3
0
Fork 0
mirror of https://github.com/Z3Prover/z3 synced 2025-04-13 04:28:17 +00:00
z3/src/util/lp/sparse_matrix.hpp
Nikolaj Bjorner 23ff580a67 get rid of timeb dependencies, pull request #1040
Signed-off-by: Nikolaj Bjorner <nbjorner@microsoft.com>
2017-05-23 12:16:43 -07:00

1257 lines
44 KiB
C++

/*
Copyright (c) 2017 Microsoft Corporation
Author: Lev Nachmanson
*/
#include "util/vector.h"
#include "util/lp/sparse_matrix.h"
#include <set>
#include <queue>
namespace lean {
template <typename T, typename X>
void sparse_matrix<T, X>::copy_column_from_static_matrix(unsigned col, static_matrix<T, X> const &A, unsigned col_index_in_the_new_matrix) {
vector<column_cell> const & A_col_vector = A.m_columns[col];
unsigned size = static_cast<unsigned>(A_col_vector.size());
vector<indexed_value<T>> & new_column_vector = m_columns[col_index_in_the_new_matrix].m_values;
for (unsigned l = 0; l < size; l++) {
column_cell const & col_cell = A_col_vector[l];
unsigned col_offset = static_cast<unsigned>(new_column_vector.size());
vector<indexed_value<T>> & row_vector = m_rows[col_cell.m_i];
unsigned row_offset = static_cast<unsigned>(row_vector.size());
const T & val = A.get_val(col_cell);
new_column_vector.push_back(indexed_value<T>(val, col_cell.m_i, row_offset));
row_vector.push_back(indexed_value<T>(val, col_index_in_the_new_matrix, col_offset));
m_n_of_active_elems++;
}
}
template <typename T, typename X>
void sparse_matrix<T, X>::copy_B(static_matrix<T, X> const &A, vector<unsigned> & basis) {
unsigned m = A.row_count(); // this should be the size of basis
for (unsigned j = m; j-- > 0;) {
copy_column_from_static_matrix(basis[j], A, j);
}
}
// constructor that copies columns of the basis from A
template <typename T, typename X>
sparse_matrix<T, X>::sparse_matrix(static_matrix<T, X> const &A, vector<unsigned> & basis) :
m_n_of_active_elems(0),
m_pivot_queue(A.row_count()),
m_row_permutation(A.row_count()),
m_column_permutation(A.row_count()),
m_work_pivot_vector(A.row_count(), -1),
m_processed(A.row_count()) {
init_row_headers();
init_column_headers();
copy_B(A, basis);
}
template <typename T, typename X>
void sparse_matrix<T, X>::set_with_no_adjusting_for_row(unsigned row, unsigned col, T val) { // should not be used in efficient code
vector<indexed_value<T>> & row_vec = m_rows[row];
for (auto & iv : row_vec) {
if (iv.m_index == col) {
iv.set_value(val);
return;
}
}
// have not found the column between the indices
row_vec.push_back(indexed_value<T>(val, col)); // what about m_other ???
}
template <typename T, typename X>
void sparse_matrix<T, X>::set_with_no_adjusting_for_col(unsigned row, unsigned col, T val) { // should not be used in efficient code
vector<indexed_value<T>> & col_vec = m_columns[col].m_values;
for (auto & iv : col_vec) {
if (iv.m_index == row) {
iv.set_value(val);
return;
}
}
// have not found the column between the indices
col_vec.push_back(indexed_value<T>(val, row)); // what about m_other ???
}
template <typename T, typename X>
void sparse_matrix<T, X>::set_with_no_adjusting(unsigned row, unsigned col, T val) { // should not be used in efficient code
set_with_no_adjusting_for_row(row, col, val);
set_with_no_adjusting_for_col(row, col, val);
}
template <typename T, typename X>
void sparse_matrix<T, X>::set(unsigned row, unsigned col, T val) { // should not be used in efficient code
lean_assert(row < dimension() && col < dimension());
// m_dense.set_elem(row, col, val);
row = adjust_row(row);
col = adjust_column(col);
set_with_no_adjusting(row, col, val);
// lean_assert(*this == m_dense);
}
template <typename T, typename X>
T const & sparse_matrix<T, X>::get_not_adjusted(unsigned row, unsigned col) const {
for (indexed_value<T> const & iv : m_rows[row]) {
if (iv.m_index == col) {
return iv.m_value;
}
}
return numeric_traits<T>::zero();
}
template <typename T, typename X>
T const & sparse_matrix<T, X>::get(unsigned row, unsigned col) const { // should not be used in efficient code
row = adjust_row(row);
auto & row_chunk = m_rows[row];
col = adjust_column(col);
for (indexed_value<T> const & iv : row_chunk) {
if (iv.m_index == col) {
return iv.m_value;
}
}
return numeric_traits<T>::zero();
}
// constructor creating a zero matrix of dim*dim
template <typename T, typename X>
sparse_matrix<T, X>::sparse_matrix(unsigned dim) :
m_pivot_queue(dim), // dim will be the initial size of the queue
m_row_permutation(dim),
m_column_permutation(dim),
m_work_pivot_vector(dim, -1),
m_processed(dim) {
init_row_headers();
init_column_headers();
}
template <typename T, typename X>
void sparse_matrix<T, X>::init_row_headers() {
for (unsigned l = 0; l < m_row_permutation.size(); l++) {
m_rows.push_back(vector<indexed_value<T>>());
}
}
template <typename T, typename X>
