cost_generic_resub.hpp
No OneTemporary
Actions

Size

13 KB

Referenced Files

None

Subscribers

None

cost_generic_resub.hpp
View Options

	/* mockturtle: C++ logic network library
	* Copyright (C) 2018-2022 EPFL
	*
	* Permission is hereby granted, free of charge, to any person
	* obtaining a copy of this software and associated documentation
	* files (the "Software"), to deal in the Software without
	* restriction, including without limitation the rights to use,
	* copy, modify, merge, publish, distribute, sublicense, and/or sell
	* copies of the Software, and to permit persons to whom the
	* Software is furnished to do so, subject to the following
	* conditions:
	*
	* The above copyright notice and this permission notice shall be
	* included in all copies or substantial portions of the Software.
	*
	* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
	* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
	* OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
	* NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
	* HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
	* WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
	* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
	* OTHER DEALINGS IN THE SOFTWARE.
	*/

	/*!
	\file cost_generic_resub.hpp
	\brief generic widnowing algorithm with customized cost function

	\author Hanyu Wang
	*/

	#pragma once

	#include "../../networks/aig.hpp"
	#include "../../networks/xag.hpp"
	#include "../../traits.hpp"
	#include "../../utils/index_list.hpp"
	#include "../../utils/stopwatch.hpp"
	#include "../../views/cost_view.hpp"
	#include "../../views/depth_view.hpp"
	#include "../../views/fanout_view.hpp"
	#include "../../views/topo_view.hpp"
	#include "../detail/resub_utils.hpp"
	#include "../dont_cares.hpp"
	#include "../reconv_cut.hpp"
	#include "../simulation.hpp"
	#include "boolean_optimization.hpp"
	#include "cost_resyn.hpp"
	#include <kitty/kitty.hpp>

	#include <functional>
	#include <optional>
	#include <vector>

	namespace mockturtle::experimental
	{

	/*! \brief Parameters for cost.
	*/
	struct costfn_windowing_params
	{
	/! \brief Maximum number of PIs of reconvergence-driven cuts. /
	uint32_t max_pis{ 8 };

	/! \brief Maximum number of divisors to consider. /
	uint32_t max_divisors{ 150 };

	/! \brief Maximum number of nodes added by resubstitution. /
	uint32_t max_inserts{ 2 };

	/! \brief Maximum fanout of a node to be considered as root. /
	uint32_t skip_fanout_limit_for_roots{ 1000 };

	/! \brief Maximum fanout of a node to be considered as divisor. /
	uint32_t skip_fanout_limit_for_divisors{ 100 };

	/! \brief Use don't cares for optimization. /
	bool use_dont_cares{ false };

	/! \brief Window size for don't cares calculation. /
	uint32_t window_size{ 12u };

	/*! \brief Whether to normalize the truth tables.
	*
	* For some enumerative resynthesis engines, if the truth tables
	* are normalized, some cases can be eliminated and thus improves
	* efficiency. When this option is turned off, be sure to use an
	* implementation of resynthesis that does not make this assumption;
	* otherwise, quality degradation may be observed.
	*
	* Normalization is typically only useful for enumerative methods
	* and for smaller solutions (i.e. when `max_inserts` < 2). Turning
	* on normalization may result in larger runtime overhead when there
	* are many divisors or when the truth tables are long.
	*/
	bool normalize{ false };
	};

	struct costfn_windowing_stats
	{
	/! \brief Total runtime. /
	stopwatch<>::duration time_total{ 0 };

	/! \brief Accumulated runtime for cut computation. /
	stopwatch<>::duration time_cuts{ 0 };

	/! \brief Accumulated runtime for mffc computation. /
	stopwatch<>::duration time_mffc{ 0 };

	/! \brief Accumulated runtime for divisor collection. /
	stopwatch<>::duration time_divs{ 0 };

	/! \brief Accumulated runtime for simulation. /
	stopwatch<>::duration time_sim{ 0 };

	/! \brief Accumulated runtime for don't care computation. /
	stopwatch<>::duration time_dont_care{ 0 };

	/! \brief Total number of leaves. /
	uint64_t num_leaves{ 0u };

