✨ Post-Layout Optimization (#260)

* post placement optimization * Some improvements * align with python implementation * refactoring * fix bug * comments * refactor * reformat * comments * fix ouput bug * fix bug * move code to header file * rename * reformat code * use tuple instead of pair * fix windows ci error * add basic unit test * fix warnings * add docstrings * add unittests for detail functions * also delete excess wiring in columns * disable failing test case * reformat * 🎨 Refactor physical design optimization methods for general gate-level layout The physical design optimization methods have been significantly refactored to work with general gate-level layouts. This includes making various functions template-based, replacing several types to be layout-agnostic, and adding static assertions to ensure the layouts are of appropriate types. Cleaning physical design optimization helps to reduce redundancy and improve flexibility. As a result, a broader range of gate level layouts can be optimized. This should provide performance improvements and allow for more advanced optimization techniques in future development. * resolve first batch of todos * optimize layout in-place * use own types for complex structures * add optimization algorithm to cli * add optimiztaion algorithm to docs * fix small bug and increase coverage * clang-format * simplify code * update clang-tidy-review version * more expressive docstrings * reformat code * typo * 🎨 Improved upon the structure of the code even further and added more TODO notes * resolve ToDos * remove todo * remove debug output * clang-tidy * update docs * rewrite delete wires algorithm to delete all rows and columns at once * clang-tidy * only calculate fanins for used coordinates * increase coverage to 100% * fix bug * ✨ Added a deep-copy `clone()` function to `gate_level_layout` * 🎨 Some renaming and docstring updates * 🎨 Some renaming and slight code cleanup * resolve todos in experiments * Apply suggestions from code review Co-authored-by: Marcel Walter <[email protected]> * resolve todos * fix failing tests * clang-tidy * further improvements * ✨ Added `clone` functions to all core layouts * clone layout * ✨ Added working `clone` functions to all core layouts * 🐛 Fix compiler issue * 🎨 ClangFormat * 🚨 Suppress `clang-tidy` warning * 🚨 Suppress `clang-tidy` warning * 🎨 Final polishing * ✅ Enabled commented-out test case * 🔀 Fix merge conflicts --------- Co-authored-by: Marcel Walter <[email protected]> Co-authored-by: Marcel Walter <[email protected]>
cda-tum · Sep 15, 2023 · 44d0530 · 44d0530
1 parent eabe4cc
commit 44d0530
Show file tree

