FLA_Error FLA_Svd_uv_var2_components( dim_t n_iter_max, dim_t k_accum, dim_t b_alg,
FLA_Obj A, FLA_Obj s, FLA_Obj U, FLA_Obj V,
double* dtime_bred, double* dtime_bsvd, double* dtime_appq,
double* dtime_qrfa, double* dtime_gemm )
{
FLA_Error r_val = FLA_SUCCESS;
FLA_Datatype dt;
FLA_Datatype dt_real;
FLA_Datatype dt_comp;
FLA_Obj T, S, rL, rR, d, e, G, H, RG, RH, W;
dim_t m_A, n_A;
dim_t min_m_n;
dim_t n_GH;
double crossover_ratio = 17.0 / 9.0;
double dtime_temp;
n_GH = k_accum;
m_A = FLA_Obj_length( A );
n_A = FLA_Obj_width( A );
min_m_n = FLA_Obj_min_dim( A );
dt = FLA_Obj_datatype( A );
dt_real = FLA_Obj_datatype_proj_to_real( A );
dt_comp = FLA_Obj_datatype_proj_to_complex( A );
// If the matrix is a scalar, then the SVD is easy.
if ( min_m_n == 1 )
{
FLA_Copy( A, s );
FLA_Set_to_identity( U );
FLA_Set_to_identity( V );
return FLA_SUCCESS;
}
// Create matrices to hold block Householder transformations.
FLA_Bidiag_UT_create_T( A, &T, &S );
// Create vectors to hold the realifying scalars.
FLA_Obj_create( dt, min_m_n, 1, 0, 0, &rL );
FLA_Obj_create( dt, min_m_n, 1, 0, 0, &rR );
// Create vectors to hold the diagonal and sub-diagonal.
FLA_Obj_create( dt_real, min_m_n, 1, 0, 0, &d );
FLA_Obj_create( dt_real, min_m_n-1, 1, 0, 0, &e );
// Create matrices to hold the left and right Givens scalars.
FLA_Obj_create( dt_comp, min_m_n-1, n_GH, 0, 0, &G );
FLA_Obj_create( dt_comp, min_m_n-1, n_GH, 0, 0, &H );
// Create matrices to hold the left and right Givens matrices.
FLA_Obj_create( dt_real, min_m_n, min_m_n, 0, 0, &RG );
FLA_Obj_create( dt_real, min_m_n, min_m_n, 0, 0, &RH );
FLA_Obj_create( dt, m_A, n_A, 0, 0, &W );
if ( m_A >= n_A )
{
if ( m_A < crossover_ratio * n_A )
{
dtime_temp = FLA_Clock();
{
// Reduce the matrix to bidiagonal form.
// Apply scalars to rotate elements on the sub-diagonal to the real domain.
// Extract the diagonal and sub-diagonal from A.
FLA_Bidiag_UT( A, T, S );
FLA_Bidiag_UT_realify( A, rL, rR );
FLA_Bidiag_UT_extract_diagonals( A, d, e );
}
*dtime_bred = FLA_Clock() - dtime_temp;
dtime_temp = FLA_Clock();
{
// Form U and V.
FLA_Bidiag_UT_form_U( A, T, U );
FLA_Bidiag_UT_form_V( A, S, V );
}
*dtime_appq = FLA_Clock() - dtime_temp;
// Apply the realifying scalars in rL and rR to U and V, respectively.
{
FLA_Obj UL, UR;
FLA_Obj VL, VR;
FLA_Part_1x2( U, &UL, &UR, min_m_n, FLA_LEFT );
FLA_Part_1x2( V, &VL, &VR, min_m_n, FLA_LEFT );
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_CONJUGATE, rL, UL );
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_NO_CONJUGATE, rR, VL );
}
dtime_temp = FLA_Clock();
{
// Perform a singular value decomposition on the bidiagonal matrix.
r_val = FLA_Bsvd_v_opt_var2( n_iter_max, d, e, G, H, RG, RH, W, U, V, b_alg );
}
*dtime_bsvd = FLA_Clock() - dtime_temp;
}
else // if ( crossover_ratio * n_A <= m_A )
{
FLA_Obj TQ, R;
FLA_Obj AT,
AB;
FLA_Obj UL, UR;
//FLA_QR_UT_create_T( A, &TQ );
FLA_Obj_create( dt, 32, n_A, 0, 0, &TQ );
dtime_temp = FLA_Clock();
{
// Perform a QR factorization on A and form Q in U.
