Merge pull request #2806 from AlexandreSinger/feature-ap-global-place…

…ment

[AP] Global Placer

vpr/src/analytical_place/analytical_placement_flow.cpp

@@ -12,6 +12,7 @@
 #include "atom_netlist.h"
 #include "full_legalizer.h"
 #include "gen_ap_netlist_from_atoms.h"
+#include "global_placer.h"
 #include "globals.h"
 #include "partial_legalizer.h"
 #include "partial_placement.h"
@@ -58,7 +59,6 @@ static void print_ap_netlist_stats(const APNetlist& netlist) {
 }
 
 void run_analytical_placement_flow(t_vpr_setup& vpr_setup) {
-    (void)vpr_setup;
     // Start an overall timer for the Analytical Placement flow.
     vtr::ScopedStartFinishTimer timer("Analytical Placement");
 
@@ -79,16 +79,9 @@ void run_analytical_placement_flow(t_vpr_setup& vpr_setup) {
     print_ap_netlist_stats(ap_netlist);
 
     // Run the Global Placer
-    //  For now, just runs the solver and partial legalizer 10 times arbitrarily.
-    PartialPlacement p_placement(ap_netlist);
-    std::unique_ptr<AnalyticalSolver> solver = make_analytical_solver(e_analytical_solver::QP_HYBRID,
-                                                                      ap_netlist);
-    std::unique_ptr<PartialLegalizer> legalizer = make_partial_legalizer(e_partial_legalizer::FLOW_BASED,
-                                                                         ap_netlist);
-    for (size_t i = 0; i < 10; i++) {
-        solver->solve(i, p_placement);
-        legalizer->legalize(p_placement);
-    }
+    std::unique_ptr<GlobalPlacer> global_placer = make_global_placer(e_global_placer::SimPL,
+                                                                     ap_netlist);
+    PartialPlacement p_placement = global_placer->place();
 
     // Verify that the partial placement is valid before running the full
     // legalizer.

vpr/src/analytical_place/full_legalizer.cpp

@@ -394,8 +394,6 @@ void FullLegalizer::place_clusters(const ClusteredNetlist& clb_nlist,
 
     // FIXME: Allocate and load moveable blocks?
     //      - This may be needed to perform SA. Not needed right now.
-
-    // TODO: Check initial placement legality
 }
 