Hide file tree

Showing 23 changed files with 2,020 additions and 15 deletions.
diff --git a/cli/cmd/physical_design/optimize.hpp b/cli/cmd/physical_design/optimize.hpp
@@ -0,0 +1,104 @@
+//
+// Created by Simon Hofmann on 02.08.23.
+//
+
+#ifndef FICTION_CMD_OPTIMIZE_HPP
+#define FICTION_CMD_OPTIMIZE_HPP
+
+#include <fiction/algorithms/physical_design/post_layout_optimization.hpp>
+#include <fiction/layouts/clocked_layout.hpp>
+#include <fiction/layouts/gate_level_layout.hpp>
+#include <fiction/traits.hpp>
+#include <fiction/types.hpp>
+
+#include <alice/alice.hpp>
+
+#include <iostream>
+#include <optional>
+#include <variant>
+
+namespace alice
+{
+/**
+ * Optimizes a 2DDWave-clocked Cartesian layout.
+ */
+class optimize_command : public command
+{
+  public:
+    /**
+     * Standard constructor. Adds descriptive information, options, and flags.
+     *
+     * @param e alice::environment that specifies stores etc.
+     */
+    explicit optimize_command(const environment::ptr& e) :
+            command(e, "Optimizes a 2DDWave-clocked Cartesian layout with respect to area. It achieves this objective "
+                       "by strategically repositioning gates within the layout, removing excess wiring, and "
+                       "effectively relocating outputs to more favorable positions.")
+    {}
+
+  protected:
+    /**
+     * Statistics.
+     */
+    fiction::post_layout_optimization_stats st{};
+
+    /**
+     * Optimizes a 2DDWave-clocked Cartesian layout.
+     */
+    void execute() override
+    {
+        auto& gls = store<fiction::gate_layout_t>();
+
+        // error case: empty gate-level layout store
+        if (gls.empty())
+        {
+            env->out() << "[w] no gate layout in store" << std::endl;
+            return;
+        }
+
+        const auto& lyt = gls.current();
+
+        const auto check_clocking_scheme = [](auto&& lyt_ptr)
+        { return lyt_ptr->is_clocking_scheme(fiction::clock_name::TWODDWAVE); };
+
+        // error case: layout is not 2DDWave-clocked
+        if (const auto is_twoddwave_clocked = std::visit(check_clocking_scheme, lyt); !is_twoddwave_clocked)
+        {
+            env->out() << "[e] layout has to be 2DDWave-clocked" << std::endl;
+            return;
+        }
+
+        const auto apply_optimization = [&](auto&& lyt_ptr)
+        {
+            using Lyt               = typename std::decay_t<decltype(lyt_ptr)>::element_type;
+            auto       lyt_copy     = lyt_ptr->clone();
+            const auto lyt_copy_ptr = std::make_shared<Lyt>(lyt_copy);
+
+            if constexpr (fiction::is_cartesian_layout_v<Lyt>)
+            {
+                fiction::post_layout_optimization(*lyt_copy_ptr, &st);
+                fiction::restore_names(*lyt_ptr, *lyt_copy_ptr);
+                gls.extend() = lyt_copy_ptr;
+            }
+            else
+            {
+                std::cout << "[e] layout has to be Cartesian" << std::endl;
+            }
+        };
+
+        try
+        {
+            std::visit(apply_optimization, lyt);
+        }
+        catch (...)
+        {
+            env->out() << "[e] an error occurred while optimizing" << std::endl;
+        }
+    }
+};
+
+ALICE_ADD_COMMAND(optimize, "Physical Design")
+
+}  // namespace alice
+
+#endif  // FICTION_CMD_OPTIMIZE_HPP
diff --git a/cli/commands.hpp b/cli/commands.hpp
@@ -26,6 +26,7 @@
 #include "cmd/physical_design/exact.hpp"
 #include "cmd/physical_design/hex.hpp"
 #include "cmd/physical_design/onepass.hpp"
+#include "cmd/physical_design/optimize.hpp"
 #include "cmd/physical_design/ortho.hpp"
 #include "cmd/technology/area.hpp"
 #include "cmd/technology/cell.hpp"

diff --git a/docs/algorithms/algorithms.rst b/docs/algorithms/algorithms.rst
@@ -53,6 +53,7 @@ Physical Design
    color_routing.rst
    apply_gate_library.rst
    hexagonalization.rst
+   post_layout_optimization.rst
 
