/*++ Copyright (c) 2016 Microsoft Corporation Module Name: model_based_opt.cpp Abstract: Model-based optimization and projection for linear real, integer arithmetic. Author: Nikolaj Bjorner (nbjorner) 2016-27-4 Revision History: --*/ #include "math/simplex/model_based_opt.h" #include "util/uint_set.h" #include "util/z3_exception.h" std::ostream& operator<<(std::ostream& out, opt::ineq_type ie) { switch (ie) { case opt::t_eq: return out << " = "; case opt::t_lt: return out << " < "; case opt::t_le: return out << " <= "; case opt::t_mod: return out << " mod "; } return out; } namespace opt { /** * Convert a row ax + coeffs + coeff = value into a definition for x * x = (value - coeffs - coeff)/a * as backdrop we have existing assignments to x and other variables that * satisfy the equality with value, and such that value satisfies * the row constraint ( = , <= , < , mod) */ model_based_opt::def::def(row const& r, unsigned x) { for (var const & v : r.m_vars) { if (v.m_id != x) { m_vars.push_back(v); } else { m_div = -v.m_coeff; } } m_coeff = r.m_coeff; switch (r.m_type) { case opt::t_lt: m_coeff += m_div; break; case opt::t_le: // for: ax >= t, then x := (t + a - 1) div a if (m_div.is_pos()) { m_coeff += m_div; m_coeff -= rational::one(); } break; default: break; } normalize(); SASSERT(m_div.is_pos()); } model_based_opt::def model_based_opt::def::operator+(def const& other) const { def result; vector const& vs1 = m_vars; vector const& vs2 = other.m_vars; vector & vs = result.m_vars; rational c1(1), c2(1); if (m_div != other.m_div) { c1 = other.m_div; c2 = m_div; } unsigned i = 0, j = 0; while (i < vs1.size() || j < vs2.size()) { unsigned v1 = UINT_MAX, v2 = UINT_MAX; if (i < vs1.size()) v1 = vs1[i].m_id; if (j < vs2.size()) v2 = vs2[j].m_id; if (v1 == v2) { vs.push_back(vs1[i]); vs.back().m_coeff *= c1; vs.back().m_coeff += c2 * vs2[j].m_coeff; ++i; ++j; if (vs.back().m_coeff.is_zero()) { vs.pop_back(); } } else if (v1 < v2) { vs.push_back(vs1[i]); vs.back().m_coeff *= c1; } else { vs.push_back(vs2[j]); vs.back().m_coeff *= c2; } } result.m_div = c1*m_div; result.m_coeff = (m_coeff*c1) + (other.m_coeff*c2); result.normalize(); return result; } model_based_opt::def model_based_opt::def::operator/(rational const& r) const { def result(*this); result.m_div *= r; result.normalize(); return result; } model_based_opt::def model_based_opt::def::operator*(rational const& n) const { def result(*this); for (var& v : result.m_vars) { v.m_coeff *= n; } result.m_coeff *= n; result.normalize(); return result; } model_based_opt::def model_based_opt::def::operator+(rational const& n) const { def result(*this); result.m_coeff += n * result.m_div; result.normalize(); return result; } void model_based_opt::def::normalize() { SASSERT(m_div.is_int()); if (m_div.is_neg()) { for (var& v : m_vars) v.m_coeff.neg(); m_coeff.neg(); m_div.neg(); } if (m_div.is_one()) return; rational g(m_div); if (!m_coeff.is_int()) return; g = gcd(g, m_coeff); for (var const& v : m_vars) { if (!v.m_coeff.is_int()) return; g = gcd(g, abs(v.m_coeff)); if (g.is_one()) break; } if (!g.is_one()) { for (var& v : m_vars) v.m_coeff /= g; m_coeff /= g; m_div /= g; } } model_based_opt::model_based_opt() { m_rows.push_back(row()); } bool model_based_opt::invariant() { for (unsigned i = 0; i < m_rows.size(); ++i) { if (!invariant(i, m_rows[i])) { return false; } } return true; } #define PASSERT(_e_) { CTRACE("qe", !