default=1, help='timing freq')
+    parser.add_argument('--calibration', type=str2bool, default=True, help='perform timer calibration') 
+    parser.add_argument('--calibration_step', type=float, default=0.1, help='timing calibration step')
+    parser.add_argument('--timing_start_iter', type=int, default=100, help='start iteration of timing optimization')
+    parser.add_argument('--timing_init_weight', type=float, default=0.05, help='initial timing wirelength weight')
+    parser.add_argument('--decay_factor', type=float, default=0.3, help='decay factor of timing weight')
+    parser.add_argument('--decay_boost', type=float, default=3, help='dynamic decay boost factor')
+    parser.add_argument('--wire_resistance_per_micron', type=float, default=2.535, help='unit wire resistance, normalized across all layers')
+    parser.add_argument('--wire_capacitance_per_micron', type=float, default=0.16e-15, help='unit wire capacitance, normalized across all layers')
+
     # detailed placement and evaluation
     parser.add_argument('--legalization', type=str2bool, default=True, help='perform lg') 
     parser.add_argument('--detail_placement', type=str2bool, default=True, help='perform dp') 
@@ -77,6 +89,7 @@ def get_option():
     args = parser.parse_args()
 
     args.exp_id = datetime.datetime.now().strftime('%Y-%m-%d-%H:%M:%S') + args.exp_id
+    args.exp_id = "{}_{}".format(args.exp_id, args.design_name)
 
     if args.dataset == "ispd2015":
         print("We haven't yet support fence region in ispd2015, use ispd2015_fix instead")

src/calculator.py

@@ -1,7 +1,6 @@
 import torch
 from .param_scheduler import ParamScheduler
-from .core import merged_wl_loss_grad
+from .core import merged_wl_loss_grad, merged_wl_loss_grad_timing
-
 
 def apply_precond(mov_node_pos: torch.Tensor, ps: ParamScheduler, args):
     if not args.use_precond:
@@ -51,6 +50,17 @@ def calc_obj_and_grad(
             data.hpwl_scale, ps.wa_coeff, args.deterministic
         )
         mov_node_pos.grad[mov_lhs:mov_rhs] += conn_node_grad_by_wl[mov_lhs:mov_rhs]
+
+        if ps.enable_timing:
+            wl_loss_timing, conn_node_grad_by_timing = merged_wl_loss_grad_timing(
+                conn_node_pos, data.gputimer.timing_pin_weight,
+                data.pin_id2node_id, data.pin_rel_cpos,
+                data.node2pin_list, data.node2pin_list_end, data.hyperedge_list, data.hyperedge_list_end,
+                data.net_mask, data.net_weight, data.hpwl_scale, ps.wa_coeff, args.deterministic
+            )
+            mov_node_pos.grad[mov_lhs:mov_rhs] += conn_node_grad_by_timing[mov_lhs:mov_rhs]
+            wl_loss += wl_loss_timing
+            
         if ps.enable_sample_force:
             if ps.iter > 3 and ps.iter % 20 == 0:
                 # ps.iter > 3 for warmup

src/core/__init__.py

@@ -1,4 +1,5 @@
 from .flute import Flute, get_flute_wl
 from .electronic_density_layer import ElectronicDensityLayer
 from .wa_wirelength_hpwl import masked_scale_hpwl, merged_wl_loss_grad
-from .route_force import get_route_force, run_gr_and_fft, run_gr_and_fft_main, route_inflation, route_inflation_roll_back
+from .route_force import get_route_force, run_gr_and_fft, run_gr_and_fft_main, route_inflation, route_inflation_roll_back
+from .timing_opt import GPUTimer, merged_wl_loss_grad_timing

