Merge pull request #3160 from AlexandreSinger/feature-ap-anneal_init_t

[AP][DP] Added Option to Scale Init T in Annealer

Author: AlexandreSinger

doc/src/vpr/command_line_usage.rst

@@ -810,6 +810,22 @@ If any of init_t, exit_t or alpha_t is specified, the user schedule, with a fixe
 
     **Default:** ``circuit``
 
+.. option:: --anneal_auto_init_t_scale <float>
+
+    A scale on the starting temperature of the anneal for the automatic annealing
+    schedule.
+
+    When in the automatic annealing schedule, the annealer will select a good
+    initial temperature based on the quality of the initial placement. This option
+    allows you to scale that initial temperature up or down by multiplying the
+    initial temperature by the given scale. Increasing this number
+    will increase the initial temperature which will have the annealer potentially
+    explore more of the space at the expense of run time. Depending on the quality
+    of the initial placement, this may improve or hurt the quality of the final
+    placement.
+
+    **Default:** ``1.0``
+
 .. option:: --init_t <float>
 
     The starting temperature of the anneal for the manual annealing schedule.

vpr/src/analytical_place/detailed_placer.cpp

@@ -75,6 +75,19 @@ AnnealerDetailedPlacer::AnnealerDetailedPlacer(const BlkLocRegistry& curr_cluste
         }
     }
 
+    // The solution produced by the AP flow is significantly better than the
+    // initial placement solution produced by the initial placer in the default
+    // flow. Even though the annealer auto-selects its initial temperature based
+    // on the quality of the placement, we found that the initial temperatute was
+    // still too high and it was hurting quality. This scales down the initial
+    // temperature auto-selected by the annealer to prevent this and improve
+    // runtime.
+    // Here we still scale by whatever the user passed in to scale the initial
+    // temperature by, but we also scale it down further. This allows AP to scale
+    // the initial temperature down by default, while still allowing the user
+    // some control.
+    float anneal_auto_init_t_scale = vpr_setup.PlacerOpts.place_auto_init_t_scale * 0.5f;
+
     placer_ = std::make_unique<Placer>((const Netlist<>&)clustered_netlist,
                                        curr_clustered_placement,
                                        vpr_setup.PlacerOpts,
@@ -84,6 +97,7 @@ AnnealerDetailedPlacer::AnnealerDetailedPlacer(const BlkLocRegistry& curr_cluste
                                        netlist_pin_lookup_,
                                        FlatPlacementInfo(),
                                        place_delay_model,
+                                       anneal_auto_init_t_scale,
                                        g_vpr_ctx.placement().cube_bb,
                                        false /*is_flat*/,
                                        false /*quiet*/);

vpr/src/base/CheckSetup.cpp

@@ -69,6 +69,11 @@ void CheckSetup(const t_packer_opts& packer_opts,
                         NUM_PL_MOVE_TYPES);
     }
 
+    if (placer_opts.place_auto_init_t_scale < 0.0) {
+        VPR_FATAL_ERROR(VPR_ERROR_OTHER,
+                        "Cannot have negative annealer auto initial temperature scale.\n");
+    }
+
     // Rules for doing Analytical Placement
     if (ap_opts.doAP) {
         // Make sure that the --place option was not set.

vpr/src/base/SetupVPR.cpp

@@ -703,6 +703,8 @@ static void SetupPlacerOpts(const t_options& Options, t_placer_opts* PlacerOpts)
 
     PlacerOpts->placer_debug_block = Options.placer_debug_block;
     PlacerOpts->placer_debug_net = Options.placer_debug_net;
+
+    PlacerOpts->place_auto_init_t_scale = Options.place_auto_init_t_scale.value();
 }
 
 static void SetupAnalysisOpts(const t_options& Options, t_analysis_opts& analysis_opts) {

vpr/src/base/read_options.cpp

@@ -2200,6 +2200,22 @@ argparse::ArgumentParser create_arg_parser(const std::string& prog_name, t_optio
         .default_value("circuit")
         .show_in(argparse::ShowIn::HELP_ONLY);
 
+    place_grp.add_argument(args.place_auto_init_t_scale, "--anneal_auto_init_t_scale")
+        .help(
+            "A scale on the starting temperature of the anneal for the automatic annealing "
+            "schedule.\n"
+            "\n"
+            "When in the automatic annealing schedule, the annealer will select a good "
+            "initial temperature based on the quality of the initial placement. This option "
+            "allows you to scale that initial temperature up or down by multiplying the "
+            "initial temperature by the given scale. Increasing this number "
+            "will increase the initial temperature which will have the annealer potentially "
+            "explore more of the space at the expense of run time. Depending on the quality "
+            "of the initial placement, this may improve or hurt the quality of the final "
+            "placement.")
+        .default_value("1.0")
+        .show_in(argparse::ShowIn::HELP_ONLY);
+
     place_grp.add_argument(args.PlaceInitT, "--init_t")
         .help("Initial temperature for manual annealing schedule")
         .default_value("100.0")

