template <class ZT, class FT> void MatHouseholder<ZT, FT>::update_R_naively(int i)
{
// This function try to strictly respect Algorithm 2 of [MSV'09].
FPLLL_DEBUG_CHECK(i <= n_known_rows_naively);
int j;
// Set B in R_naively.
if (enable_row_expo)
{
long max_expo = LONG_MIN;
for (j = 0; j < n; j++)
{
b(i, j).get_f_exp(R_naively(i, j), tmp_col_expo[j]);
max_expo = max(max_expo, tmp_col_expo[j]);
}
for (j = 0; j < n; j++)
R_naively(i, j).mul_2si(R_naively(i, j), tmp_col_expo[j] - max_expo);
row_expo_naively[i] = max_expo;
FPLLL_DEBUG_CHECK(row_expo_naively[i] >= 0);
}
else
{
for (j = 0; j < n; j++)
R_naively(i, j).set_z(b(i, j));
}
for (j = 0; j < i; j++)
{
// vj * ri[j..n]^T
V_naively[j].dot_product(ftmp0, R_naively[i], j, n);
//-vj * ri[j..n]^T
ftmp0.neg(ftmp0);
// ri[j..n] = ri[j..n] - (vj * ri[j..n]^T) * vj
R_naively[i].addmul(V_naively[j], ftmp0, j, n);
// ri[j] = sigma_naively[j] * ri[j]
R_naively(i, j).mul(sigma_naively[j], R_naively(i, j));
}
// Here, ftmp2 is equal to s in [MSV, ISSAC'09].
// Copy R_naively[i][i..n] in V_naively
for (j = i; j < n; j++)
V_naively(i, j) = R_naively(i, j);
// sigma_naively[i] = sign(r[1])
sigma_naively[i] = (R_naively(i, i).cmp(0.0) < 0) ? -1.0 : 1.0;
R_naively[i].dot_product(ftmp2, R_naively[i], i, n);
ftmp2.sqrt(ftmp2);
ftmp2.mul(ftmp2, sigma_naively[i]);
// Here, ftmp2 = sigma_naively[i] * ||r||
// ftmp0 = (r[1] + ftmp2)
ftmp0.add(R_naively(i, i), ftmp2);
if (ftmp0.cmp(0.0) != 0)
{
if (i + 1 == n)
ftmp1 = 0.0;
else
R_naively[i].dot_product(ftmp1, R_naively[i], i + 1, n);
if (ftmp1.cmp(0.0) != 0)
{
// ftmp1 = (-sum(r[j]^2, j, 2, n-i+1)
ftmp1.neg(ftmp1);
// vi[1] = (-sum(r[j]^2, j, 2, n-i+1) / (r[1] + ftmp2)
V_naively(i, i).div(ftmp1, ftmp0);
ftmp2.neg(ftmp2);
ftmp0.mul(ftmp2, V_naively(i, i));
// Here, ftmp0 = -ftmp2 * vi[1]
// ftmp0 = sqrt(-ftmp2 * vi[1])
ftmp0.sqrt(ftmp0);
V_naively[i].div(V_naively[i], i, n, ftmp0);
R_naively(i, i).abs(ftmp2);
for (j = i + 1; j < n; j++)
R_naively(i, j) = 0.0;
}
else
{
if (R_naively(i, i).cmp(0.0) < 0)
R_naively(i, i).neg(R_naively(i, i));
V_naively(i, i) = 0.0;
for (int k = i + 1; k < n; k++)
{
V_naively(i, k) = 0.0;
R_naively(i, k) = 0.0;
}
}
}
else
{
for (int k = i; k < n; k++)
{
V_naively(i, k) = 0.0;
R_naively(i, k) = 0.0;
}
}
n_known_rows_naively++;
}
// TODO: maybe add a variable available on DEBUG to verify in update_R(i, false) was done
// This function corresponds more or less to householder_v in hplll.
