mirrors/z3
3
0
Fork 0
mirror of https://github.com/Z3Prover/z3 synced 2025-04-24 09:35:32 +00:00
Code Activity

move parallel-tactic to solver level

Browse source 
Signed-off-by: Nikolaj Bjorner <nbjorner@microsoft.com>
This commit is contained in:
Nikolaj Bjorner 2018-04-16 08:21:21 -07:00
parent cd35caff52
commit a37303a045
16 changed files with 89 additions and 86 deletions
Download patch file Download diff file Expand all files Collapse all files

3
src/solver/CMakeLists.txt
View file

@ -3,6 +3,7 @@ z3_add_component(solver
check_sat_result.cpp
combined_solver.cpp
mus.cpp
parallel_tactic.cpp
smt_logics.cpp
solver.cpp
solver_na2as.cpp
@ -14,4 +15,6 @@ z3_add_component(solver
tactic
PYG_FILES
combined_solver_params.pyg
parallel_params.pyg
)

717
src/solver/parallel_tactic.cpp Normal file
View file

@ -0,0 +1,717 @@
/*++
Copyright (c) 2017 Microsoft Corporation
Module Name:
parallel_tactic.cpp
Abstract:
Parallel tactic based on cubing.
Author:
Nikolaj Bjorner (nbjorner) 2017-10-9
Miguel Neves
Notes:
A task comprises of a non-empty sequence of cubes, a type and parameters
It invokes the following procedure:
1. Clone the state with the remaining cubes if there is more than one cube. Re-enqueue the remaining cubes.
2. Apply simplifications and pre-processing according to configuration.
3. Cube using the parameter settings prescribed in m_params.
4. Optionally pass the cubes as assumptions and solve each sub-cube with a prescribed resource bound.
5. Assemble cubes that could not be solved and create a cube state.
--*/
#include <thread>
#include <mutex>
#include <condition_variable>
#include "util/scoped_ptr_vector.h"
#include "ast/ast_util.h"
#include "ast/ast_translation.h"
#include "solver/solver.h"
#include "solver/solver2tactic.h"
#include "tactic/tactic.h"
#include "tactic/tactical.h"
#include "tactic/portfolio/fd_solver.h"
#include "solver/parallel_tactic.h"
#include "solver/parallel_params.hpp"
#include "smt/tactic/smt_tactic.h"
#include "smt/smt_solver.h"
#include "sat/sat_solver/inc_sat_solver.h"
#include "sat/tactic/sat_tactic.h"
class parallel_tactic : public tactic {
class solver_state;
class task_queue {
std::mutex m_mutex;
std::condition_variable m_cond;
ptr_vector<solver_state> m_tasks;
ptr_vector<solver_state> m_active;
unsigned m_num_waiters;
volatile bool m_shutdown;
void inc_wait() {
std::lock_guard<std::mutex> lock(m_mutex);
++m_num_waiters;
}
void dec_wait() {
std::lock_guard<std::mutex> lock(m_mutex);
--m_num_waiters;
}
solver_state* try_get_task() {
solver_state* st = nullptr;
std::lock_guard<std::mutex> lock(m_mutex);
if (!m_tasks.empty()) {
st = m_tasks.back();
m_tasks.pop_back();
m_active.push_back(st);
}
return st;
}
public:
task_queue():
m_num_waiters(0),
m_shutdown(false) {}
~task_queue() { reset(); }
void shutdown() {
if (!m_shutdown) {
m_shutdown = true;
m_cond.notify_all();
std::lock_guard<std::mutex> lock(m_mutex);
for (solver_state* st : m_active) {
st->m().limit().cancel();
}
}
}
void add_task(solver_state* task) {
std::lock_guard<std::mutex> lock(m_mutex);
m_tasks.push_back(task);
if (m_num_waiters > 0) {
m_cond.notify_one();
}
}
bool is_idle() {
std::lock_guard<std::mutex> lock(m_mutex);
return m_tasks.empty() && m_num_waiters > 0;
}
solver_state* get_task() {
while (!m_shutdown) {
inc_wait();
solver_state* st = try_get_task();
if (st) {
dec_wait();
return st;
}
{
std::unique_lock<std::mutex> lock(m_mutex);
m_cond.wait(lock);
}
dec_wait();
}
return nullptr;
}
void task_done(solver_state* st) {
std::lock_guard<std::mutex> lock(m_mutex);
m_active.erase(st);
if (m_tasks.empty() && m_active.empty()) {
m_shutdown = true;
m_cond.notify_all();
}
}
void reset() {
for (auto* t : m_tasks) dealloc(t);
for (auto* t : m_active) dealloc(t);
m_tasks.reset();
m_active.reset();
}
std::ostream& display(std::ostream& out) {
std::lock_guard<std::mutex> lock(m_mutex);
out << "num_tasks " << m_tasks.size() << " active: " << m_active.size() << "\n";
for (solver_state* st : m_tasks) {
st->display(out);
}
return out;
}
};
class cube_var {
expr_ref_vector m_vars;
expr_ref_vector m_cube;
public:
cube_var(expr_ref_vector& c, expr_ref_vector& vs):
m_vars(vs), m_cube(c) {}
cube_var operator()(ast_translation& tr) {