src/core/timing_opt.py

@@ -0,0 +1,271 @@
+import torch
+from cpp_to_py import gputimer, wirelength_timing_cuda
+from src.param_scheduler import MetricRecorder
+from utils import *
+
+class GPUTimer():
+    def __init__(self, data, rawdb, gpdb, params, args):
+        self.metrics = [
+            "wns",
+            "tns",
+        ]
+        self.recorder = MetricRecorder(**{m: [] for m in self.metrics})
+        
+        self.data = data
+        self.net_names = data.net_names
+        self.pin_names = data.pin_names
+        
+        self.microns = data.microns
+        self.wire_resistance_per_micron = args.wire_resistance_per_micron
+        self.wire_capacitance_per_micron = args.wire_capacitance_per_micron
+
+        self.node_size = data.node_size.detach().clone()
+        node_lpos = data.node_pos.detach() - self.node_size / 2
+        self.pin_rel_lpos = data.pin_rel_lpos.detach() + data.pin_size / 2
+
+        die_info = data.die_info
+        xl, xh, yl, yh = die_info.cpu().numpy()
+
+        self.mov_lhs, self.mov_rhs = data.movable_index
+        fix_lhs, fix_rhs = data.fixed_connected_index
+        num_movable_nodes = self.mov_rhs - self.mov_lhs
+        self.fix_conn_node_lpos = node_lpos[fix_lhs:fix_rhs]
+        self.conn_node_lpos = torch.cat([
+            node_lpos[self.mov_lhs:self.mov_rhs], self.fix_conn_node_lpos
+        ], dim=0)
+
+        scale_factor = 1.0 / data.site_width
+        self.node_weight = data.node_special_type == 2
+        
+        self.timing_raw_db = gputimer.create_timing_rawdb(
+            self.conn_node_lpos,
+            data.node_size,
+            self.pin_rel_lpos,
+            data.pin_id2node_id,
+            data.pin_id2net_id.int(),
+            data.node2pin_list,
+            data.node2pin_list_end,
+            data.hyperedge_list.int(),
+            data.hyperedge_list_end.int(),
+            data.net_mask,
+            num_movable_nodes,
+            scale_factor,
+            self.microns,
+            self.wire_resistance_per_micron,
+            self.wire_capacitance_per_micron
+        )
+        
+        self.timer = gputimer.create_gputimer(params, rawdb, gpdb, self.timing_raw_db)
+        
+        ## Timing optimization
+        self.timer.init()
+        self.timer.levelize()
+        self.pin_slack = torch.zeros(data.num_pins, dtype=torch.float32, device=data.device)
+        self.timing_pin_weight = torch.ones(data.num_pins, dtype=torch.float32, device=data.device)
+        self.history_x = None
+        
+        self.tns_record = []
+        self.wns_record = []
+        self.wns_max = []
+        self.delay_K_max = []
+        self.delay_1_max = []
+        self.w_2 = None
+        self.w_1 = None
+        self.a_1 = None
+        self.decay = args.decay_factor
+        self.decay_boost = args.decay_boost
+        self.init_alpha = 1.05
+        self.beta = 0
+        self.alpha = torch.ones(data.num_pins, dtype=torch.float32, device=data.device) * self.init_alpha
+        self.global_weight = 1
+
+        self.target_wns = 0
+        
+    def update_timing(self, node_pos):
+        node_lpos = (node_pos.detach() - self.node_size / 2).to(self.data.device)
+        self.conn_node_lpos = torch.cat([
+            node_lpos[self.mov_lhs:self.mov_rhs], self.fix_conn_node_lpos
+        ], dim=0)
+        
+        self.timer.update_states()
+        self.timer.update_rc(node_lpos, False, False, False)
+        self.timer.update_timing()
+    
+    def update_timing_eval(self, node_pos):
+        node_lpos = (node_pos.detach() - self.node_size / 2).to(self.data.device)
+        self.conn_node_lpos = torch.cat([
+            node_lpos[self.mov_lhs:self.mov_rhs], self.fix_conn_node_lpos
+        ], dim=0)
+        
+        self.timer.update_states()
+        self.timer.update_rc_flute(node_lpos, False)
+        self.timer.update_timing()
+    
+    def update_timing_calibrated(self, node_pos, record=False):
+        node_lpos = (node_pos.detach() - self.node_size / 2).to(self.data.device)
+        self.conn_node_lpos = torch.cat([
+            node_lpos[self.mov_lhs:self.mov_rhs], self.fix_conn_node_lpos
+        ], dim=0)
+        
+        if record:
+            self.timer.update_states()
+            self.timer.update_rc_flute(node_lpos, True)
+            self.timer.update_states()
+            self.timer.update_rc(node_lpos, True, True, True)
+        else:
+            self.timer.update_states()
+            self.timer.update_rc(node_lpos, False, True, True)
+        self.timer.update_timing()
+        
+        
+    def report_timing_slack(self):
+        time_unit = self.timer.time_unit()
+        self.timer.update_endpoints()
+        wns_early, tns_early, wns_late, tns_late = self.timer.report_wns_and_tns()
+        wns_early = (wns_early.item() * (time_unit * 1e9))
+        wns_late = (wns_late.item() * (time_unit * 1e9))
+        tns_early = (tns_early.item() * (time_unit * 1e9))
+        tns_late = (tns_late.item() * (time_unit * 1e9))
+        self.push_metric(-wns_late, -tns_late)
+        return wns_early, tns_early, wns_late, tns_late
+
+    def report_pin_slack(self):
+        self.pin_slack = self.timer.report_pin_slack()
+        return self.pin_slack
+    
+    def report_path(self, ep_name=None, el = -1, verbose=False):
