[AP] Added APPack to the AP Flow as a Full Legalizer

doc/src/vpr/command_line_usage.rst

@@ -89,6 +89,8 @@ VPR runs all stages of (pack, place, route, and analysis) if none of :option:`--
     as such, the :option:`--pack` and :option:`--place` options should not be set when this option is set.
     This flow requires that the device has a fixed size and some of the primitive blocks are fixed somewhere on the device grid.
 
+    .. seealso:: See :ref:`analytical_placement_options` for the options for this flow.
+
     .. seealso:: See :ref:`Fixed FPGA Grid Layout <fixed_arch_grid_layout>` and :option:`--device` for how to fix the device size.
 
     .. seealso:: See :ref:`VPR Placement Constraints <placement_constraints>` for how to fix primitive blocks in a design to the device grid.
@@ -1163,6 +1165,40 @@ The following options are only used when FPGA device and netlist contain a NoC r
 
     **Default:** ``vpr_noc_placement_output.txt``
 
+
+.. _analytical_placement_options:
+
+Analytical Placement Options
+^^^^^^^^^^^^^^^
+Instead of Packing atoms into clusters and placing the clusters into valid tile
+sites on the FPGA, Analytical Placement uses analytical techniques to place atoms
+on the FPGA device by relaxing the constraints on where they can be placed. This
+atom-level placement is then legalized into a clustered placement and passed into
+the router in VPR.
+
+Analytical Placement is generally split into three stages:
+
+* Global Placement: Uses analytical techniques to place atoms on the FPGA grid.
+
+* Full Legalization: Legalizes a flat (atom) placement into legal clusters placed on the FPGA grid.
+
+* Detailed Placement: While keeping the clusters legal, performs optimizations on the clustered placement.
+
+.. warning::
+
+    Analytical Placement is experimental and under active development.
+
+.. option:: --ap_full_legalizer {naive | appack}
+
+    Controls which Full Legalizer to use in the AP Flow.
+
+    * ``naive`` Use a Naive Full Legalizer which will try to create clusters exactly where their atoms are placed.
+
+    * ``appack`` Use APPack, which takes the Packer in VPR and uses the flat atom placement to create better clusters.
+
+    **Default:** ``appack``
+
+
 .. _router_options:
 
 Router Options
@@ -1179,7 +1215,7 @@ VPR uses a negotiated congestion algorithm (based on Pathfinder) to perform rout
     This means that during the routing stage, all nets, both intra- and inter-cluster, are routed directly from one primitive pin to another primitive pin.
     This increases routing time but can improve routing quality by re-arranging LUT inputs and exposing additional optimization opportunities in architectures with local intra-cluster routing that is not a full crossbar.
 
-    **Default:** ``OFF`
+    **Default:** ``off``
 
 .. option:: --max_router_iterations <int>
 

vpr/src/analytical_place/analytical_placement_flow.cpp

@@ -93,17 +93,15 @@ void run_analytical_placement_flow(t_vpr_setup& vpr_setup) {
                                   device_ctx.grid.get_num_layers()));
 
     // Run the Full Legalizer.
-    FullLegalizer full_legalizer(ap_netlist,
-                                 vpr_setup,
-                                 device_ctx.grid,
-                                 device_ctx.arch,
-                                 atom_nlist,
-                                 prepacker,
-                                 device_ctx.logical_block_types,
-                                 vpr_setup.PackerRRGraph,
-                                 device_ctx.arch->models,
-                                 device_ctx.arch->model_library,
-                                 vpr_setup.PackerOpts);
-    full_legalizer.legalize(p_placement);
+    const t_ap_opts& ap_opts = vpr_setup.APOpts;
+    std::unique_ptr<FullLegalizer> full_legalizer = make_full_legalizer(ap_opts.full_legalizer_type,
+                                                                        ap_netlist,
+                                                                        atom_nlist,
+                                                                        prepacker,
+                                                                        vpr_setup,
+                                                                        *device_ctx.arch,
+                                                                        device_ctx.grid,
+                                                                        device_ctx.logical_block_types);
+    full_legalizer->legalize(p_placement);
 }
 

vpr/src/analytical_place/ap_flow_enums.h

@@ -0,0 +1,20 @@
+/**
+ * @file
+ * @author  Alex Singer
+ * @date    February 2025
+ * @brief   Enumerations used by the Analytical Placement Flow.
+ */
+
+#pragma once
+
+/**
+ * @brief The type of a Full Legalizer.
+ *
+ * The Analytical Placement flow may implement different Full Legalizers. This
+ * enum can select between these different Full Legalizers.
+ */
+enum class e_ap_full_legalizer {
+    Naive,  ///< The Naive Full Legalizer, which clusters atoms placed in the same tile and tries to place them in that tile according to the flat placement.
+    APPack  ///< The APPack Full Legalizer, which uses the flat placement to improve the Packer and Placer.
+};
+

vpr/src/analytical_place/full_legalizer.cpp

@@ -10,9 +10,13 @@
 #include "full_legalizer.h"
 