return cube_var(tr(m_cube), tr(m_vars));
}
expr_ref_vector const& cube() const { return m_cube; }
expr_ref_vector const& vars() const { return m_vars; }
};
class solver_state {
scoped_ptr<ast_manager> m_manager; // ownership handle to ast_manager
vector<cube_var> m_cubes; // set of cubes to process by task
expr_ref_vector m_asserted_cubes; // set of cubes asserted on the current solver
params_ref m_params; // configuration parameters
ref<solver> m_solver; // solver state
unsigned m_depth; // number of nested calls to cubing
double m_width; // estimate of fraction of problem handled by state
public:
solver_state(ast_manager* m, solver* s, params_ref const& p):
m_manager(m),
m_asserted_cubes(s->get_manager()),
m_params(p),
m_solver(s),
m_depth(0),
m_width(1.0)
{
}
ast_manager& m() { return m_solver->get_manager(); }
solver& get_solver() { return *m_solver; }
solver* copy_solver() { return m_solver->translate(m_solver->get_manager(), m_params); }
solver const& get_solver() const { return *m_solver; }
solver_state* clone() {
SASSERT(!m_cubes.empty());
ast_manager& m = m_solver->get_manager();
ast_manager* new_m = alloc(ast_manager, m, m.proof_mode());
ast_translation tr(m, *new_m);
solver* s = m_solver.get()->translate(*new_m, m_params);
solver_state* st = alloc(solver_state, new_m, s, m_params);
for (auto & c : m_cubes) st->m_cubes.push_back(c(tr));
for (expr* c : m_asserted_cubes) st->m_asserted_cubes.push_back(tr(c));
st->m_depth = m_depth;
st->m_width = m_width;
return st;
}
vector<cube_var> const& cubes() const { return m_cubes; }
// remove up to n cubes from list of cubes.
vector<cube_var> split_cubes(unsigned n) {
vector<cube_var> result;
while (n-- > 0 && !m_cubes.empty()) {
result.push_back(m_cubes.back());
m_cubes.pop_back();
}
return result;
}
void set_cubes(vector<cube_var>& c) {
m_cubes.reset();
m_cubes.append(c);
}
void inc_depth(unsigned inc) { m_depth += inc; }
void inc_width(unsigned w) { m_width *= w; }
double get_width() const { return m_width; }
unsigned get_depth() const { return m_depth; }
lbool simplify() {
lbool r = l_undef;
IF_VERBOSE(2, verbose_stream() << "(parallel.tactic simplify-1)\n";);
set_simplify_params(true); // retain blocked
r = get_solver().check_sat(0,0);
if (r != l_undef) return r;
IF_VERBOSE(2, verbose_stream() << "(parallel.tactic simplify-2)\n";);
set_simplify_params(false); // remove blocked
r = get_solver().check_sat(0,0);
return r;
}
void assert_cube(expr_ref_vector const& cube) {
get_solver().assert_expr(cube);
m_asserted_cubes.append(cube);
}
void set_cube_params() {
}
void set_conquer_params() {
set_conquer_params(get_solver());
}
void set_conquer_params(solver& s) {
parallel_params pp(m_params);
params_ref p;
p.copy(m_params);
p.set_bool("gc.burst", true); // apply eager gc
p.set_uint("simplify.delay", 1000); // delay simplification by 1000 conflicts
p.set_bool("lookahead_simplify", false);
p.set_uint("restart.max", pp.conquer_restart_max());
p.set_uint("inprocess.max", UINT_MAX); // base bounds on restart.max
s.updt_params(p);
}
void set_simplify_params(bool retain_blocked) {
parallel_params pp(m_params);
params_ref p;
p.copy(m_params);
double exp = pp.simplify_exp();
exp = std::max(exp, 1.0);
unsigned mult = static_cast<unsigned>(pow(exp, m_depth - 1));
p.set_uint("inprocess.max", pp.simplify_inprocess_max() * mult);
p.set_uint("restart.max", pp.simplify_restart_max() * mult);
p.set_bool("lookahead_simplify", true);
p.set_bool("retain_blocked_clauses", retain_blocked);
get_solver().updt_params(p);
}
bool canceled() {
return m().canceled();
}
std::ostream& display(std::ostream& out) {
out << ":depth " << m_depth << " :width " << m_width << "\n";
out << ":asserted " << m_asserted_cubes.size() << "\n";
return out;
}
};
private:
solver_ref m_solver;
ast_manager& m_manager;
params_ref m_params;
sref_vector<model> m_models;
unsigned m_num_threads;
statistics m_stats;
task_queue m_queue;
std::mutex m_mutex;
double m_progress;
unsigned m_branches;
unsigned m_backtrack_frequency;
unsigned m_conquer_delay;
volatile bool m_has_undef;
bool m_allsat;
unsigned m_num_unsat;
int m_exn_code;
std::string m_exn_msg;
void init() {
parallel_params pp(m_params);
m_num_threads = std::min((unsigned)omp_get_num_procs(), pp.threads_max());