(_e_), display(tout, r); display(tout);); SASSERT(_e_); } bool model_based_opt::invariant(unsigned index, row const& r) { vector const& vars = r.m_vars; for (unsigned i = 0; i < vars.size(); ++i) { // variables in each row are sorted and have non-zero coefficients PASSERT(i + 1 == vars.size() || vars[i].m_id < vars[i+1].m_id); PASSERT(!vars[i].m_coeff.is_zero()); PASSERT(index == 0 || m_var2row_ids[vars[i].m_id].contains(index)); } PASSERT(r.m_value == eval(r)); PASSERT(r.m_type != t_eq || r.m_value.is_zero()); // values satisfy constraints PASSERT(index == 0 || r.m_type != t_lt || r.m_value.is_neg()); PASSERT(index == 0 || r.m_type != t_le || !r.m_value.is_pos()); PASSERT(index == 0 || r.m_type != t_mod || (mod(r.m_value, r.m_mod).is_zero())); return true; } // a1*x + obj // a2*x + t2 <= 0 // a3*x + t3 <= 0 // a4*x + t4 <= 0 // a1 > 0, a2 > 0, a3 > 0, a4 < 0 // x <= -t2/a2 // x <= -t2/a3 // determine lub among these. // then resolve lub with others // e.g., -t2/a2 <= -t3/a3, then // replace inequality a3*x + t3 <= 0 by -t2/a2 + t3/a3 <= 0 // mark a4 as invalid. // // a1 < 0, a2 < 0, a3 < 0, a4 > 0 // x >= t2/a2 // x >= t3/a3 // determine glb among these // the resolve glb with others. // e.g. t2/a2 >= t3/a3 // then replace a3*x + t3 by t3/a3 - t2/a2 <= 0 // inf_eps model_based_opt::maximize() { SASSERT(invariant()); unsigned_vector bound_trail, bound_vars; TRACE("opt", display(tout << "tableau\n");); while (!objective().m_vars.empty()) { var v = objective().m_vars.back(); unsigned x = v.m_id; rational const& coeff = v.m_coeff; unsigned bound_row_index; rational bound_coeff; if (find_bound(x, bound_row_index, bound_coeff, coeff.is_pos())) { SASSERT(!bound_coeff.is_zero()); TRACE("opt", display(tout << "update: " << v << " ", objective()); for (unsigned above : m_above) { display(tout << "resolve: ", m_rows[above]); }); for (unsigned above : m_above) { resolve(bound_row_index, bound_coeff, above, x); } for (unsigned below : m_below) { resolve(bound_row_index, bound_coeff, below, x); } // coeff*x + objective <= ub // a2*x + t2 <= 0 // => coeff*x <= -t2*coeff/a2 // objective + t2*coeff/a2 <= ub mul_add(false, m_objective_id, - coeff/bound_coeff, bound_row_index); retire_row(bound_row_index); bound_trail.push_back(bound_row_index); bound_vars.push_back(x); } else { TRACE("opt", display(tout << "unbound: " << v << " ", objective());); update_values(bound_vars, bound_trail); return inf_eps::infinity(); } } // // update the evaluation of variables to satisfy the bound. // update_values(bound_vars, bound_trail); rational value = objective().m_value; if (objective().m_type == t_lt) { return inf_eps(inf_rational(value, rational(-1))); } else { return inf_eps(inf_rational(value)); } } void model_based_opt::update_value(unsigned x, rational const& val) { rational old_val = m_var2value[x]; m_var2value[x] = val; SASSERT(val.is_int() || !is_int(x)); unsigned_vector const& row_ids = m_var2row_ids[x]; for (unsigned row_id : row_ids) { rational coeff = get_coefficient(row_id, x); if (coeff.is_zero()) { continue; } row & r = m_rows[row_id]; rational delta = coeff * (val - old_val); r.m_value += delta; SASSERT(invariant(row_id, r)); } } void model_based_opt::update_values(unsigned_vector const& bound_vars, unsigned_vector const& bound_trail) { for (unsigned i = bound_trail.size(); i-- > 0; ) { unsigned x = bound_vars[i]; row& r = m_rows[bound_trail[i]]; rational val = r.m_coeff; rational old_x_val = m_var2value[x]; rational new_x_val; rational x_coeff, eps(0); vector const& vars = r.m_vars; for (var const& v : vars) { if (x == v.m_id) { x_coeff = v.m_coeff; } else { val += m_var2value[v.m_id]*v.m_coeff; } } SASSERT(!x_coeff.is_zero()); new_x_val = -val/x_coeff; if (r.m_type == t_lt) { eps = abs(old_x_val - new_x_val)/rational(2); eps = std::min(rational::one(), eps); SASSERT(!eps.is_zero()); // // ax + t < 0 // <=> x < -t/a // <=> x := -t/a - epsilon // if (x_coeff.is_pos()) { new_x_val -= eps; } // // -ax + t < 0 // <=> -ax < -t // <=> -x < -t/a // <=> x > t/a // <=> x := t/a + epsilon // else { new_x_val += eps; } } TRACE("opt", display(tout << "v" << x << " coeff_x: " << x_coeff << " old_x_val: " << old_x_val << " new_x_val: " << new_x_val << " eps: " << eps << " ", r); ); m_var2value[x] = new_x_val; r.m_value = eval(r); SASSERT(invariant(bound_trail[i], r)); } // update and check bounds for all other affected rows. for (unsigned i = bound_trail.size(); i-- > 0; ) { unsigned x = bound_vars[i]; unsigned_vector const& row_ids = m_var2row_ids[x]; for (unsigned row_id : row_ids) { row & r = m_rows[row_id]; r.m_value = eval(r); SASSERT(invariant(row_id, r)); } } SASSERT(invariant()); } bool model_based_opt::find_bound(unsigned x, unsigned& bound_row_index, rational& bound_coeff, bool is_pos) { bound_row_index = UINT_MAX; rational lub_val; rational const& x_val = m_var2value[x]; unsigned_vector const& row_ids = m_var2row_ids[x]; uint_set visited; m_above.reset(); m_below.reset(); for (unsigned row_id : row_ids) { SASSERT(row_id != m_objective_id); if (visited.contains(row_id)) { continue; } visited.insert(row_id); row& r = m_rows[row_id]; if (r.m_alive) { rational a = get_coefficient(row_id, x); if (a.is_zero()) { // skip } else if (a.is_pos() == is_pos || r.m_type == t_eq) { rational value = x_val - (r.m_value/a); if (bound_row_index == UINT_MAX) { lub_val = value; bound_row_index = row_id; bound_coeff = a; } else if ((value == lub_val && r.m_type == opt::t_lt) || (is_pos && value < lub_val) || (!is_pos && value > lub_val)) { m_above.push_back(bound_row_index); lub_val = value; bound_row_index = row_id; bound_coeff = a; } else { m_above.push_back(row_id); } } else { m_below.push_back(row_id); } } } return bound_row_index != UINT_MAX; } void model_based_opt::retire_row(unsigned row_id) { m_rows[row_id].m_alive = false; m_retired_rows.push_back(row_id); } rational model_based_opt::eval(unsigned x) const { return m_var2value[x]; } rational model_based_opt::eval(def const& d) const { vector const& vars = d.m_vars; rational val = d.m_coeff; for (var const& v : vars) { val += v.m_coeff * eval(v.m_id); } val /= d.m_div; return val; } rational model_based_opt::eval(row const& r) const { vector const& vars = r.m_vars; rational val = r.m_coeff; for (var const& v : vars) { val += v.m_coeff * eval(v.m_id); } return val; } rational model_based_opt::eval(vector const& coeffs) const { rational val(0); for (var const& v : coeffs) val += v.m_coeff * eval(v.m_id); return val; } rational model_based_opt::get_coefficient(unsigned row_id, unsigned var_id) const { return m_rows[row_id].get_coefficient(var_id); } rational model_based_opt::row::get_coefficient(unsigned var_id) const { if (m_vars.empty()) { return rational::zero(); } unsigned lo = 0, hi = m_vars.size(); while (lo < hi) { unsigned mid = lo + (hi - lo)/2; SASSERT(mid < hi); unsigned id = m_vars[mid].m_id; if (id == var_id) { lo = mid; break; } if (id < var_id) { lo = mid + 1; } else { hi = mid; } } if (lo == m_vars.size()) { return rational::zero(); } unsigned id = m_vars[lo].m_id; if (id == var_id) { return m_vars[lo].m_coeff; } else { return rational::zero(); } } // // Let // row1: t1 + a1*x <= 0 // row2: t2 + a2*x <= 0 // // assume a1, a2 have the same signs: // (t2 + a2*x) <= (t1 + a1*x)*a2/a1 // <=> t2*a1/a2 - t1 <= 0 // <=> t2 - t1*a2/a1 <= 0 // // assume a1 > 0, -a2 < 0: // t1 + a1*x <= 0, t2 - a2*x <= 0 // t2/a2 <= -t1/a1 // t2 + t1*a2/a1 <= 0 // assume -a1 < 0, a2 > 0: // t1 - a1*x <= 0, t2 + a2*x <= 0 // t1/a1 <= -t2/a2 // t2 + t1*a2/a1 <= 0 // // the resolvent is the same in all cases (simpler proof should exist) // void model_based_opt::resolve(unsigned row_src, rational const& a1, unsigned row_dst, unsigned x) { SASSERT(a1 == get_coefficient(row_src, x)); SASSERT(!a1.is_zero()); SASSERT(row_src != row_dst); if (m_rows[row_dst].m_alive) { rational a2 = get_coefficient(row_dst, x); if (is_int(x)) { TRACE("opt", tout << a1 << " " << a2 << ": "; display(tout, m_rows[row_dst]); display(tout, m_rows[row_src]);); if (a1.is_pos() != a2.is_pos() || m_rows[row_src].m_type == opt::t_eq) { mul_add(x, a1, row_src, a2, row_dst); } else { mul(row_dst, abs(a1)); mul_add(false, row_dst, -abs(a2), row_src); } TRACE("opt", display(tout, m_rows[row_dst]);); normalize(row_dst); } else { mul_add(row_dst != m_objective_id && a1.is_pos() == a2.is_pos(), row_dst, -a2/a1, row_src); } } } void model_based_opt::solve(unsigned row_src, rational const& a1, unsigned row_dst, unsigned x) { SASSERT(a1 == get_coefficient(row_src, x)); SASSERT(a1.is_pos()); SASSERT(row_src != row_dst); if (!m_rows[row_dst].m_alive) return; rational a2 = get_coefficient(row_dst, x); mul(row_dst, a1); mul_add(false, row_dst, -a2, row_src); SASSERT(get_coefficient(row_dst, x).is_zero()); } // resolution for integer rows. void model_based_opt::mul_add( unsigned x, rational const& src_c, unsigned row_src, rational const& dst_c, unsigned row_dst) { row& dst = m_rows[row_dst]; row const& src = m_rows[row_src]; SASSERT(is_int(x)); SASSERT(t_le == dst.m_type && t_le == src.m_type); SASSERT(src_c.is_int()); SASSERT(dst_c.is_int()); SASSERT(m_var2value[x].is_int()); rational abs_src_c = abs(src_c); rational abs_dst_c = abs(dst_c); rational x_val = m_var2value[x]; rational slack = (abs_src_c - rational::one()) * (abs_dst_c - rational::one()); rational dst_val = dst.m_value - x_val*dst_c; rational src_val = src.m_value - x_val*src_c; rational distance = abs_src_c * dst_val + abs_dst_c * src_val + slack; bool use_case1 = distance.is_nonpos() || abs_src_c.is_one() || abs_dst_c.is_one(); #if 0 if (distance.is_nonpos() && !abs_src_c.is_one() && !abs_dst_c.is_one()) { unsigned r = copy_row(row_src); mul_add(false, r, rational::one(), row_dst); del_var(r, x); add(r, slack); TRACE("qe", tout << m_rows[r];); SASSERT(!m_rows[r].m_value.is_pos()); } #endif if (use_case1) { TRACE("opt", tout << "slack: " << slack << " " << src_c << " " << dst_val << " " << dst_c << " " << src_val << "\n";); // dst <- abs_src_c*dst + abs_dst_c*src + slack mul(row_dst, abs_src_c); add(row_dst, slack); mul_add(false, row_dst, abs_dst_c, row_src); return; } // // create finite disjunction for |b|. // exists x, z in [0 .. |b|-2] . b*x + s + z = 0 && ax + t <= 0 && bx + s <= 0 // <=> // exists x, z in [0 .. |b|-2] . b*x = -z - s && ax + t <= 0 && bx + s <= 0 // <=> // exists x, z in [0 .. |b|-2] . b*x = -z - s && a|b|x + |b|t <= 0 && bx + s <= 0 // <=> // exists x, z in [0 .. |b|-2] . b*x = -z - s && a|b|x + |b|t <= 0 && -z - s + s <= 0 // <=> // exists x, z in [0 .. |b|-2] . b*x = -z - s && a|b|x + |b|t <= 0 && -z <= 0 // <=> // exists x, z in [0 .. |b|-2] . b*x = -z - s && a|b|x + |b|t <= 0 // <=> // exists x, z in [0 .. |b|-2] . b*x = -z - s && a*n_sign(b)(s + z) + |b|t <= 0 // <=> // exists z in [0 .. |b|-2] . |b| | (z + s) && a*n_sign(b)(s + z) + |b|t <= 0 // TRACE("qe", tout << "finite disjunction " << distance << " " << src_c << " " << dst_c << "\n";); vector coeffs; if (abs_dst_c <= abs_src_c) { rational z = mod(dst_val, abs_dst_c); if (!z.is_zero()) z = abs_dst_c - z; mk_coeffs_without(coeffs, dst.m_vars, x); add_divides(coeffs, dst.m_coeff + z, abs_dst_c); add(row_dst, z); mul(row_dst, src_c * n_sign(dst_c)); mul_add(false, row_dst, abs_dst_c, row_src); } else { // z := b - (s + bx) mod b // := b - s mod b // b | s + z <=> b | s + b - s mod b <=> b | s - s mod b rational z = mod(src_val, abs_src_c); if (!z.is_zero()) z = abs_src_c - z; mk_coeffs_without(coeffs, src.m_vars, x); add_divides(coeffs, src.m_coeff + z, abs_src_c); mul(row_dst, abs_src_c); add(row_dst, z * dst_c * n_sign(src_c)); mul_add(false, row_dst, dst_c * n_sign(src_c), row_src); } } void model_based_opt::mk_coeffs_without(vector& dst, vector const& src, unsigned x) { for (var const & v : src) { if (v.m_id != x) dst.push_back(v); } } rational model_based_opt::n_sign(rational const& b) const { return rational(b.is_pos()?