template <class ZT, class FT> void MatHouseholder<ZT, FT>::update_R_last(int i)
{
// sigma[i] = sign(r[1])
sigma[i] = (R(i, i).cmp(0.0) < 0) ? -1.0 : 1.0;
// V(i, i) is used as a temporary variable. In the following, V(i, i) is always modified.
if (i + 1 == n)
ftmp3 = 0.0;
else
{
// r^T * r (with first coefficient ignored)
R[i].dot_product(ftmp3, R[i], i + 1, n);
}
ftmp1.mul(R(i, i), R(i, i));
// ftmp1 = r^T * r
ftmp1.add(ftmp1, ftmp3);
if (ftmp1.cmp(0.0) != 0)
{
// ||r||
ftmp2.sqrt(ftmp1);
// s = sigma[i] * ||r|| = sigma[i] * sqrt(r * r^T)
ftmp0.mul(sigma[i], ftmp2);
ftmp1.add(R(i, i), ftmp0);
// ftmp3 = (-sum(r[j]^2, j, 2, n-i+1)
ftmp3.neg(ftmp3);
// ftmp3 = (-sum(r[j]^2, j, 2, n-i+1) / (r[1] + s)
ftmp3.div(ftmp3, ftmp1);
// If ftmp3 = 0, then ftmp0 = 0 and then, divide by ftmp0 is not allowed
if (ftmp3.cmp(0.0) != 0)
{
// ftmp0 = -sigma[i] * ||r||
ftmp0.neg(ftmp0);
// ftmp0 = -sigma[i] * ||r|| * (-sum(r[j]^2, j, 2, n-i+1) / (r[1] + s)
ftmp0.mul(ftmp0, ftmp3);
// ftmp0 = sqrt(-sigma[i] * ||r|| * (-sum(r[j]^2, j, 2, n-i+1) / (r[1] + s))
ftmp0.sqrt(ftmp0);
#ifndef HOUSEHOLDER_PRECOMPUTE_INVERSE
// V(i, i) =
// ((-sum(r[j]^2, j, 2, n-i+1) / (r[1] + s)) / sqrt(-sigma[i] * ||r|| * (-sum(r[j]^2, j, 2,
// n-i+1) / (r[1] + s))
V(i, i).div(ftmp3, ftmp0);
// R(i, i) = ||r||
R(i, i) = ftmp2;
V[i].div(R[i], i + 1, n, ftmp0);
#else // HOUSEHOLDER_PRECOMPUTE_INVERSE
ftmp0.div(1.0, ftmp0);
// Here, ftmp0 = 1 / sqrt(-s * vi[1])
V(i, i).mul(ftmp3, ftmp0);
R(i, i) = ftmp2;
// FIXME: not tested.
V[i].mul(R[i], i + 1, n, ftmp0);
R_inverse_diag[i].div(1.0, ftmp2);
#endif // HOUSEHOLDER_PRECOMPUTE_INVERSE
// Here, vi = vi / ftmp0 and ri[i..n] = (||r||, 0, 0, ..., 0)
#ifdef DEBUG
// Setting R(i, k) to 0 is not neccessary since this value will be not used in HLLL.
// However, in DEBUG, we must set to do not break test.
for (int k = i + 1; k < n; k++)
R(i, k) = 0.0;
#endif // DEBUG
}
else
{
V(i, i) = 0.0;
if (R(i, i).cmp(0.0) < 0)
R(i, i).neg(R(i, i));
// TODO: this is a row operation (from i + 1 to n).
for (int k = i + 1; k < n; k++)
{
// if enable_row_expo, R can be not correct at some point of the computation
FPLLL_DEBUG_CHECK(enable_row_expo ? 1 : R(i, k).is_zero());
V(i, k) = 0.0;
// Setting R(i, k) to 0 is not neccessary since this value will be not used in HLLL.
// However, in DEBUG, we must set to do not break test.