 void FullLegalizer::legalize(const PartialPlacement& p_placement) {

vpr/src/analytical_place/global_placer.cpp

@@ -0,0 +1,134 @@
+/**
+ * @file
+ * @author  Alex Singer
+ * @date    October 2024
+ * @brief   The definitions of the global placers used in the AP flow and their
+ *          base class.
+ */
+
+#include "global_placer.h"
+#include <cstdio>
+#include <memory>
+#include "analytical_solver.h"
+#include "ap_netlist.h"
+#include "partial_legalizer.h"
+#include "partial_placement.h"
+#include "vpr_error.h"
+#include "vtr_log.h"
+#include "vtr_time.h"
+
+std::unique_ptr<GlobalPlacer> make_global_placer(e_global_placer placer_type,
+                                                 const APNetlist& netlist) {
+    // Based on the placer type passed in, build the global placer.
+    switch (placer_type) {
+        case e_global_placer::SimPL:
+            return std::make_unique<SimPLGlobalPlacer>(netlist);
+        default:
+            VPR_FATAL_ERROR(VPR_ERROR_AP,
+                            "Unrecognized global placer type");
+
+    }
+}
+
+SimPLGlobalPlacer::SimPLGlobalPlacer(const APNetlist& netlist) : GlobalPlacer(netlist) {
+    // This can be a long method. Good to time this to see how long it takes to
+    // construct the global placer.
+    vtr::ScopedStartFinishTimer global_placer_building_timer("Constructing Global Placer");
+    // Build the solver.
+    solver_ = make_analytical_solver(e_analytical_solver::QP_HYBRID,
+                                     netlist);
+    // Build the partial legalizer
+    partial_legalizer_ = make_partial_legalizer(e_partial_legalizer::FLOW_BASED,
+                                                netlist);
+}
+
+/**
+ * @brief Helper method to print the header of the per-iteration status updates
+ *        of the global placer.
+ */
+static void print_SimPL_status_header() {
+    VTR_LOG("----  ----------------  ----------------  -----------  --------------  ----------\n");
+    VTR_LOG("Iter  Lower Bound HPWL  Upper Bound HPWL  Solver Time  Legalizer Time  Total Time\n");
+    VTR_LOG("                                                (sec)           (sec)       (sec)\n");
+    VTR_LOG("----  ----------------  ----------------  -----------  --------------  ----------\n");
+}
+
+/**
+ * @brief Helper method to print the per-iteration status of the global placer.
+ */
+static void print_SimPL_status(size_t iteration,
+                               double lb_hpwl,
+                               double ub_hpwl,
+                               float solver_time,
+                               float legalizer_time,
+                               float total_time) {
+    // Iteration
+    VTR_LOG("%4zu", iteration);
+
+    // Lower Bound HPWL
+    VTR_LOG("  %16.2f", lb_hpwl);
+
+    // Upper Bound HPWL
+    VTR_LOG("  %16.2f", ub_hpwl);
+
+    // Solver runtime
+    VTR_LOG("  %11.3f", solver_time);
+
+    // Legalizer runtime
+    VTR_LOG("  %14.3f", legalizer_time);
+
+    // Total runtime
+    VTR_LOG("  %10.3f", total_time);
+
+    VTR_LOG("\n");
+
+    fflush(stdout);
+}
+
+PartialPlacement SimPLGlobalPlacer::place() {
+    // Create a timer to time the entire global placement time.
+    vtr::ScopedStartFinishTimer global_placer_time("AP Global Placer");
+    // Create a timer to keep track of how long the solver and legalizer take.
+    vtr::Timer runtime_timer;
+    // Print the status header.
+    if (log_verbosity_ >= 1)
+        print_SimPL_status_header();
+    // Initialialize the partial placement object.
+    PartialPlacement p_placement(netlist_);
+    // Run the global placer.
+    for (size_t i = 0; i < max_num_iterations_; i++) {
+        float iter_start_time = runtime_timer.elapsed_sec();
+
+        // Run the solver.
+        float solver_start_time = runtime_timer.elapsed_sec();
+        solver_->solve(i, p_placement);
+        float solver_end_time = runtime_timer.elapsed_sec();
+        double lb_hpwl = p_placement.get_hpwl(netlist_);
+
+        // Run the legalizer.
+        float legalizer_start_time = runtime_timer.elapsed_sec();
+        partial_legalizer_->legalize(p_placement);
+        float legalizer_end_time = runtime_timer.elapsed_sec();
+        double ub_hpwl = p_placement.get_hpwl(netlist_);
+
+        // Print some stats
+        if (log_verbosity_ >= 1) {
+            float iter_end_time = runtime_timer.elapsed_sec();
+            print_SimPL_status(i, lb_hpwl, ub_hpwl,
+                               solver_end_time - solver_start_time,
+                               legalizer_end_time - legalizer_start_time,
+                               iter_end_time - iter_start_time);
+        }
+
+        // Exit condition: If the upper-bound and lower-bound HPWLs are
+        // sufficiently close together then stop.
+        double hpwl_relative_gap = (ub_hpwl - lb_hpwl) / ub_hpwl;
+        if (hpwl_relative_gap < target_hpwl_relative_gap_)
+            break;
+    }
+    // Return the placement from the final iteration.
+    // TODO: investigate saving the best solution found so far. It should be
+    //       cheap to save a copy of the PartialPlacement object.
+    return p_placement;
+}
+