void sparse_matrix<T, X>::init_column_headers() { // we alway have only square sparse_matrix
for (unsigned l = 0; l < m_row_permutation.size(); l++) {
m_columns.push_back(col_header());
}
}
template <typename T, typename X>
unsigned sparse_matrix<T, X>::lowest_row_in_column(unsigned j) {
auto & mc = get_column_values(adjust_column(j));
unsigned ret = 0;
for (auto & iv : mc) {
unsigned row = adjust_row_inverse(iv.m_index);
if (row > ret) {
ret = row;
}
}
return ret;
}
template <typename T, typename X>
void sparse_matrix<T, X>::remove_element(vector<indexed_value<T>> & row_vals, unsigned row_offset, vector<indexed_value<T>> & column_vals, unsigned column_offset) {
if (column_offset != column_vals.size() - 1) {
auto & column_iv = column_vals[column_offset] = column_vals.back(); // copy from the tail
column_iv_other(column_iv).m_other = column_offset;
if (row_offset != row_vals.size() - 1) {
auto & row_iv = row_vals[row_offset] = row_vals.back(); // copy from the tail
row_iv_other(row_iv).m_other = row_offset;
}
} else if (row_offset != row_vals.size() - 1) {
auto & row_iv = row_vals[row_offset] = row_vals.back(); // copy from the tail
row_iv_other(row_iv).m_other = row_offset;
}
// do nothing - just decrease the sizes
column_vals.pop_back();
row_vals.pop_back();
m_n_of_active_elems--; // the value is correct only when refactoring
}
template <typename T, typename X>
void sparse_matrix<T, X>::remove_element(vector<indexed_value<T>> & row_chunk, indexed_value<T> & row_el_iv) {
auto & column_chunk = get_column_values(row_el_iv.m_index);
indexed_value<T> & col_el_iv = column_chunk[row_el_iv.m_other];
remove_element(row_chunk, col_el_iv.m_other, column_chunk, row_el_iv.m_other);
}
template <typename T, typename X>
void sparse_matrix<T, X>::put_max_index_to_0(vector<indexed_value<T>> & row_vals, unsigned max_index) {
if (max_index == 0) return;
indexed_value<T> * max_iv = & row_vals[max_index];
indexed_value<T> * start_iv = & row_vals[0];
// update the "other" columns elements which are bound to the start_iv and max_iv
m_columns[max_iv->m_index].m_values[max_iv->m_other].m_other = 0;
m_columns[start_iv->m_index].m_values[start_iv->m_other].m_other = max_index;
// swap the elements
indexed_value<T> t = * max_iv;
* max_iv = * start_iv;
* start_iv = t;
}
template <typename T, typename X>
void sparse_matrix<T, X>::set_max_in_row(vector<indexed_value<T>> & row_vals) {
if (row_vals.size() == 0)
return;
T max_val = abs(row_vals[0].m_value);
unsigned max_index = 0;
for (unsigned i = 1; i < row_vals.size(); i++) {
T iabs = abs(row_vals[i].m_value);
if (iabs > max_val) {
max_val = iabs;
max_index = i;
}
}
put_max_index_to_0(row_vals, max_index);
}
template <typename T, typename X>
bool sparse_matrix<T, X>::pivot_with_eta(unsigned i, eta_matrix<T, X> *eta_matrix, lp_settings & settings) {
const T& pivot = eta_matrix->get_diagonal_element();
for (auto & it : eta_matrix->m_column_vector.m_data) {
if (!pivot_row_to_row(i, it.second, it.first, settings)) {
return false;
}
}
divide_row_by_constant(i, pivot, settings);
if (!shorten_active_matrix(i, eta_matrix)) {
return false;
}
return true;
}
// returns the offset of the pivot column in the row
template <typename T, typename X>
void sparse_matrix<T, X>::scan_row_to_work_vector_and_remove_pivot_column(unsigned row, unsigned pivot_column) {
auto & rvals = m_rows[row];
unsigned size = rvals.size();
for (unsigned j = 0; j < size; j++) {
auto & iv = rvals[j];
if (iv.m_index != pivot_column) {
m_work_pivot_vector[iv.m_index] = j;
} else {
remove_element(rvals, iv);
j--;
size--;
}
}
}
#ifdef LEAN_DEBUG
template <typename T, typename X>
vector<T> sparse_matrix<T, X>::get_full_row(unsigned i) const {
vector<T> r;
for (unsigned j = 0; j < column_count(); j++)
r.push_back(get(i, j));
return r;
}
#endif
// This method pivots row i to row i0 by muliplying row i by
// alpha and adding it to row i0.
// After pivoting the row i0 has a max abs value set correctly at the beginning of m_start,
// Returns false if the resulting row is all zeroes, and true otherwise
template <typename T, typename X>
bool sparse_matrix<T, X>::pivot_row_to_row(unsigned i, const T& alpha, unsigned i0, lp_settings & settings ) {
lean_assert(i < dimension() && i0 < dimension());
lean_assert(i != i0);
unsigned pivot_col = adjust_column(i);
i = adjust_row(i);
i0 = adjust_row(i0);
vector<unsigned> became_zeros;
// the offset of element of the pivot column in row i0
scan_row_to_work_vector_and_remove_pivot_column(i0, pivot_col);
auto & i0_row_vals = m_rows[i0];
// run over the pivot row and update row i0
unsigned prev_size_i0 = i0_row_vals.size();
for (const auto & iv : m_rows[i]) {
unsigned j = iv.m_index;
if (j == pivot_col) continue;
T alv = alpha * iv.m_value;
int j_offs = m_work_pivot_vector[j];