m_progress = 0;
m_has_undef = false;
m_allsat = false;
m_branches = 0;
m_num_unsat = 0;
m_backtrack_frequency = pp.conquer_backtrack_frequency();
m_conquer_delay = pp.conquer_delay();
m_exn_code = 0;
m_params.set_bool("override_incremental", true);
}
void log_branches(lbool status) {
IF_VERBOSE(0, verbose_stream() << "(tactic.parallel :progress " << m_progress << "% ";
if (status == l_true) verbose_stream() << ":status sat ";
if (status == l_undef) verbose_stream() << ":status unknown ";
verbose_stream() << ":closed " << m_num_unsat << " :open " << m_branches << ")\n";);
}
void add_branches(unsigned b) {
if (b == 0) return;
{
std::lock_guard<std::mutex> lock(m_mutex);
m_branches += b;
}
log_branches(l_false);
}
void dec_branch() {
std::lock_guard<std::mutex> lock(m_mutex);
--m_branches;
}
void close_branch(solver_state& s, lbool status) {
double f = 100.0 / s.get_width();
{
std::lock_guard<std::mutex> lock(m_mutex);
m_progress += f;
--m_branches;
}
log_branches(status);
}
void report_sat(solver_state& s) {
close_branch(s, l_true);
model_ref mdl;
s.get_solver().get_model(mdl);
if (mdl) {
std::lock_guard<std::mutex> lock(m_mutex);
if (&s.m() != &m_manager) {
ast_translation tr(s.m(), m_manager);
mdl = mdl->translate(tr);
}
m_models.push_back(mdl.get());
}
if (!m_allsat) {
m_queue.shutdown();
}
}
void inc_unsat() {
std::lock_guard<std::mutex> lock(m_mutex);
++m_num_unsat;
}
void report_unsat(solver_state& s) {
inc_unsat();
close_branch(s, l_false);
}
void report_undef(solver_state& s) {
m_has_undef = true;
close_branch(s, l_undef);
}
void cube_and_conquer(solver_state& s) {
ast_manager& m = s.m();
vector<cube_var> cube, hard_cubes, cubes;
expr_ref_vector vars(m);
cube_again:
// extract up to one cube and add it.
cube.reset();
cube.append(s.split_cubes(1));
SASSERT(cube.size() <= 1);
IF_VERBOSE(2, verbose_stream() << "(tactic.parallel :split-cube " << cube.size() << ")\n";);
if (!s.cubes().empty()) m_queue.add_task(s.clone());
if (!cube.empty()) {
s.assert_cube(cube.get(0).cube());
vars.reset();
vars.append(cube.get(0).vars());
}
simplify_again:
s.inc_depth(1);
// simplify
if (canceled(s)) return;
switch (s.simplify()) {
case l_undef: break;
case l_true: report_sat(s); return;
case l_false: report_unsat(s); return;
}
if (canceled(s)) return;
if (memory_pressure()) {
goto simplify_again;
}
// extract cubes.
cubes.reset();
s.set_cube_params();
solver_ref conquer;
unsigned cutoff = UINT_MAX;
bool first = true;
unsigned num_backtracks = 0, width = 0;
while (cutoff > 0 && !canceled(s)) {
expr_ref_vector c = s.get_solver().cube(vars, cutoff);
if (c.empty()) {
goto simplify_again;
}
if (m.is_false(c.back())) {
break;
}
lbool is_sat = l_undef;
if (width >= m_conquer_delay && !conquer) {
conquer = s.copy_solver();
s.set_conquer_params(*conquer.get());
}
if (conquer) {
is_sat = conquer->check_sat(c);
}
switch (is_sat) {
case l_false:
cutoff = c.size();
backtrack(*conquer.get(), c, (num_backtracks++) % m_backtrack_frequency == 0);
if (cutoff != c.size()) {
IF_VERBOSE(0, verbose_stream() << "(tactic.parallel :backtrack " << cutoff << " -> " << c.size() << ")\n");
cutoff = c.size();
}
inc_unsat();
log_branches(l_false);
break;
case l_true:
report_sat(s);
if (conquer) {
collect_statistics(*conquer.get());
}
return;
case l_undef:
++width;
IF_VERBOSE(2, verbose_stream() << "(tactic.parallel :cube " << c.size() << " :vars " << vars.size() << ")\n");
cubes.push_back(cube_var(c, vars));
cutoff = UINT_MAX;
break;
}
if (cubes.size() >= conquer_batch_size() || (!cubes.empty() && m_queue.is_idle())) {
spawn_cubes(s, std::max(2u, width), cubes);
first = false;
cubes.reset();
}
}
if (conquer) {
collect_statistics(*conquer.get());
}
if (cubes.empty() && first) {
report_unsat(s);
}
else if (cubes.empty()) {
dec_branch();
}
else {
s.inc_width(width);
add_branches(cubes.size()-1);
s.set_cubes(cubes);
goto cube_again;
}
}
void spawn_cubes(solver_state& s, unsigned width, vector<cube_var>& cubes) {
if (cubes.empty()) return;
add_branches(cubes.size());
s.set_cubes(cubes);
solver_state* s1 = s.clone();
s1->inc_width(width);
m_queue.add_task(s1);
}
/*
* \brief backtrack from unsatisfiable core
*/
void backtrack(solver& s, expr_ref_vector& asms, bool full) {