FLA_QR_UT( A, TQ );
}
*dtime_qrfa = FLA_Clock() - dtime_temp;
dtime_temp = FLA_Clock();
{
FLA_QR_UT_form_Q( A, TQ, U );
}
*dtime_appq = FLA_Clock() - dtime_temp;
FLA_Obj_free( &TQ );
// Set the lower triangle of R to zero and then copy the upper
// triangle of A to R.
FLA_Part_2x1( A, &AT,
&AB, n_A, FLA_TOP );
FLA_Obj_create( dt, n_A, n_A, 0, 0, &R );
FLA_Setr( FLA_LOWER_TRIANGULAR, FLA_ZERO, R );
FLA_Copyr( FLA_UPPER_TRIANGULAR, AT, R );
dtime_temp = FLA_Clock();
{
// Reduce the matrix to bidiagonal form.
// Apply scalars to rotate elements on the superdiagonal to the real domain.
// Extract the diagonal and superdiagonal from A.
FLA_Bidiag_UT( R, T, S );
FLA_Bidiag_UT_realify( R, rL, rR );
FLA_Bidiag_UT_extract_diagonals( R, d, e );
}
*dtime_bred = FLA_Clock() - dtime_temp;
dtime_temp = FLA_Clock();
{
// Form V from right Householder vectors in upper triangle of R.
FLA_Bidiag_UT_form_V( R, S, V );
// Form U in R.
FLA_Bidiag_UT_form_U( R, T, R );
}
*dtime_appq += FLA_Clock() - dtime_temp;
// Apply the realifying scalars in rL and rR to U and V, respectively.
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_CONJUGATE, rL, R );
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_NO_CONJUGATE, rR, V );
dtime_temp = FLA_Clock();
{
// Perform a singular value decomposition on the bidiagonal matrix.
r_val = FLA_Bsvd_v_opt_var2( n_iter_max, d, e, G, H, RG, RH, W, R, V, b_alg );
}
*dtime_bsvd = FLA_Clock() - dtime_temp;
dtime_temp = FLA_Clock();
{
// Multiply R into U, storing the result in A and then copying back
// to U.
FLA_Part_1x2( U, &UL, &UR, n_A, FLA_LEFT );
FLA_Gemm( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_ONE, UL, R, FLA_ZERO, A );
FLA_Copy( A, UL );
}
*dtime_gemm = FLA_Clock() - dtime_temp;
FLA_Obj_free( &R );
}
}
else // if ( m_A < n_A )
{
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
}
// Copy the converged eigenvalues to the output vector.
FLA_Copy( d, s );
// Sort the singular values and singular vectors in descending order.
FLA_Sort_svd( FLA_BACKWARD, s, U, V );
FLA_Obj_free( &T );
FLA_Obj_free( &S );
FLA_Obj_free( &rL );
FLA_Obj_free( &rR );
FLA_Obj_free( &d );
FLA_Obj_free( &e );
FLA_Obj_free( &G );
FLA_Obj_free( &H );
FLA_Obj_free( &RG );
FLA_Obj_free( &RH );
FLA_Obj_free( &W );
return r_val;
}
FLA_Error FLA_Svd_uv_unb_var1( dim_t n_iter_max, FLA_Obj A, FLA_Obj s, FLA_Obj U, FLA_Obj V, dim_t k_accum, dim_t b_alg )
{
FLA_Error r_val = FLA_SUCCESS;
FLA_Datatype dt;
FLA_Datatype dt_real;
FLA_Datatype dt_comp;
FLA_Obj scale, T, S, rL, rR, d, e, G, H;
dim_t m_A, n_A;
dim_t min_m_n;
dim_t n_GH;
double crossover_ratio = 17.0 / 9.0;
n_GH = k_accum;
m_A = FLA_Obj_length( A );
n_A = FLA_Obj_width( A );
min_m_n = FLA_Obj_min_dim( A );
dt = FLA_Obj_datatype( A );
dt_real = FLA_Obj_datatype_proj_to_real( A );
dt_comp = FLA_Obj_datatype_proj_to_complex( A );
// Create matrices to hold block Householder transformations.
FLA_Bidiag_UT_create_T( A, &T, &S );
// Create vectors to hold the realifying scalars.
FLA_Obj_create( dt, min_m_n, 1, 0, 0, &rL );
FLA_Obj_create( dt, min_m_n, 1, 0, 0, &rR );
// Create vectors to hold the diagonal and sub-diagonal.