if (j_offs == -1) { // it is a new element
if (!settings.abs_val_is_smaller_than_drop_tolerance(alv)) {
add_new_element(i0, j, alv);
}
}
else {
auto & row_el_iv = i0_row_vals[j_offs];
row_el_iv.m_value += alv;
if (settings.abs_val_is_smaller_than_drop_tolerance(row_el_iv.m_value)) {
became_zeros.push_back(j_offs);
ensure_increasing(became_zeros);
}
else {
m_columns[j].m_values[row_el_iv.m_other].set_value(row_el_iv.m_value);
}
}
}
// clean the work vector
for (unsigned k = 0; k < prev_size_i0; k++) {
m_work_pivot_vector[i0_row_vals[k].m_index] = -1;
}
for (unsigned k = became_zeros.size(); k-- > 0; ) {
unsigned j = became_zeros[k];
remove_element(i0_row_vals, i0_row_vals[j]);
if (i0_row_vals.empty())
return false;
}
if (numeric_traits<T>::precise() == false)
set_max_in_row(i0_row_vals);
return !i0_row_vals.empty();
}
// set the max val as well
// returns false if the resulting row is all zeroes, and true otherwise
template <typename T, typename X>
bool sparse_matrix<T, X>::set_row_from_work_vector_and_clean_work_vector_not_adjusted(unsigned i0, indexed_vector<T> & work_vec,
lp_settings & settings) {
remove_zero_elements_and_set_data_on_existing_elements_not_adjusted(i0, work_vec, settings);
// all non-zero elements in m_work_pivot_vector are new
for (unsigned j : work_vec.m_index) {
if (numeric_traits<T>::is_zero(work_vec[j])) {
continue;
}
lean_assert(!settings.abs_val_is_smaller_than_drop_tolerance(work_vec[j]));
add_new_element(i0, adjust_column(j), work_vec[j]);
work_vec[j] = numeric_traits<T>::zero();
}
work_vec.m_index.clear();
auto & row_vals = m_rows[i0];
if (row_vals.size() == 0) {
return false;
}
set_max_in_row(row_vals); // it helps to find larger pivots
return true;
}
template <typename T, typename X>
void sparse_matrix<T, X>::remove_zero_elements_and_set_data_on_existing_elements(unsigned row) {
auto & row_vals = m_rows[row];
for (unsigned k = static_cast<unsigned>(row_vals.size()); k-- > 0;) { // we cannot simply run the iterator since we are removing
// elements from row_vals
auto & row_el_iv = row_vals[k];
unsigned j = row_el_iv.m_index;
T & wj = m_work_pivot_vector[j];
if (is_zero(wj)) {
remove_element(row_vals, row_el_iv);
} else {
m_columns[j].m_values[row_el_iv.m_other].set_value(wj);
row_el_iv.set_value(wj);
wj = zero_of_type<T>();
}
}
}
// work_vec here has not adjusted column indices
template <typename T, typename X>
void sparse_matrix<T, X>::remove_zero_elements_and_set_data_on_existing_elements_not_adjusted(unsigned row, indexed_vector<T> & work_vec, lp_settings & settings) {
auto & row_vals = m_rows[row];
for (unsigned k = static_cast<unsigned>(row_vals.size()); k-- > 0;) { // we cannot simply run the iterator since we are removing
// elements from row_vals
auto & row_el_iv = row_vals[k];
unsigned j = row_el_iv.m_index;
unsigned rj = adjust_column_inverse(j);
T val = work_vec[rj];
if (settings.abs_val_is_smaller_than_drop_tolerance(val)) {
remove_element(row_vals, row_el_iv);
lean_assert(numeric_traits<T>::is_zero(val));
} else {
m_columns[j].m_values[row_el_iv.m_other].set_value(row_el_iv.m_value = val);
work_vec[rj] = numeric_traits<T>::zero();
}
}
}
// adding delta columns at the end of the matrix
template <typename T, typename X>
void sparse_matrix<T, X>::add_columns_at_the_end(unsigned delta) {
for (unsigned i = 0; i < delta; i++) {
col_header col_head;
m_columns.push_back(col_head);
}
m_column_permutation.enlarge(delta);
}
template <typename T, typename X>
void sparse_matrix<T, X>::delete_column(int i) {
lean_assert(i < dimension());
for (auto cell = m_columns[i].m_head; cell != nullptr;) {
auto next_cell = cell->m_down;
kill_cell(cell);
cell = next_cell;
}
}
template <typename T, typename X>
void sparse_matrix<T, X>::divide_row_by_constant(unsigned i, const T & t, lp_settings & settings) {
lean_assert(!settings.abs_val_is_smaller_than_zero_tolerance(t));
i = adjust_row(i);
for (auto & iv : m_rows[i]) {
T &v = iv.m_value;
v /= t;
if (settings.abs_val_is_smaller_than_drop_tolerance(v)){
v = numeric_traits<T>::zero();
}
m_columns[iv.m_index].m_values[iv.m_other].set_value(v);
}
}
// solving x * this = y, and putting the answer into y
// the matrix here has to be upper triangular
template <typename T, typename X>
void sparse_matrix<T, X>::solve_y_U(vector<T> & y) const { // works by rows
#ifdef LEAN_DEBUG
// T * rs = clone_vector<T>(y, dimension());
#endif
unsigned end = dimension();
for (unsigned i = 0; i + 1 < end; i++) {
// all y[i] has correct values already
const T & yv = y[i];
if (numeric_traits<T>::is_zero(yv)) continue;
auto & mc = get_row_values(adjust_row(i));
for (auto & c : mc) {
unsigned col = adjust_column_inverse(c.m_index);
if (col != i) {
y[col] -= c.m_value * yv;
}
}
}
#ifdef LEAN_DEBUG
// dense_matrix<T> deb(*this);