ast_manager& m = s.get_manager();
expr_ref_vector core(m);
obj_hashtable<expr> hcore;
s.get_unsat_core(core);
while (!asms.empty() && !core.contains(asms.back())) asms.pop_back();
if (!full) return;
//s.assert_expr(m.mk_not(mk_and(core)));
if (!asms.empty()) {
expr* last = asms.back();
expr_ref not_last(mk_not(m, last), m);
asms.pop_back();
asms.push_back(not_last);
lbool r = s.check_sat(asms);
asms.pop_back();
if (r != l_false) {
asms.push_back(last);
return;
}
core.reset();
s.get_unsat_core(core);
if (core.contains(not_last)) {
//s.assert_expr(m.mk_not(mk_and(core)));
r = s.check_sat(asms);
}
if (r == l_false) {
backtrack(s, asms, full);
}
}
}
bool canceled(solver_state& s) {
if (s.canceled()) {
m_has_undef = true;
return true;
}
else {
return false;
}
}
bool memory_pressure() {
return memory::above_high_watermark();
}
void run_solver() {
try {
while (solver_state* st = m_queue.get_task()) {
cube_and_conquer(*st);
collect_statistics(*st);
m_queue.task_done(st);
if (st->m().canceled()) m_queue.shutdown();
IF_VERBOSE(1, display(verbose_stream()););
dealloc(st);
}
}
catch (z3_exception& ex) {
IF_VERBOSE(1, verbose_stream() << ex.msg() << "\n";);
m_queue.shutdown();
std::lock_guard<std::mutex> lock(m_mutex);
if (ex.has_error_code()) {
m_exn_code = ex.error_code();
}
else {
m_exn_msg = ex.msg();
m_exn_code = -1;
}
}
}
void collect_statistics(solver_state& s) {
collect_statistics(s.get_solver());
}
void collect_statistics(solver& s) {
std::lock_guard<std::mutex> lock(m_mutex);
s.collect_statistics(m_stats);
}
lbool solve(model_ref& mdl) {
add_branches(1);
vector<std::thread> threads;
for (unsigned i = 0; i < m_num_threads; ++i)
threads.push_back(std::thread([this]() { run_solver(); }));
for (std::thread& t : threads)
t.join();
m_manager.limit().reset_cancel();
if (m_exn_code == -1)
throw default_exception(m_exn_msg);
if (m_exn_code != 0)
throw z3_error(m_exn_code);
if (!m_models.empty()) {
mdl = m_models.back();
return l_true;
}
if (m_has_undef)
return l_undef;
return l_false;
}
std::ostream& display(std::ostream& out) {
unsigned n_models, n_unsat;
double n_progress;
{
std::lock_guard<std::mutex> lock(m_mutex);
n_models = m_models.size();
n_unsat = m_num_unsat;
n_progress = m_progress;
}
m_stats.display(out);
m_queue.display(out);
out << "(tactic.parallel :unsat " << n_unsat << " :progress " << n_progress << "% :models " << n_models << ")\n";
return out;
}
public:
parallel_tactic(solver* s, params_ref const& p) :
m_solver(s),
m_manager(s->get_manager()),
m_params(p) {
init();
}
void operator ()(const goal_ref & g,goal_ref_buffer & result) {
ast_manager& m = g->m();
solver* s = m_solver->translate(m, m_params);
solver_state* st = alloc(solver_state, 0, s, m_params);
m_queue.add_task(st);
expr_ref_vector clauses(m);
ptr_vector<expr> assumptions;
obj_map<expr, expr*> bool2dep;
ref<generic_model_converter> fmc;
extract_clauses_and_dependencies(g, clauses, assumptions, bool2dep, fmc);
for (expr * clause : clauses) {
s->assert_expr(clause);
}
SASSERT(assumptions.empty());
model_ref mdl;
lbool is_sat = solve(mdl);
switch (is_sat) {
case l_true:
if (g->models_enabled()) {
g->add(concat(fmc.get(), model2model_converter(mdl.get())));
}
g->reset();
break;
case l_false:
SASSERT(!g->proofs_enabled());
SASSERT(!g->unsat_core_enabled());
g->assert_expr(m.mk_false(), nullptr, nullptr);
break;
case l_undef:
if (m.canceled()) {
throw tactic_exception(Z3_CANCELED_MSG);
}
break;
}
result.push_back(g.get());
}
unsigned conquer_batch_size() const {
parallel_params pp(m_params);
return pp.conquer_batch_size();
}
void cleanup() {
m_queue.reset();
}
tactic* translate(ast_manager& m) {
solver* s = m_solver->translate(m, m_params);
return alloc(parallel_tactic, s, m_params);
}
virtual void updt_params(params_ref const & p) {
m_params.copy(p);
parallel_params pp(p);
m_conquer_delay = pp.conquer_delay();
}
virtual void collect_statistics(statistics & st) const {
st.copy(m_stats);
st.update("par unsat", m_num_unsat);
st.update("par models", m_models.size());
st.update("par progress", m_progress);
}
virtual void reset_statistics() {
m_stats.reset();
}
};
tactic * mk_parallel_tactic(solver* s, params_ref const& p) {
return alloc(parallel_tactic, s, p);
}