	/! \brief Total number of divisors. /
	uint64_t num_divisors{ 0u };

	/! \brief Number of constructed windows. /
	uint32_t num_windows{ 0u };

	/! \brief Total number of MFFC nodes. /
	uint64_t sum_mffc_size{ 0u };

	void report() const
	{
	// clang-format off
	fmt::print( "[i] costfn_windowing report\n" );
	fmt::print( " tot. #leaves = {:5d}, tot. #divs = {:5d}, sum \|MFFC\| = {:5d}\n", num_leaves, num_divisors, sum_mffc_size );
	fmt::print( " avg. #leaves = {:>5.2f}, avg. #divs = {:>5.2f}, avg. \|MFFC\| = {:>5.2f}\n", float( num_leaves ) / float( num_windows ), float( num_divisors ) / float( num_windows ), float( sum_mffc_size ) / float( num_windows ) );
	fmt::print( " ===== Runtime Breakdown =====\n" );
	fmt::print( " Total : {:>5.2f} secs\n", to_seconds( time_total ) );
	fmt::print( " Cut : {:>5.2f} secs\n", to_seconds( time_cuts ) );
	fmt::print( " MFFC : {:>5.2f} secs\n", to_seconds( time_mffc ) );
	fmt::print( " Divs : {:>5.2f} secs\n", to_seconds( time_divs ) );
	fmt::print( " Simulation: {:>5.2f} secs\n", to_seconds( time_sim ) );
	fmt::print( " Dont cares: {:>5.2f} secs\n", to_seconds( time_dont_care ) );
	// clang-format on
	}
	};

	namespace detail
	{
	template<class Ntk, class TT>
	struct cost_aware_problem
	{
	using node = typename Ntk::node;
	using signal = typename Ntk::signal;

	signal root;
	std::vector<signal> divs;
	std::vector<uint32_t> div_ids; /* positions of divisor truth tables in `tts` */
	std::vector<node> div_id_to_node; /* maps IDs in `div_ids` to the corresponding node */
	std::vector<TT> tts;
	TT care;
	uint32_t mffc_size;
	uint32_t max_cost{ std::numeric_limits<uint32_t>::max() };
	};

	template<class Ntk, class TT = kitty::dynamic_truth_table>
	class costfn_windowing
	{
	public:
	using problem_t = cost_aware_problem<Ntk, TT>;
	using params_t = costfn_windowing_params;
	using stats_t = costfn_windowing_stats;

	using node = typename Ntk::node;
	using signal = typename Ntk::signal;

	explicit costfn_windowing( Ntk& ntk, params_t const& ps, stats_t& st )
	: ntk( ntk ), ps( ps ), st( st ), cps( { ps.max_pis } ), mffc_mgr( ntk ),
	divs_mgr( ntk, divisor_collector_params( { ps.max_divisors, ps.max_divisors, ps.skip_fanout_limit_for_divisors } ) ),
	sim( ntk, win.tts, ps.max_pis )
	{
	static_assert( has_fanout_size_v<Ntk>, "Ntk does not implement the fanout_size method" );
	static_assert( has_set_value_v<Ntk>, "Ntk does not implement the set_value method" );
	static_assert( has_value_v<Ntk>, "Ntk does not implement the value method" );
	static_assert( has_get_node_v<Ntk>, "Ntk does not implement the get_node method" );
	static_assert( has_make_signal_v<Ntk>, "Ntk does not implement the make_signal method" );
	static_assert( has_is_complemented_v<Ntk>, "Ntk does not implement the is_complemented method" );
	static_assert( has_substitute_node_v<Ntk>, "Ntk does not implement the substitute_node method" );
	}

	void init()
	{
	}

	std::optional<std::reference_wrapper<problem_t>> operator()( node const& n )
	{
	stopwatch t( st.time_total );
	if ( ntk.fanout_size( n ) > ps.skip_fanout_limit_for_roots )
	{
	return std::nullopt; /* skip nodes with too many fanouts */
	}