vpr/src/base/read_options.h

@@ -129,6 +129,7 @@ struct t_options {
     argparse::ArgValue<int> Seed;
     argparse::ArgValue<bool> ShowPlaceTiming;
     argparse::ArgValue<float> PlaceInnerNum;
+    argparse::ArgValue<float> place_auto_init_t_scale;
     argparse::ArgValue<float> PlaceInitT;
     argparse::ArgValue<float> PlaceExitT;
     argparse::ArgValue<float> PlaceAlphaT;

vpr/src/base/vpr_types.h

@@ -1004,8 +1004,9 @@ enum class e_move_type;
  *   @param place_constraint_subtile
  *              True if subtiles should be specified when printing floorplan
  *              constraints. False if not.
- *
- *
+ *   @param place_auto_init_t_scale
+ *              When the annealer is using the automatic schedule, this option
+ *              scales the initial temperature selected.
  */
 struct t_placer_opts {
     t_place_algorithm place_algorithm;
@@ -1077,6 +1078,8 @@ struct t_placer_opts {
     std::string allowed_tiles_for_delay_model;
 
     e_place_delta_delay_algorithm place_delta_delay_matrix_calculation_method;
+
+    float place_auto_init_t_scale;
 };
 
 /******************************************************************

vpr/src/place/annealer.cpp

@@ -206,6 +206,7 @@ PlacementAnnealer::PlacementAnnealer(const t_placer_opts& placer_opts,
                                      PlacerSetupSlacks* setup_slacks,
                                      SetupTimingInfo* timing_info,
                                      NetPinTimingInvalidator* pin_timing_invalidator,
+                                     float auto_init_t_scale,
                                      int move_lim)
     : placer_opts_(placer_opts)
     , placer_state_(placer_state)
@@ -285,7 +286,8 @@ PlacementAnnealer::PlacementAnnealer(const t_placer_opts& placer_opts,
     move_type_stats_.rejected_moves.resize({device_ctx.logical_block_types.size(), (int)e_move_type::NUMBER_OF_AUTO_MOVES}, 0);
 
     // Update the starting temperature for placement annealing to a more appropriate value
-    annealing_state_.t = estimate_starting_temperature_();
+    VTR_ASSERT_SAFE_MSG(auto_init_t_scale >= 0, "Initial temperature scale cannot be negative.");
+    annealing_state_.t = estimate_starting_temperature_() * auto_init_t_scale;
 }
 
 float PlacementAnnealer::estimate_starting_temperature_() {

vpr/src/place/annealer.h

@@ -185,6 +185,7 @@ class PlacementAnnealer {
                       PlacerSetupSlacks* setup_slacks,
                       SetupTimingInfo* timing_info,
                       NetPinTimingInvalidator* pin_timing_invalidator,
+                      float auto_init_t_scale,
                       int move_lim);
 
     /**

vpr/src/place/place.cpp

@@ -112,7 +112,8 @@ void try_place(const Netlist<>& net_list,
     ClusteredPinAtomPinsLookup netlist_pin_lookup(cluster_ctx.clb_nlist, atom_ctx.netlist(), pb_gpin_lookup);
 
     Placer placer(net_list, {}, placer_opts, analysis_opts, noc_opts, pb_gpin_lookup, netlist_pin_lookup,
-                  flat_placement_info, place_delay_model, mutable_placement.cube_bb, is_flat, /*quiet=*/false);
+                  flat_placement_info, place_delay_model, placer_opts.place_auto_init_t_scale,
+                  mutable_placement.cube_bb, is_flat, /*quiet=*/false);
 
     placer.place();
 

vpr/src/place/placer.cpp

@@ -35,6 +35,7 @@ Placer::Placer(const Netlist<>& net_list,
                const ClusteredPinAtomPinsLookup& netlist_pin_lookup,
                const FlatPlacementInfo& flat_placement_info,
                std::shared_ptr<PlaceDelayModel> place_delay_model,
+               float anneal_auto_init_t_scale,
                bool cube_bb,
                bool is_flat,
                bool quiet)
@@ -152,6 +153,7 @@ Placer::Placer(const Netlist<>& net_list,
     annealer_ = std::make_unique<PlacementAnnealer>(placer_opts_, placer_state_, place_macros, costs_, net_cost_handler_, noc_cost_handler_,
                                                     noc_opts_, rng_, std::move(move_generator), std::move(move_generator2), place_delay_model_.get(),
                                                     placer_criticalities_.get(), placer_setup_slacks_.get(), timing_info_.get(), pin_timing_invalidator_.get(),
+                                                    anneal_auto_init_t_scale,
                                                     move_lim);
 }
 

vpr/src/place/placer.h

@@ -49,6 +49,7 @@ class Placer {
            const ClusteredPinAtomPinsLookup& netlist_pin_lookup,
            const FlatPlacementInfo& flat_placement_info,
            std::shared_ptr<PlaceDelayModel> place_delay_model,
+           float anneal_auto_init_t_scale,
            bool cube_bb,
            bool is_flat,
            bool quiet);