FLA_Obj_create( dt_real, min_m_n, 1, 0, 0, &d );
FLA_Obj_create( dt_real, min_m_n-1, 1, 0, 0, &e );
// Create matrices to hold the left and right Givens scalars.
FLA_Obj_create( dt_comp, min_m_n-1, n_GH, 0, 0, &G );
FLA_Obj_create( dt_comp, min_m_n-1, n_GH, 0, 0, &H );
// Create a real scaling factor.
FLA_Obj_create( dt_real, 1, 1, 0, 0, &scale );
// Compute a scaling factor; If none is needed, sigma will be set to one.
FLA_Svd_compute_scaling( A, scale );
// Scale the matrix if scale is non-unit.
if ( !FLA_Obj_equals( scale, FLA_ONE ) )
FLA_Scal( scale, A );
if ( m_A < crossover_ratio * n_A )
{
// Reduce the matrix to bidiagonal form.
// Apply scalars to rotate elements on the superdiagonal to the real domain.
// Extract the diagonal and superdiagonal from A.
FLA_Bidiag_UT( A, T, S );
FLA_Bidiag_UT_realify( A, rL, rR );
FLA_Bidiag_UT_extract_real_diagonals( A, d, e );
// Form U and V.
FLA_Bidiag_UT_form_U( A, T, U );
FLA_Bidiag_UT_form_V( A, S, V );
// Apply the realifying scalars in rL and rR to U and V, respectively.
{
FLA_Obj UL, UR;
FLA_Obj VL, VR;
FLA_Part_1x2( U, &UL, &UR, min_m_n, FLA_LEFT );
FLA_Part_1x2( V, &VL, &VR, min_m_n, FLA_LEFT );
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_CONJUGATE, rL, UL );
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_NO_CONJUGATE, rR, VL );
}
// Perform a singular value decomposition on the bidiagonal matrix.
r_val = FLA_Bsvd_v_opt_var1( n_iter_max, d, e, G, H, U, V, b_alg );
}
else // if ( crossover_ratio * n_A <= m_A )
{
FLA_Obj TQ, R;
FLA_Obj AT,
AB;
FLA_Obj UL, UR;
// Perform a QR factorization on A and form Q in U.
FLA_QR_UT_create_T( A, &TQ );
FLA_QR_UT( A, TQ );
FLA_QR_UT_form_Q( A, TQ, U );
FLA_Obj_free( &TQ );
// Set the lower triangle of R to zero and then copy the upper
// triangle of A to R.
FLA_Part_2x1( A, &AT,
&AB, n_A, FLA_TOP );
FLA_Obj_create( dt, n_A, n_A, 0, 0, &R );
FLA_Setr( FLA_LOWER_TRIANGULAR, FLA_ZERO, R );
FLA_Copyr( FLA_UPPER_TRIANGULAR, AT, R );
// Reduce the matrix to bidiagonal form.
// Apply scalars to rotate elements on the superdiagonal to the real domain.
// Extract the diagonal and superdiagonal from A.
FLA_Bidiag_UT( R, T, S );
FLA_Bidiag_UT_realify( R, rL, rR );
FLA_Bidiag_UT_extract_real_diagonals( R, d, e );
// Form V from right Householder vectors in upper triangle of R.
FLA_Bidiag_UT_form_V( R, S, V );
// Form U in R.
FLA_Bidiag_UT_form_U( R, T, R );
// Apply the realifying scalars in rL and rR to U and V, respectively.
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_CONJUGATE, rL, R );
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_NO_CONJUGATE, rR, V );
// Perform a singular value decomposition on the bidiagonal matrix.
r_val = FLA_Bsvd_v_opt_var1( n_iter_max, d, e, G, H, R, V, b_alg );
// Multiply R into U, storing the result in A and then copying back
// to U.
FLA_Part_1x2( U, &UL, &UR, n_A, FLA_LEFT );
FLA_Gemm( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_ONE, UL, R, FLA_ZERO, A );
FLA_Copy( A, UL );
FLA_Obj_free( &R );
}
// Copy the converged eigenvalues to the output vector.
FLA_Copy( d, s );
// Sort the singular values and singular vectors in descending order.
