int productQ(char side, char trans, ScalarArray<T>* qr, T* tau, ScalarArray<T>* c) {
DECLARE_CONTEXT;
int m = c->rows;
int n = c->cols;
int k = qr->cols;
T* a = qr->m;
assert((side == 'L') ? qr->rows == m : qr->rows == n);
int ldq = qr->lda;
int ldc = c->lda;
int info;
int workSize;
T workSize_req;
{
size_t _m = m, _n = n, _k = k;
size_t muls = 2 * _m * _n * _k - _n * _k * _k + 2 * _n * _k;
size_t adds = 2 * _m * _n * _k - _n * _k * _k + _n * _k;
increment_flops(Multipliers<T>::mul * muls + Multipliers<T>::add * adds);
}
info = proxy_lapack_convenience::or_un_mqr(side, trans, m, n, k, a, ldq, tau, c->m, ldc, &workSize_req, -1);
HMAT_ASSERT(!info);
workSize = (int) hmat::real(workSize_req) + 1;
T* work = new T[workSize];
HMAT_ASSERT(work);
info = proxy_lapack_convenience::or_un_mqr(side, trans, m, n, k, a, ldq, tau, c->m, ldc, work, workSize);
HMAT_ASSERT(!info);
delete[] work;
return 0;
}
template<> int sddCall<hmat::C_t>(char jobz, int m, int n, hmat::C_t* a, int lda,
double* sigma, hmat::C_t* u, int ldu, hmat::C_t* vt, int ldvt) {
int result;
int workSize;
hmat::C_t workSize_C;
int p = min(m, n);
float* sigmaFloat = new float[p];
int* iwork = new int[8*p];
// We request the right size for WORK
result = proxy_lapack::gesdd(jobz, m, n, a, lda, sigmaFloat, u, ldu, vt, ldvt, &workSize_C, -1, iwork);
HMAT_ASSERT(!result);
workSize = (int) workSize_C.real() + 1;
hmat::C_t* work = new hmat::C_t[workSize];
HMAT_ASSERT(work) ;
result = proxy_lapack::gesdd(jobz, m, n, a, lda, sigmaFloat, u, ldu, vt, ldvt, work, workSize, iwork);
HMAT_ASSERT(!result);
delete[] work;
delete[] iwork;
for (int i = 0; i < p; i++) {
sigma[i] = sigmaFloat[i];
}
delete[] sigmaFloat;
return result;
}
template<typename T> T* qrDecomposition(ScalarArray<T>* m) {
DECLARE_CONTEXT;
// SUBROUTINE DGEQRF( M, N, A, LDA, TAU, WORK, LWORK, INFO )
int rows = m->rows;
int cols = m->cols;
T* tau = (T*) calloc(min(rows, cols), sizeof(T));
{
size_t mm = max(rows, cols);
size_t n = min(rows, cols);
size_t multiplications = mm * n * n - (n * n * n) / 3 + mm * n + (n * n) / 2 + (29 * n) / 6;
size_t additions = mm * n * n + (n * n * n) / 3 + 2 * mm * n - (n * n) / 2 + (5 * n) / 6;
increment_flops(Multipliers<T>::mul * multiplications + Multipliers<T>::add * additions);
}
int info;
int workSize;
T workSize_S;
// int info = LAPACKE_sgeqrf(LAPACK_COL_MAJOR, rows, cols, m->m, rows, *tau);
info = proxy_lapack::geqrf(rows, cols, m->m, rows, tau, &workSize_S, -1);
HMAT_ASSERT(!info);
workSize = (int) hmat::real(workSize_S) + 1;
T* work = new T[workSize];// TODO Mettre dans la pile ??
HMAT_ASSERT(work) ;
info = proxy_lapack::geqrf(rows, cols, m->m, rows, tau, work, workSize);
delete[] work;
HMAT_ASSERT(!info);
return tau;
}
void ClusterData::moveDoF(int index, ClusterData* right)
{
HMAT_ASSERT(offset_ + size_ == right->offset_ );
HMAT_ASSERT(index >= offset_);
HMAT_ASSERT(index < offset_ + size_);
// Swap degree of freedom and the last element of this bucket
std::swap(indices()[index], indices()[offset_ + size_ - 1]);
// Make current index set smaller ...
