Merge pull request #2965 from AlexandreSinger/feature-pack-timing-manager

AlexandreSinger · web-flow · commit 433a91749597 · 2025-04-10T14:54:01.000-04:00
[Pack][Timing] Abstracted How Timing is Used in the Packer
diff --git a/vpr/src/analytical_place/full_legalizer.cpp b/vpr/src/analytical_place/full_legalizer.cpp
@@ -58,7 +58,7 @@ std::unique_ptr<FullLegalizer> make_full_legalizer(e_ap_full_legalizer full_lega
                                                    const APNetlist& ap_netlist,
                                                    const AtomNetlist& atom_netlist,
                                                    const Prepacker& prepacker,
-                                                   t_vpr_setup& vpr_setup,
+                                                   const t_vpr_setup& vpr_setup,
                                                    const t_arch& arch,
                                                    const DeviceGrid& device_grid) {
     switch (full_legalizer_type) {
@@ -513,8 +513,8 @@ void APPack::legalize(const PartialPlacement& p_placement) {
     }
 
     // Run the Packer stage with the flat placement as a hint.
-    try_pack(&vpr_setup_.PackerOpts,
-             &vpr_setup_.AnalysisOpts,
+    try_pack(vpr_setup_.PackerOpts,
+             vpr_setup_.AnalysisOpts,
              arch_,
              vpr_setup_.RoutingArch,
              vpr_setup_.PackerRRGraph,
diff --git a/vpr/src/analytical_place/full_legalizer.h b/vpr/src/analytical_place/full_legalizer.h
@@ -37,7 +37,7 @@ class FullLegalizer {
     FullLegalizer(const APNetlist& ap_netlist,
                   const AtomNetlist& atom_netlist,
                   const Prepacker& prepacker,
-                  t_vpr_setup& vpr_setup,
+                  const t_vpr_setup& vpr_setup,
                   const t_arch& arch,
                   const DeviceGrid& device_grid)
         : ap_netlist_(ap_netlist)
@@ -68,7 +68,7 @@ class FullLegalizer {
 
     /// @brief The VPR setup options passed into the VPR flow. This must be
     ///        mutable since some parts of packing modify the options.
-    t_vpr_setup& vpr_setup_;
+    const t_vpr_setup& vpr_setup_;
 
     /// @brief Information on the architecture of the FPGA.
     const t_arch& arch_;
@@ -84,7 +84,7 @@ std::unique_ptr<FullLegalizer> make_full_legalizer(e_ap_full_legalizer full_lega
                                                    const APNetlist& ap_netlist,
                                                    const AtomNetlist& atom_netlist,
                                                    const Prepacker& prepacker,
-                                                   t_vpr_setup& vpr_setup,
+                                                   const t_vpr_setup& vpr_setup,
                                                    const t_arch& arch,
                                                    const DeviceGrid& device_grid);
 
diff --git a/vpr/src/base/SetupVPR.cpp b/vpr/src/base/SetupVPR.cpp
@@ -591,10 +591,6 @@ void SetupPackerOpts(const t_options& Options,
     PackerOpts->feasible_block_array_size = Options.pack_feasible_block_array_size;
     PackerOpts->use_attraction_groups = Options.use_attraction_groups;
 
-    //TODO: document?
-    PackerOpts->inter_cluster_net_delay = 1.0; /* DEFAULT */
-    PackerOpts->auto_compute_inter_cluster_net_delay = true;
-
     PackerOpts->device_layout = Options.device_layout;
 
     PackerOpts->timing_update_type = Options.timing_update_type;
diff --git a/vpr/src/base/ShowSetup.cpp b/vpr/src/base/ShowSetup.cpp
@@ -757,7 +757,6 @@ static void ShowPackerOpts(const t_packer_opts& PackerOpts) {
     }
     VTR_LOG("PackerOpts.connection_driven: %s", (PackerOpts.connection_driven ? "true\n" : "false\n"));
     VTR_LOG("PackerOpts.global_clocks: %s", (PackerOpts.global_clocks ? "true\n" : "false\n"));
-    VTR_LOG("PackerOpts.inter_cluster_net_delay: %f\n", PackerOpts.inter_cluster_net_delay);
     VTR_LOG("PackerOpts.timing_driven: %s", (PackerOpts.timing_driven ? "true\n" : "false\n"));
     VTR_LOG("PackerOpts.target_external_pin_util: %s", vtr::join(PackerOpts.target_external_pin_util, " ").c_str());
     VTR_LOG("\n");
diff --git a/vpr/src/base/vpr_api.cpp b/vpr/src/base/vpr_api.cpp
@@ -620,7 +620,7 @@ bool vpr_pack(t_vpr_setup& vpr_setup, const t_arch& arch) {
     const Prepacker prepacker(g_vpr_ctx.atom().netlist(),
                               g_vpr_ctx.device().logical_block_types);
 