27
src/solver/parallel_tactic.h Normal file
View file

@ -0,0 +1,27 @@
/*++
Copyright (c) 2017 Microsoft Corporation
Module Name:
parallel_tactic.h
Abstract:
Parallel tactic in the style of Treengeling.
Author:
Nikolaj Bjorner (nbjorner) 2017-10-9
Notes:
--*/
#ifndef PARALLEL_TACTIC_H_
#define PARALLEL_TACTIC_H_
class tactic;
class solver;
tactic * mk_parallel_tactic(solver* s, params_ref const& p);
#endif

15
src/solver/parallel_tactic_params.pyg Normal file
View file

@ -0,0 +1,15 @@
def_module_params('parallel',
description='parameters for parallel solver',
class_name='parallel_tactic_params',
export=True,
params=(
('enable', BOOL, False, 'enable parallel solver by default on selected tactics (for QF_BV)'),
('threads.max', UINT, 10000, 'caps maximal number of threads below the number of processors'),
('conquer.batch_size', UINT, 1000, 'number of cubes to batch together for fast conquer'),
('conquer.restart.max', UINT, 5, 'maximal number of restarts during conquer phase'),
('conquer.delay', UINT, 10, 'delay of cubes until applying conquer'),
('conquer.backtrack_frequency', UINT, 10, 'frequency to apply core minimization during conquer'),
('simplify.exp', DOUBLE, 1, 'restart and inprocess max is multipled by simplify.exp ^ depth'),
('simplify.restart.max', UINT, 5000, 'maximal number of restarts during simplification phase'),
('simplify.inprocess.max', UINT, 2, 'maximal number of inprocessing steps during simplification'),
))