 #include <list>
+#include <memory>
 #include <unordered_set>
 #include <vector>
 
+#include "FlatPlacementInfo.h"
+#include "ap_flow_enums.h"
+#include "device_grid.h"
 #include "partial_placement.h"
 #include "ShowSetup.h"
 #include "ap_netlist_fwd.h"
@@ -25,6 +29,7 @@
 #include "logic_types.h"
 #include "pack.h"
 #include "physical_types.h"
+#include "place.h"
 #include "place_and_route.h"
 #include "place_constraints.h"
 #include "place_macro.h"
@@ -42,6 +47,38 @@
 #include "vtr_time.h"
 #include "vtr_vector.h"
 
+
+std::unique_ptr<FullLegalizer> make_full_legalizer(e_ap_full_legalizer full_legalizer_type,
+                                                   const APNetlist& ap_netlist,
+                                                   const AtomNetlist& atom_netlist,
+                                                   const Prepacker& prepacker,
+                                                   t_vpr_setup& vpr_setup,
+                                                   const t_arch& arch,
+                                                   const DeviceGrid& device_grid,
+                                                   const std::vector<t_logical_block_type>& logical_block_types) {
+    switch (full_legalizer_type) {
+        case e_ap_full_legalizer::Naive:
+            return std::make_unique<NaiveFullLegalizer>(ap_netlist,
+                                                        atom_netlist,
+                                                        prepacker,
+                                                        vpr_setup,
+                                                        arch,
+                                                        device_grid,
+                                                        logical_block_types);
+        case e_ap_full_legalizer::APPack:
+            return std::make_unique<APPack>(ap_netlist,
+                                            atom_netlist,
+                                            prepacker,
+                                            vpr_setup,
+                                            arch,
+                                            device_grid,
+                                            logical_block_types);
+        default:
+             VPR_FATAL_ERROR(VPR_ERROR_AP,
+                             "Unrecognized full legalizer type");
+    }
+}
+
 namespace {
 
 /// @brief A unique ID for each root tile on the device.
@@ -239,24 +276,24 @@ static LegalizationClusterId create_new_cluster(PackMoleculeId seed_molecule_id,
     return LegalizationClusterId();
 }
 
-void FullLegalizer::create_clusters(const PartialPlacement& p_placement) {
+void NaiveFullLegalizer::create_clusters(const PartialPlacement& p_placement) {
     // PACKING:
     // Initialize the cluster legalizer (Packing)
     // FIXME: The legalization strategy is currently set to full. Should handle
     //        this better to make it faster.
-    t_pack_high_fanout_thresholds high_fanout_thresholds(packer_opts_.high_fanout_threshold);
+    t_pack_high_fanout_thresholds high_fanout_thresholds(vpr_setup_.PackerOpts.high_fanout_threshold);
     ClusterLegalizer cluster_legalizer(atom_netlist_,
                                        prepacker_,
                                        logical_block_types_,
-                                       lb_type_rr_graphs_,
-                                       user_models_,
-                                       library_models_,
-                                       packer_opts_.target_external_pin_util,
+                                       vpr_setup_.PackerRRGraph,
+                                       arch_.models,
+                                       arch_.model_library,
+                                       vpr_setup_.PackerOpts.target_external_pin_util,
                                        high_fanout_thresholds,
                                        ClusterLegalizationStrategy::FULL,
-                                       packer_opts_.enable_pin_feasibility_filter,
-                                       packer_opts_.feasible_block_array_size,
-                                       packer_opts_.pack_verbosity);
+                                       vpr_setup_.PackerOpts.enable_pin_feasibility_filter,
+                                       vpr_setup_.PackerOpts.feasible_block_array_size,
+                                       vpr_setup_.PackerOpts.pack_verbosity);
     // Create clusters for each tile.
     //  Start by giving each root tile a unique ID.
     size_t grid_width = device_grid_.width();
@@ -330,24 +367,24 @@ void FullLegalizer::create_clusters(const PartialPlacement& p_placement) {
 
     // Check and output the clustering.
     std::unordered_set<AtomNetId> is_clock = alloc_and_load_is_clock();
-    check_and_output_clustering(cluster_legalizer, packer_opts_, is_clock, arch_);
+    check_and_output_clustering(cluster_legalizer, vpr_setup_.PackerOpts, is_clock, &arch_);
     // Reset the cluster legalizer. This is required to load the packing.
     cluster_legalizer.reset();
     // Regenerate the clustered netlist from the file generated previously.
     // FIXME: This writing and loading from a file is wasteful. Should generate
     //        the clusters directly from the cluster legalizer.
-    vpr_load_packing(vpr_setup_, *arch_);
+    vpr_load_packing(vpr_setup_, arch_);
     load_cluster_constraints();
     const ClusteredNetlist& clb_nlist = g_vpr_ctx.clustering().clb_nlist;
 
