int phonopy_pinv_libflame(double *matrix,
double *eigvals,
const int size,
const double cutoff)
{
FLA_Obj A, B, l;
/* FLA_Obj C; */
double *inv_eigvals;
int i;
inv_eigvals = (double*)malloc(sizeof(double) * size);
FLA_Init();
FLA_Obj_create_without_buffer(FLA_DOUBLE, size, size, &A);
FLA_Obj_attach_buffer(matrix, 0, 0, &A);
FLA_Obj_create_without_buffer(FLA_DOUBLE, size, 1, &l);
FLA_Obj_attach_buffer(eigvals, 0, 0, &l);
/* Eigensolver */
FLA_Obj_create_copy_of(FLA_NO_TRANSPOSE, A, &B);
FLA_Hevd(FLA_EVD_WITH_VECTORS, FLA_LOWER_TRIANGULAR, B, l);
/* SVD */
/* FLA_Obj_create(FLA_DOUBLE, size, size, 0, 0, &B); */
/* use U */
/* FLA_Svd(FLA_SVD_VECTORS_ALL, FLA_SVD_VECTORS_NONE, A, l, B, C); */
/* use V */
/* FLA_Svd(FLA_SVD_VECTORS_NONE, FLA_SVD_VECTORS_ALL, A, l, C, B); */
FLA_Obj_free_without_buffer(&l);
for (i = 0; i < size; i++) {
if (eigvals[i] < cutoff) {
inv_eigvals[i] = 0;
} else {
inv_eigvals[i] = 1.0 / sqrt(eigvals[i]);
}
}
FLA_Obj_create_without_buffer(FLA_DOUBLE, size, 1, &l);
FLA_Obj_attach_buffer(inv_eigvals, 0, 0, &l);
FLA_Apply_diag_matrix(FLA_RIGHT, FLA_NO_CONJUGATE, l, B);
FLA_Syrk(FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_ONE, B, FLA_ZERO, A);
FLA_Symmetrize(FLA_LOWER_TRIANGULAR, A);
FLA_Obj_free_without_buffer(&A);
FLA_Obj_free_without_buffer(&l);
FLA_Obj_free(&B);
FLA_Finalize();
free(inv_eigvals);
return 0;
}
FLA_Error FLA_Hevd_lv_var4_components( dim_t n_iter_max, FLA_Obj A, FLA_Obj l, dim_t k_accum, dim_t b_alg,
double* dtime_tred, double* dtime_tevd, double* dtime_appq )
{
FLA_Error r_val = FLA_SUCCESS;
FLA_Uplo uplo = FLA_LOWER_TRIANGULAR;
FLA_Datatype dt;
FLA_Datatype dt_real;
FLA_Datatype dt_comp;
FLA_Obj T, r, d, e, G, R, W;
FLA_Obj d0, e0, ls, pu;
dim_t mn_A;
dim_t n_G = k_accum;
double dtime_temp;
mn_A = FLA_Obj_length( A );
dt = FLA_Obj_datatype( A );
dt_real = FLA_Obj_datatype_proj_to_real( A );
dt_comp = FLA_Obj_datatype_proj_to_complex( A );
*dtime_tred = 1;
*dtime_tevd = 1;
*dtime_appq = 1;
// If the matrix is a scalar, then the EVD is easy.
if ( mn_A == 1 )
{
FLA_Copy( A, l );
FLA_Set( FLA_ONE, A );
return FLA_SUCCESS;
}
// Create a matrix to hold block Householder transformations.
FLA_Tridiag_UT_create_T( A, &T );
// Create a vector to hold the realifying scalars.
FLA_Obj_create( dt, mn_A, 1, 0, 0, &r );
// Create vectors to hold the diagonal and sub-diagonal.
FLA_Obj_create( dt_real, mn_A, 1, 0, 0, &d );
FLA_Obj_create( dt_real, mn_A-1, 1, 0, 0, &e );
FLA_Obj_create( dt_real, mn_A, 1, 0, 0, &d0 );
FLA_Obj_create( dt_real, mn_A-1, 1, 0, 0, &e0 );
FLA_Obj_create( dt_real, mn_A, 1, 0, 0, &pu );
FLA_Obj_create( FLA_INT, mn_A, 1, 0, 0, &ls );
FLA_Obj_create( dt_comp, mn_A-1, n_G, 0, 0, &G );
FLA_Obj_create( dt_real, mn_A, mn_A, 0, 0, &R );
FLA_Obj_create( dt, mn_A, mn_A, 0, 0, &W );
dtime_temp = FLA_Clock();
{
// Reduce the matrix to tridiagonal form.
FLA_Tridiag_UT( uplo, A, T );
}
*dtime_tred = FLA_Clock() - dtime_temp;
// Apply scalars to rotate elements on the sub-diagonal to the real domain.