+        if ep_name is not None:
+            ep_idx = self.pin_names.index(ep_name)
+            path, at, delay = self.timer.report_path(ep_idx, el, verbose)
+        else:
+            path, at, delay = self.timer.report_path(-1, el, verbose)
+        return path, at, delay
+    
+    def report_arrival(self, pin_name):
+        pin_idx = self.pin_names.index(pin_name)
+        return self.timer.report_pin_at()[pin_idx]
+    
+    def report_slew(self, pin_name):
+        pin_idx = self.pin_names.index(pin_name)
+        return self.timer.report_pin_slew()[pin_idx]
+    
+    def report_load(self, pin_name):
+        pin_idx = self.pin_names.index(pin_name)
+        return self.timer.report_pin_load()[pin_idx]
+
+    def report_required(self, pin_name):
+        pin_idx = self.pin_names.index(pin_name)
+        return self.timer.report_pin_rat()[pin_idx]
+    
+    def report_slack(self, pin_name):
+        pin_idx = self.pin_names.index(pin_name)
+        return self.timer.report_pin_slack()[pin_idx]
+
+    def get_node_critocality(self):
+        pin_slacks, _ = torch.min((torch.nan_to_num(self.report_pin_slack()) * (1e-9 / self.timer.time_unit())).clamp(max=0), 1)
+        endpoints_index = self.timer.endpoints_index().long()
+        endpoints_index = torch.unique(endpoints_index)
+        ep_id2node_id = self.data.pin_id2node_id[endpoints_index]
+        ep_slacks = pin_slacks[endpoints_index]
+        node_slacks = torch.zeros(self.node_weight.size(0), dtype=torch.float32, device=self.data.device)
+        node_slacks.scatter_add_(0, ep_id2node_id, ep_slacks)
+        node_critocality = torch.abs(node_slacks) / (torch.abs(node_slacks)).max()
+        return node_critocality
+
+    def step(self, ps, node_pos, data):
+        slacks, _ = torch.min(torch.nan_to_num(self.report_pin_slack()).clamp(max=0), 1)
+        delay_k, _ = self.timer.report_criticality_threshold(0.75, False, True)
+        delay_1, pin_visited = self.timer.report_criticality_threshold(0.99, False, True)
+        
+        self.tns_record.append(self.recorder.tns[-1])
+        self.wns_record.append(self.recorder.wns[-1])
+        self.wns_max.append(slacks.min().item())
+        self.delay_K_max.append(delay_k.max().item())
+        self.delay_1_max.append(delay_1.max().item())
+        
+        window = min(10, len(self.wns_max))
+        x_wns = torch.tensor(self.wns_max[-window:], dtype=torch.float32)
+        x_delay_k = torch.tensor(self.delay_K_max[-window:], dtype=torch.float32)
+        x_delay_1 = torch.tensor(self.delay_1_max[-window:], dtype=torch.float32)
+
+        wns_mean = x_wns[-1]
+        delay_k_mean = x_delay_k[-1]
+        delay_1_mean = x_delay_1[-1]
+        
+        pin_weight = slacks.abs() / (np.abs(wns_mean)) * self.beta
+        pin_weight += (delay_k / delay_k_mean.clamp(min=1)) * self.beta * 2
+        pin_weight += torch.pow(2, (delay_1 / delay_1_mean.clamp(min=1))) * pin_visited.clamp(max=1)
+        
+        w_0 = pin_weight
+        delta_w_0 = None
+        delta_w_1 = None
+        if self.w_1 is not None:
+            delta_w_0 = w_0 - self.w_1
+        if self.w_2 is not None:
+            delta_w_1 = self.w_1 - self.w_2
+
+        if delta_w_0 is not None and delta_w_1 is not None:
+            decay = (self.decay * torch.pow(5, delta_w_0.clamp(min=0)) / self.decay_boost).clamp(max=0.5)
+        else:
+            decay = 1
+            
+        self.w_2 = self.w_1
+        self.w_1 = pin_weight.clone()
+        if self.history_x is None:
+            self.history_x = self.timing_pin_weight.clone()
+        self.timing_pin_weight = pin_weight.clamp(min=ps.timing_wl_weight)
+        self.timing_pin_weight = (decay * self.timing_pin_weight + (1 - decay) * self.history_x) * self.global_weight
+        self.timing_pin_weight = self.timing_pin_weight.contiguous().to(node_pos.device)
+        self.history_x = self.timing_pin_weight.clone()
+        
+    def push_metric(self, wns, tns):
+        metrics_dict = {
+            "wns": wns,
+            "tns": tns,
+        }
+        self.recorder.push(**metrics_dict)
+         
+    def visualize(self, args):
+        file_prefix = "%s_" % args.design_name
+        res_root = os.path.join(args.result_dir, args.exp_id)
+        prefix = os.path.join(res_root, args.eval_dir, file_prefix)
+        if not os.path.exists(os.path.dirname(prefix)):
+            os.makedirs(os.path.dirname(prefix))
+        self.recorder.visualize(prefix, True)
+
+def merged_wl_loss_grad_timing(
+    node_pos,
+    timing_pin_grads,
+    pin_id2node_id,
+    pin_rel_cpos,
+    node2pin_list,
+    node2pin_list_end,
+    hyperedge_list,
+    hyperedge_list_end,
+    net_mask,
+    net_weight,
+    hpwl_scale,
+    gamma,
+    deterministic,
+    cache_hpwl=True,
+):
+    (
+        partial_wa_wl,
+        node_grad,
+        partial_hpwl,
+    ) = wirelength_timing_cuda.merged_wl_loss_grad_timing(
+        node_pos,
+        timing_pin_grads,
+        pin_id2node_id,
+        pin_rel_cpos,
+        node2pin_list,
+        node2pin_list_end,
+        hyperedge_list,
+        hyperedge_list_end,
+        net_mask,
+        net_weight,
+        hpwl_scale,
+        gamma,
+        deterministic,
+    )
+    return torch.sum(partial_wa_wl), node_grad