	/* compute a cut and collect supported nodes */
	std::vector<node> leaves = call_with_stopwatch( st.time_cuts, [&]() {
	return reconvergence_driven_cut<Ntk, false, has_level_v<Ntk>>( ntk, { n }, cps ).first;
	} );
	std::vector<node> supported;
	call_with_stopwatch( st.time_divs, [&]() {
	divs_mgr.collect_supported_nodes( n, leaves, supported );
	} );

	/* simulate */
	call_with_stopwatch( st.time_sim, [&]() {
	sim.simulate( leaves, supported );
	} );

	/* mark MFFC nodes and collect divisors */
	++mffc_marker;
	win.mffc_size = call_with_stopwatch( st.time_mffc, [&]() {
	return mffc_mgr.call_on_mffc_and_count( n, leaves, [&]( node const& n ) {
	ntk.set_value( n, mffc_marker );
	} );
	} );
	call_with_stopwatch( st.time_divs, [&]() {
	collect_divisors( leaves, supported );
	} );

	/* normalize */
	call_with_stopwatch( st.time_sim, [&]() {
	if ( ps.normalize )
	{
	win.root = normalize_truth_tables() ? !ntk.make_signal( n ) : ntk.make_signal( n );
	}
	else
	{
	win.root = ntk.make_signal( n );
	}
	} );

	/* compute don't cares */
	call_with_stopwatch( st.time_dont_care, [&]() {
	if ( ps.use_dont_cares )
	{
	win.care = ~satisfiability_dont_cares( ntk, leaves, ps.window_size );
	}
	else
	{
	win.care = ~kitty::create<TT>( ps.max_pis );
	}
	} );

	/* compute cost */
	win.max_cost = ntk.get_cost( n, win.divs );

	st.num_windows++;
	st.num_leaves += leaves.size();
	st.num_divisors += win.divs.size();
	st.sum_mffc_size += win.mffc_size;

	return win;
	}

	template<typename res_t>
	uint32_t gain( problem_t const& prob, res_t const& res ) const
	{
	static_assert( is_index_list_v<res_t>, "res_t is not an index_list (windowing engine and resynthesis engine do not match)" );
	return 1; /* cannot predict the final cost */
	}

	template<typename res_t>
	bool update_ntk( problem_t const& prob, res_t const& res )
	{
	static_assert( is_index_list_v<res_t>, "res_t is not an index_list (windowing engine and resynthesis engine do not match)" );
	assert( res.num_pos() == 1 );
	insert( ntk, std::begin( prob.divs ), std::end( prob.divs ), res, [&]( signal const& g ) {
	ntk.substitute_node( ntk.get_node( prob.root ), ntk.is_complemented( prob.root ) ? !g : g );
	} );
	return true; /* continue optimization */
	}

	template<typename res_t>
	bool report( problem_t const& prob, res_t const& res )
	{
	static_assert( is_index_list_v<res_t>, "res_t is not an index_list (windowing engine and resynthesis engine do not match)" );
	assert( res.num_pos() == 1 );
	fmt::print( "[i] found solution {} for root signal {}{}\n", to_index_list_string( res ), ntk.is_complemented( prob.root ) ? "!" : "", ntk.get_node( prob.root ) );
	return true;
	}

	private:
	void collect_divisors( std::vector<node> const& leaves, std::vector<node> const& supported )
	{
	win.divs.clear();
	win.div_ids.clear();

	uint32_t i{ 1 };
	for ( auto const& l : leaves )
	{
	win.div_ids.emplace_back( i++ );
	win.divs.emplace_back( ntk.make_signal( l ) );
	}

	i = ps.max_pis + 1;
	for ( auto const& n : supported )
	{
	if ( ntk.value( n ) != mffc_marker ) /* not in MFFC, not root */
	{
	win.div_ids.emplace_back( i );
	win.divs.emplace_back( ntk.make_signal( n ) );
	}
	++i;
	}
	assert( i == win.tts.size() );
	}

	bool normalize_truth_tables()
	{
	assert( win.divs.size() == win.div_ids.size() );
	for ( auto i = 0u; i < win.divs.size(); ++i )
	{
	if ( kitty::get_bit( win.tts.at( win.div_ids.at( i ) ), 0 ) )
	{
	win.tts.at( win.div_ids.at( i ) ) = ~win.tts.at( win.div_ids.at( i ) );
	win.divs.at( i ) = !win.divs.at( i );
	}
	}