FLA_Tridiag_UT_realify( uplo, A, r );
// Extract the diagonal and sub-diagonal from A.
FLA_Tridiag_UT_extract_diagonals( uplo, A, d, e );
dtime_temp = FLA_Clock();
{
// Form Q, overwriting A.
FLA_Tridiag_UT_form_Q( uplo, A, T );
}
*dtime_appq = FLA_Clock() - dtime_temp;
// Apply the scalars in r to Q.
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_CONJUGATE, r, A );
// Find the eigenvalues only.
FLA_Copy( d, d0 ); FLA_Copy( e, e0 );
//r_val = FLA_Tevd_n_opt_var1( n_iter_max, d0, e0, G, A );
{
int info;
double* buff_d = FLA_DOUBLE_PTR( d0 );
double* buff_e = FLA_DOUBLE_PTR( e0 );
dsterf_( &mn_A, buff_d, buff_e, &info );
}
FLA_Sort( FLA_FORWARD, d0 );
FLA_Set( FLA_ZERO, ls );
FLA_Set( FLA_ZERO, pu );
dtime_temp = FLA_Clock();
{
// Perform an eigenvalue decomposition on the tridiagonal matrix.
r_val = FLA_Tevd_v_opt_var4( n_iter_max, d, e, d0, ls, pu, G, R, W, A, b_alg );
}
*dtime_tevd = FLA_Clock() - dtime_temp;
// Copy the converged eigenvalues to the output vector.
FLA_Copy( d, l );
// Sort the eigenvalues and eigenvectors in ascending order.
FLA_Sort_evd( FLA_FORWARD, l, A );
FLA_Obj_free( &T );
FLA_Obj_free( &r );
FLA_Obj_free( &d );
FLA_Obj_free( &e );
FLA_Obj_free( &d0 );
FLA_Obj_free( &pu );
FLA_Obj_free( &e0 );
FLA_Obj_free( &ls );
FLA_Obj_free( &G );
FLA_Obj_free( &R );
FLA_Obj_free( &W );
return r_val;
}
int main(int argc, char *argv[])
{
int
m_input,
m,
p_first, p_last, p_inc,
p,
k_accum,
b_alg,
n_iter_max,
variant,
n_repeats,
i,
n_variants = 2;
char *colors = "brkgmcbrkg";
char *ticks = "o+*xso+*xs";
char m_dim_desc[14];
char m_dim_tag[10];
double max_gflops=6.0;
double
dtime,
gflops,
diff1, diff2;
FLA_Datatype datatype, dt_real;
FLA_Obj
A, l, Q, Ql, TT, r, d, e, A_orig, G, R, W2, de, alpha;
FLA_Init();
fprintf( stdout, "%c number of repeats:", '%' );
scanf( "%d", &n_repeats );
fprintf( stdout, "%c %d\n", '%', n_repeats );
fprintf( stdout, "%c enter n_iter_max (per eigenvalue): ", '%' );
scanf( "%d", &n_iter_max );
fprintf( stdout, "%c %d\n", '%', n_iter_max );
fprintf( stdout, "%c enter number of sets of Givens rotations to accumulate:", '%' );
scanf( "%d", &k_accum );
fprintf( stdout, "%c %d\n", '%', k_accum );
fprintf( stdout, "%c enter blocking size for application of G:", '%' );
scanf( "%d", &b_alg );
fprintf( stdout, "%c %d\n", '%', b_alg );
fprintf( stdout, "%c enter problem size first, last, inc:", '%' );
scanf( "%d%d%d", &p_first, &p_last, &p_inc );
fprintf( stdout, "%c %d %d %d\n", '%', p_first, p_last, p_inc );
fprintf( stdout, "%c enter m (-1 means bind to problem size): ", '%' );
scanf( "%d", &m_input );
fprintf( stdout, "%c %d\n", '%', m_input );
fprintf( stdout, "\n" );
if ( m_input > 0 ) {
sprintf( m_dim_desc, "m = %d", m_input );
sprintf( m_dim_tag, "m%dc", m_input);
}
else if( m_input < -1 ) {
sprintf( m_dim_desc, "m = p/%d", -m_input );
sprintf( m_dim_tag, "m%dp", -m_input );
}
else if( m_input == -1 ) {
sprintf( m_dim_desc, "m = p" );
sprintf( m_dim_tag, "m%dp", 1 );
}
for ( p = p_first, i = 1; p <= p_last; p += p_inc, i += 1 )
{
m = m_input;
if( m < 0 ) m = p / abs(m_input);
//datatype = FLA_FLOAT;
//datatype = FLA_DOUBLE;
//datatype = FLA_COMPLEX;
datatype = FLA_DOUBLE_COMPLEX;
FLA_Obj_create( datatype, m, m, 0, 0, &A );
FLA_Obj_create( datatype, m, m, 0, 0, &A_orig );
FLA_Obj_create( datatype, m, m, 0, 0, &Q );
FLA_Obj_create( datatype, m, m, 0, 0, &Ql );