FLA_Sort_svd( FLA_BACKWARD, s, U, V );
// If the matrix was scaled, rescale the singular values.
if ( !FLA_Obj_equals( scale, FLA_ONE ) )
FLA_Inv_scal( scale, s );
FLA_Obj_free( &scale );
FLA_Obj_free( &T );
FLA_Obj_free( &S );
FLA_Obj_free( &rL );
FLA_Obj_free( &rR );
FLA_Obj_free( &d );
FLA_Obj_free( &e );
FLA_Obj_free( &G );
FLA_Obj_free( &H );
return r_val;
}
int main( int argc, char** argv ) {
FLA_Datatype datatype = TESTTYPE;
FLA_Obj A, A_flame, A_lapack, C;
int m;
FLA_Error init_result;
FLA_Obj TU, TV, U_flame, V_flame, d_flame, e_flame, B_flame;
FLA_Obj tauq, taup, d_lapack, e_lapack, U_lapack, V_lapack, W, B_lapack;
testtype *buff_tauq, *buff_taup, *buff_d_lapack, *buff_e_lapack,
*buff_W, *buff_A_lapack, *buff_U_lapack, *buff_V_lapack;
int lwork, info, is_flame;
if ( argc == 3 ) {
m = atoi(argv[1]);
is_flame = atoi(argv[2]);
} else {
fprintf(stderr, " \n");
fprintf(stderr, "Usage: %s m is_flame\n", argv[0]);
fprintf(stderr, " m : matrix length\n");
fprintf(stderr, " is_flame : 1 yes, 0 no\n");
fprintf(stderr, " \n");
return -1;
}
if ( m == 0 )
return 0;
FLA_Init_safe( &init_result );
fprintf( stdout, "lapack2flame: %d x %d: \n", m, m);
FLA_Obj_create( datatype, m, m, 0, 0, &A );
FLA_Random_matrix( A );
FLA_Obj_create_copy_of( FLA_NO_TRANSPOSE, A, &A_flame );
FLA_Obj_create_copy_of( FLA_NO_TRANSPOSE, A, &A_lapack );
FLA_Obj_create( datatype, m, m, 0, 0, &C );
FLA_Random_matrix( C );
if ( is_flame ) {
fprintf( stdout, " flame executed\n");
FLA_Bidiag_UT_create_T( A_flame, &TU, &TV );
FLA_Bidiag_UT( A_flame, TU, TV );
FLA_Obj_create_copy_of( FLA_NO_TRANSPOSE, A_flame, &U_flame );
FLA_Obj_create_copy_of( FLA_NO_TRANSPOSE, A_flame, &V_flame );
FLA_Bidiag_UT_form_U( U_flame, TU, U_flame );
FLA_Bidiag_UT_form_V( V_flame, TV, V_flame );
FLA_Obj_create( datatype, m, 1, 0, 0, &d_flame );
FLA_Obj_create( datatype, m - 1, 1, 0, 0, &e_flame );
FLA_Bidiag_UT_extract_diagonals( A_flame, d_flame, e_flame );
FLA_Obj_create( datatype, m, m, 0, 0, &B_flame ); FLA_Set( FLA_ZERO, B_flame );
{
FLA_Obj BTL, BTR, BBL, BBR;
FLA_Part_2x2( B_flame, &BTL, &BTR, &BBL, &BBR, 1,1, FLA_BL );
FLA_Set_diagonal_matrix( d_flame, B_flame );
FLA_Set_diagonal_matrix( e_flame, BTR );
}
if (1) {
fprintf( stdout, " - FLAME ----------\n");
FLA_Obj_fshow( stdout, " - Given A - ", A, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - A - ", A_flame, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - U - ", U_flame, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - V - ", V_flame, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - d - ", d_flame, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - e - ", e_flame, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - B - ", B_flame, "% 6.4e", "------");
}
} else {
fprintf( stdout, " lapack executed\n");
FLA_Obj_create( datatype, m, 1, 0, 0, &tauq );
FLA_Obj_create( datatype, m, 1, 0, 0, &taup );
FLA_Obj_create( datatype, m, 1, 0, 0, &d_lapack );
FLA_Obj_create( datatype, m - 1, 1, 0, 0, &e_lapack );
buff_A_lapack = (testtype*)FLA_Obj_buffer_at_view( A_lapack );
buff_tauq = (testtype*)FLA_Obj_buffer_at_view( tauq );
buff_taup = (testtype*)FLA_Obj_buffer_at_view( taup );
buff_d_lapack = (testtype*)FLA_Obj_buffer_at_view( d_lapack );
buff_e_lapack = (testtype*)FLA_Obj_buffer_at_view( e_lapack );
lwork = 32*m;
FLA_Obj_create( datatype, lwork, 1, 0, 0, &W );
buff_W = (testtype*)FLA_Obj_buffer_at_view( W );
sgebrd_( &m, &m,
buff_A_lapack, &m,
buff_d_lapack,
buff_e_lapack,
buff_tauq,
buff_taup,
buff_W,
&lwork,