// T * clone_y = clone_vector<T>(y, dimension());
// deb.apply_from_right(clone_y);
// lean_assert(vectors_are_equal(rs, clone_y, dimension()));
// delete [] clone_y;
// delete [] rs;
#endif
}
// solving x * this = y, and putting the answer into y
// the matrix here has to be upper triangular
template <typename T, typename X>
void sparse_matrix<T, X>::solve_y_U_indexed(indexed_vector<T> & y, const lp_settings & settings) {
#if 0 && LEAN_DEBUG
vector<T> ycopy(y.m_data);
if (numeric_traits<T>::precise() == false)
solve_y_U(ycopy);
#endif
vector<unsigned> sorted_active_columns;
extend_and_sort_active_rows(y.m_index, sorted_active_columns);
for (unsigned k = sorted_active_columns.size(); k-- > 0; ) {
unsigned j = sorted_active_columns[k];
auto & yj = y[j];
for (auto & c: m_columns[adjust_column(j)].m_values) {
unsigned i = adjust_row_inverse(c.m_index);
if (i == j) continue;
yj -= y[i] * c.m_value;
}
}
y.m_index.clear();
for (auto j : sorted_active_columns) {
if (!settings.abs_val_is_smaller_than_drop_tolerance(y[j]))
y.m_index.push_back(j);
else if (!numeric_traits<T>::precise())
y.m_data[j] = zero_of_type<T>();
}
lean_assert(y.is_OK());
#if 0 && LEAN_DEBUG
if (numeric_traits<T>::precise() == false)
lean_assert(vectors_are_equal(ycopy, y.m_data));
#endif
}
// fills the indices for such that y[i] can be not a zero
// sort them so the smaller indices come first
// void fill_reachable_indices(std::set<unsigned> & rset, T *y) {
// std::queue<unsigned> q;
// int m = dimension();
// for (int i = m - 1; i >= 0; i--) {
// if (!numeric_traits<T>::is_zero(y[i])){
// for (cell * c = m_columns[adjust_column(i)].m_head; c != nullptr; c = c->m_down) {
// unsigned row = adjust_row_inverse(c->m_i);
// q.push(row);
// }
// }
// }
// while (!q.empty()) {
// unsigned i = q.front();
// q.pop();
// for (cell * c = m_columns[adjust_column(i)].m_head; c != nullptr; c = c->m_down) {
// unsigned row = adjust_row_inverse(c->m_i);
// if (rset.find(row) == rset.end()){
// rset.insert(row);
// q.push(row);
// }
// }
// }
// }
template <typename T, typename X>
template <typename L>
void sparse_matrix<T, X>::find_error_in_solution_U_y(vector<L>& y_orig, vector<L> & y) {
unsigned i = dimension();
while (i--) {
y_orig[i] -= dot_product_with_row(i, y);
}
}
template <typename T, typename X>
template <typename L>
void sparse_matrix<T, X>::find_error_in_solution_U_y_indexed(indexed_vector<L>& y_orig, indexed_vector<L> & y, const vector<unsigned>& sorted_active_rows) {
for (unsigned i: sorted_active_rows)
y_orig.add_value_at_index(i, -dot_product_with_row(i, y)); // cannot round up here!!!
// y_orig can contain very small values
}
template <typename T, typename X>
template <typename L>
void sparse_matrix<T, X>::add_delta_to_solution(const vector<L>& del, vector<L> & y) {
unsigned i = dimension();
while (i--) {
y[i] += del[i];
}
}
template <typename T, typename X>
template <typename L>
void sparse_matrix<T, X>::add_delta_to_solution(const indexed_vector<L>& del, indexed_vector<L> & y) {
// lean_assert(del.is_OK());
// lean_assert(y.is_OK());
for (auto i : del.m_index) {
y.add_value_at_index(i, del[i]);
}
}
template <typename T, typename X>
template <typename L>
void sparse_matrix<T, X>::double_solve_U_y(indexed_vector<L>& y, const lp_settings & settings){
lean_assert(y.is_OK());
indexed_vector<L> y_orig(y); // copy y aside
vector<unsigned> active_rows;
solve_U_y_indexed_only(y, settings, active_rows);
lean_assert(y.is_OK());
find_error_in_solution_U_y_indexed(y_orig, y, active_rows);
// y_orig contains the error now
if (y_orig.m_index.size() * ratio_of_index_size_to_all_size<T>() < 32 * dimension()) {
active_rows.clear();
solve_U_y_indexed_only(y_orig, settings, active_rows);
add_delta_to_solution(y_orig, y);
y.clean_up();
} else { // the dense version
solve_U_y(y_orig.m_data);
add_delta_to_solution(y_orig.m_data, y.m_data);
y.restore_index_and_clean_from_data();
}
lean_assert(y.is_OK());
}
template <typename T, typename X>
template <typename L>
void sparse_matrix<T, X>::double_solve_U_y(vector<L>& y){
vector<L> y_orig(y); // copy y aside
solve_U_y(y);
find_error_in_solution_U_y(y_orig, y);
// y_orig contains the error now
solve_U_y(y_orig);
add_delta_to_solution(y_orig, y);
}
// solving this * x = y, and putting the answer into y
// the matrix here has to be upper triangular
template <typename T, typename X>
template <typename L>
void sparse_matrix<T, X>::solve_U_y(vector<L> & y) { // it is a column wise version
#ifdef LEAN_DEBUG
// T * rs = clone_vector<T>(y, dimension());
#endif
for (unsigned j = dimension(); j--; ) {
const L & yj = y[j];
if (is_zero(yj)) continue;
for (const auto & iv : m_columns[adjust_column(j)].m_values) {
unsigned i = adjust_row_inverse(iv.m_index);
if (i != j) {
y[i] -= iv.m_value * yj;
