mirrors/z3
3
0
Fork 0
mirror of https://github.com/Z3Prover/z3 synced 2025-10-31 11:42:28 +00:00
Code Activity

some comments and change to how parameter variants are stored

Browse source 
Signed-off-by: Nikolaj Bjorner <nbjorner@microsoft.com>
This commit is contained in:
Nikolaj Bjorner 2025-10-30 15:59:13 -07:00
parent d3e2527899
commit 8901f8f44c
2 changed files with 71 additions and 66 deletions
Download patch file Download diff file Expand all files Collapse all files

111
src/smt/smt_parallel.cpp
View file

@ -83,7 +83,7 @@ namespace smt {
return r;
}
unsigned parallel::param_generator::replay_proof_prefixes(vector<smt_params> candidate_param_states, unsigned max_conflicts_epsilon=200) {
unsigned parallel::param_generator::replay_proof_prefixes(vector<param_values> const& candidate_param_states, unsigned max_conflicts_epsilon=200) {
unsigned conflict_budget = m_max_prefix_conflicts + max_conflicts_epsilon;
unsigned best_param_state_idx;
double best_score;
@ -95,11 +95,11 @@ namespace smt {
double score = 0.0;
// apply the ith param state to probe_ctx
smt_params params = candidate_param_states[i];
params_ref p;
params.updt_params(p);
params_ref p = apply_param_values(candidate_param_states[i]);
probe_ctx->updt_params(p);
// todo: m_recorded_cubes as a expr_ref_vector
for (auto const& clause : probe_ctx->m_recorded_clauses) {
expr_ref_vector negated_lits(probe_ctx->m);
for (literal lit : clause) {
@ -111,17 +111,13 @@ namespace smt {
}
// Replay the negated clause
lbool r = probe_ctx->check(negated_lits.size(), negated_lits.data());
::statistics st;
probe_ctx->collect_statistics(st);
unsigned conflicts = 0, decisions = 0, rlimit = 0;
unsigned conflicts = probe_ctx->m_stats.m_num_conflicts;
unsigned decisions = probe_ctx->m_stats.m_num_decisions;
// I can't figure out how to access the statistics fields, I only see an update method
// st.get_uint("conflicts", conflicts);
// st.get_uint("decisions", decisions);
// st.get_uint("rlimit count", rlimit);
score += conflicts + decisions + rlimit;
score += conflicts + decisions;
}
if (i == 0 || score < best_score) {
@ -134,49 +130,40 @@ namespace smt {
}
void parallel::param_generator::init_param_state() {
// param_descrs smt_desc;
// smt_params_helper::collect_param_descrs(smt_desc);
smt_params_helper smtp(m_p);
m_my_param_state.insert(symbol("smt.arith.nl.branching"), smtp.arith_nl_branching());
m_my_param_state.insert(symbol("smt.arith.nl.cross_nested"), smtp.arith_nl_cross_nested());
m_my_param_state.insert(symbol("smt.arith.nl.delay"), smtp.arith_nl_delay());
m_my_param_state.insert(symbol("smt.arith.nl.expensive_patching"), smtp.arith_nl_expensive_patching());
m_my_param_state.insert(symbol("smt.arith.nl.gb"), smtp.arith_nl_gb());
m_my_param_state.insert(symbol("smt.arith.nl.horner"), smtp.arith_nl_horner());
m_my_param_state.insert(symbol("smt.arith.nl.horner_frequency"), smtp.arith_nl_horner_frequency());
m_my_param_state.insert(symbol("smt.arith.nl.optimize_bounds"), smtp.arith_nl_optimize_bounds());
m_my_param_state.insert(symbol("smt.arith.nl.propagate_linear_monomials"), smtp.arith_nl_propagate_linear_monomials());
m_my_param_state.insert(symbol("smt.arith.nl.tangents"), smtp.arith_nl_tangents());
m_param_state.push_back({symbol("smt.arith.nl.branching"), smtp.arith_nl_branching()});
m_param_state.push_back({symbol("smt.arith.nl.cross_nested"), smtp.arith_nl_cross_nested()});
m_param_state.push_back({symbol("smt.arith.nl.delay"), unsigned_value({smtp.arith_nl_delay(), 5, 10})});
m_param_state.push_back({symbol("smt.arith.nl.expensive_patching"), smtp.arith_nl_expensive_patching()});
m_param_state.push_back({symbol("smt.arith.nl.gb"), smtp.arith_nl_grobner()});
m_param_state.push_back({symbol("smt.arith.nl.horner"), smtp.arith_nl_horner()});
m_param_state.push_back({symbol("smt.arith.nl.horner_frequency"), unsigned_value({smtp.arith_nl_horner_frequency(), 2, 6})
});
m_param_state.push_back({symbol("smt.arith.nl.optimize_bounds"), smtp.arith_nl_optimize_bounds()});
m_param_state.push_back(
{symbol("smt.arith.nl.propagate_linear_monomials"), smtp.arith_nl_propagate_linear_monomials()});