FLA_Obj_create( datatype, m, 1, 0, 0, &r );
FLA_Obj_create( datatype, m, m, 0, 0, &W2 );
FLA_Obj_create( datatype, m-1, k_accum, 0, 0, &G );
dt_real = FLA_Obj_datatype_proj_to_real( A );
FLA_Obj_create( dt_real, m, 1, 0, 0, &l );
FLA_Obj_create( dt_real, m, 1, 0, 0, &d );
FLA_Obj_create( dt_real, m-1, 1, 0, 0, &e );
FLA_Obj_create( dt_real, m, m, 0, 0, &R );
FLA_Obj_create( dt_real, 1, 1, 0, 0, &alpha );
*FLA_DOUBLE_PTR( alpha ) = 1.0 / ( sqrt( sqrt( (double) m ) ) );
FLA_Random_unitary_matrix( Q );
//FLA_Fill_with_uniform_dist( FLA_ONE, l );
//FLA_Fill_with_inverse_dist( FLA_ONE, l );
FLA_Fill_with_geometric_dist( alpha, l );
{
FLA_Copy( Q, Ql );
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_NO_CONJUGATE, l, Ql );
FLA_Gemm( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
FLA_ONE, Ql, Q, FLA_ZERO, A );
FLA_Triangularize( FLA_LOWER_TRIANGULAR, FLA_NONUNIT_DIAG, A );
FLA_Copy( A, A_orig );
}
FLA_Set( FLA_ZERO, l );
FLA_Set( FLA_ZERO, Q );
FLA_Tridiag_UT_create_T( A, &TT );
FLA_Tridiag_UT( FLA_LOWER_TRIANGULAR, A, TT );
FLA_Tridiag_UT_realify( FLA_LOWER_TRIANGULAR, A, r );
FLA_Tridiag_UT_extract_diagonals( FLA_LOWER_TRIANGULAR, A, d, e );
FLA_Tridiag_UT_form_Q( FLA_LOWER_TRIANGULAR, A, TT );
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_CONJUGATE, r, A );
FLA_Obj_free( &TT );
time_Tevd_v( 0, FLA_ALG_REFERENCE, n_repeats, m, k_accum, b_alg, n_iter_max,
A_orig, d, e, G, R, W2, A, l, &dtime, &diff1, &diff2, &gflops );
fprintf( stdout, "data_REFq( %d, 1:3 ) = [ %d %6.3lf %9.2e %6.2le %6.2le ]; \n", i, p, gflops, dtime, diff1, diff2 );
fflush( stdout );
for ( variant = 1; variant <= n_variants; variant++ ){
fprintf( stdout, "data_var%d( %d, 1:3 ) = [ %d ", variant, i, p );
fflush( stdout );
time_Tevd_v( variant, FLA_ALG_UNB_OPT, n_repeats, m, k_accum, b_alg, n_iter_max,
A_orig, d, e, G, R, W2, A, l, &dtime, &diff1, &diff2, &gflops );
fprintf( stdout, "%6.3lf %9.2e %6.2le %6.2le ", gflops, dtime, diff1, diff2 );
fflush( stdout );
fprintf( stdout, "];\n" );
fflush( stdout );
}
fprintf( stdout, "\n" );
FLA_Obj_free( &A );
FLA_Obj_free( &A_orig );
FLA_Obj_free( &Q );
FLA_Obj_free( &Ql );
FLA_Obj_free( &G );
FLA_Obj_free( &W2 );
FLA_Obj_free( &r );
FLA_Obj_free( &l );
FLA_Obj_free( &d );
FLA_Obj_free( &e );
FLA_Obj_free( &R );
FLA_Obj_free( &alpha );
}
/*
fprintf( stdout, "figure;\n" );
fprintf( stdout, "plot( data_REF( :,1 ), data_REF( :, 2 ), '-' ); \n" );
fprintf( stdout, "hold on;\n" );
for ( i = 1; i <= n_variants; i++ ) {
fprintf( stdout, "plot( data_var%d( :,1 ), data_var%d( :, 2 ), '%c:%c' ); \n",
i, i, colors[ i-1 ], ticks[ i-1 ] );
fprintf( stdout, "plot( data_var%d( :,1 ), data_var%d( :, 4 ), '%c-.%c' ); \n",
i, i, colors[ i-1 ], ticks[ i-1 ] );
}
fprintf( stdout, "legend( ... \n" );
fprintf( stdout, "'Reference', ... \n" );
for ( i = 1; i < n_variants; i++ )
fprintf( stdout, "'unb\\_var%d', 'blk\\_var%d', ... \n", i, i );
fprintf( stdout, "'unb\\_var%d', 'blk\\_var%d' ); \n", i, i );
fprintf( stdout, "xlabel( 'problem size p' );\n" );
fprintf( stdout, "ylabel( 'GFLOPS/sec.' );\n" );
fprintf( stdout, "axis( [ 0 %d 0 %.2f ] ); \n", p_last, max_gflops );
fprintf( stdout, "title( 'FLAME Hevd_lv performance (%s, %s)' );\n",
m_dim_desc, n_dim_desc );
fprintf( stdout, "print -depsc tridiag_%s_%s.eps\n", m_dim_tag, n_dim_tag );
fprintf( stdout, "hold off;\n");
fflush( stdout );
*/