}
}
}
#ifdef LEAN_DEBUG
// dense_matrix<T> deb(*this);
// T * clone_y = clone_vector<T>(y, dimension());
// deb.apply_from_left(clone_y);
// lean_assert(vectors_are_equal(rs, clone_y, dimension()));
#endif
}
template <typename T, typename X>
void sparse_matrix<T, X>::process_index_recursively_for_y_U(unsigned j, vector<unsigned> & sorted_active_rows) {
lean_assert(m_processed[j] == false);
m_processed[j]=true;
auto & row = m_rows[adjust_row(j)];
for (auto & c : row) {
unsigned i = adjust_column_inverse(c.m_index);
if (i == j) continue;
if (!m_processed[i]) {
process_index_recursively_for_y_U(i, sorted_active_rows);
}
}
sorted_active_rows.push_back(j);
}
template <typename T, typename X>
void sparse_matrix<T, X>::process_column_recursively(unsigned j, vector<unsigned> & sorted_active_rows) {
lean_assert(m_processed[j] == false);
auto & mc = m_columns[adjust_column(j)].m_values;
for (auto & iv : mc) {
unsigned i = adjust_row_inverse(iv.m_index);
if (i == j) continue;
if (!m_processed[i]) {
process_column_recursively(i, sorted_active_rows);
}
}
m_processed[j]=true;
sorted_active_rows.push_back(j);
}
template <typename T, typename X>
void sparse_matrix<T, X>::create_graph_G(const vector<unsigned> & index_or_right_side, vector<unsigned> & sorted_active_rows) {
for (auto i : index_or_right_side) {
if (m_processed[i]) continue;
process_column_recursively(i, sorted_active_rows);
}
for (auto i : sorted_active_rows) {
m_processed[i] = false;
}
}
template <typename T, typename X>
void sparse_matrix<T, X>::extend_and_sort_active_rows(const vector<unsigned> & index_or_right_side, vector<unsigned> & sorted_active_rows) {
for (auto i : index_or_right_side) {
if (m_processed[i]) continue;
process_index_recursively_for_y_U(i, sorted_active_rows);
}
for (auto i : sorted_active_rows) {
m_processed[i] = false;
}
}
template <typename T, typename X>
template <typename L>
void sparse_matrix<T, X>::solve_U_y_indexed_only(indexed_vector<L> & y, const lp_settings & settings, vector<unsigned> & sorted_active_rows) { // it is a column wise version
create_graph_G(y.m_index, sorted_active_rows);
for (auto k = sorted_active_rows.size(); k-- > 0;) {
unsigned j = sorted_active_rows[k];
const L & yj = y[j];
if (is_zero(yj)) continue;
auto & mc = m_columns[adjust_column(j)].m_values;
for (auto & iv : mc) {
unsigned i = adjust_row_inverse(iv.m_index);
if (i != j) {
y[i] -= iv.m_value * yj;
}
}
}
y.m_index.clear();
for (auto j : sorted_active_rows) {
if (!settings.abs_val_is_smaller_than_drop_tolerance(y[j]))
y.m_index.push_back(j);
else if (!numeric_traits<L>::precise())
y[j] = zero_of_type<L>();
}
lean_assert(y.is_OK());
#ifdef LEAN_DEBUG
// dense_matrix<T,X> deb(this);
// vector<T> clone_y(y.m_data);
// deb.apply_from_left(clone_y);
// lean_assert(vectors_are_equal(rs, clone_y));
#endif
}
template <typename T, typename X>
template <typename L>
L sparse_matrix<T, X>::dot_product_with_row (unsigned row, const vector<L> & y) const {
L ret = zero_of_type<L>();
auto & mc = get_row_values(adjust_row(row));
for (auto & c : mc) {
unsigned col = m_column_permutation[c.m_index];
ret += c.m_value * y[col];
}
return ret;
}
template <typename T, typename X>
template <typename L>
L sparse_matrix<T, X>::dot_product_with_row (unsigned row, const indexed_vector<L> & y) const {
L ret = zero_of_type<L>();
auto & mc = get_row_values(adjust_row(row));
for (auto & c : mc) {
unsigned col = m_column_permutation[c.m_index];
ret += c.m_value * y[col];
}
return ret;
}
template <typename T, typename X>
unsigned sparse_matrix<T, X>::get_number_of_nonzeroes() const {
unsigned ret = 0;
for (unsigned i = dimension(); i--; ) {
ret += number_of_non_zeroes_in_row(i);
}
return ret;
}
template <typename T, typename X>
bool sparse_matrix<T, X>::get_non_zero_column_in_row(unsigned i, unsigned *j) const {
// go over the i-th row
auto & mc = get_row_values(adjust_row(i));
if (mc.size() > 0) {
*j = m_column_permutation[mc[0].m_index];
return true;
}
return false;
}
template <typename T, typename X>
void sparse_matrix<T, X>::remove_element_that_is_not_in_w(vector<indexed_value<T>> & column_vals, indexed_value<T> & col_el_iv) {
auto & row_chunk = m_rows[col_el_iv.m_index];
indexed_value<T> & row_el_iv = row_chunk[col_el_iv.m_other];
unsigned index_in_row = col_el_iv.m_other;
remove_element(row_chunk, col_el_iv.m_other, column_vals, row_el_iv.m_other);
if (index_in_row == 0)
set_max_in_row(row_chunk);
}
// w contains the new column
// the old column inside of the matrix has not been changed yet
template <typename T, typename X>
void sparse_matrix<T, X>::remove_elements_that_are_not_in_w_and_update_common_elements(unsigned column_to_replace, indexed_vector<T> & w) {
// --------------------------------
// column_vals represents the old column