-1:1); } void model_based_opt::mul(unsigned dst, rational const& c) { if (c.is_one()) return; row& r = m_rows[dst]; for (auto & v : r.m_vars) { v.m_coeff *= c; } r.m_coeff *= c; r.m_value *= c; } void model_based_opt::add(unsigned dst, rational const& c) { row& r = m_rows[dst]; r.m_coeff += c; r.m_value += c; } void model_based_opt::sub(unsigned dst, rational const& c) { row& r = m_rows[dst]; r.m_coeff -= c; r.m_value -= c; } void model_based_opt::del_var(unsigned dst, unsigned x) { row& r = m_rows[dst]; unsigned j = 0; for (var & v : r.m_vars) { if (v.m_id == x) { r.m_value -= eval(x)*r.m_coeff; } else { r.m_vars[j++] = v; } } r.m_vars.shrink(j); } void model_based_opt::normalize(unsigned row_id) { row& r = m_rows[row_id]; if (r.m_vars.empty()) { retire_row(row_id); return; } if (r.m_type == t_mod) return; rational g(abs(r.m_vars[0].m_coeff)); bool all_int = g.is_int(); for (unsigned i = 1; all_int && !g.is_one() && i < r.m_vars.size(); ++i) { rational const& coeff = r.m_vars[i].m_coeff; if (coeff.is_int()) { g = gcd(g, abs(coeff)); } else { all_int = false; } } if (all_int && !r.m_coeff.is_zero()) { if (r.m_coeff.is_int()) { g = gcd(g, abs(r.m_coeff)); } else { all_int = false; } } if (all_int && !g.is_one()) { SASSERT(!g.is_zero()); mul(row_id, rational::one()/g); } } // // set row1 <- row1 + c*row2 // void model_based_opt::mul_add(bool same_sign, unsigned row_id1, rational const& c, unsigned row_id2) { if (c.is_zero()) { return; } m_new_vars.reset(); row& r1 = m_rows[row_id1]; row const& r2 = m_rows[row_id2]; unsigned i = 0, j = 0; while (i < r1.m_vars.size() || j < r2.m_vars.size()) { if (j == r2.m_vars.size()) { m_new_vars.append(r1.m_vars.size() - i, r1.m_vars.data() + i); break; } if (i == r1.m_vars.size()) { for (; j < r2.m_vars.size(); ++j) { m_new_vars.push_back(r2.m_vars[j]); m_new_vars.back().m_coeff *= c; if (row_id1 != m_objective_id) { m_var2row_ids[r2.m_vars[j].m_id].push_back(row_id1); } } break; } unsigned v1 = r1.m_vars[i].m_id; unsigned v2 = r2.m_vars[j].m_id; if (v1 == v2) { m_new_vars.push_back(r1.m_vars[i]); m_new_vars.back().m_coeff += c*r2.m_vars[j].m_coeff; ++i; ++j; if (m_new_vars.back().m_coeff.is_zero()) { m_new_vars.pop_back(); } } else if (v1 < v2) { m_new_vars.push_back(r1.m_vars[i]); ++i; } else { m_new_vars.push_back(r2.m_vars[j]); m_new_vars.back().m_coeff *= c; if (row_id1 != m_objective_id) { m_var2row_ids[r2.m_vars[j].m_id].push_back(row_id1); } ++j; } } r1.m_coeff += c*r2.m_coeff; r1.m_vars.swap(m_new_vars); r1.m_value += c*r2.m_value; if (!same_sign && r2.m_type == t_lt) { r1.m_type = t_lt; } else if (same_sign && r1.m_type == t_lt && r2.m_type == t_lt) { r1.m_type = t_le; } SASSERT(invariant(row_id1, r1)); } void model_based_opt::display(std::ostream& out) const { for (auto const& r : m_rows) { display(out, r); } for (unsigned i = 0; i < m_var2row_ids.size(); ++i) { unsigned_vector const& rows = m_var2row_ids[i]; out << i << ": "; for (auto const& r : rows) { out << r << " "; } out << "\n"; } } void model_based_opt::display(std::ostream& out, vector const& vars, rational const& coeff) { unsigned i = 0; for (var const& v : vars) { if (i > 0 && v.m_coeff.is_pos()) { out << "+ "; } ++i; if (v.m_coeff.is_one()) { out << "v" << v.m_id << " "; } else { out << v.m_coeff << "*v" << v.m_id << " "; } } if (coeff.is_pos()) { out << " + " << coeff << " "; } else if (coeff.is_neg()) { out << coeff << " "; } } std::ostream& model_based_opt::display(std::ostream& out, row const& r) { out << (r.m_alive?"a":"d") << " "; display(out, r.m_vars, r.m_coeff); if (r.m_type == opt::t_mod) { out << r.m_type << " " << r.m_mod << " = 0; value: " << r.m_value << "\n"; } else { out << r.m_type << " 0; value: " << r.m_value << "\n"; } return out; } std::ostream& model_based_opt::display(std::ostream& out, def const& r) { display(out, r.m_vars, r.m_coeff); if (!r.m_div.is_one()) { out << " / " << r.m_div; } return out; } unsigned model_based_opt::add_var(rational const& value, bool is_int) { unsigned v = m_var2value.size(); m_var2value.push_back(value); m_var2is_int.push_back(is_int); SASSERT(value.is_int() || !is_int); m_var2row_ids.push_back(unsigned_vector()); return v; } rational model_based_opt::get_value(unsigned var) { return m_var2value[var]; } void model_based_opt::set_row(unsigned row_id, vector const& coeffs, rational const& c, rational const& m, ineq_type rel) { row& r = m_rows[row_id]; rational val(c); SASSERT(r.m_vars.empty()); r.m_vars.append(coeffs.size(), coeffs.data()); bool is_int_row = !coeffs.empty(); std::sort(r.m_vars.begin(), r.m_vars.end(), var::compare()); for (auto const& c : coeffs) { val += m_var2value[c.m_id] * c.m_coeff; SASSERT(!is_int(c.m_id) || c.m_coeff.is_int()); is_int_row &= is_int(c.m_id); } r.m_alive = true; r.m_coeff = c; r.m_value = val; r.m_type = rel; r.m_mod = m; if (is_int_row && rel == t_lt) { r.m_type = t_le; r.m_coeff += rational::one(); r.m_value += rational::one(); } } unsigned model_based_opt::new_row() { unsigned row_id = 0; if (m_retired_rows.empty()) { row_id = m_rows.size(); m_rows.push_back(row()); } else { row_id = m_retired_rows.back(); m_retired_rows.pop_back(); m_rows[row_id].reset(); m_rows[row_id].m_alive = true; } return row_id; } unsigned model_based_opt::copy_row(unsigned src) { unsigned dst = new_row(); row const& r = m_rows[src]; set_row(dst, r.m_vars, r.m_coeff, r.m_mod, r.m_type); for (auto const& v : r.m_vars) { m_var2row_ids[v.m_id].push_back(dst); } SASSERT(invariant(dst, m_rows[dst])); return dst; } void model_based_opt::add_constraint(vector const& coeffs, rational const& c, ineq_type rel) { add_constraint(coeffs, c, rational::zero(), rel); } void model_based_opt::add_divides(vector const& coeffs, rational const& c, rational const& m) { add_constraint(coeffs, c, m, t_mod); } void model_based_opt::add_constraint(vector const& coeffs, rational const& c, rational const& m, ineq_type rel) { unsigned row_id = new_row(); set_row(row_id, coeffs, c, m, rel); for (var const& coeff : coeffs) { m_var2row_ids[coeff.m_id].push_back(row_id); } SASSERT(invariant(row_id, m_rows[row_id])); } void model_based_opt::set_objective(vector const& coeffs, rational const& c) { set_row(m_objective_id, coeffs, c, rational::zero(), t_le); } void model_based_opt::get_live_rows(vector& rows) { for (row const& r : m_rows) { if (r.m_alive) { rows.push_back(r); } } } // // pick glb and lub representative. // The representative is picked such that it // represents the fewest inequalities. // The constraints that enforce a glb or lub are not forced. // The constraints that separate the glb from ub or the lub from lb // are not forced. // In other words, suppose there are // . N inequalities of the form t <= x // . M inequalities of the form s >= x // . t0 is glb among N under valuation. // . s0 is lub among M under valuation. // If N < M // create the inequalities: // t <= t0 for each t other than t0 (N-1 inequalities). // t0 <= s for each s (M inequalities). // If N >= M the construction is symmetric. // model_based_opt::def model_based_opt::project(unsigned x, bool compute_def) { unsigned_vector& lub_rows = m_lub; unsigned_vector& glb_rows = m_glb; unsigned_vector& mod_rows = m_mod; unsigned lub_index = UINT_MAX, glb_index = UINT_MAX; bool lub_strict = false, glb_strict = false; rational lub_val, glb_val; rational const& x_val = m_var2value[x]; unsigned_vector const& row_ids = m_var2row_ids[x]; uint_set visited; lub_rows.reset(); glb_rows.reset(); mod_rows.reset(); bool lub_is_unit = false, glb_is_unit = false; unsigned eq_row = UINT_MAX; // select the lub and glb. for (unsigned row_id : row_ids) { if (visited.contains(row_id)) { continue; } visited.insert(row_id); row& r = m_rows[row_id]; if (!r.m_alive) { continue; } rational a = get_coefficient(row_id, x); if (a.is_zero()) { continue; } if (r.m_type == t_eq) { eq_row = row_id; continue; } if (r.m_type == t_mod) { mod_rows.push_back(row_id); } else if (a.is_pos()) { rational lub_value = x_val - (r.m_value/a); if (lub_rows.empty() || lub_value < lub_val || (lub_value == lub_val && r.m_type == t_lt && !lub_strict)) { lub_val = lub_value; lub_index = row_id; lub_strict = r.m_type == t_lt; } lub_rows.push_back(row_id); lub_is_unit &= a.is_one(); } else { SASSERT(a.is_neg()); rational glb_value = x_val - (r.m_value/a); if (glb_rows.empty() || glb_value > glb_val || (glb_value == glb_val && r.m_type == t_lt && !glb_strict)) { glb_val = glb_value; glb_index = row_id; glb_strict = r.m_type == t_lt; } glb_rows.push_back(row_id); glb_is_unit &= a.is_minus_one(); } } if (!mod_rows.empty()) { return solve_mod(x, mod_rows, compute_def); } if (eq_row != UINT_MAX) { return solve_for(eq_row, x, compute_def); } def result; unsigned lub_size = lub_rows.size(); unsigned glb_size = glb_rows.size(); unsigned row_index = (lub_size <= glb_size) ? lub_index : glb_index; // There are only upper or only lower bounds. if (row_index == UINT_MAX) { if (compute_def) { if (lub_index != UINT_MAX) { result = solve_for(lub_index, x, true); } else if (glb_index != UINT_MAX) { result = solve_for(glb_index, x, true); } else { result = def() + m_var2value[x]; } SASSERT(eval(result) == eval(x)); } else { for (unsigned row_id : lub_rows) retire_row(row_id); for (unsigned row_id : glb_rows) retire_row(row_id); } return result; } SASSERT(lub_index != UINT_MAX); SASSERT(glb_index != UINT_MAX); if (compute_def) { if (lub_size <= glb_size) { result = def(m_rows[lub_index], x); } else { result = def(m_rows[glb_index], x); } } // The number of matching lower and upper bounds is small. if ((lub_size <= 2 || glb_size <= 2) && (lub_size <= 3 && glb_size <= 3) && (!is_int(x) || lub_is_unit || glb_is_unit)) { for (unsigned i = 0; i < lub_size; ++i) { unsigned row_id1 = lub_rows[i]; bool last = i + 1 == lub_size; rational coeff = get_coefficient(row_id1, x); for (unsigned row_id2 : glb_rows) { if (last) { resolve(row_id1, coeff, row_id2, x); } else { unsigned row_id3 = copy_row(row_id2); resolve(row_id1, coeff, row_id3, x); } } } for (unsigned row_id : lub_rows) retire_row(row_id); return result; } // General case. rational coeff = get_coefficient(row_index, x); for (unsigned row_id : lub_rows) { if (row_id != row_index) { resolve(row_index, coeff, row_id, x); } } for (unsigned row_id : glb_rows) { if (row_id != row_index) { resolve(row_index, coeff, row_id, x); } } retire_row(row_index); return result; } // // compute D and u. // // D = lcm(d1, d2) // u = eval(x) mod D // // d1 | (a1x + t1) & d2 | (a2x + t2) // = // d1 | (a1(D*x' + u) + t1) & d2 | (a2(D*x' + u) + t2) // = // d1 | (a1*u + t1) & d2 | (a2*u + t2) // // x := D*x' + u // model_based_opt::def model_based_opt::solve_mod(unsigned x, unsigned_vector const& mod_rows, bool compute_def) { SASSERT(!mod_rows.empty()); rational D(1); for (unsigned idx : mod_rows) { D = lcm(D, m_rows[idx].m_mod); } if (D.is_zero()) { throw default_exception("modulo 0 is not defined"); } if (D.is_neg()) D = abs(D); TRACE("opt1", display(tout << "lcm: " << D << " x: v" << x << " tableau\n");); rational val_x = m_var2value[x]; rational u = mod(val_x, D); SASSERT(u.is_nonneg() && u < D); for (unsigned idx : mod_rows) { replace_var(idx, x, u); SASSERT(invariant(idx, m_rows[idx])); normalize(idx); } TRACE("opt1", display(tout << "tableau after replace x under mod\n");); // // update inequalities such that u is added to t and // D is multiplied to coefficient of x. // the interpretation of the new version of x is (x-u)/D // // a*x + t <= 0 // a*(D*x' + u) + t <= 0 // a*D*x' + a*u + t <= 0 // rational new_val = (val_x - u) / D; SASSERT(new_val.is_int()); unsigned y = add_var(new_val, true); unsigned_vector const& row_ids = m_var2row_ids[x]; uint_set visited; for (unsigned row_id : row_ids) { if (!visited.contains(row_id)) { // x |-> D*y + u replace_var(row_id, x, D, y, u); visited.insert(row_id); normalize(row_id); } } TRACE("opt1", display(tout << "tableau after replace x by y := v" << y << "\n");); def result = project(y, compute_def); if (compute_def) { result = (result * D) + u; m_var2value[x] = eval(result); } TRACE("opt1", display(tout << "tableau after project y" << y << "\n");); return result; } // update row with: x |-> C void model_based_opt::replace_var(unsigned row_id, unsigned x, rational const& C) { row& r = m_rows[row_id]; SASSERT(!get_coefficient(row_id, x).is_zero()); unsigned sz = r.m_vars.size(); unsigned i = 0, j = 0; rational coeff(0); for (; i < sz; ++i) { if (r.m_vars[i].m_id == x) { coeff = r.m_vars[i].m_coeff; } else { if (i != j) { r.m_vars[j] = r.m_vars[i]; } ++j; } } if (j != sz) { r.m_vars.shrink(j); } r.m_coeff += coeff*C; r.m_value += coeff*(C - m_var2value[x]); } // update row with: x |-> A*y + B void model_based_opt::replace_var(unsigned row_id, unsigned x, rational const& A, unsigned y, rational const& B) { row& r = m_rows[row_id]; rational coeff = get_coefficient(row_id, x); if (coeff.is_zero()) return; if (!r.m_alive) return; replace_var(row_id, x, B); r.m_vars.push_back(var(y, coeff*A)); r.m_value += coeff*A*m_var2value[y]; if (!r.m_vars.empty() && r.m_vars.back().m_id > y) { std::sort(r.m_vars.begin(), r.m_vars.end(), var::compare()); } m_var2row_ids[y].push_back(row_id); SASSERT(invariant(row_id, r)); } // 3x + t = 0 & 7 | (c*x + s) & ax <= u // 3 | -t & 21 | (-ct + 3s) & a-t <= 3u model_based_opt::def model_based_opt::solve_for(unsigned row_id1, unsigned x, bool compute_def) { TRACE("opt", tout << "v" << x << " := " << eval(x) << "\n" << m_rows[row_id1] << "\n";); rational a = get_coefficient(row_id1, x), b; ineq_type ty = m_rows[row_id1].m_type; SASSERT(!a.is_zero()); SASSERT(m_rows[row_id1].m_alive); if (a.is_neg()) { a.neg(); m_rows[row_id1].neg(); } SASSERT(a.is_pos()); if (ty == t_lt) { SASSERT(compute_def); m_rows[row_id1].m_coeff += a; m_rows[row_id1].m_type = t_le; m_rows[row_id1].m_value += a; } if (m_var2is_int[x] && !a.is_one()) { row& r1 = m_rows[row_id1]; vector coeffs; mk_coeffs_without(coeffs, r1.m_vars, x); rational c = mod(-eval(coeffs), a); add_divides(coeffs, c, a); } unsigned_vector const& row_ids = m_var2row_ids[x]; uint_set visited; visited.insert(row_id1); for (unsigned row_id2 : row_ids) { if (!visited.contains(row_id2)) { visited.insert(row_id2); b = get_coefficient(row_id2, x); if (b.is_zero()) continue; row& dst = m_rows[row_id2]; switch (dst.m_type) { case t_eq: case t_lt: case t_le: solve(row_id1, a, row_id2, x); break; case t_mod: // mod reduction already done. UNREACHABLE(); break; } } } def result; if (compute_def) { result = def(m_rows[row_id1], x); m_var2value[x] = eval(result); TRACE("opt1", tout << "updated eval " << x << " := " << eval(x) << "\n";); } retire_row(row_id1); return result; } vector model_based_opt::project(unsigned num_vars, unsigned const* vars, bool compute_def) { vector result; for (unsigned i = 0; i < num_vars; ++i) { result.push_back(project(vars[i], compute_def)); TRACE("opt", display(tout << "After projecting: v" << vars[i] << "\n");); } return result; } }