-    return try_pack(&vpr_setup.PackerOpts, &vpr_setup.AnalysisOpts,
+    return try_pack(vpr_setup.PackerOpts, vpr_setup.AnalysisOpts,
                     arch, vpr_setup.RoutingArch,
                     vpr_setup.PackerRRGraph,
                     prepacker,
diff --git a/vpr/src/base/vpr_types.h b/vpr/src/base/vpr_types.h
@@ -717,9 +717,7 @@ struct t_packer_opts {
     enum e_cluster_seed cluster_seed_type;
     float alpha;
     float beta;
-    float inter_cluster_net_delay;
     float target_device_utilization;
-    bool auto_compute_inter_cluster_net_delay;
     e_unrelated_clustering allow_unrelated_clustering;
     bool connection_driven;
     int pack_verbosity;
diff --git a/vpr/src/pack/cluster_util.cpp b/vpr/src/pack/cluster_util.cpp
@@ -2,17 +2,13 @@
 #include <algorithm>
 #include <unordered_set>
 
-#include "PreClusterTimingGraphResolver.h"
-#include "PreClusterDelayCalculator.h"
 #include "atom_netlist.h"
 #include "attraction_groups.h"
 #include "cluster_legalizer.h"
 #include "clustered_netlist.h"
-#include "concrete_timing_info.h"
+#include "globals.h"
 #include "output_clustering.h"
 #include "prepack.h"
-#include "tatum/TimingReporter.hpp"
-#include "tatum/echo_writer.hpp"
 #include "vpr_context.h"
 
 /*Print the contents of each cluster to an echo file*/
@@ -67,58 +63,6 @@ static void echo_clusters(char* filename, const ClusterLegalizer& cluster_legali
     fclose(fp);
 }
 
-void calc_init_packing_timing(const t_packer_opts& packer_opts,
-                              const t_analysis_opts& analysis_opts,
-                              const Prepacker& prepacker,
-                              std::shared_ptr<PreClusterDelayCalculator>& clustering_delay_calc,
-                              std::shared_ptr<SetupTimingInfo>& timing_info,
-                              vtr::vector<AtomBlockId, float>& atom_criticality) {
-    const AtomContext& atom_ctx = g_vpr_ctx.atom();
-
-    /*
-     * Initialize the timing analyzer
-     */
-    clustering_delay_calc = std::make_shared<PreClusterDelayCalculator>(atom_ctx.netlist(), atom_ctx.lookup(), packer_opts.inter_cluster_net_delay, prepacker);
-    timing_info = make_setup_timing_info(clustering_delay_calc, packer_opts.timing_update_type);
-
-    //Calculate the initial timing
-    timing_info->update();
-
-    if (isEchoFileEnabled(E_ECHO_PRE_PACKING_TIMING_GRAPH)) {
-        auto& timing_ctx = g_vpr_ctx.timing();
-        tatum::write_echo(getEchoFileName(E_ECHO_PRE_PACKING_TIMING_GRAPH),
-                          *timing_ctx.graph, *timing_ctx.constraints, *clustering_delay_calc, timing_info->analyzer());
-
-        tatum::NodeId debug_tnode = id_or_pin_name_to_tnode(analysis_opts.echo_dot_timing_graph_node);
-        write_setup_timing_graph_dot(getEchoFileName(E_ECHO_PRE_PACKING_TIMING_GRAPH) + std::string(".dot"),
-                                     *timing_info, debug_tnode);
-    }
-
-    {
-        auto& timing_ctx = g_vpr_ctx.timing();
-        PreClusterTimingGraphResolver resolver(atom_ctx.netlist(),
-                                               atom_ctx.lookup(), *timing_ctx.graph, *clustering_delay_calc);
-        resolver.set_detail_level(analysis_opts.timing_report_detail);
-
-        tatum::TimingReporter timing_reporter(resolver, *timing_ctx.graph,
-                                              *timing_ctx.constraints);
-
-        timing_reporter.report_timing_setup(
-            "pre_pack.report_timing.setup.rpt",
-            *timing_info->setup_analyzer(),
-            analysis_opts.timing_report_npaths);
-    }
-
-    //Calculate true criticalities of each block
-    for (AtomBlockId blk : atom_ctx.netlist().blocks()) {
-        for (AtomPinId in_pin : atom_ctx.netlist().block_input_pins(blk)) {
-            //Max criticality over incoming nets
-            float crit = timing_info->setup_pin_criticality(in_pin);
-            atom_criticality[blk] = std::max(atom_criticality[blk], crit);
-        }
-    }
-}
-
 void check_and_output_clustering(ClusterLegalizer& cluster_legalizer,
                                  const t_packer_opts& packer_opts,
                                  const std::unordered_set<AtomNetId>& is_clock,
diff --git a/vpr/src/pack/cluster_util.h b/vpr/src/pack/cluster_util.h
@@ -11,10 +11,6 @@ class AttractionInfo;
 class ClusterBlockId;
 class ClusterLegalizer;
 class ClusteredNetlist;
-class PreClusterDelayCalculator;
-class Prepacker;
-class SetupTimingInfo;
-class t_pack_molecule;
 struct AtomContext;
 