size_ -= 1;
// ... and expand right sibling
right->offset_ -= 1;
right->size_ += 1;
}
template<> int svdCall<hmat::C_t>(char jobu, char jobv, int m, int n, hmat::C_t* a,
int lda, double* sigma, hmat::C_t* u, int ldu, hmat::C_t* vt,
int ldvt) {
int result;
int workSize;
hmat::C_t workSize_C;
int p = min(m, n);
float* sigmaFloat = new float[p];
// We request the right size for WORK
result = proxy_lapack::gesvd(jobu, jobv, m, n, a, lda, sigmaFloat, u, ldu, vt, ldvt, &workSize_C, -1);
if(result != 0)
throw hmat::LapackException("gesvd", result);
workSize = (int) workSize_C.real() + 1;
hmat::C_t* work = new hmat::C_t[workSize];
HMAT_ASSERT(work) ;
result = proxy_lapack::gesvd(jobu, jobv, m, n, a, lda, sigmaFloat, u, ldu, vt, ldvt, work, workSize);
if(result != 0)
throw hmat::LapackException("gesvd", result);
delete[] work;
for (int i = 0; i < p; i++) {
sigma[i] = sigmaFloat[i];
}
delete[] sigmaFloat;
return result;
}
template<> int sddCall<hmat::Z_t>(char jobz, int m, int n, hmat::Z_t* a, int lda,
double* sigma, hmat::Z_t* u, int ldu, hmat::Z_t* vt, int ldvt) {
int result;
int workSize;
hmat::Z_t workSize_Z;
int* iwork = new int[8*min(m,n)];
// We request the right size for WORK
result = proxy_lapack::gesdd(jobz, m, n, a, lda, sigma, u, ldu, vt, ldvt, &workSize_Z, -1, iwork);
HMAT_ASSERT(!result);
workSize = (int) workSize_Z.real() + 1;
hmat::Z_t* work = new hmat::Z_t[workSize];
HMAT_ASSERT(work) ;
result = proxy_lapack::gesdd(jobz, m, n, a, lda, sigma, u, ldu, vt, ldvt, work, workSize, iwork);
HMAT_ASSERT(!result);
delete[] work;
delete[] iwork;
return result;
}
template<typename T> int truncatedSvd(ScalarArray<T>* m, ScalarArray<T>** u, Vector<double>** sigma, ScalarArray<T>** vt) {
DECLARE_CONTEXT;
static char * useGESSD = getenv("HMAT_GESSD");
if(useGESSD)
return truncatedSdd(m, u, sigma, vt);
// Allocate free space for U, S, V
int rows = m->rows;
int cols = m->cols;
int p = min(rows, cols);
*u = new ScalarArray<T>(rows, p);
*sigma = new Vector<double>(p);
*vt = new ScalarArray<T>(p, cols);
assert(m->lda >= m->rows);
char jobz = 'S';
int mm = rows;
int n = cols;
T* a = m->m;
int lda = rows;
int info;
{
const size_t _m = mm, _n = n;
// Warning: These quantities are a rough approximation.
// What's wrong with these estimates:
// - Golub only gives 14 * M*N*N + 8 N*N*N
// - This is for real numbers
// - We assume the same number of * and +
size_t adds = 7 * _m * _n * _n + 4 * _n * _n * _n;
size_t muls = 7 * _m * _n * _n + 4 * _n * _n * _n;
increment_flops(Multipliers<T>::add * adds + Multipliers<T>::mul * muls);
}
info = svdCall<T>(jobz, jobz, mm, n, a, lda, (*sigma)->m, (*u)->m,
(*u)->lda, (*vt)->m, (*vt)->lda);
if (info) {
cerr << "Erreur dans xGESVD: " << info << endl;
}
HMAT_ASSERT(!info);
return info;
}
template<> int svdCall<hmat::D_t>(char jobu, char jobv, int m, int n, hmat::D_t* a,
int lda, double* sigma, hmat::D_t* u, int ldu, hmat::D_t* vt,
int ldvt) {
int workSize;
hmat::D_t workSize_D;
int result;
// We request the right size for WORK
result = proxy_lapack::gesvd(jobu, jobv, m, n, a, lda, sigma, u, ldu, vt, ldvt, &workSize_D, -1);
if(result != 0)
throw hmat::LapackException("gesvd", result);
workSize = (int) workSize_D + 1;
hmat::D_t* work = new hmat::D_t[workSize];
HMAT_ASSERT(work) ;
result = proxy_lapack::gesvd(jobu, jobv, m, n, a, lda, sigma, u, ldu, vt, ldvt, work, workSize);
if(result != 0)
throw hmat::LapackException("gesvd", result);
delete[] work;
return result;
}