m_param_state.push_back({symbol("smt.arith.nl.tangents"), smtp.arith_nl_tangents()});
};
// TODO: this should mutate only one field at a time an mutate it based on m_my_param_state to keep it generic.
parallel::param_generator::param_values parallel::param_generator::mutate_param_state() {
smt_params parallel::param_generator::mutate_param_state() {
smt_params p = m_param_state;
random_gen m_rand;
auto flip_bool = [&](bool &x) {
if (m_rand(2) == 0)
x = !x;
};
auto mutate_uint = [&](unsigned &x, unsigned lo, unsigned hi) {
if ((m_rand() % 2) == 0)
x = lo + (m_rand((hi - lo + 1)));
};
flip_bool(p.m_nl_arith_branching);
flip_bool(p.m_nl_arith_cross_nested);
mutate_uint(p.m_nl_arith_delay, 5, 20);
flip_bool(p.m_nl_arith_expensive_patching);
flip_bool(p.m_nl_arith_gb);
flip_bool(p.m_nl_arith_horner);
mutate_uint(p.m_nl_arith_horner_frequency, 2, 6);
flip_bool(p.m_nl_arith_optimize_bounds);
flip_bool(p.m_nl_arith_propagate_linear_monomials);
flip_bool(p.m_nl_arith_tangents);
return p;
param_values new_param_values(m_param_state);
unsigned index = ctx->get_random_value() % new_param_values.size();
auto &param = new_param_values[index];
if (std::holds_alternative<bool>(param.second)) {
bool value = *std::get_if<bool>(&param.second);
param.second = !value;
}
else if (std::holds_alternative<unsigned_value>(param.second)) {
auto [value, lo, hi] = *std::get_if<unsigned_value>(&param.second);
unsigned new_value = value;
while (new_value == value) {
new_value = lo + ctx->get_random_value() % (hi - lo + 1);
}
std::get<unsigned_value>(param.second).value = new_value;
}
return new_param_values;
}
void parallel::param_generator::protocol_iteration() {
@ -185,6 +172,8 @@ namespace smt {
// copy current param state to all param probe contexts, before running the next prefix step
// this ensures that each param probe context replays the prefix from the same configuration
// instead just one one context and reset it each time before copy.
for (unsigned i = 0; i < m_param_probe_contexts.size(); ++i) {
context::copy(*ctx, *m_param_probe_contexts[i], true);
}
@ -195,8 +184,12 @@ namespace smt {
case l_undef: {
// TODO, change from smt_params to a generic param state representation based on params_ref
// only params_ref have effect on updates.
smt_params best_param_state = m_param_state;
vector<smt_params> candidate_param_states;
param_values best_param_state = m_param_state;
vector<param_values> candidate_param_states;
// you can create the mutations on the fly and get the scores
// you don't have to copy all over each tester.
candidate_param_states.push_back(best_param_state); // first candidate param state is current best
while (candidate_param_states.size() <= N) {
@ -207,7 +200,8 @@ namespace smt {
if (best_param_state_idx != 0) {
m_param_state = candidate_param_states[best_param_state_idx];
b.set_param_state(m_param_state);
auto p = apply_param_values(m_param_state);
b.set_param_state(p);
IF_VERBOSE(1, verbose_stream() << " PARAM TUNER found better param state at index " << best_param_state_idx << "\n");
} else {
IF_VERBOSE(1, verbose_stream() << " PARAM TUNER retained current param state\n");
@ -318,12 +312,12 @@ namespace smt {
}
parallel::param_generator::param_generator(parallel& p)
: p(p), b(p.m_batch_manager), m_param_state(p.ctx.get_fparams()), m_p(p.ctx.get_params()), m_l2g(m, p.ctx.m) {
ctx = alloc(context, m, m_param_state, m_p);
: p(p), b(p.m_batch_manager), m_p(p.ctx.get_params()), m_l2g(m, p.ctx.m) {
ctx = alloc(context, m, p.ctx.get_fparams(), m_p);
context::copy(p.ctx, *ctx, true);
for (unsigned i = 0; i < N; ++i) {
m_param_probe_contexts.push_back(alloc(context, m, m_param_state, m_p));
m_param_probe_contexts.push_back(alloc(context, m, ctx->get_fparams(), m_p));
}
// don't share initial units
@ -483,7 +477,8 @@ namespace smt {
}
}
smt_params parallel::batch_manager::get_best_param_state() {
// todo make this thread safe by not using reference counts implicit in params ref but instead copying the entire structure.
params_ref parallel::batch_manager::get_best_param_state() {
std::scoped_lock lock(mux);
return m_param_state;
}

