[Packer] Created GreedyClusterer Class

Began encapsulating the different parts of the packer into classes. This
will help organize the packer better, which will make it easier to
modify in the future.

My plan is to clean up the different parts of the packer so I can add
flat placement information into the gain calculation so it can be
integrated into the AP flow.

Author: AlexandreSinger

vpr/src/analytical_place/full_legalizer.cpp

@@ -17,7 +17,6 @@
 #include "ShowSetup.h"
 #include "ap_netlist_fwd.h"
 #include "check_netlist.h"
-#include "cluster.h"
 #include "cluster_legalizer.h"
 #include "cluster_util.h"
 #include "clustered_netlist.h"

vpr/src/base/vpr_api.cpp

@@ -64,7 +64,6 @@
 #include "check_route.h"
 #include "constant_nets.h"
 #include "atom_netlist_utils.h"
-#include "cluster.h"
 #include "output_clustering.h"
 #include "vpr_constraints_reader.h"
 #include "place_constraints.h"

vpr/src/pack/cluster.h

@@ -1879,6 +1879,35 @@ void print_pb_type_count_recurr(t_pb_type* pb_type, size_t max_name_chars, size_
     }
 }
 
+/**
+ * Print the total number of used physical blocks for each pb type in the architecture
+ */
+void print_pb_type_count(const ClusteredNetlist& clb_nlist) {
+    auto& device_ctx = g_vpr_ctx.device();
+
+    std::map<t_pb_type*, int> pb_type_count;
+
+    size_t max_depth = 0;
+    for (ClusterBlockId blk : clb_nlist.blocks()) {
+        size_t pb_max_depth = update_pb_type_count(clb_nlist.block_pb(blk), pb_type_count, 0);
+
+        max_depth = std::max(max_depth, pb_max_depth);
+    }
+
+    size_t max_pb_type_name_chars = 0;
+    for (auto& pb_type : pb_type_count) {
+        max_pb_type_name_chars = std::max(max_pb_type_name_chars, strlen(pb_type.first->name));
+    }
+
+    VTR_LOG("\nPb types usage...\n");
+    for (const auto& logical_block_type : device_ctx.logical_block_types) {
+        if (!logical_block_type.pb_type) continue;
+
+        print_pb_type_count_recurr(logical_block_type.pb_type, max_pb_type_name_chars + max_depth, 0, pb_type_count);
+    }
+    VTR_LOG("\n");
+}
+
 t_logical_block_type_ptr identify_logic_block_type(std::map<const t_model*, std::vector<t_logical_block_type_ptr>>& primitive_candidate_block_types) {
     std::string lut_name = ".names";
 

vpr/src/pack/cluster_util.cpp

@@ -481,6 +481,8 @@ void update_le_count(const t_pb* pb, const t_logical_block_type_ptr logic_block_
 
 void print_pb_type_count_recurr(t_pb_type* type, size_t max_name_chars, size_t curr_depth, std::map<t_pb_type*, int>& pb_type_count);
 
+void print_pb_type_count(const ClusteredNetlist& clb_nlist);
+
 /*
  * @brief This function identifies the logic block type which is defined by the
  *        block type which has a lut primitive.