src/database.py

@@ -110,11 +110,12 @@ class PlaceData(object):
         # TODO: more cases, hardcode?
         self.node_special_type = torch.zeros(len(node_id2celltype_name), dtype=torch.int32)
         ## too slow...
-        # for node_id, celltype_name in enumerate(node_id2celltype_name):
+        for node_id, celltype_name in enumerate(node_id2celltype_name):
-        #     if celltype_name.startswith("CORE/BUF"):
+            if celltype_name.startswith("CORE/BUF"):
-        #         self.node_special_type[node_id] = 1
+                self.node_special_type[node_id] = 1
-        #     if celltype_name.startswith("CORE/DFF"):
+            if celltype_name.startswith("CORE/DFF"):
-        #         self.node_special_type[node_id] = 2
+                self.node_special_type[node_id] = 2
+
 
         dataset_format = ""
         if "aux" in dataset_path.keys():
@@ -506,10 +507,16 @@ class PlaceData(object):
         has_dict = any([isinstance(item, dict) for _, item in self])
 
         if not has_dict:
-            info = [size_repr(key, item) for key, item in self]
+            info = [
+                size_repr(key, item) for key, item in self if type(item) != dict
+            ]
             return "{}({}, {})".format(cls, self.design_name, ", ".join(info))
         else:
-            info = [size_repr(key, item, indent=2) for key, item in self]
+            info = [
+                size_repr(key, item, indent=2)
+                for key, item in self
+                if type(item) != dict
+            ]
             return "{}({}, \n{}\n)".format(cls, self.design_name, ",\n".join(info))
 
     def backup_ori_var(self):
@@ -606,6 +613,7 @@ class PlaceData(object):
         self.net_mask = torch.logical_and(
             self.net_to_num_pins <= args.ignore_net_degree, self.net_to_num_pins >= 2
         )  # 0: ignore, 1: consider in wirelength calculation
+        self.net_weight = torch.ones(self.num_nets, device=device, dtype=dtype)
         # macros -> all mov nodes has ultra-large areas with >= 3 row height and all fixed nodes
         # But nodes with zero width or zero height are not considered as macros
         mov_lhs, mov_rhs = self.movable_index
@@ -882,6 +890,7 @@ class PlaceData(object):
             scale = (self.die_ur - self.die_ll) * 0.001
             loc = (self.die_ur + self.die_ll) * 0.5
             mov_node_pos = torch.randn_like(mov_node_pos) * scale + loc
+            # mov_node_pos = torch.randn(mov_node_pos.shape).to(mov_node_pos.device) * scale + loc
         elif init_method == "randn_center_lxly":
             # TODO: Mixed-size placement is very sensitive to the initial location.
             #       An elegant yet effective initialization may be needed.

src/initializer.py

@@ -1,7 +1,7 @@
 import torch
 from .database import PlaceData
 from cpp_to_py import density_map_cuda
-from .core import merged_wl_loss_grad
+from .core import merged_wl_loss_grad, merged_wl_loss_grad_timing
 
 
 def get_init_density_map(rawdb, gpdb, data: PlaceData, args, logger, ps=None):
@@ -23,6 +23,9 @@ def get_init_density_map(rawdb, gpdb, data: PlaceData, args, logger, ps=None):
         node_pos = torch.cat([data.node_pos[data.is_mov_macro].contiguous(), node_pos])
         node_size = torch.cat([data.node_size[data.is_mov_macro].contiguous(), node_size])
         node_weight = node_size.new_ones(node_size.shape[0])
+    if args.timing_opt: # TODO: sideline
+        node_size = node_size.clone()
+        node_size[:, 0] *= 1.05
     init_density_map = density_map_cuda.forward_naive(
         node_pos, node_size, node_weight, data.unit_len, zeros_density_map,
         data.num_bin_x, data.num_bin_y, node_pos.shape[0], -1.0, -1.0, 1e-4, False,
@@ -126,11 +129,11 @@ def init_params(
     (mov_node_pos * 0.0).sum().backward()
     optimizer.zero_grad(set_to_none=False)
 