	if ( kitty::get_bit( win.tts.back(), 0 ) )
	{
	win.tts.back() = ~win.tts.back();
	return true;
	}
	else
	{
	return false;
	}
	}

	private:
	Ntk& ntk;
	problem_t win;
	params_t const& ps;
	stats_t& st;
	reconvergence_driven_cut_parameters const cps;
	typename mockturtle::detail::node_mffc_inside<Ntk> mffc_mgr; // TODO: namespaces can be removed when we move out of experimental::
	divisor_collector<Ntk> divs_mgr;
	window_simulator<Ntk, TT> sim;
	uint32_t mffc_marker{ 0u };
	std::shared_ptr<typename network_events<Ntk>::modified_event_type> lazy_update_event;
	}; /* costfn_windowing */

	template<class Ntk, class TT, class ResynEngine>
	class costfn_resynthesis
	{
	public:
	using problem_t = cost_aware_problem<Ntk, TT>;
	using res_t = typename ResynEngine::index_list_t;
	using params_t = typename ResynEngine::params;
	using stats_t = typename ResynEngine::stats;

	explicit costfn_resynthesis( Ntk const& ntk, params_t const& ps, stats_t& st )
	: ntk( ntk ), engine( ntk, ps, st )
	{
	static_assert( has_cost_v<Ntk>, "Ntk does not implement the get_cost method" );
	}

	void init()
	{
	}

	std::optional<res_t> operator()( problem_t& prob )
	{
	return engine( prob.tts.back(), prob.care, prob.divs, std::begin( prob.div_ids ), std::end( prob.div_ids ), prob.tts, prob.max_cost );
	}

	private:
	Ntk const& ntk;
	typename ResynEngine::stats rst;
	ResynEngine engine;
	}; /* costfn_resynthesis */

	} /* namespace detail */

	using cost_generic_resub_params = boolean_optimization_params<costfn_windowing_params, cost_resyn_params>;
	using cost_generic_resub_stats = boolean_optimization_stats<costfn_windowing_stats, cost_resyn_stats>;

	/*! \brief Cost-generic resubstitution algorithm.
	*
	* This algorithm creates a reconvergence-driven window for each node in the
	* network, collects divisors, and builds the resynthesis problem. A search core
	* then collects all the resubstitution candidates with the same functionality as
	* the target. The candidate with the lowest cost will then replace the MFFC
	* of the window.
	*
	* \param ntk Network
	* \param cost_fn Customized cost function
	* \param ps Optimization params
	* \param pst Optimization statistics
	*/
	template<class Ntk, class CostFn>
	void cost_generic_resub( Ntk& ntk, CostFn cost_fn, cost_generic_resub_params const& ps, cost_generic_resub_stats* pst = nullptr )
	{
	fanout_view fntk( ntk );
	cost_view viewed( fntk, cost_fn );
	using Viewed = decltype( viewed );
	using TT = typename kitty::dynamic_truth_table;
	using windowing_t = typename detail::costfn_windowing<Viewed, TT>;
	using engine_t = cost_resyn<Viewed, TT>;
	using resyn_t = typename detail::costfn_resynthesis<Viewed, TT, engine_t>;
	using opt_t = typename detail::boolean_optimization_impl<Viewed, windowing_t, resyn_t>;

	cost_generic_resub_stats st;
	opt_t p( viewed, ps, st );
	p.run();
	if ( ps.verbose )
	{
	st.report();
	}
	if ( pst )
	{
	*pst = st;
	}
	}

	} // namespace mockturtle::experimental

File Metadata

Mime Type: text/x-c++
Expires: Sun, Jul 12, 7:18 PM (11 h, 37 m)
Storage Engine: blob
Storage Format: Raw Data
Storage Handle: 351473
Default Alt Text: cost_generic_resub.hpp (13 KB)

cost_generic_resub.hppNo OneTemporaryActions

cost_generic_resub.hppView Options

File Metadata

Event Timeline

cost_generic_resub.hpp
No OneTemporary
Actions

cost_generic_resub.hpp
View Options