 /**
@@ -26,16 +22,6 @@ struct AtomContext;
 /*   Clustering helper functions   */
 /***********************************/
 
-/*
- * @brief Calculate the initial timing at the start of packing stage.
- */
-void calc_init_packing_timing(const t_packer_opts& packer_opts,
-                              const t_analysis_opts& analysis_opts,
-                              const Prepacker& prepacker,
-                              std::shared_ptr<PreClusterDelayCalculator>& clustering_delay_calc,
-                              std::shared_ptr<SetupTimingInfo>& timing_info,
-                              vtr::vector<AtomBlockId, float>& atom_criticality);
-
 /*
  * @brief Check clustering legality and output it.
  */
diff --git a/vpr/src/pack/greedy_candidate_selector.cpp b/vpr/src/pack/greedy_candidate_selector.cpp
@@ -10,6 +10,7 @@
 #include <cmath>
 #include <queue>
 #include <vector>
+#include "PreClusterTimingManager.h"
 #include "appack_context.h"
 #include "flat_placement_types.h"
 #include "flat_placement_utils.h"
@@ -90,7 +91,7 @@ GreedyCandidateSelector::GreedyCandidateSelector(
     const std::unordered_set<AtomNetId>& is_clock,
     const std::unordered_set<AtomNetId>& is_global,
     const std::unordered_set<AtomNetId>& net_output_feeds_driving_block_input,
-    const SetupTimingInfo& timing_info,
+    const PreClusterTimingManager& pre_cluster_timing_manager,
     const APPackContext& appack_ctx,
     int log_verbosity)
     : atom_netlist_(atom_netlist)
@@ -103,7 +104,7 @@ GreedyCandidateSelector::GreedyCandidateSelector(
     , is_clock_(is_clock)
     , is_global_(is_global)
     , net_output_feeds_driving_block_input_(net_output_feeds_driving_block_input)
-    , timing_info_(timing_info)
+    , pre_cluster_timing_manager_(pre_cluster_timing_manager)
     , appack_ctx_(appack_ctx)
     , rng_(0) {
 
@@ -544,12 +545,15 @@ void GreedyCandidateSelector::update_timing_gain_values(
     if (net_output_feeds_driving_block_input_.count(net_id) != 0)
         pins = atom_netlist_.net_sinks(net_id);
 
+    // Get the setup timing info used to compute timing gain terms.
+    const SetupTimingInfo& timing_info = pre_cluster_timing_manager_.get_timing_info();
+
     if (net_relation_to_clustered_block == e_net_relation_to_clustered_block::OUTPUT
         && !is_global_.count(net_id)) {
         for (AtomPinId pin_id : pins) {
             AtomBlockId blk_id = atom_netlist_.pin_block(pin_id);
             if (!cluster_legalizer.is_atom_clustered(blk_id)) {
-                double timing_gain = timing_info_.setup_pin_criticality(pin_id);
+                double timing_gain = timing_info.setup_pin_criticality(pin_id);
 
                 if (cluster_gain_stats.timing_gain.count(blk_id) == 0) {
                     cluster_gain_stats.timing_gain[blk_id] = 0;
@@ -569,7 +573,7 @@ void GreedyCandidateSelector::update_timing_gain_values(
 
         if (!cluster_legalizer.is_atom_clustered(new_blk_id)) {
             for (AtomPinId pin_id : atom_netlist_.net_sinks(net_id)) {
-                double timing_gain = timing_info_.setup_pin_criticality(pin_id);
+                double timing_gain = timing_info.setup_pin_criticality(pin_id);
 