vpr/src/pack/cluster_util.h

@@ -0,0 +1,123 @@
+/**
+ * @file
+ * @author  Alex Singer
+ * @date    November 2024
+ * @brief   The declarations of the Greedy Clusterer class which is used to
+ *          encapsulate the process of greedy clustering.
+ */
+
+#pragma once
+
+#include <map>
+#include <unordered_set>
+#include "physical_types.h"
+
+// Forward declarations
+class AtomNetId;
+class AtomNetlist;
+class AttractionInfo;
+class ClusterLegalizer;
+class Prepacker;
+struct t_analysis_opts;
+struct t_clustering_data;
+struct t_pack_high_fanout_thresholds;
+struct t_packer_opts;
+
+/**
+ * @brief A clusterer that generates clusters by greedily choosing the clusters
+ *        which appear to have the best gain for a given neighbor.
+ *
+ * This clusterer generates one cluster at a time by finding candidate molecules
+ * and selecting the molecule with the highest gain.
+ */
+class GreedyClusterer {
+public:
+    /**
+     * @brief Constructor of the Greedy Clusterer class.
+     *
+     * The clusterer may be invoked many times during the packing flow. This
+     * constructor will pre-compute information before clustering which can
+     * improve the performance of the clusterer.
+     *
+     *  @param packer_opts
+     *              Options passed by the user to configure the packing and
+     *              clustering algorithms.
+     *  @param analysis_opts
+     *              Options passed by the user to configure timing analysis in
+     *              the clusterer.
+     *  @param atom_netlist
+     *              The atom netlist to cluster over.
+     *  @param arch
+     *              The architecture to cluster over.
+     *  @param high_fanout_thresholds
+     *              The thresholds for what to consider as a high-fanout net
+     *              for each logical block type.
+     *  @param is_clock
+     *              The set of clock nets in the Atom Netlist.
+     *  @param is_global
+     *              The set of global nets in the Atom Netlist.
+     */
+    GreedyClusterer(const t_packer_opts& packer_opts,
+                    const t_analysis_opts& analysis_opts,
+                    const AtomNetlist& atom_netlist,
+                    const t_arch* arch,
+                    const t_pack_high_fanout_thresholds& high_fanout_thresholds,
+                    const std::unordered_set<AtomNetId>& is_clock,
+                    const std::unordered_set<AtomNetId>& is_global);
+
+    /**
+     * @brief Performs clustering on the pack molecules formed by the prepacker.
+     *
+     * The clustering is contained within the Cluster Legalizer.
+     *
+     *  @param cluster_legalizer
+     *              The cluster legalizer which is used to create clusters and
+     *              grow clusters by adding molecules to a cluster.
+     *  @param prepacker
+     *              The prepacker object which contains the pack molecules that
+     *              atoms are pre-packed into before clustering.
+     *  @param allow_unrelated_clustering
+     *              Allows primitives which have no attraction to the given
+     *              cluster to be packed into it.
+     *  @param balance_block_type_utilization
+     *              When true, tries to create clusters that balance the logical
+     *              block type utilization.
+     *  @param attraction_groups
+     *              Information on the attraction groups used during the
+     *              clustering process.
+     *
+     *  @return num_used_type_instances
+     *              The number of used logical block types by the clustering.
+     *              This information may be useful when detecting if the
+     *              clustering can fit on the device.
+     */
+    std::map<t_logical_block_type_ptr, size_t>
+    do_clustering(ClusterLegalizer& cluster_legalizer,
+                  Prepacker& prepacker,
+                  bool allow_unrelated_clustering,
+                  bool balance_block_type_utilization,
+                  AttractionInfo& attraction_groups);
+
+private:
+    /*
+     * When attraction groups are created, the purpose is to pack more densely by adding more molecules
+     * from the cluster's attraction group to the cluster. In a normal flow, (when attraction groups are
+     * not on), the cluster keeps being packed until the get_molecule routines return either a repeated
+     * molecule or a nullptr. When attraction groups are on, we want to keep exploring molecules for the
+     * cluster until a nullptr is returned. So, the number of repeated molecules is changed from 1 to 500,
+     * effectively making the clusterer pack a cluster until a nullptr is returned.
+     */
+    static constexpr int attraction_groups_max_repeated_molecules_ = 500;
+
+    const t_packer_opts& packer_opts_;
+    const t_analysis_opts& analysis_opts_;
+    const AtomNetlist& atom_netlist_;
+    const t_arch* arch_ = nullptr;
+    const t_pack_high_fanout_thresholds& high_fanout_thresholds_;
+    const std::unordered_set<AtomNetId>& is_clock_;
+    const std::unordered_set<AtomNetId>& is_global_;
+
+    /// @brief Pre-computed logical block types for each model in the architecture.
+    std::map<const t_model*, std::vector<t_logical_block_type_ptr>> primitive_candidate_block_types_;
+};
+