 
 Verification

diff --git a/docs/algorithms/post_layout_optimization.rst b/docs/algorithms/post_layout_optimization.rst
@@ -0,0 +1,15 @@
+.. _post_layout_optimization:
+
+Optimizing 2DDWave-clocked Cartesian Layouts
+--------------------------------------------
+
+**Header:** ``fiction/algorithms/physical_design/post_layout_optimization.hpp``
+
+This algorithm aims to decrease the overall layout area of a given 2DDWave-clocked Cartesian layout that has been
+generated using either heuristic methods or machine learning techniques. It achieves this objective by strategically
+repositioning gates within the layout, removing excess wiring, and effectively relocating outputs to more favorable
+positions.
+
+.. doxygenstruct:: fiction::post_layout_optimization_stats
+   :members:
+.. doxygenfunction:: fiction::post_layout_optimization
diff --git a/experiments/post_layout_optimization/post_layout_optimization.cpp b/experiments/post_layout_optimization/post_layout_optimization.cpp
@@ -0,0 +1,108 @@
+#include "fiction_experiments.hpp"
+
+#include <fiction/algorithms/physical_design/orthogonal.hpp>  // scalable heuristic for physical design of FCN layouts
+#include <fiction/algorithms/physical_design/post_layout_optimization.hpp>  // scalable heuristic for physical design of FCN layouts
+#include <fiction/algorithms/properties/critical_path_length_and_throughput.hpp>  // critical path and throughput calculations
+#include <fiction/algorithms/verification/equivalence_checking.hpp>               // SAT-based equivalence checking
+
+#include <fmt/format.h>                      // output formatting
+#include <lorina/lorina.hpp>                 // Verilog/BLIF/AIGER/... file parsing
+#include <mockturtle/io/verilog_reader.hpp>  // call-backs to read Verilog files into networks
+
+#include <cassert>
+#include <chrono>
+#include <cstdlib>
+#include <string>
+
+int main()  // NOLINT
+{
+    using gate_lyt =
+        fiction::gate_level_layout<fiction::clocked_layout<fiction::tile_based_layout<fiction::cartesian_layout<>>>>;
+
+    experiments::experiment<std::string, uint32_t, uint32_t, uint32_t, uint64_t, uint64_t, uint64_t, uint64_t, uint64_t,
+                            uint64_t, uint32_t, uint32_t, uint64_t, uint64_t, double, double, float, std::string>
+        optimization_exp{"optimization",
+                         "benchmark",
+                         "inputs",
+                         "outputs",
+                         "initial nodes",
+                         "ortho layout width (in tiles)",
+                         "ortho layout height (in tiles)",
+                         "ortho layout area (in tiles)",
+                         "optimized layout width (in tiles)",
+                         "optimized layout height (in tiles)",
+                         "optimized layout area (in tiles)",
+                         "gates",
+                         "wires",
+                         "critical path",
+                         "throughput",
+                         "runtime ortho (in sec)",
+                         "runtime optimization (in sec)",
+                         "improvement (%)",
+                         "equivalent"};
+
+    // stats for SMT-based physical design
+    fiction::orthogonal_physical_design_stats orthogonal_stats{};
+    fiction::post_layout_optimization_stats   post_layout_optimization_stats{};
+
+    static constexpr const uint64_t bench_select =
+        fiction_experiments::all & ~fiction_experiments::epfl & ~fiction_experiments::iscas85;
+
+    for (const auto& benchmark : fiction_experiments::all_benchmarks(bench_select))
+    {
+        fmt::print("[i] processing {}\n", benchmark);
+
+        fiction::technology_network network{};
+
+        const auto read_verilog_result =
+            lorina::read_verilog(fiction_experiments::benchmark_path(benchmark), mockturtle::verilog_reader(network));
+        assert(read_verilog_result == lorina::return_code::success);
+
+        // perform layout generation with an OGD-based heuristic algorithm
+        auto gate_level_layout = fiction::orthogonal<gate_lyt>(network, {}, &orthogonal_stats);
+
+        //  compute critical path and throughput
+        fiction::critical_path_length_and_throughput_stats cp_tp_stats{};
+        fiction::critical_path_length_and_throughput(gate_level_layout, &cp_tp_stats);
+
+        // calculate bounding box
+        const auto bounding_box_before_optimization = fiction::bounding_box_2d(gate_level_layout);
+
+        const auto width_before_optimization  = bounding_box_before_optimization.get_x_size() + 1;
+        const auto height_before_optimization = bounding_box_before_optimization.get_y_size() + 1;
+        const auto area_before_optimization   = width_before_optimization * height_before_optimization;
+
+        // perform post-layout optimization
+        fiction::post_layout_optimization<gate_lyt>(gate_level_layout, &post_layout_optimization_stats);
+
+        // check equivalence
+        fiction::equivalence_checking_stats eq_stats{};
+        fiction::equivalence_checking<fiction::technology_network, gate_lyt>(network, gate_level_layout, &eq_stats);
+
+        const std::string eq_result = eq_stats.eq == fiction::eq_type::STRONG ? "STRONG" :
+                                      eq_stats.eq == fiction::eq_type::WEAK   ? "WEAK" :
+                                                                                "NO";
+
+        // calculate bounding box
+        const auto bounding_box_after_optimization = fiction::bounding_box_2d(gate_level_layout);
+
+        const auto width_after_optimization  = bounding_box_after_optimization.get_x_size() + 1;
+        const auto height_after_optimization = bounding_box_after_optimization.get_y_size() + 1;
+        const auto area_after_optimization   = width_after_optimization * height_after_optimization;
+
+        const float improv = 100 * static_cast<float>((area_before_optimization - area_after_optimization)) /
+                             static_cast<float>(area_before_optimization);
+        // log results
+        optimization_exp(benchmark, network.num_pis(), network.num_pos(), network.num_gates(),
+                         width_before_optimization, height_before_optimization, area_before_optimization,
+                         width_after_optimization, height_after_optimization, area_after_optimization,
+                         gate_level_layout.num_gates(), gate_level_layout.num_wires(), cp_tp_stats.critical_path_length,
+                         cp_tp_stats.throughput, mockturtle::to_seconds(orthogonal_stats.time_total),
+                         mockturtle::to_seconds(post_layout_optimization_stats.time_total), improv, eq_result);
+
+        optimization_exp.save();
+        optimization_exp.table();
+    }
+
+    return EXIT_SUCCESS;
+}