vpr/src/analytical_place/global_placer.h

@@ -0,0 +1,143 @@
+/**
+ * @file
+ * @author  Alex Singer
+ * @date    October 2024
+ * @brief   The declarations of the Global Placer base class which is used to
+ *          define the functionality of all global placers in the AP flow.
+ *
+ * A Global Placer creates a Partial Placement given only the netlist and the
+ * architecture. It uses analytical techniques (i.e. efficient numerical
+ * minimization of an objective function of a placement) to find a placement
+ * that optimizes for objectives subject to some of the constraints of the FPGA
+ * architecture.
+ */
+
+#pragma once
+
+#include <memory>
+
+// Forward declarations
+class APNetlist;
+class AnalyticalSolver;
+class PartialPlacement;
+class PartialLegalizer;
+
+/**
+ * @brief Enumeration of all of the global placers currently implemented in VPR.
+ */
+enum class e_global_placer {
+    SimPL       // Global placer based on the SimPL paper.
+};
+
+/**
+ * @brief The Global Placer base class
+ *
+ * This declares the functionality that all Global Placers will use. This
+ * provides a standard interface for the global placers so they can be used
+ * interchangably. This makes it very easy to test and compare different global
+ * placers.
+ */
+class GlobalPlacer {
+public:
+    virtual ~GlobalPlacer() {}
+
+    /**
+     * @brief Constructor of the base GlobalPlacer class
+     *
+     *  @param netlist      Netlist of the design at some abstraction level;
+     *                      typically this would have some atoms and groups of
+     *                      atoms (in a pack pattern).
+     *  @param log_verbosity    The verbosity of log messages in the Global
+     *                          Placer.
+     */
+    GlobalPlacer(const APNetlist& netlist, int log_verbosity = 1)
+                    : netlist_(netlist),
+                      log_verbosity_(log_verbosity) {}
+
+    /**
+     * @brief Perform global placement on the given netlist.
+     *
+     * The role of a global placer is to try and find a placement for the given
+     * netlist which optimizes some objective function and is mostly legal.
+     */
+    virtual PartialPlacement place() = 0;
+
+protected:
+
+    /// @brief The APNetlist the global placer is placing.
+    const APNetlist& netlist_;
+
+    /// @brief The setting of how verbose the log messages should be in the
+    ///        global placer. Anything larger than zero will display per
+    ///        iteration status messages.
+    int log_verbosity_;
+};
+
+/**
+ * @brief A factory method which creates a Global Placer of the given type.
+ */
+std::unique_ptr<GlobalPlacer> make_global_placer(e_global_placer placer_type,
+                                                 const APNetlist& netlist);
+
+/**
+ * @brief A Global Placer based on the SimPL work for analytical ASIC placement.
+ *          https://doi.org/10.1145/2461256.2461279
+ *
+ * This placement technique uses a solver to generate a placement that optimizes
+ * over some objective function and is likely very illegal (has many overlapping
+ * blocks and blocks in the wrong places). This solution represents the "lower-
+ * bound" on the solution quality.
+ *
+ * This technique passes this "lower-bound" solution into a legalizer, which
+ * tries to find the closest legal solution to the lower-bound solution (by
+ * spreading out blocks and placing them in legal positions). This often
+ * destroys the quality of the lower-bound solution, and is considered an
+ * "upper-bound" on the solution quality.
+ *
+ * Each iteration of this global placer, the upper-bound solution is fed into
+ * the solver as a "hint" to what a legal solution looks like. This allows the
+ * solver to produce another placement which will make decisions knowing where
+ * the blocks will end-up in the legal solution. This worstens the quality of
+ * the lower-bound solution; however, after passing this solution back into
+ * the legalizer, this will likely improve the quality of the upper-bound
+ * solution.
+ *
+ * Over several iterations the upper-bound and lower-bound solutions will
+ * approach each other until a good quality, mostly-legal solution is found.
+ */
+class SimPLGlobalPlacer : public GlobalPlacer {
+private:
+
+    /// @brief The maximum number of iterations the global placer can perform.
+    static constexpr size_t max_num_iterations_ = 100;
+
+    /// @brief The target relative gap between the HPWL of the upper-bound and
+    ///        lower-bound placements. The placer will stop if the difference
+    ///        between the two bounds, normalized to the upper-bound, is smaller
+    ///        than this number.
+    static constexpr double target_hpwl_relative_gap_ = 0.10;
+
+    /// @brief The solver which generates the lower-bound placement.
+    std::unique_ptr<AnalyticalSolver> solver_;
+
+    /// @brief The legalizer which generates the upper-bound placement.
+    std::unique_ptr<PartialLegalizer> partial_legalizer_;
+
+public:
+
+    /**
+     * @brief Constructor for the SimPL Global Placer
+     *
+     * Constructs the solver and partial legalizer.
+     */
+    SimPLGlobalPlacer(const APNetlist& netlist);
+
+    /**
+     * @brief Run a SimPL-like global placement algorithm
+     *
+     * This iteratively runs the solver and legalizer until a good quality and
+     * mostly-legal placement is found.
+     */
+    PartialPlacement place() final;
+};
+

vtr_flow/tasks/regression_tests/vtr_reg_strong/basic_ap/config/golden_results.txt