auto & column_vals = m_columns[column_to_replace].m_values;
for (unsigned k = static_cast<unsigned>(column_vals.size()); k-- > 0;) {
indexed_value<T> & col_el_iv = column_vals[k];
unsigned i = col_el_iv.m_index;
T &w_data_at_i = w[adjust_row_inverse(i)];
if (numeric_traits<T>::is_zero(w_data_at_i)) {
remove_element_that_is_not_in_w(column_vals, col_el_iv);
} else {
auto& row_chunk = m_rows[i];
unsigned index_in_row = col_el_iv.m_other;
if (index_in_row == 0) {
bool look_for_max = abs(w_data_at_i) < abs(row_chunk[0].m_value);
row_chunk[0].set_value(col_el_iv.m_value = w_data_at_i);
if (look_for_max)
set_max_in_row(i);
} else {
row_chunk[index_in_row].set_value(col_el_iv.m_value = w_data_at_i);
if (abs(w_data_at_i) > abs(row_chunk[0].m_value))
put_max_index_to_0(row_chunk, index_in_row);
}
w_data_at_i = numeric_traits<T>::zero();
}
}
}
template <typename T, typename X>
void sparse_matrix<T, X>::add_new_element(unsigned row, unsigned col, const T& val) {
auto & row_vals = m_rows[row];
auto & col_vals = m_columns[col].m_values;
unsigned row_el_offs = static_cast<unsigned>(row_vals.size());
unsigned col_el_offs = static_cast<unsigned>(col_vals.size());
row_vals.push_back(indexed_value<T>(val, col, col_el_offs));
col_vals.push_back(indexed_value<T>(val, row, row_el_offs));
m_n_of_active_elems++;
}
// w contains the "rest" of the new column; all common elements of w and the old column has been zeroed
// the old column inside of the matrix has not been changed yet
template <typename T, typename X>
void sparse_matrix<T, X>::add_new_elements_of_w_and_clear_w(unsigned column_to_replace, indexed_vector<T> & w, lp_settings & settings) {
for (unsigned i : w.m_index) {
T w_at_i = w[i];
if (numeric_traits<T>::is_zero(w_at_i)) continue; // was dealt with already
if (!settings.abs_val_is_smaller_than_drop_tolerance(w_at_i)) {
unsigned ai = adjust_row(i);
add_new_element(ai, column_to_replace, w_at_i);
auto & row_chunk = m_rows[ai];
lean_assert(row_chunk.size() > 0);
if (abs(w_at_i) > abs(row_chunk[0].m_value))
put_max_index_to_0(row_chunk, static_cast<unsigned>(row_chunk.size()) - 1);
}
w[i] = numeric_traits<T>::zero();
}
w.m_index.clear();
}
template <typename T, typename X>
void sparse_matrix<T, X>::replace_column(unsigned column_to_replace, indexed_vector<T> & w, lp_settings &settings) {
column_to_replace = adjust_column(column_to_replace);
remove_elements_that_are_not_in_w_and_update_common_elements(column_to_replace, w);
add_new_elements_of_w_and_clear_w(column_to_replace, w, settings);
}
template <typename T, typename X>
unsigned sparse_matrix<T, X>::pivot_score(unsigned i, unsigned j) {
// It goes like this (rnz-1)(cnz-1) is the Markovitz number, that is the max number of
// new non zeroes we can obtain after the pivoting.
// In addition we will get another cnz - 1 elements in the eta matrix created for this pivot,
// which gives rnz(cnz-1). For example, is 0 for a column singleton, but not for
// a row singleton ( which is not a column singleton).
auto col_header = m_columns[j];
return static_cast<unsigned>(get_row_values(i).size() * (col_header.m_values.size() - col_header.m_shortened_markovitz - 1));
}
template <typename T, typename X>
void sparse_matrix<T, X>::enqueue_domain_into_pivot_queue() {
lean_assert(m_pivot_queue.size() == 0);
for (unsigned i = 0; i < dimension(); i++) {
auto & rh = m_rows[i];
unsigned rnz = static_cast<unsigned>(rh.size());
for (auto iv : rh) {
unsigned j = iv.m_index;
m_pivot_queue.enqueue(i, j, rnz * (static_cast<unsigned>(m_columns[j].m_values.size()) - 1));
}
}
}
template <typename T, typename X>
void sparse_matrix<T, X>::set_max_in_rows() {
unsigned i = dimension();
while (i--)
set_max_in_row(i);
}
template <typename T, typename X>
void sparse_matrix<T, X>::zero_shortened_markovitz_numbers() {
for (auto & ch : m_columns)
ch.zero_shortened_markovitz();
}
template <typename T, typename X>
void sparse_matrix<T, X>::prepare_for_factorization() {
zero_shortened_markovitz_numbers();
set_max_in_rows();
enqueue_domain_into_pivot_queue();
}
template <typename T, typename X>
void sparse_matrix<T, X>::recover_pivot_queue(vector<upair> & rejected_pivots) {
for (auto p : rejected_pivots) {
m_pivot_queue.enqueue(p.first, p.second, pivot_score(p.first, p.second));
}
}
template <typename T, typename X>
int sparse_matrix<T, X>::elem_is_too_small(unsigned i, unsigned j, int c_partial_pivoting) {
auto & row_chunk = m_rows[i];
if (j == row_chunk[0].m_index) {
return 0; // the max element is at the head
}
T max = abs(row_chunk[0].m_value);
for (unsigned k = 1; k < row_chunk.size(); k++) {
auto &iv = row_chunk[k];
if (iv.m_index == j)
return abs(iv.m_value) * c_partial_pivoting < max ? 1: 0;
}
return 2; // the element became zero but it still sits in the active pivots?