26
src/smt/smt_parallel.h
View file

@ -81,7 +81,7 @@ namespace smt {
std::mutex mux;
state m_state = state::is_running;
stats m_stats;
smt_params m_param_state;
params_ref m_param_state;
using node = search_tree::node<cube_config>;
search_tree::tree<cube_config> m_search_tree;
@ -106,10 +106,10 @@ namespace smt {
void set_sat(ast_translation& l2g, model& m);
void set_exception(std::string const& msg);
void set_exception(unsigned error_code);
void set_param_state(smt_params const& p) { m_param_state = p; }
void set_param_state(params_ref const& p) { m_param_state.copy(p); }
void collect_statistics(::statistics& st) const;
smt_params get_best_param_state();
params_ref get_best_param_state();
bool get_cube(ast_translation& g2l, unsigned id, expr_ref_vector& cube, node*& n);
void backtrack(ast_translation& l2g, expr_ref_vector const& core, node* n);
void split(ast_translation& l2g, unsigned id, node* n, expr* atom);
@ -139,22 +139,32 @@ namespace smt {
scoped_ptr<context> m_prefix_solver;
scoped_ptr_vector<context> m_param_probe_contexts;
smt_params m_param_state;
params_ref m_p;
using param_value = std::variant<unsigned, bool, double>;
symbol_table<param_value> m_my_param_state;
struct unsigned_value {
unsigned value;
unsigned min_value;
unsigned max_value;
};
using param_value = std::variant<unsigned_value, bool>;
using param_values = vector<std::pair<symbol, param_value>>;
param_values m_param_state;
params_ref apply_param_values(param_values const &pv) {
return m_p;
}
// todo
private:
void init_param_state();
smt_params mutate_param_state();
param_values mutate_param_state();
public:
param_generator(parallel &p);
lbool run_prefix_step();
void protocol_iteration();
unsigned replay_proof_prefixes(vector<smt_params> candidate_param_states, unsigned max_conflicts_epsilon);
unsigned replay_proof_prefixes(vector<param_values> const& candidate_param_states, unsigned max_conflicts_epsilon);
reslimit &limit() {
return m.limit();