                 if (cluster_gain_stats.timing_gain.count(new_blk_id) == 0) {
                     cluster_gain_stats.timing_gain[new_blk_id] = 0;
diff --git a/vpr/src/pack/greedy_candidate_selector.h b/vpr/src/pack/greedy_candidate_selector.h
@@ -26,8 +26,8 @@
 class AtomNetlist;
 class AttractionInfo;
 class FlatPlacementInfo;
+class PreClusterTimingManager;
 class Prepacker;
-class SetupTimingInfo;
 class t_pack_high_fanout_thresholds;
 struct t_model;
 struct t_molecule_stats;
@@ -225,9 +225,10 @@ class GreedyCandidateSelector {
      *              The set of nets whose output feeds the block that drives
      *              itself. This may cause double-counting in the gain
      *              calculations and needs special handling.
-     *  @param timing_info
-     *              Setup timing info for this Atom Netlist. Used to incorporate
-     *              timing / criticality into the gain calculation.
+     *  @param pre_cluster_timing_manager
+     *              Timing manager that holds the information on timing of
+     *              different connections in the circuit. Used for computing
+     *              the timing gain terms.
      *  @param appack_ctx
      *              The APPack context which contains options for the flat
      *              placement guided packing.
@@ -244,7 +245,7 @@ class GreedyCandidateSelector {
                             const std::unordered_set<AtomNetId>& is_clock,
                             const std::unordered_set<AtomNetId>& is_global,
                             const std::unordered_set<AtomNetId>& net_output_feeds_driving_block_input,
-                            const SetupTimingInfo& timing_info,
+                            const PreClusterTimingManager& pre_cluster_timing_manager,
                             const APPackContext& appack_ctx,
                             int log_verbosity);
 
@@ -565,8 +566,9 @@ class GreedyCandidateSelector {
     ///        drive them.
     const std::unordered_set<AtomNetId>& net_output_feeds_driving_block_input_;
 
-    /// @brief Setup timing info used to help select critical candidates to pack.
-    const SetupTimingInfo& timing_info_;
+    /// @brief The pre-clustering timing manager which holds the timing information
+    ///        of the primitive netlist.
+    const PreClusterTimingManager& pre_cluster_timing_manager_;
 
     /// @brief Inter-block nets within a finalized cluster. Used for finding
     ///        transitive candidates.
diff --git a/vpr/src/pack/greedy_clusterer.cpp b/vpr/src/pack/greedy_clusterer.cpp
@@ -79,6 +79,7 @@ GreedyClusterer::GreedyClusterer(const t_packer_opts& packer_opts,
                                  const t_pack_high_fanout_thresholds& high_fanout_thresholds,
                                  const std::unordered_set<AtomNetId>& is_clock,
                                  const std::unordered_set<AtomNetId>& is_global,
+                                 const PreClusterTimingManager& pre_cluster_timing_manager,
                                  const APPackContext& appack_ctx)
     : packer_opts_(packer_opts)
     , analysis_opts_(analysis_opts)
@@ -87,6 +88,7 @@ GreedyClusterer::GreedyClusterer(const t_packer_opts& packer_opts,
     , high_fanout_thresholds_(high_fanout_thresholds)
     , is_clock_(is_clock)
     , is_global_(is_global)
+    , pre_cluster_timing_manager_(pre_cluster_timing_manager)
     , appack_ctx_(appack_ctx)
     , primitive_candidate_block_types_(identify_primitive_candidate_block_types())
     , log_verbosity_(packer_opts.pack_verbosity)
@@ -113,18 +115,6 @@ GreedyClusterer::do_clustering(ClusterLegalizer& cluster_legalizer,
     t_cluster_progress_stats clustering_stats;
     clustering_stats.num_molecules = prepacker.molecules().size();
 
-    // TODO: Create a ClusteringTimingManager class.
-    //       This code relies on the prepacker, once the prepacker is moved to
-    //       the constructor, this code can also move to the constructor.
-    std::shared_ptr<PreClusterDelayCalculator> clustering_delay_calc;
-    std::shared_ptr<SetupTimingInfo> timing_info;
-    // Default criticalities set to zero (e.g. if not timing driven)
-    vtr::vector<AtomBlockId, float> atom_criticality(atom_netlist_.blocks().size(), 0.f);
-    if (packer_opts_.timing_driven) {
-        calc_init_packing_timing(packer_opts_, analysis_opts_, prepacker,
-                                 clustering_delay_calc, timing_info, atom_criticality);
-    }
-
     // Calculate the max molecule stats, which is used for gain calculation.
     const t_molecule_stats max_molecule_stats = prepacker.calc_max_molecule_stats(atom_netlist_);
 