}
template <typename T, typename X>
bool sparse_matrix<T, X>::remove_row_from_active_pivots_and_shorten_columns(unsigned row) {
unsigned arow = adjust_row(row);
for (auto & iv : m_rows[arow]) {
m_pivot_queue.remove(arow, iv.m_index);
m_n_of_active_elems--; // the value is correct only when refactoring
if (adjust_column_inverse(iv.m_index) <= row)
continue; // this column will be removed anyway
auto & col = m_columns[iv.m_index];
col.shorten_markovich_by_one();
if (col.m_values.size() <= col.m_shortened_markovitz)
return false; // got a zero column
}
return true;
}
template <typename T, typename X>
void sparse_matrix<T, X>::remove_pivot_column(unsigned row) {
unsigned acol = adjust_column(row);
for (const auto & iv : m_columns[acol].m_values)
if (adjust_row_inverse(iv.m_index) >= row)
m_pivot_queue.remove(iv.m_index, acol);
}
template <typename T, typename X>
void sparse_matrix<T, X>::update_active_pivots(unsigned row) {
unsigned arow = adjust_row(row);
for (const auto & iv : m_rows[arow]) {
col_header & ch = m_columns[iv.m_index];
int cols = static_cast<int>(ch.m_values.size()) - ch.m_shortened_markovitz - 1;
lean_assert(cols >= 0);
for (const auto &ivc : ch.m_values) {
unsigned i = ivc.m_index;
if (adjust_row_inverse(i) <= row) continue; // the i is not an active row
m_pivot_queue.enqueue(i, iv.m_index, static_cast<unsigned>(m_rows[i].size())*cols);
}
}
}
template <typename T, typename X>
bool sparse_matrix<T, X>::shorten_active_matrix(unsigned row, eta_matrix<T, X> *eta_matrix) {
if (!remove_row_from_active_pivots_and_shorten_columns(row))
return false;
remove_pivot_column(row);
// need to know the max priority of the queue here
update_active_pivots(row);
if (eta_matrix == nullptr) return true;
// it looks like double work, but the pivot scores have changed for all rows
// touched by eta_matrix
for (auto & it : eta_matrix->m_column_vector.m_data) {
unsigned row = adjust_row(it.first);
const auto & row_values = m_rows[row];
unsigned rnz = static_cast<unsigned>(row_values.size());
for (auto & iv : row_values) {
const col_header& ch = m_columns[iv.m_index];
int cnz = static_cast<int>(ch.m_values.size()) - ch.m_shortened_markovitz - 1;
lean_assert(cnz >= 0);
m_pivot_queue.enqueue(row, iv.m_index, rnz * cnz);
}
}
return true;
}
template <typename T, typename X>
unsigned sparse_matrix<T, X>::pivot_score_without_shortened_counters(unsigned i, unsigned j, unsigned k) {
auto &cols = m_columns[j].m_values;
unsigned cnz = cols.size();
for (auto & iv : cols) {
if (adjust_row_inverse(iv.m_index) < k)
cnz--;
}
lean_assert(cnz > 0);
return m_rows[i].m_values.size() * (cnz - 1);
}
#ifdef LEAN_DEBUG
template <typename T, typename X>
bool sparse_matrix<T, X>::can_improve_score_for_row(unsigned row, unsigned score, T const & c_partial_pivoting, unsigned k) {
unsigned arow = adjust_row(row);
auto & row_vals = m_rows[arow].m_values;
auto & begin_iv = row_vals[0];
T row_max = abs(begin_iv.m_value);
lean_assert(adjust_column_inverse(begin_iv.m_index) >= k);
if (pivot_score_without_shortened_counters(arow, begin_iv.m_index, k) < score) {
print_active_matrix(k);
return true;
}
for (unsigned jj = 1; jj < row_vals.size(); jj++) {
auto & iv = row_vals[jj];
lean_assert(adjust_column_inverse(iv.m_index) >= k);
lean_assert(abs(iv.m_value) <= row_max);
if (c_partial_pivoting * abs(iv.m_value) < row_max) continue;
if (pivot_score_without_shortened_counters(arow, iv.m_index, k) < score) {
print_active_matrix(k);
return true;
}
}
return false;
}
template <typename T, typename X>
bool sparse_matrix<T, X>::really_best_pivot(unsigned i, unsigned j, T const & c_partial_pivoting, unsigned k) {
unsigned queue_pivot_score = pivot_score_without_shortened_counters(i, j, k);
for (unsigned ii = k; ii < dimension(); ii++) {
lean_assert(!can_improve_score_for_row(ii, queue_pivot_score, c_partial_pivoting, k));
}
return true;
}
template <typename T, typename X>
void sparse_matrix<T, X>::print_active_matrix(unsigned k, std::ostream & out) {
out << "active matrix for k = " << k << std::endl;
if (k >= dimension()) {
out << "empty" << std::endl;
return;
}
unsigned dim = dimension() - k;
dense_matrix<T, X> b(dim, dim);
for (unsigned i = 0; i < dim; i++)
for (unsigned j = 0; j < dim; j++ )
b.set_elem(i, j, zero_of_type<T>());
for (int i = k; i < dimension(); i++) {
unsigned col = adjust_column(i);
for (auto &iv : get_column_values(col)) {
unsigned row = iv.m_index;
unsigned row_ex = this->adjust_row_inverse(row);
if (row_ex < k) continue;
auto v = this->get_not_adjusted(row, col);
b.set_elem(row_ex - k, i -k, v);
}
}
print_matrix(b, out);
}
template <typename T, typename X>
bool sparse_matrix<T, X>::pivot_queue_is_correct_for_row(unsigned i, unsigned k) {
unsigned arow = adjust_row(i);
for (auto & iv : m_rows[arow].m_values) {
lean_assert(pivot_score_without_shortened_counters(arow, iv.m_index, k + 1) ==
m_pivot_queue.get_priority(arow, iv.m_index));
}
return true;
}
template <typename T, typename X>
bool sparse_matrix<T, X>::pivot_queue_is_correct_after_pivoting(int k) {
for (unsigned i = k + 1; i < dimension(); i++ )
lean_assert(pivot_queue_is_correct_for_row(i, k));
lean_assert(m_pivot_queue.is_correct());
return true;
}
#endif
template <typename T, typename X>
bool sparse_matrix<T, X>::get_pivot_for_column(unsigned &i, unsigned &j, int c_partial_pivoting, unsigned k) {
vector<upair> pivots_candidates_that_are_too_small;
while (!m_pivot_queue.is_empty()) {
m_pivot_queue.dequeue(i, j);
unsigned i_inv = adjust_row_inverse(i);
if (i_inv < k) continue;
unsigned j_inv = adjust_column_inverse(j);
if (j_inv < k) continue;
int _small = elem_is_too_small(i, j, c_partial_pivoting);
if (!_small) {
#ifdef LEAN_DEBUG
// if (!really_best_pivot(i, j, c_partial_pivoting, k)) {
// print_active_matrix(k);
// lean_assert(false);
// }
#endif
recover_pivot_queue(pivots_candidates_that_are_too_small);
i = i_inv;
j = j_inv;
return true;
}
if (_small != 2) { // 2 means that the pair is not in the matrix
pivots_candidates_that_are_too_small.push_back(std::make_pair(i, j));
}
}
recover_pivot_queue(pivots_candidates_that_are_too_small);
return false;
}
template <typename T, typename X>
bool sparse_matrix<T, X>::elem_is_too_small(vector<indexed_value<T>> & row_chunk, indexed_value<T> & iv, int c_partial_pivoting) {
if (&iv == &row_chunk[0]) {
return false; // the max element is at the head
}
T val = abs(iv.m_value);
T max = abs(row_chunk[0].m_value);
return val * c_partial_pivoting < max;
}
template <typename T, typename X>
bool sparse_matrix<T, X>::shorten_columns_by_pivot_row(unsigned i, unsigned pivot_column) {
vector<indexed_value<T>> & row_chunk = get_row_values(i);
for (indexed_value<T> & iv : row_chunk) {
unsigned j = iv.m_index;
if (j == pivot_column) {
lean_assert(!col_is_active(j));
continue;
}
m_columns[j].shorten_markovich_by_one();
if (m_columns[j].m_shortened_markovitz >= get_column_values(j).size()) { // got the zero column under the row!