@@ -1,5 +1,5 @@
 arch	circuit	script_params	vtr_flow_elapsed_time	vtr_max_mem_stage	vtr_max_mem	error	odin_synth_time	max_odin_mem	parmys_synth_time	max_parmys_mem	abc_depth	abc_synth_time	abc_cec_time	abc_sec_time	max_abc_mem	ace_time	max_ace_mem	num_clb	num_io	num_memories	num_mult	vpr_status	vpr_revision	vpr_build_info	vpr_compiler	vpr_compiled	hostname	rundir	max_vpr_mem	num_primary_inputs	num_primary_outputs	num_pre_packed_nets	num_pre_packed_blocks	num_netlist_clocks	num_post_packed_nets	num_post_packed_blocks	device_width	device_height	device_grid_tiles	device_limiting_resources	device_name	pack_mem	pack_time	placed_wirelength_est	total_swap	accepted_swap	rejected_swap	aborted_swap	place_mem	place_time	place_quench_time	placed_CPD_est	placed_setup_TNS_est	placed_setup_WNS_est	placed_geomean_nonvirtual_intradomain_critical_path_delay_est	place_delay_matrix_lookup_time	place_quench_timing_analysis_time	place_quench_sta_time	place_total_timing_analysis_time	place_total_sta_time	ap_mem	ap_time	ap_full_legalizer_mem	ap_full_legalizer_time	min_chan_width	routed_wirelength	min_chan_width_route_success_iteration	logic_block_area_total	logic_block_area_used	min_chan_width_routing_area_total	min_chan_width_routing_area_per_tile	min_chan_width_route_time	min_chan_width_total_timing_analysis_time	min_chan_width_total_sta_time	crit_path_num_rr_graph_nodes	crit_path_num_rr_graph_edges	crit_path_collapsed_nodes	crit_path_routed_wirelength	crit_path_route_success_iteration	crit_path_total_nets_routed	crit_path_total_connections_routed	crit_path_total_heap_pushes	crit_path_total_heap_pops	critical_path_delay	geomean_nonvirtual_intradomain_critical_path_delay	setup_TNS	setup_WNS	hold_TNS	hold_WNS	crit_path_routing_area_total	crit_path_routing_area_per_tile	router_lookahead_computation_time	crit_path_route_time	crit_path_create_rr_graph_time	crit_path_create_intra_cluster_rr_graph_time	crit_path_tile_lookahead_computation_time	crit_path_router_lookahead_computation_time	crit_path_total_timing_analysis_time	crit_path_total_sta_time	
-fixed_k6_frac_N8_22nm.xml	single_wire.v	common	4.25	vpr	70.91 MiB		-1	-1	0.18	16276	1	0.39	-1	-1	29812	-1	-1	0	1	0	0	success	v8.0.0-11571-g5eb3aa508	release VTR_ASSERT_LEVEL=3	GNU 9.4.0 on Linux-4.15.0-213-generic x86_64	2024-10-18T00:28:35	betzgrp-wintermute.eecg.utoronto.ca	/home/singera8/vtr-verilog-to-routing/vtr_flow/tasks	72608	1	1	0	2	0	1	2	17	17	289	-1	unnamed_device	-1	-1	-1	-1	-1	-1	-1	70.9 MiB	0.14	-1	-1	-1	-1	-1	-1	-1	-1	-1	-1	70.9 MiB	0.14	70.9 MiB	0.10	8	14	1	6.79088e+06	0	166176.	575.005	0.36	0.00138004	0.00129992	20206	45088	-1	19	1	1	1	194	45	0.7726	nan	-0.7726	-0.7726	0	0	202963.	702.294	0.09	0.00	0.07	-1	-1	0.09	0.00122838	0.00119736	