@@ -140,7 +130,7 @@ GreedyClusterer::do_clustering(ClusterLegalizer& cluster_legalizer,
                                                is_clock_,
                                                is_global_,
                                                net_output_feeds_driving_block_input_,
-                                               *timing_info,
+                                               pre_cluster_timing_manager_,
                                                appack_ctx_,
                                                log_verbosity_);
 
@@ -149,7 +139,7 @@ GreedyClusterer::do_clustering(ClusterLegalizer& cluster_legalizer,
                                      prepacker,
                                      packer_opts_.cluster_seed_type,
                                      max_molecule_stats,
-                                     atom_criticality);
+                                     pre_cluster_timing_manager_);
 
     // Pick the first seed molecule.
     PackMoleculeId seed_mol_id = seed_selector.get_next_seed(prepacker,
diff --git a/vpr/src/pack/greedy_clusterer.h b/vpr/src/pack/greedy_clusterer.h
@@ -22,7 +22,7 @@ class AtomNetlist;
 class AttractionInfo;
 class DeviceContext;
 class GreedyCandidateSelector;
-class SetupTimingInfo;
+class PreClusterTimingManager;
 class t_pack_high_fanout_thresholds;
 struct t_analysis_opts;
 struct t_clustering_data;
@@ -76,6 +76,11 @@ class GreedyClusterer {
      *              The set of global nets in the Atom Netlist. These will be
      *              routed on special dedicated networks, and hence are less
      *              relavent to locality / attraction.
+     *  @param pre_cluster_timing_manager
+     *              Timing manager class which holds the timing information of
+     *              the primitive netlist. Used by the seed selector to select
+     *              critical seeds and the candidate selector to select
+     *              timing critical candidates.
      *  @param appack_ctx
      *              The APPack state. This contains the options used to
      *              configure APPack and the flat placement.
@@ -87,6 +92,7 @@ class GreedyClusterer {
                     const t_pack_high_fanout_thresholds& high_fanout_thresholds,
                     const std::unordered_set<AtomNetId>& is_clock,
                     const std::unordered_set<AtomNetId>& is_global,
+                    const PreClusterTimingManager& pre_cluster_timing_manager,
                     const APPackContext& appack_ctx);
 
     /**
@@ -233,6 +239,9 @@ class GreedyClusterer {
     /// @brief A set of atom nets which are considered as global nets.
     const std::unordered_set<AtomNetId>& is_global_;
 
+    /// @brief Timing manager class which holds the primitive-level timing information.
+    const PreClusterTimingManager& pre_cluster_timing_manager_;
+
     /// @brief The APPack state. This is used by the candidate selector to try
     ///        and propose better candidates based on a flat placement.
     const APPackContext& appack_ctx_;
diff --git a/vpr/src/pack/greedy_seed_selector.cpp b/vpr/src/pack/greedy_seed_selector.cpp
diff --git a/vpr/src/pack/greedy_seed_selector.h b/vpr/src/pack/greedy_seed_selector.h
diff --git a/vpr/src/pack/pack.cpp b/vpr/src/pack/pack.cpp
diff --git a/vpr/src/pack/pack.h b/vpr/src/pack/pack.h
diff --git a/vpr/src/timing/PreClusterTimingManager.cpp b/vpr/src/timing/PreClusterTimingManager.cpp
diff --git a/vpr/src/timing/PreClusterTimingManager.h b/vpr/src/timing/PreClusterTimingManager.h

Original file line number	Diff line number	Diff line change
`@@ -757,7 +757,6 @@ static void ShowPackerOpts(const t_packer_opts& PackerOpts) {`
`757`	`757`	`}`
`758`	`758`	`VTR_LOG("PackerOpts.connection_driven: %s", (PackerOpts.connection_driven ? "true\n" : "false\n"));`
`759`	`759`	`VTR_LOG("PackerOpts.global_clocks: %s", (PackerOpts.global_clocks ? "true\n" : "false\n"));`
`760`		`- VTR_LOG("PackerOpts.inter_cluster_net_delay: %f\n", PackerOpts.inter_cluster_net_delay);`
`761`	`760`	`VTR_LOG("PackerOpts.timing_driven: %s", (PackerOpts.timing_driven ? "true\n" : "false\n"));`
`762`	`761`	`VTR_LOG("PackerOpts.target_external_pin_util: %s", vtr::join(PackerOpts.target_external_pin_util, " ").c_str());`
`763`	`762`	`VTR_LOG("\n");`