&info );
FLA_Obj_create( datatype, m, m, 0, 0, &U_lapack );
FLA_Obj_create( datatype, m, m, 0, 0, &V_lapack );
FLA_Copy( A_lapack, U_lapack );
FLA_Copy( A_lapack, V_lapack );
buff_U_lapack = (testtype*)FLA_Obj_buffer_at_view( U_lapack );
buff_V_lapack = (testtype*)FLA_Obj_buffer_at_view( V_lapack );
sorgbr_( "Q", &m, &m, &m,
buff_U_lapack, &m,
buff_tauq,
buff_W,
&lwork,
&info );
sorgbr_( "P", &m, &m, &m,
buff_V_lapack, &m,
buff_taup,
buff_W,
&lwork,
&info );
FLA_Obj_create( datatype, m, m, 0, 0, &B_lapack ); FLA_Set( FLA_ZERO, B_lapack );
{
FLA_Obj BTL, BTR, BBL, BBR;
FLA_Part_2x2( B_lapack, &BTL, &BTR, &BBL, &BBR, 1,1, FLA_BL );
FLA_Set_diagonal_matrix( d_lapack, B_lapack );
FLA_Set_diagonal_matrix( e_lapack, BTR );
}
FLA_Obj_free( &W );
if (1) {
fprintf( stdout, " - LAPACK ----------\n");
FLA_Obj_fshow( stdout, " - Given A - ", A, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - A - ", A_lapack, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - U - ", U_lapack, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - V - ", V_lapack, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - d - ", d_lapack, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - e - ", e_lapack, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - B - ", B_lapack, "% 6.4e", "------");
}
}
{
testtype dummy;
int zero = 0, one = 1;
FLA_Obj D_lapack;
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, A, &D_lapack ); FLA_Set( FLA_ZERO, D_lapack );
if ( is_flame ) {
buff_d_lapack = (testtype*)FLA_Obj_buffer_at_view( d_flame );
buff_e_lapack = (testtype*)FLA_Obj_buffer_at_view( e_flame );
buff_U_lapack = (testtype*)FLA_Obj_buffer_at_view( U_flame );
buff_V_lapack = (testtype*)FLA_Obj_buffer_at_view( V_flame );
}
FLA_Obj_create( datatype, 4*m, 1, 0, 0, &W );
buff_W = (testtype*)FLA_Obj_buffer_at_view( W );
sbdsqr_( "U", &m, &m, &m, &zero,
buff_d_lapack, buff_e_lapack,
buff_V_lapack, &m,
buff_U_lapack, &m,
&dummy, &one,
buff_W, &info );
FLA_Obj_free( &W );
if (info != 0)
printf( " Error info = %d\n", info );
if ( is_flame )
FLA_Set_diagonal_matrix( d_flame, D_lapack );
else
FLA_Set_diagonal_matrix( d_lapack, D_lapack );
if ( is_flame ) {
fprintf( stdout, " - FLAME ----------\n");
FLA_Obj_fshow( stdout, " - U - ", U_flame, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - V - ", V_flame, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - d - ", d_flame, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - e - ", e_flame, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - D - ", D_lapack, "% 6.4e", "------");
} else {
fprintf( stdout, " - LAPACK ----------\n");
FLA_Obj_fshow( stdout, " - U - ", U_lapack, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - V - ", V_lapack, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - d - ", d_lapack, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - e - ", e_lapack, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - D - ", D_lapack, "% 6.4e", "------");
}
FLA_Obj_free( &D_lapack );
}
if ( is_flame ) {
FLA_Obj_free( &TU );
FLA_Obj_free( &TV );
FLA_Obj_free( &U_flame );
FLA_Obj_free( &V_flame );
FLA_Obj_free( &d_flame );
FLA_Obj_free( &e_flame );
FLA_Obj_free( &B_flame );
} else {
FLA_Obj_free( &tauq );
FLA_Obj_free( &taup );
FLA_Obj_free( &d_lapack );
FLA_Obj_free( &e_lapack );
FLA_Obj_free( &U_lapack );
FLA_Obj_free( &V_lapack );
FLA_Obj_free( &B_lapack );
}
FLA_Obj_free( &A );
FLA_Obj_free( &A_flame );
FLA_Obj_free( &A_lapack );
FLA_Obj_free( &C );
FLA_Finalize_safe( init_result );
}