-fixed_k6_frac_N8_22nm.xml	single_ff.v	common	4.68	vpr	71.03 MiB		-1	-1	0.20	16236	1	0.39	-1	-1	29696	-1	-1	1	2	0	0	success	v8.0.0-11571-g5eb3aa508	release VTR_ASSERT_LEVEL=3	GNU 9.4.0 on Linux-4.15.0-213-generic x86_64	2024-10-18T00:28:35	betzgrp-wintermute.eecg.utoronto.ca	/home/singera8/vtr-verilog-to-routing/vtr_flow/tasks	72736	2	1	3	3	1	3	4	17	17	289	-1	unnamed_device	-1	-1	-1	-1	-1	-1	-1	71.0 MiB	0.14	-1	-1	-1	-1	-1	-1	-1	-1	-1	-1	71.0 MiB	0.14	71.0 MiB	0.10	20	31	1	6.79088e+06	13472	414966.	1435.87	0.63	0.00135413	0.0012936	22510	95286	-1	35	1	2	2	213	52	0.942216	0.942216	-1.68896	-0.942216	0	0	503264.	1741.40	0.17	0.00	0.14	-1	-1	0.17	0.00127341	0.00123431	
-fixed_k6_frac_N8_22nm.xml	ch_intrinsics.v	common	6.69	vpr	71.67 MiB		-1	-1	0.46	18220	3	0.40	-1	-1	33084	-1	-1	40	99	3	0	success	v8.0.0-11571-g5eb3aa508	release VTR_ASSERT_LEVEL=3	GNU 9.4.0 on Linux-4.15.0-213-generic x86_64	2024-10-18T00:28:35	betzgrp-wintermute.eecg.utoronto.ca	/home/singera8/vtr-verilog-to-routing/vtr_flow/tasks	73388	99	130	240	229	1	247	272	17	17	289	-1	unnamed_device	-1	-1	-1	-1	-1	-1	-1	71.7 MiB	0.28	-1	-1	-1	-1	-1	-1	-1	-1	-1	-1	71.7 MiB	0.28	71.7 MiB	0.21	32	3122	15	6.79088e+06	2.18288e+06	586450.	2029.24	1.84	0.271358	0.247517	24814	144142	-1	2952	30	711	1121	349988	188928	2.0466	2.0466	-154.346	-2.0466	-0.04337	-0.04337	744469.	2576.02	0.25	0.25	0.22	-1	-1	0.25	0.102379	0.0937273	
-fixed_k6_frac_N8_22nm.xml	diffeq1.v	common	32.17	vpr	74.24 MiB		-1	-1	0.75	23104	15	0.61	-1	-1	34204	-1	-1	74	162	0	5	success	v8.0.0-11571-g5eb3aa508	release VTR_ASSERT_LEVEL=3	GNU 9.4.0 on Linux-4.15.0-213-generic x86_64	2024-10-18T00:28:35	betzgrp-wintermute.eecg.utoronto.ca	/home/singera8/vtr-verilog-to-routing/vtr_flow/tasks	76024	162	96	817	258	1	797	337	17	17	289	-1	unnamed_device	-1	-1	-1	-1	-1	-1	-1	74.2 MiB	1.01	-1	-1	-1	-1	-1	-1	-1	-1	-1	-1	74.2 MiB	1.01	74.2 MiB	0.60	60	15916	46	6.79088e+06	2.97693e+06	1.01997e+06	3529.29	24.68	3.56948	3.3772	29998	257685	-1	13617	19	4413	11558	1499556	342325	21.9657	21.9657	-1806.56	-21.9657	0	0	1.27783e+06	4421.56	0.40	0.79	0.44	-1	-1	0.40	0.334496	0.31821	