-    if not ps.rerun_route or route_fn is None:
+    if (not ps.rerun_route or route_fn is None) and not args.timing_opt:
         init_density_weight = (wl_grad.norm(p=1) / density_grad.norm(p=1)).detach()
         # init_density_weight = (wl_grad.norm(p=1) / grad_mat.norm(p=1)).detach()
         ps.set_init_param(init_density_weight, data)
-    else:
+    elif ps.rerun_route and route_fn is not None:
         _, filler_lhs = data.movable_connected_index
         filler_rhs = mov_node_pos.shape[0]
 
@@ -147,7 +150,22 @@ def init_params(
         ps.set_route_init_param(
             init_density_weight, init_route_weight, init_congest_weight, data, args
         )
-
+    elif args.timing_opt:
+        init_pin_weight = torch.ones(data.num_pins, dtype=torch.float32, device=data.device)
+        _, wl_grad_timing = merged_wl_loss_grad_timing(
+            conn_node_pos, init_pin_weight,
+            data.pin_id2node_id, data.pin_rel_cpos,
+            data.node2pin_list, data.node2pin_list_end, data.hyperedge_list, data.hyperedge_list_end,
+            data.net_mask, data.net_weight, data.hpwl_scale, ps.wa_coeff, args.deterministic
+        )
+        init_density_weight = ((wl_grad.norm(p=1) + wl_grad_timing.norm(p=1)) / density_grad.norm(p=1)).detach()
+        ps.set_init_param(init_density_weight, data)
+        ps.timing_wl_weight = args.timing_init_weight
+        data.gputimer.timing_pin_weight *= ps.timing_wl_weight
+        data.gputimer.beta = 5 * ps.timing_wl_weight
+        # timing mode fluctuation, larger divergence life
+        ps.max_life = 50
+        ps.life = ps.max_life
 
 # Nesterove learning rate initialization
 def estimate_initial_learning_rate(obj_and_grad_fn, constraint_fn, x_k, lr):

src/param_scheduler.py

@@ -29,10 +29,12 @@ class MetricRecorder:
                 )
             self[key].append(item)
 
-    def visualize(self, prefix):
+    def visualize(self, prefix, log = False):
         for key, value in self:
             x = list(range(len(value)))
             plt.plot(x, value, label=key)
+            if log:
+                plt.yscale("log")
             plt.legend()
             plt.savefig(prefix + "%s.png" % key)
             plt.close()
@@ -148,6 +150,13 @@ class ParamScheduler:
         self.include_macros = args.include_macros
         self.zero_macro_grad = False
 
+        # timing parameter
+        self.timing_sol_recorder = []
+        self.enable_timing = False
+        self.timing_wl_weight = 0
+        self.best_sol_timing: torch.Tensor = None
+        self.best_metric_timing = {"wns": float("inf"), "tns": float("inf"), "iter": 0}
+
     def set_init_param(self, init_density_weight, data: PlaceData):
         # init_density_weight
         self.init_iter = self.iter
@@ -230,6 +239,19 @@ class ParamScheduler:
         self.gr_sol_recorder.append((
             gr_metrics, hpwl, overflow, mov_node_pos.detach().clone()
         ))
+        
+    def push_timing_sol(self, timing_metrics, hpwl, overflow, mov_node_pos: torch.Tensor):
+        if overflow > self.stop_overflow: # skip overflow solutions
+            return
+        wns_early, tns_early, wns_late, tns_late = timing_metrics
+        wns = -min(wns_early, wns_late, 0)
+        tns = -min(tns_early, tns_late, 0)
+        if wns <= self.best_metric_timing["wns"]:
+            self.best_metric_timing["wns"] = wns
+            self.best_metric_timing["iter"] = self.iter
+            self.best_sol_timing = mov_node_pos.detach().clone()
+            if tns < self.best_metric_timing["tns"]:
+                self.best_metric_timing["tns"] = tns
 
     def step(self, hpwl, overflow, node_pos, data):
         self.update_precond_weight(data)
@@ -582,6 +604,20 @@ class ParamScheduler:
             (best_idx, numOvflNets, gr_wirelength, gr_numVias, gr_numShorts, rc_hor_mean, rc_ver_mean)
         )
         return best_sol
+    
+    def get_best_timing_sol(self):
+        logger = self.__logger__
+        if self.best_sol_timing is not None:
+            wns = self.best_metric_timing["wns"]
+            tns = self.best_metric_timing["tns"]
+            iterationn = self.best_metric_timing["iter"]
+            logger.info(
+                "Find best timing solution: WNS %.4f TNS %.4f at Iter %d" % (wns, tns, iterationn)
+            )
+        else:
+            logger.info("Cannot find best timing solution. Use the last solution.")
+            return None
+        return self.best_sol_timing.data
 
 
     def visualize(self, args, logger):

src/run_placement_nesterov.py

@@ -10,7 +10,7 @@ def get_trunc_node_pos_fn(mov_node_size, data):
         return x
     return trunc_node_pos_fn
 
