[AP][Timing] Added Basic Net Weighting

Added basic timing awareness to the AP flow by weighting nets in the AP
solver by their criticality (the max criticality of all edges through
that net). This makes the solver try to minimize the length of nets that
are more critical more than nets that are less critical (according to
the pre-clustering timing analyzer).

Added a command-line option to tradeoff between timing and wirelength in
the AP flow.

Author: AlexandreSinger

doc/src/vpr/command_line_usage.rst

@@ -1253,6 +1253,15 @@ Analytical Placement is generally split into three stages:
 
     **Default:** ``annealer``
 
+.. option:: --ap_timing_tradeoff <float>
+
+    Controls the trade-off between wirelength (HPWL) and delay minimization in the AP flow.
+
+    A value of 0.0 makes the AP flow focus completely on wirelength minimization,
+    while a value of 1.0 makes the AP flow focus completely on timing optimization.
+
+    **Default:** ``0.5``
+
 .. option:: --ap_verbosity <int>
 
     Controls the verbosity of the AP flow output.

vpr/src/analytical_place/analytical_placement_flow.cpp

@@ -7,9 +7,11 @@
 
 #include "analytical_placement_flow.h"
 #include <memory>
+#include "PreClusterTimingManager.h"
 #include "analytical_solver.h"
 #include "ap_netlist.h"
 #include "atom_netlist.h"
+#include "cluster_util.h"
 #include "detailed_placer.h"
 #include "full_legalizer.h"
 #include "gen_ap_netlist_from_atoms.h"
@@ -120,6 +122,7 @@ static PartialPlacement run_global_placer(const t_ap_opts& ap_opts,
                                           const AtomNetlist& atom_nlist,
                                           const APNetlist& ap_netlist,
                                           const Prepacker& prepacker,
+                                          const PreClusterTimingManager& pre_cluster_timing_manager,
                                           const DeviceContext& device_ctx) {
     if (g_vpr_ctx.atom().flat_placement_info().valid) {
         VTR_LOG("Flat Placement is provided in the AP flow, skipping the Global Placement.\n");
@@ -139,6 +142,8 @@ static PartialPlacement run_global_placer(const t_ap_opts& ap_opts,
                                                                          device_ctx.grid,
                                                                          device_ctx.logical_block_types,
                                                                          device_ctx.physical_tile_types,
+                                                                         pre_cluster_timing_manager,
+                                                                         ap_opts.ap_timing_tradeoff,
                                                                          ap_opts.log_verbosity);
         return global_placer->place();
     }
@@ -163,12 +168,25 @@ void run_analytical_placement_flow(t_vpr_setup& vpr_setup) {
                                                      constraints);
     print_ap_netlist_stats(ap_netlist);
 
+    // Pre-compute the pre-clustering timing delays. This object will be passed
+    // into the global placer and the full legalizer to make them timing driven.
+    PreClusterTimingManager pre_cluster_timing_manager(vpr_setup.PackerOpts.timing_driven,
+                                                       atom_nlist,
+                                                       g_vpr_ctx.atom().lookup(),
+                                                       prepacker,
+                                                       vpr_setup.PackerOpts.timing_update_type,
+                                                       *device_ctx.arch,
+                                                       vpr_setup.RoutingArch,
+                                                       vpr_setup.PackerOpts.device_layout,
+                                                       vpr_setup.AnalysisOpts);
+
     // Run the Global Placer.
     const t_ap_opts& ap_opts = vpr_setup.APOpts;
     PartialPlacement p_placement = run_global_placer(ap_opts,
                                                      atom_nlist,
                                                      ap_netlist,
                                                      prepacker,
+                                                     pre_cluster_timing_manager,
                                                      device_ctx);
 
     // Verify that the partial placement is valid before running the full
@@ -185,6 +203,7 @@ void run_analytical_placement_flow(t_vpr_setup& vpr_setup) {
                                                                         ap_netlist,
                                                                         atom_nlist,
                                                                         prepacker,
+                                                                        pre_cluster_timing_manager,
                                                                         vpr_setup,
                                                                         *device_ctx.arch,
                                                                         device_ctx.grid);