#ifdef DEBUG
R(i, k) = 0.0;
#endif // DEBUG
}
#ifdef HOUSEHOLDER_PRECOMPUTE_INVERSE
R_inverse_diag[i].div(1.0, R(i, i));
#endif // HOUSEHOLDER_PRECOMPUTE_INVERSE
}
}
else
{
R(i, i) = 0.0;
V(i, i) = 0.0;
// The for loop can for(int k = i; k < n; k++) if the setting of R(i, k) is uncommented.
// TODO: this is a row operation (from i + 1 to n).
for (int k = i + 1; k < n; k++)
{
// if enable_row_expo, R can be not correct at some point of the computation
FPLLL_DEBUG_CHECK(enable_row_expo ? 1 : R(i, k).is_zero());
V(i, k) = 0.0;
// Setting R(i, k) to 0 is not neccessary since this value will be not used in HLLL.
// However, in DEBUG, we must set to do not break test.
#ifdef DEBUG
R(i, k) = 0.0;
#endif // DEBUG
}
#ifdef HOUSEHOLDER_PRECOMPUTE_INVERSE
// Result is inf.
// TODO: set inf instead of doing the computation.
R_inverse_diag[i].div(1.0, 0.0);
#endif // HOUSEHOLDER_PRECOMPUTE_INVERSE
}
n_known_rows++;
}
vector<Strategy> load_strategies_json(const std::string &filename)
{
json js;
{
std::ifstream fs(filename);
if (fs.fail())
throw std::runtime_error("Cannot open strategies file.");
fs >> js;
}
vector<Strategy> strategies;
for(auto it = js.begin(); it != js.end(); ++it)
{
const json &j_strat = *it;
const size_t block_size = j_strat["block_size"];
FPLLL_DEBUG_CHECK(block_size < 4096); // some arbitrary upper limit
// ensure to add this strategy in the right place
while(strategies.size() <= block_size)
{
strategies.emplace_back();
}
Strategy strategy;
strategy.block_size = block_size;
if (j_strat.find("preprocessing_block_sizes") != j_strat.end())
{
for (auto p_it = j_strat["preprocessing_block_sizes"].begin(); p_it != j_strat["preprocessing_block_sizes"].end();
++p_it)
{
if ((*p_it).is_number())
{
strategy.preprocessing_block_sizes.emplace_back((*p_it).get<int>());
}
else
{
strategy.preprocessing_block_sizes.emplace_back((*p_it)["block_size"]);
}
}
}
if (j_strat.find("pruning_parameters") != j_strat.end())
{
for (auto p_it = j_strat["pruning_parameters"].begin();
p_it != j_strat["pruning_parameters"].end(); ++p_it)
{
const json &j_prun = *p_it;
Pruning pruning;
pruning.radius_factor = j_prun[0];
// fplll enforces that the first pruning coefficient is 1.0
FPLLL_DEBUG_CHECK((double) j_prun[1][0] == 1.0);
for (auto c_it = j_prun[1].begin(); c_it != j_prun[1].end(); ++c_it)
{
double c = (*c_it);
pruning.coefficients.emplace_back(c);
}
pruning.probability = j_prun[2];
FPLLL_DEBUG_CHECK(pruning.probability > 0.0 && pruning.probability <= 1.0);
strategy.pruning_parameters.emplace_back(pruning);
}
}
strategies[block_size] = std::move(strategy);
}
// finally, we make sure all strategies are sound
for(auto it = strategies.begin(); it != strategies.end(); ++it) {
if(it->pruning_parameters.size() == 0)
it->pruning_parameters.emplace_back(Pruning());
}
return strategies;
}
template <class ZT, class FT> void MatGSOGram<ZT, FT>::move_row(int old_r, int new_r)
{
FPLLL_DEBUG_CHECK(!cols_locked);
if (new_r < old_r)
{
FPLLL_DEBUG_CHECK(old_r < n_known_rows && !cols_locked);
for (int i = new_r; i < n_known_rows; i++)
{
invalidate_gso_row(i, new_r);
}