-def global_placement_main(gpdb, rawdb, ps: ParamScheduler, data: PlaceData, args, logger):
+def global_placement_main(gpdb, rawdb, ps: ParamScheduler, data: PlaceData, args, logger, params, gputimer=None):
     init_density_map = data.init_density_map
     if not args.global_placement:
         logger.info("Global placement is switched off. Please make sure the input "
@@ -135,12 +135,36 @@ def global_placement_main(gpdb, rawdb, ps: ParamScheduler, data: PlaceData, args
     terminate_signal = False
     route_early_terminate_signal = False
     log_info = False
+    timing_cali_thrs_overflow = 0.5
+    timing_calibration = False
     for iteration in range(args.inner_iter):
         # optimizer.zero_grad() # zero grad inside obj_and_grad_fn
         obj = optimizer.step(obj_and_grad_fn)
         hpwl, overflow = evaluator_fn(mov_node_pos)
         # update parameters
         ps.step(hpwl, overflow, mov_node_pos, data)
+
+        # Perform timing-opt.
+        if args.timing_opt and iteration > args.timing_start_iter and iteration % 1 == 0:
+            ps.enable_timing = True
+            node_pos = torch.cat([mov_node_pos[mov_lhs:mov_rhs].clone(), data.node_pos[mov_rhs:]], dim=0)
+            
+            if args.calibration and ps.recorder.overflow[-1] < timing_cali_thrs_overflow:
+                gputimer.update_timing_calibrated(node_pos, record=True)
+                timing_cali_thrs_overflow -= args.calibration_step
+                timing_calibration = True
+            elif timing_calibration:
+                gputimer.update_timing_calibrated(node_pos)
+            else:
+                gputimer.update_timing(node_pos)
+            
+            timing_metrics = gputimer.report_timing_slack()
+            wns_early, tns_early, wns_late, tns_late = timing_metrics
+            ps.push_timing_sol(timing_metrics, hpwl, overflow, mov_node_pos)
+            
+            if iteration % args.timing_freq == 0:
+                gputimer.step(ps, node_pos, data)
+
         if ps.need_to_early_stop():
             terminate_signal = True
             log_info = True
@@ -317,6 +341,10 @@ def global_placement_main(gpdb, rawdb, ps: ParamScheduler, data: PlaceData, args
                     ps.wa_coeff,
                 )
             )
+            if ps.enable_timing:
+                log_str += " | early WNS/TNS: %.4f %.4f (ns) | late WNS/TNS: %.4f %.4f (ns)" % (
+                    wns_early, tns_early, wns_late, tns_late
+                )
             logger.info(log_str)
             if args.draw_placement:
                 info = (iteration, hpwl, data.design_name)
@@ -365,6 +393,10 @@ def global_placement_main(gpdb, rawdb, ps: ParamScheduler, data: PlaceData, args
             ps.push_gr_sol(gr_metrics, hpwl, overflow, mov_node_pos)
         best_sol_gr = ps.get_best_gr_sol()
         mov_node_pos[mov_lhs:mov_rhs].data.copy_(best_sol_gr[mov_lhs:mov_rhs])
+    if ps.enable_timing and not ps.enable_route:
+        best_sol_timing = ps.get_best_timing_sol()
+        if best_sol_timing is not None:
+            mov_node_pos[mov_lhs:mov_rhs].data.copy_(best_sol_timing[mov_lhs:mov_rhs])
 
     node_pos = mov_node_pos[mov_lhs:mov_rhs]
     node_pos = torch.cat([node_pos, data.node_pos[mov_rhs:]], dim=0)
@@ -423,8 +455,11 @@ def run_placement_main_nesterov(args, logger):
             
     # global placement
     node_pos, iteration, gp_hpwl, overflow, gp_time, gp_per_iter = global_placement_main(
-        gpdb, rawdb, ps, data, args, logger
+        gpdb, rawdb, ps, data, args, logger, params, gputimer
     )
+    if args.timing_opt:
+        wns_early_gp, tns_early_gp, wns_late_gp, tns_late_gp = timing_eval_func(node_pos)
+
     # detail placement
     node_pos, dp_hpwl, top5overflow, lg_time, dp_time = detail_placement_main(
         node_pos, gpdb, rawdb, ps, data, args, logger