vpr/src/analytical_place/analytical_solver.cpp

@@ -13,6 +13,9 @@
 #include <memory>
 #include <utility>
 #include <vector>
+#include "PreClusterTimingManager.h"
+#include "atom_netlist.h"
+#include "atom_netlist_fwd.h"
 #include "device_grid.h"
 #include "flat_placement_types.h"
 #include "partial_placement.h"
@@ -42,12 +45,20 @@
 std::unique_ptr<AnalyticalSolver> make_analytical_solver(e_ap_analytical_solver solver_type,
                                                          const APNetlist& netlist,
                                                          const DeviceGrid& device_grid,
+                                                         const AtomNetlist& atom_netlist,
+                                                         const PreClusterTimingManager& pre_cluster_timing_manager,
+                                                         float ap_timing_tradeoff,
                                                          int log_verbosity) {
     // Based on the solver type passed in, build the solver.
     switch (solver_type) {
         case e_ap_analytical_solver::QP_Hybrid:
 #ifdef EIGEN_INSTALLED
-            return std::make_unique<QPHybridSolver>(netlist, device_grid, log_verbosity);
+            return std::make_unique<QPHybridSolver>(netlist,
+                                                    device_grid,
+                                                    atom_netlist,
+                                                    pre_cluster_timing_manager,
+                                                    ap_timing_tradeoff,
+                                                    log_verbosity);
 #else
             (void)netlist;
             (void)device_grid;
@@ -58,7 +69,12 @@ std::unique_ptr<AnalyticalSolver> make_analytical_solver(e_ap_analytical_solver
 #endif // EIGEN_INSTALLED
         case e_ap_analytical_solver::LP_B2B:
 #ifdef EIGEN_INSTALLED
-            return std::make_unique<B2BSolver>(netlist, device_grid, log_verbosity);
+            return std::make_unique<B2BSolver>(netlist,
+                                               device_grid,
+                                               atom_netlist,
+                                               pre_cluster_timing_manager,
+                                               ap_timing_tradeoff,
+                                               log_verbosity);
 #else
             VPR_FATAL_ERROR(VPR_ERROR_AP,
                             "LP B2B Solver requires the Eigen library");
@@ -72,10 +88,15 @@ std::unique_ptr<AnalyticalSolver> make_analytical_solver(e_ap_analytical_solver
     return nullptr;
 }
 
-AnalyticalSolver::AnalyticalSolver(const APNetlist& netlist, int log_verbosity)
+AnalyticalSolver::AnalyticalSolver(const APNetlist& netlist,
+                                   const AtomNetlist& atom_netlist,
+                                   const PreClusterTimingManager& pre_cluster_timing_manager,
+                                   float ap_timing_tradeoff,
+                                   int log_verbosity)
     : netlist_(netlist)
     , blk_id_to_row_id_(netlist.blocks().size(), APRowId::INVALID())
     , row_id_to_blk_id_(netlist.blocks().size(), APBlockId::INVALID())
+    , net_weights_(netlist.nets().size(), 1.0f)
     , log_verbosity_(log_verbosity) {
     // Get the number of moveable blocks in the netlist and create a unique
     // row ID from [0, num_moveable_blocks) for each moveable block in the
@@ -94,6 +115,21 @@ AnalyticalSolver::AnalyticalSolver(const APNetlist& netlist, int log_verbosity)
         current_row_id++;
         num_moveable_blocks_++;
     }
+
+    if (pre_cluster_timing_manager.is_valid()) {
+        for (APNetId net_id : netlist.nets()) {
+            // Get the atom net associated with the given AP net. When
+            // constructing the AP netlist, we happen to set the name of each
+            // AP net to the same name as the atom net that generated them!
+            // TODO: Create a proper lookup structure to go from the AP Netlist
+            //       back to the Atom Netlist.
+            AtomNetId atom_net_id = atom_netlist.find_net(netlist.net_name(net_id));
+            VTR_ASSERT(atom_net_id.is_valid());
+            float crit = pre_cluster_timing_manager.calc_net_setup_criticality(atom_net_id, atom_netlist);
+
+            net_weights_[net_id] = ap_timing_tradeoff * crit + (1.0f - ap_timing_tradeoff);
+        }
+    }
 }
 