+fixed_k6_frac_N8_22nm.xml	single_wire.v	common	2.23	vpr	70.88 MiB		-1	-1	0.13	15888	1	0.16	-1	-1	29628	-1	-1	0	1	0	0	success	v8.0.0-11573-g5ea68eac9	release VTR_ASSERT_LEVEL=3	GNU 9.4.0 on Linux-4.15.0-213-generic x86_64	2024-10-18T14:02:52	betzgrp-wintermute.eecg.utoronto.ca	/home/singera8/vtr-verilog-to-routing/vtr_flow/tasks	72576	1	1	0	2	0	1	2	17	17	289	-1	unnamed_device	-1	-1	-1	-1	-1	-1	-1	70.9 MiB	0.13	-1	-1	-1	-1	-1	-1	-1	-1	-1	-1	70.9 MiB	0.13	70.9 MiB	0.09	8	14	1	6.79088e+06	0	166176.	575.005	0.37	0.001421	0.00133975	20206	45088	-1	19	1	1	1	194	45	0.7726	nan	-0.7726	-0.7726	0	0	202963.	702.294	0.09	0.00	0.07	-1	-1	0.09	0.00128701	0.00125353	
+fixed_k6_frac_N8_22nm.xml	single_ff.v	common	2.68	vpr	70.80 MiB		-1	-1	0.14	16336	1	0.16	-1	-1	29640	-1	-1	1	2	0	0	success	v8.0.0-11573-g5ea68eac9	release VTR_ASSERT_LEVEL=3	GNU 9.4.0 on Linux-4.15.0-213-generic x86_64	2024-10-18T14:02:52	betzgrp-wintermute.eecg.utoronto.ca	/home/singera8/vtr-verilog-to-routing/vtr_flow/tasks	72500	2	1	3	3	1	3	4	17	17	289	-1	unnamed_device	-1	-1	-1	-1	-1	-1	-1	70.8 MiB	0.14	-1	-1	-1	-1	-1	-1	-1	-1	-1	-1	70.8 MiB	0.14	70.8 MiB	0.09	20	31	1	6.79088e+06	13472	414966.	1435.87	0.65	0.00147526	0.00140277	22510	95286	-1	35	1	2	2	213	52	0.942216	0.942216	-1.68896	-0.942216	0	0	503264.	1741.40	0.18	0.00	0.15	-1	-1	0.18	0.00127433	0.00123146	
+fixed_k6_frac_N8_22nm.xml	ch_intrinsics.v	common	4.51	vpr	71.77 MiB		-1	-1	0.47	18200	3	0.09	-1	-1	33188	-1	-1	41	99	3	0	success	v8.0.0-11573-g5ea68eac9	release VTR_ASSERT_LEVEL=3	GNU 9.4.0 on Linux-4.15.0-213-generic x86_64	2024-10-18T14:02:52	betzgrp-wintermute.eecg.utoronto.ca	/home/singera8/vtr-verilog-to-routing/vtr_flow/tasks	73496	99	130	240	229	1	247	273	17	17	289	-1	unnamed_device	-1	-1	-1	-1	-1	-1	-1	71.8 MiB	0.30	-1	-1	-1	-1	-1	-1	-1	-1	-1	-1	71.8 MiB	0.30	71.8 MiB	0.23	32	3145	19	6.79088e+06	2.19635e+06	586450.	2029.24	1.54	0.232941	0.21265	24814	144142	-1	2897	11	635	1003	105957	24691	2.0466	2.0466	-155.681	-2.0466	-0.21204	-0.16867	744469.	2576.02	0.27	0.09	0.22	-1	-1	0.27	0.046062	0.042439	
+fixed_k6_frac_N8_22nm.xml	diffeq1.v	common	30.51	vpr	74.25 MiB		-1	-1	0.73	23000	15	0.35	-1	-1	34316	-1	-1	65	162	0	5	success	v8.0.0-11573-g5ea68eac9	release VTR_ASSERT_LEVEL=3	GNU 9.4.0 on Linux-4.15.0-213-generic x86_64	2024-10-18T14:02:52	betzgrp-wintermute.eecg.utoronto.ca	/home/singera8/vtr-verilog-to-routing/vtr_flow/tasks	76028	162	96	817	258	1	792	328	17	17	289	-1	unnamed_device	-1	-1	-1	-1	-1	-1	-1	74.2 MiB	1.81	-1	-1	-1	-1	-1	-1	-1	-1	-1	-1	74.2 MiB	1.81	74.2 MiB	0.61	68	16708	27	6.79088e+06	2.85568e+06	1.14541e+06	3963.36	22.39	3.52495	3.33579	31438	289477	-1	14397	19	4042	10567	1489293	328107	22.3059	22.3059	-1909.12	-22.3059	0	0	1.42693e+06	4937.46	0.51	0.78	0.51	-1	-1	0.51	0.316463	0.300694