thirdparty/flute_mp/CMakeLists.txt

@@ -0,0 +1,43 @@
+# cmake_minimum_required(VERSION 2.8.12)
+
+# set(CMAKE_CXX_STANDARD 17)
+
+# set(FLUTE_INCLUDE_DIR "${CMAKE_CURRENT_LIST_DIR}" CACHE INTERNAL "Directory where flute.h is located")
+
+# set(CMAKE_CXX_STANDARD_REQUIRED ON)
+
+# project(flute)
+# if(NOT CMAKE_BUILD_TYPE)
+#   set(CMAKE_BUILD_TYPE "Release" CACHE STRING
+#     "Choose the type of build, options are: Debug Release."
+#     FORCE)
+# endif(NOT CMAKE_BUILD_TYPE)
+
+# set(SOURCES ${CMAKE_CURRENT_SOURCE_DIR}/flute.cpp)
+
+# include_directories("${CMAKE_CURRENT_SOURCE_DIR}")
+# set(CMAKE_POSITION_INDEPENDENT_CODE ON)
+# add_library(${PROJECT_NAME} STATIC ${SOURCES})
+
+# install(TARGETS ${PROJECT_NAME} DESTINATION thirdparty/${PROJECT_NAME})
+# install(DIRECTORY lut.ICCAD2015 DESTINATION thirdparty/${PROJECT_NAME})
+
+
+CMAKE_MINIMUM_REQUIRED (VERSION 3.1)
+PROJECT(flute_mp)
+
+SET(CMAKE_CXX_STANDARD 17)
+
+set(FLUTE_MP_INCLUDE_DIR "${CMAKE_CURRENT_LIST_DIR}" CACHE INTERNAL "Directory where flute.h is located")
+
+FILE(GLOB_RECURSE SRC_FILES_FLUTE_MP ${CMAKE_CURRENT_SOURCE_DIR}/*.cpp)
+
+ADD_LIBRARY(flute_mp
+  SHARED
+  ${SRC_FILES_FLUTE_MP}
+  )
+
+TARGET_INCLUDE_DIRECTORIES(flute_mp PUBLIC ${FLUTE_MP_INCLUDE_DIR})
+TARGET_COMPILE_OPTIONS(flute_mp PRIVATE -fPIC)
+
+INSTALL(TARGETS flute_mp DESTINATION ${XPLACE_LIB_DIR})

thirdparty/flute_mp/LICENSE

@@ -0,0 +1,50 @@
+READ THIS LICENSE AGREEMENT CAREFULLY BEFORE USING THIS PRODUCT. BY USING
+THIS PRODUCT YOU INDICATE YOUR ACCEPTANCE OF THE TERMS OF THE FOLLOWING
+AGREEMENT. THESE TERMS APPLY TO YOU AND ANY SUBSEQUENT LICENSEE OF THIS
+PRODUCT.
+
+License Agreement for FLUTE
+
+Copyright (c) 2004 by Dr. Chris C. N. Chu
+All rights reserved
+
+ATTRIBUTION ASSURANCE LICENSE (adapted from the original BSD license)
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the conditions below are
+met. These conditions require a modest attribution to Dr. Chris C. N. Chu
+(the "Author").
+
+1. Redistributions of the source code, with or without modification (the
+   "Code"), must be accompanied by any documentation and, each time
+   the resulting executable program or a program dependent thereon is
+   launched, a prominent display (e.g., splash screen or banner text) of
+   the Author's attribution information, which includes:
+   (a) Dr. Chris C. N. Chu ("AUTHOR"),
+   (b) Iowa State University ("PROFESSIONAL IDENTIFICATION"), and
+   (c) http://class.ee.iastate.edu/cnchu/ ("URL").
+
+2. Users who intend to use the Code for commercial purposes will notify
+   Author prior to such commercial use.
+
+3. Neither the name nor any trademark of the Author may be used to
+   endorse or promote products derived from this software without
+   specific prior written permission.
+
+4. Users are entirely responsible, to the exclusion of the Author and any
+   other persons, for compliance with (1) regulations set by owners or
+   administrators of employed equipment, (2) licensing terms of any other
+   software, and (3) local, national, and international regulations
+   regarding use, including those regarding import, export, and use of
+   encryption software.
+
+THIS FREE SOFTWARE IS PROVIDED BY THE AUTHOR "AS IS" AND ANY EXPRESS OR
+IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
+OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
+IN NO EVENT SHALL THE AUTHOR OR ANY CONTRIBUTOR BE LIABLE FOR ANY DIRECT,
+INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, EFFECTS OF UNAUTHORIZED OR MALICIOUS
+NETWORK ACCESS; PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
+THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