vpr/src/pack/cluster.cpp renamed to vpr/src/pack/greedy_clusterer.cpp

@@ -1,10 +1,11 @@
 #include <unordered_set>
 
 #include "SetupGrid.h"
-#include "cluster.h"
 #include "cluster_legalizer.h"
 #include "cluster_util.h"
+#include "constraints_report.h"
 #include "globals.h"
+#include "greedy_clusterer.h"
 #include "pack.h"
 #include "prepack.h"
 #include "vpr_context.h"
@@ -29,7 +30,6 @@ bool try_pack(t_packer_opts* packer_opts,
     const DeviceContext& device_ctx = g_vpr_ctx.device();
 
     std::unordered_set<AtomNetId> is_clock, is_global;
-    t_clustering_data clustering_data;
     VTR_LOG("Begin packing '%s'.\n", packer_opts->circuit_file_name.c_str());
 
     is_clock = alloc_and_load_is_clock();
@@ -91,7 +91,6 @@ bool try_pack(t_packer_opts* packer_opts,
     }
 
     int pack_iteration = 1;
-    bool floorplan_regions_overfull = false;
 
     // Initialize the cluster legalizer.
     ClusterLegalizer cluster_legalizer(atom_ctx.nlist,
@@ -110,27 +109,24 @@ bool try_pack(t_packer_opts* packer_opts,
     VTR_LOG("Packing with pin utilization targets: %s\n", cluster_legalizer.get_target_external_pin_util().to_string().c_str());
     VTR_LOG("Packing with high fanout thresholds: %s\n", high_fanout_thresholds.to_string().c_str());
 
-    while (true) {
-        free_clustering_data(*packer_opts, clustering_data);
-
+    // Initialize the greedy clusterer.
+    GreedyClusterer clusterer(*packer_opts,
+                              *analysis_opts,
+                              arch,
+                              high_fanout_thresholds,
+                              is_clock,
+                              is_global);
 
+    while (true) {
         //Cluster the netlist
         //  num_used_type_instances: A map used to save the number of used
         //                           instances from each logical block type.
         std::map<t_logical_block_type_ptr, size_t> num_used_type_instances;
-        num_used_type_instances = do_clustering(*packer_opts,
-                                                *analysis_opts,
-                                                arch,
-                                                prepacker,
-                                                cluster_legalizer,
-                                                is_clock,
-                                                is_global,
-                                                allow_unrelated_clustering,
-                                                balance_block_type_util,
-                                                attraction_groups,
-                                                floorplan_regions_overfull,
-                                                high_fanout_thresholds,
-                                                clustering_data);
+        num_used_type_instances = clusterer.do_clustering(cluster_legalizer,
+                                                          prepacker,
+                                                          allow_unrelated_clustering,
+                                                          balance_block_type_util,
+                                                          attraction_groups);
 
         //Try to size/find a device
         bool fits_on_device = try_size_device_grid(*arch, num_used_type_instances, packer_opts->target_device_utilization, packer_opts->device_layout);
@@ -139,6 +135,7 @@ bool try_pack(t_packer_opts* packer_opts,
          * is not dense enough and there are floorplan constraints, it is presumed that the constraints are the cause
          * of the floorplan not fitting, so attraction groups are turned on for later iterations.
          */
+        bool floorplan_regions_overfull = floorplan_constraints_regions_overfull(cluster_legalizer);
         bool floorplan_not_fitting = (floorplan_regions_overfull || g_vpr_ctx.floorplanning().constraints.get_num_partitions() > 0);
 
         if (fits_on_device && !floorplan_regions_overfull) {
@@ -261,9 +258,6 @@ bool try_pack(t_packer_opts* packer_opts,
     //check clustering and output it
     check_and_output_clustering(cluster_legalizer, *packer_opts, is_clock, arch);
 
-    // Free Data Structures
-    free_clustering_data(*packer_opts, clustering_data);
-
     VTR_LOG("\n");
     VTR_LOG("Netlist conversion complete.\n");
     VTR_LOG("\n");