 #ifdef EIGEN_INSTALLED
@@ -201,12 +237,15 @@ void QPHybridSolver::init_linear_system() {
     for (APNetId net_id : netlist_.nets()) {
         size_t num_pins = netlist_.net_pins(net_id).size();
         VTR_ASSERT_DEBUG(num_pins > 1);
+
+        double net_weight = net_weights_[net_id];
+
         if (num_pins > star_num_pins_threshold) {
             // Create a star node and connect each block in the net to the star
             // node.
             // Using the weight from FastPlace
             // TODO: Investigate other weight terms.
-            double w = static_cast<double>(num_pins) / static_cast<double>(num_pins - 1);
+            double w = net_weight * static_cast<double>(num_pins) / static_cast<double>(num_pins - 1);
             size_t star_node_id = num_moveable_blocks_ + star_node_offset;
             for (APPinId pin_id : netlist_.net_pins(net_id)) {
                 APBlockId blk_id = netlist_.pin_block(pin_id);
@@ -220,7 +259,7 @@ void QPHybridSolver::init_linear_system() {
             // exactly once to every other block in the net.
             // Using the weight from FastPlace
             // TODO: Investigate other weight terms.
-            double w = 1.0 / static_cast<double>(num_pins - 1);
+            double w = net_weight * 1.0 / static_cast<double>(num_pins - 1);
             for (size_t ipin_idx = 0; ipin_idx < num_pins; ipin_idx++) {
                 APPinId first_pin_id = netlist_.net_pin(net_id, ipin_idx);
                 APBlockId first_blk_id = netlist_.pin_block(first_pin_id);
@@ -638,6 +677,7 @@ static inline APNetBounds get_unique_net_bounds(APNetId net_id,
 void B2BSolver::add_connection_to_system(APBlockId first_blk_id,
                                          APBlockId second_blk_id,
                                          size_t num_pins,
+                                         double net_w,
                                          const vtr::vector<APBlockId, double>& blk_locs,
                                          std::vector<Eigen::Triplet<double>>& triplet_list,
                                          Eigen::VectorXd& b) {
@@ -660,7 +700,7 @@ void B2BSolver::add_connection_to_system(APBlockId first_blk_id,
     // The denominator of weight is zero, which causes infinity term in the matrix. Another way of
     // interpreting epsilon is the minimum distance two nodes are considered to be in placement.
     double dist = std::max(std::abs(blk_locs[first_blk_id] - blk_locs[second_blk_id]), distance_epsilon_);
-    double w = (2.0 / static_cast<double>(num_pins - 1)) * (1.0 / dist);
+    double w = net_w * (2.0 / static_cast<double>(num_pins - 1)) * (1.0 / dist);
 
     // Update the connectivity matrix and the constant vector.
     // This is similar to how connections are added for the quadratic formulation.
@@ -696,6 +736,8 @@ void B2BSolver::init_linear_system(PartialPlacement& p_placement) {
         size_t num_pins = netlist_.net_pins(net_id).size();
         VTR_ASSERT_SAFE_MSG(num_pins > 1, "net must have at least 2 pins");
 
+        double net_w = net_weights_[net_id];
+
         // Find the bounding blocks
         APNetBounds net_bounds = get_unique_net_bounds(net_id, p_placement, netlist_);
 
@@ -706,19 +748,19 @@ void B2BSolver::init_linear_system(PartialPlacement& p_placement) {
         for (APPinId pin_id : netlist_.net_pins(net_id)) {
             APBlockId blk_id = netlist_.pin_block(pin_id);
             if (blk_id != net_bounds.max_x_blk && blk_id != net_bounds.min_x_blk) {
-                add_connection_to_system(blk_id, net_bounds.max_x_blk, num_pins, p_placement.block_x_locs, triplet_list_x, b_x);
-                add_connection_to_system(blk_id, net_bounds.min_x_blk, num_pins, p_placement.block_x_locs, triplet_list_x, b_x);
+                add_connection_to_system(blk_id, net_bounds.max_x_blk, num_pins, net_w, p_placement.block_x_locs, triplet_list_x, b_x);
+                add_connection_to_system(blk_id, net_bounds.min_x_blk, num_pins, net_w, p_placement.block_x_locs, triplet_list_x, b_x);
             }
             if (blk_id != net_bounds.max_y_blk && blk_id != net_bounds.min_y_blk) {
-                add_connection_to_system(blk_id, net_bounds.max_y_blk, num_pins, p_placement.block_y_locs, triplet_list_y, b_y);
-                add_connection_to_system(blk_id, net_bounds.min_y_blk, num_pins, p_placement.block_y_locs, triplet_list_y, b_y);
+                add_connection_to_system(blk_id, net_bounds.max_y_blk, num_pins, net_w, p_placement.block_y_locs, triplet_list_y, b_y);
+                add_connection_to_system(blk_id, net_bounds.min_y_blk, num_pins, net_w, p_placement.block_y_locs, triplet_list_y, b_y);
             }
         }
 