rotate(gso_valid_cols.begin() + new_r, gso_valid_cols.begin() + old_r,
gso_valid_cols.begin() + old_r + 1);
mu.rotate_right(new_r, old_r);
r.rotate_right(new_r, old_r);
if (enable_transform)
{
u.rotate_right(new_r, old_r);
if (enable_inverse_transform)
u_inv_t.rotate_right(new_r, old_r);
}
if (enable_int_gram)
{
if (gptr == nullptr)
{
throw std::runtime_error("Error: gptr is equal to the nullpointer.");
}
gptr->rotate_gram_right(new_r, old_r, d);
}
else
{
}
}
else if (new_r > old_r)
{
for (int i = old_r; i < n_known_rows; i++)
{
invalidate_gso_row(i, old_r);
}
rotate(gso_valid_cols.begin() + old_r, gso_valid_cols.begin() + old_r + 1,
gso_valid_cols.begin() + new_r + 1);
mu.rotate_left(old_r, new_r);
r.rotate_left(old_r, new_r);
if (enable_transform)
{
u.rotate_left(old_r, new_r);
if (enable_inverse_transform)
u_inv_t.rotate_left(old_r, new_r);
}
if (enable_int_gram)
{
if (old_r < n_known_rows - 1)
{
if (gptr == nullptr)
{
throw std::runtime_error("Error: gptr is equal to the nullpointer.");
}
// gptr->rotate_gram_left(old_r, new_r, n_known_rows);
gptr->rotate_gram_left(old_r, min(new_r, n_known_rows - 1), d);
}
}
else
{
}
if (new_r >= n_known_rows)
{
if (old_r < n_known_rows)
{
n_known_rows--;
n_source_rows = n_known_rows;
}
}
}
}
vector<Strategy> load_strategies_json(const std::string &filename)
{
json js;
{
std::ifstream fs(filename);
if (fs.fail())
throw std::runtime_error("Cannot open strategies file.");
fs >> js;
}
vector<Strategy> strategies;
for (auto it = js.begin(); it != js.end(); ++it)
{
const json &j_strat = *it;
const size_t block_size = j_strat["block_size"];
FPLLL_DEBUG_CHECK(block_size < 4096); // some arbitrary upper limit
// ensure to add this strategy in the right place
while (strategies.size() <= block_size)
{
strategies.emplace_back();
}
Strategy strategy;
strategy.block_size = block_size;
if (j_strat.find("preprocessing_block_sizes") != j_strat.end())
{
for (auto p_it = j_strat["preprocessing_block_sizes"].begin();
p_it != j_strat["preprocessing_block_sizes"].end(); ++p_it)
{
if ((*p_it).is_number())
{
strategy.preprocessing_block_sizes.emplace_back((*p_it).get<int>());
}
else
{
strategy.preprocessing_block_sizes.emplace_back((*p_it)["block_size"]);
}
}
}
if (j_strat.find("pruning_parameters") != j_strat.end())
{
for (auto p_it = j_strat["pruning_parameters"].begin();
p_it != j_strat["pruning_parameters"].end(); ++p_it)
{
const json &j_prun = *p_it;
PruningParams pruning;
pruning.gh_factor = j_prun[0];
for (auto c_it = j_prun[1].begin(); c_it != j_prun[1].end(); ++c_it)
{
double c = (*c_it);
pruning.coefficients.emplace_back(c);
}
pruning.expectation = j_prun[2];
// TODO: don't hardcode success probability as metric
pruning.metric = PRUNER_METRIC_PROBABILITY_OF_SHORTEST;
FPLLL_DEBUG_CHECK(pruning.expectation > 0.0 && pruning.expectation <= 1.0);
strategy.pruning_parameters.emplace_back(pruning);
}
}
strategies[block_size] = std::move(strategy);
}
// finally, we make sure all strategies are sound
for (auto it = strategies.begin(); it != strategies.end(); ++it)
{
if (it->pruning_parameters.size() == 0)
it->pruning_parameters.emplace_back(PruningParams());
}
return strategies;
}