thirdparty/flute_mp/flute.hpp

@@ -0,0 +1,73 @@
+#ifndef FLUTE_HPP_
+#define FLUTE_HPP_
+
+#include <algorithm>
+#include <climits>
+#include <cmath>
+#include <cstdio>
+#include <cstdlib>
+
+namespace Flute {
+
+#define DEGREE 9                // LUT will be used when d <= DEGREE, DEGREE <= 9
+#define FLUTEROUTING 1          // 1 to construct routing, 0 to estimate WL only
+#define REMOVE_DUPLICATE_PIN 0  // Remove dup. pin for flute_wl() & flute()
+#define ACCURACY 8              // Default accuracy is 3
+#define FLUTE_ACCURACY 8
+// #define MAXD 2008840  // max. degree of a net that can be handled
+#define MAXD 1000  // max. degree of a net that can be handled
+
+using DType = int;
+using DTYPE = int;
+
+// TODO:
+// 1. replace macros with constexpr variables
+// 2. construct a param struct
+// 3. replace all malloc-free contents
+// 4. clear inconsistent comments
+
+struct Branch {
+  DType x, y;  // starting point of the branch
+  int n;       // index of neighbor
+};
+
+struct Tree {
+  int deg;         // degree
+  DType length;    // total wirelength
+  Branch* branch;  // array of tree branches
+};
+
+// Major functions
+void readLUT(const char* powv, const char* post);
+DType flute_wl(int d, DType* x, DType* y, int acc);
+Tree flute(int d, DType* x, DType* y, int acc);
+DType wirelength(Tree t);
+void printtree(Tree t);
+
+// Other useful functions
+DType flutes_wl_LD(int d, DType* xs, DType* ys, int* s);
+DType flutes_wl_MD(int d, DType* xs, DType* ys, int* s, int acc);
+DType flutes_wl_RDP(int d, DType* xs, DType* ys, int* s, int acc);
+Tree flutes_LD(int d, DType* xs, DType* ys, int* s);
+Tree flutes_MD(int d, DType* xs, DType* ys, int* s, int acc);
+Tree flutes_RDP(int d, DType* xs, DType* ys, int* s, int acc);
+
+#if REMOVE_DUPLICATE_PIN == 1
+#define flutes_wl(d, xs, ys, s, acc) flutes_wl_RDP(d, xs, ys, s, acc)
+#define flutes(d, xs, ys, s, acc) flutes_RDP(d, xs, ys, s, acc)
+#else
+#define flutes_wl(d, xs, ys, s, acc) flutes_wl_ALLD(d, xs, ys, s, acc)
+#define flutes(d, xs, ys, s, acc) flutes_ALLD(d, xs, ys, s, acc)
+#endif
+
+#define flutes_wl_ALLD(d, xs, ys, s, acc) flutes_wl_LMD(d, xs, ys, s, acc)
+#define flutes_ALLD(d, xs, ys, s, acc) flutes_LMD(d, xs, ys, s, acc)
+
+#define flutes_wl_LMD(d, xs, ys, s, acc) \
+  (d <= DEGREE ? flutes_wl_LD(d, xs, ys, s) : flutes_wl_MD(d, xs, ys, s, acc))
+#define flutes_LMD(d, xs, ys, s, acc) \
+  (d <= DEGREE ? flutes_LD(d, xs, ys, s) : flutes_MD(d, xs, ys, s, acc))
+
+}  // namespace flute
+
+#endif  // FLUTE_HPP_

utils/io_parser.py

@@ -150,6 +150,10 @@ class IOParser(object):
         die_info = torch.tensor([dieLX, dieHX, dieLY, dieHY]).float()
         # die_shift = torch.tensor([dieLX, dieLY])
         # die_scale = torch.tensor([dieHX - dieLX, dieHY - dieLY])
+        net_names = gpdb.net_names()
+        pin_names = gpdb.pin_names()
+        node_names = gpdb.node_names()
+        microns = gpdb.microns()
 
         siteWidth = gpdb.siteWidth()
         siteHeight = gpdb.siteHeight()
@@ -202,6 +206,10 @@ class IOParser(object):
         design_info = {
             "benchmark": self.params["benchmark"],
             "dataset_path": self.params,
+            "node_names": node_names,
+            "net_names": net_names,
+            "pin_names": pin_names,
+            "microns": microns,
             "node_type_indices": node_type_indices,
             "node_id2node_name": node_id2node_name,
             "node_id2celltype_name": node_id2celltype_name,

utils/setup_dataset.py

@@ -61,6 +61,13 @@ def setup_design_args(args):
     elif args.design_name in ["bigblue3", "bigblue4"]:
         args.num_bin_x = args.num_bin_y = 2048
         args.target_density = 1.0
+    elif args.design_name in ["superblue1", "superblue3", "superblue4", "superblue5", "superblue16", "superblue18"]:
+        args.num_bin_x = args.num_bin_y = 512
+        args.target_density = 1.0
+    elif args.design_name in ["superblue7", "superblue10"]:
+        args.num_bin_x = args.num_bin_y = 1024
+        args.target_density = 1.0
+        args.start_iter = 150
     elif args.design_name in ["adaptec5"]:
         args.target_density = 0.5
         args.num_bin_x = args.num_bin_y = 1024