         // Connect the bounds to each other. Its just easier to put these here
         // instead of in the for loop above.
-        add_connection_to_system(net_bounds.max_x_blk, net_bounds.min_x_blk, num_pins, p_placement.block_x_locs, triplet_list_x, b_x);
-        add_connection_to_system(net_bounds.max_y_blk, net_bounds.min_y_blk, num_pins, p_placement.block_y_locs, triplet_list_y, b_y);
+        add_connection_to_system(net_bounds.max_x_blk, net_bounds.min_x_blk, num_pins, net_w, p_placement.block_x_locs, triplet_list_x, b_x);
+        add_connection_to_system(net_bounds.max_y_blk, net_bounds.min_y_blk, num_pins, net_w, p_placement.block_y_locs, triplet_list_y, b_y);
     }
 
     // Build the sparse connectivity matrices from the triplets.

vpr/src/analytical_place/analytical_solver.h

@@ -31,6 +31,8 @@
 // Forward declarations
 class PartialPlacement;
 class APNetlist;
+class AtomNetlist;
+class PreClusterTimingManager;
 
 /**
  * @brief A strong ID for the rows in a matrix used during solving.
@@ -60,7 +62,11 @@ class AnalyticalSolver {
      * Initializes the internal data members of the base class which are useful
      * for all solvers.
      */
-    AnalyticalSolver(const APNetlist& netlist, int log_verbosity);
+    AnalyticalSolver(const APNetlist& netlist,
+                     const AtomNetlist& atom_netlist,
+                     const PreClusterTimingManager& pre_cluster_timing_manager,
+                     float ap_timing_tradeoff,
+                     int log_verbosity);
 
     /**
      * @brief Run an iteration of the solver using the given partial placement
@@ -113,6 +119,12 @@ class AnalyticalSolver {
     ///        solver.
     vtr::vector<APRowId, APBlockId> row_id_to_blk_id_;
 
+    /// @brief The base weight of each net in the AP netlist. This weight can
+    ///        be used to make the solver more interested in some nets over
+    ///        others. These weights can be any positive value, but are often
+    ///        between 0 and 1.
+    vtr::vector<APNetId, float> net_weights_;
+
     /// @brief The verbosity of log messages in the Analytical Solver.
     int log_verbosity_;
 };
@@ -123,6 +135,9 @@ class AnalyticalSolver {
 std::unique_ptr<AnalyticalSolver> make_analytical_solver(e_ap_analytical_solver solver_type,
                                                          const APNetlist& netlist,
                                                          const DeviceGrid& device_grid,
+                                                         const AtomNetlist& atom_netlist,
+                                                         const PreClusterTimingManager& pre_cluster_timing_manager,
+                                                         float ap_timing_tradeoff,
                                                          int log_verbosity);
 
 // The Eigen library is used to solve matrix equations in the following solvers.
@@ -278,8 +293,11 @@ class QPHybridSolver : public AnalyticalSolver {
      */
     QPHybridSolver(const APNetlist& netlist,
                    const DeviceGrid& device_grid,
+                   const AtomNetlist& atom_netlist,
+                   const PreClusterTimingManager& pre_cluster_timing_manager,
+                   float ap_timing_tradeoff,
                    int log_verbosity)
-        : AnalyticalSolver(netlist, log_verbosity) {
+        : AnalyticalSolver(netlist, atom_netlist, pre_cluster_timing_manager, ap_timing_tradeoff, log_verbosity) {
         // Initializing the linear system only depends on the netlist and fixed
         // block locations. Both are provided by the netlist, allowing this to
         // be initialized in the constructor.
@@ -411,8 +429,11 @@ class B2BSolver : public AnalyticalSolver {
   public:
     B2BSolver(const APNetlist& ap_netlist,
               const DeviceGrid& device_grid,
+              const AtomNetlist& atom_netlist,
+              const PreClusterTimingManager& pre_cluster_timing_manager,
+              float ap_timing_tradeoff,
               int log_verbosity)
-        : AnalyticalSolver(ap_netlist, log_verbosity)
+        : AnalyticalSolver(ap_netlist, atom_netlist, pre_cluster_timing_manager, ap_timing_tradeoff, log_verbosity)
         , device_grid_width_(device_grid.width())
         , device_grid_height_(device_grid.height()) {}
 
@@ -503,6 +524,7 @@ class B2BSolver : public AnalyticalSolver {
     void add_connection_to_system(APBlockId first_blk_id,
                                   APBlockId second_blk_id,
                                   size_t num_pins,
+                                  double net_w,
                                   const vtr::vector<APBlockId, double>& blk_locs,
                                   std::vector<Eigen::Triplet<double>>& triplet_list,
                                   Eigen::VectorXd& b);