     // Verify the packing and print some info
-    check_netlist(packer_opts_.pack_verbosity);
+    check_netlist(vpr_setup_.PackerOpts.pack_verbosity);
     writeClusteredNetlistStats(vpr_setup_.FileNameOpts.write_block_usage);
     print_pb_type_count(clb_nlist);
 }
 
-void FullLegalizer::place_clusters(const ClusteredNetlist& clb_nlist,
-                                   const PartialPlacement& p_placement) {
+void NaiveFullLegalizer::place_clusters(const ClusteredNetlist& clb_nlist,
+                                        const PartialPlacement& p_placement) {
     // PLACING:
     // Create a lookup from the AtomBlockId to the APBlockId
     vtr::vector<AtomBlockId, APBlockId> atom_to_ap_block(atom_netlist_.blocks().size());
@@ -409,7 +446,7 @@ void FullLegalizer::place_clusters(const ClusteredNetlist& clb_nlist,
     //      - This may be needed to perform SA. Not needed right now.
 }
 
-void FullLegalizer::legalize(const PartialPlacement& p_placement) {
+void NaiveFullLegalizer::legalize(const PartialPlacement& p_placement) {
     // Create a scoped timer for the full legalizer
     vtr::ScopedStartFinishTimer full_legalizer_timer("AP Full Legalizer");
 
@@ -449,3 +486,65 @@ void FullLegalizer::legalize(const PartialPlacement& p_placement) {
     post_place_sync();
 }
 
+void APPack::legalize(const PartialPlacement& p_placement) {
+    // Create a scoped timer for the full legalizer
+    vtr::ScopedStartFinishTimer full_legalizer_timer("AP Full Legalizer");
+
+    // Convert the Partial Placement (APNetlist) to a flat placement (AtomNetlist).
+    FlatPlacementInfo flat_placement_info(atom_netlist_);
+    for (APBlockId ap_blk_id : ap_netlist_.blocks()) {
+        PackMoleculeId mol_id = ap_netlist_.block_molecule(ap_blk_id);
+        const t_pack_molecule& mol = prepacker_.get_molecule(mol_id);
+        for (AtomBlockId atom_blk_id : mol.atom_block_ids) {
+            if (!atom_blk_id.is_valid())
+                continue;
+            flat_placement_info.blk_x_pos[atom_blk_id] = p_placement.block_x_locs[ap_blk_id];
+            flat_placement_info.blk_y_pos[atom_blk_id] = p_placement.block_y_locs[ap_blk_id];
+            flat_placement_info.blk_layer[atom_blk_id] = p_placement.block_layer_nums[ap_blk_id];
+            flat_placement_info.blk_sub_tile[atom_blk_id] = p_placement.block_sub_tiles[ap_blk_id];
+        }
+    }
+
+    // Run the Packer stage with the flat placement as a hint.
+    try_pack(&vpr_setup_.PackerOpts,
+             &vpr_setup_.AnalysisOpts,
+             arch_,
+             vpr_setup_.RoutingArch,
+             vpr_setup_.user_models,
+             vpr_setup_.library_models,
+             vpr_setup_.PackerRRGraph,
+             flat_placement_info);
+
+    // The Packer stores the clusters into a .net file. Load the packing file.
+    // FIXME: This should be removed. Reading from a file is strange.
+    vpr_load_packing(vpr_setup_, arch_);
+    load_cluster_constraints();
+    const ClusteredNetlist& clb_nlist = g_vpr_ctx.clustering().clb_nlist;
+
+    // Verify the packing and print some info
+    check_netlist(vpr_setup_.PackerOpts.pack_verbosity);
+    writeClusteredNetlistStats(vpr_setup_.FileNameOpts.write_block_usage);
+    print_pb_type_count(clb_nlist);
+
+    // Pass the clustering into the Placer with the flat placement as a hint.
+    // TODO: This should only be the initial placer. Running the full SA would
+    //       be more of a Detailed Placer.
+    const auto& placement_net_list = (const Netlist<>&)clb_nlist;
+    try_place(placement_net_list,
+              vpr_setup_.PlacerOpts,
+              vpr_setup_.RouterOpts,
+              vpr_setup_.AnalysisOpts,
+              vpr_setup_.NocOpts,
+              arch_.Chans,
+              &vpr_setup_.RoutingArch,
+              vpr_setup_.Segments,
+              arch_.directs,
+              flat_placement_info,
+              false /* is_flat */);
+
+    // TODO: This was taken from vpr_api. Not sure why it is needed. Should be
+    //       made part of the placement and verify placement should check for
+    //       it.
+    post_place_sync();
+}
+