return false;
}
}
return true;
}
template <typename T, typename X>
bool sparse_matrix<T, X>::fill_eta_matrix(unsigned j, eta_matrix<T, X> ** eta) {
const vector<indexed_value<T>> & col_chunk = get_column_values(adjust_column(j));
bool is_unit = true;
for (const auto & iv : col_chunk) {
unsigned i = adjust_row_inverse(iv.m_index);
if (i > j) {
is_unit = false;
break;
}
if (i == j && iv.m_value != 1) {
is_unit = false;
break;
}
}
if (is_unit) {
*eta = nullptr;
return true;
}
#ifdef LEAN_DEBUG
*eta = new eta_matrix<T, X>(j, dimension());
#else
*eta = new eta_matrix<T, X>(j);
#endif
for (const auto & iv : col_chunk) {
unsigned i = adjust_row_inverse(iv.m_index);
if (i < j) {
continue;
}
if (i > j) {
(*eta)->push_back(i, - iv.m_value);
} else { // i == j
if ( !(*eta)->set_diagonal_element(iv.m_value)) {
delete *eta;
*eta = nullptr;
return false;
}
}
}
(*eta)->divide_by_diagonal_element();
return true;
}
#ifdef LEAN_DEBUG
template <typename T, typename X>
bool sparse_matrix<T, X>::is_upper_triangular_and_maximums_are_set_correctly_in_rows(lp_settings & settings) const {
for (unsigned i = 0; i < dimension(); i++) {
vector<indexed_value<T>> const & row_chunk = get_row_values(i);
lean_assert(row_chunk.size());
T const & max = abs(row_chunk[0].m_value);
unsigned ai = adjust_row_inverse(i);
for (auto & iv : row_chunk) {
lean_assert(abs(iv.m_value) <= max);
unsigned aj = adjust_column_inverse(iv.m_index);
if (!(ai <= aj || numeric_traits<T>::is_zero(iv.m_value)))
return false;
if (aj == ai) {
if (iv.m_value != 1) {
// std::cout << "value at diagonal = " << iv.m_value << std::endl;
return false;
}
}
if (settings.abs_val_is_smaller_than_drop_tolerance(iv.m_value) && (!is_zero(iv.m_value)))
return false;
}
}
return true;
}
template <typename T, typename X>
bool sparse_matrix<T, X>::is_upper_triangular_until(unsigned k) const {
for (unsigned j = 0; j < dimension() && j < k; j++) {
unsigned aj = adjust_column(j);
auto & col = get_column_values(aj);
for (auto & iv : col) {
unsigned row = adjust_row_inverse(iv.m_index);
if (row > j)
return false;
}
}
return true;
}
template <typename T, typename X>
void sparse_matrix<T, X>::check_column_vs_rows(unsigned col) {
auto & mc = get_column_values(col);
for (indexed_value<T> & column_iv : mc) {
indexed_value<T> & row_iv = column_iv_other(column_iv);
if (row_iv.m_index != col) {
// std::cout << "m_other in row does not belong to column " << col << ", but to column " << row_iv.m_index << std::endl;
lean_assert(false);
}
if (& row_iv_other(row_iv) != &column_iv) {
// std::cout << "row and col do not point to each other" << std::endl;
lean_assert(false);
}
if (row_iv.m_value != column_iv.m_value) {
// std::cout << "the data from col " << col << " for row " << column_iv.m_index << " is different in the column " << std::endl;
// std::cout << "in the col it is " << column_iv.m_value << ", but in the row it is " << row_iv.m_value << std::endl;
lean_assert(false);
}
}
}
template <typename T, typename X>
void sparse_matrix<T, X>::check_row_vs_columns(unsigned row) {
auto & mc = get_row_values(row);
for (indexed_value<T> & row_iv : mc) {
indexed_value<T> & column_iv = row_iv_other(row_iv);
if (column_iv.m_index != row) {
// std::cout << "col_iv does not point to correct row " << row << " but to " << column_iv.m_index << std::endl;
lean_assert(false);
}
if (& row_iv != & column_iv_other(column_iv)) {
// std::cout << "row and col do not point to each other" << std::endl;
lean_assert(false);
}
if (row_iv.m_value != column_iv.m_value) {
// std::cout << "the data from col " << column_iv.m_index << " for row " << row << " is different in the column " << std::endl;
// std::cout << "in the col it is " << column_iv.m_value << ", but in the row it is " << row_iv.m_value << std::endl;
lean_assert(false);
}
}
}
template <typename T, typename X>
void sparse_matrix<T, X>::check_rows_vs_columns() {
for (unsigned i = 0; i < dimension(); i++) {
check_row_vs_columns(i);
}
}
template <typename T, typename X>
void sparse_matrix<T, X>::check_columns_vs_rows() {
for (unsigned i = 0; i < dimension(); i++) {
check_column_vs_rows(i);
}
}
template <typename T, typename X>
void sparse_matrix<T, X>::check_matrix() {
check_rows_vs_columns();
check_columns_vs_rows();
}
#endif
}