FLA_Finalize( );
return 0;
}
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;
}
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;
}
void time_Tevd_v(
int variant, int type, int n_repeats, int m, int k_accum, int b_alg, int n_iter_max,
FLA_Obj A_orig, FLA_Obj d, FLA_Obj e, FLA_Obj G, FLA_Obj R, FLA_Obj W, FLA_Obj A, FLA_Obj l,
double *dtime, double *diff1, double* diff2, double *gflops )
{
int irep;
double
k, dtime_old = 1.0e9;
FLA_Obj
A_save, G_save, d_save, e_save;
if (
//( variant == 0 ) ||
//( variant == 1 && type == FLA_ALG_UNB_OPT ) ||
//( variant == 2 && type == FLA_ALG_UNB_OPT ) ||
FALSE
)
{
*dtime = 0.0;
*gflops = 0.0;
*diff1 = 0.0;
*diff2 = 0.0;
return;
}
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, A, &A_save );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, G, &G_save );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, d, &d_save );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, e, &e_save );
FLA_Copy_external( A, A_save );
FLA_Copy_external( G, G_save );
FLA_Copy_external( d, d_save );
FLA_Copy_external( e, e_save );
for ( irep = 0 ; irep < n_repeats; irep++ ){
FLA_Copy_external( A_save, A );
FLA_Copy_external( G_save, G );
FLA_Copy_external( d_save, d );
FLA_Copy_external( e_save, e );
*dtime = FLA_Clock();
switch( variant ){
case 0:
REF_Tevd_v( d, e, A );
break;
// Time variant 1
case 1:
{
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Tevd_v_opt_var1( n_iter_max, d, e, G, A, b_alg );
break;
}
break;
}
// Time variant 2
case 2:
{
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Tevd_v_opt_var2( n_iter_max, d, e, G, R, W, A, b_alg );
break;
}
break;
}
}
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
{
FLA_Obj V, A_rev_evd, norm, eye;
FLA_Copy( d, l );
//FLA_Obj_show( "A_save", A_save, "%9.2e + %9.2e ", "" );
//FLA_Obj_show( "A_evd", A, "%9.2e + %9.2e ", "" );
FLA_Sort_evd( FLA_FORWARD, l, A );
FLA_Obj_create_copy_of( FLA_NO_TRANSPOSE, A, &V );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, A, &A_rev_evd );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, A, &eye );
FLA_Obj_create( FLA_Obj_datatype_proj_to_real( A ), 1, 1, 0, 0, &norm );
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_NO_CONJUGATE, l, A );
FLA_Gemm( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
FLA_ONE, A, V, FLA_ZERO, A_rev_evd );
FLA_Triangularize( FLA_LOWER_TRIANGULAR, FLA_NONUNIT_DIAG, A_rev_evd );
/*
FLA_Gemm( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_ONE, A, D, FLA_ZERO, A_rev_evd );
FLA_Copy( A_rev_evd, D );
FLA_Gemm( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
FLA_ONE, D, V, FLA_ZERO, A_rev_evd );
FLA_Triangularize( FLA_LOWER_TRIANGULAR, FLA_NONUNIT_DIAG, A_rev_evd );
*/
//FLA_Obj_show( "A_rev_evd", A_rev_evd, "%9.2e + %9.2e ", "" );
FLA_Axpy( FLA_MINUS_ONE, A_orig, A_rev_evd );
FLA_Norm_frob( A_rev_evd, norm );
FLA_Obj_extract_real_scalar( norm, diff1 );
//*diff = FLA_Max_elemwise_diff( A_orig, A_rev_evd );
FLA_Set_to_identity( eye );
FLA_Copy( V, A_rev_evd );
FLA_Gemm( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
FLA_ONE, V, A_rev_evd, FLA_MINUS_ONE, eye );
FLA_Norm_frob( eye, norm );
FLA_Obj_extract_real_scalar( norm, diff2 );
/*
FLA_Obj_free( &EL );
FLA_Obj_free( &EU );
FLA_Obj_free( &D );
FLA_Obj_free( &dc );
FLA_Obj_free( &ec );
*/
FLA_Obj_free( &V );
FLA_Obj_free( &A_rev_evd );
FLA_Obj_free( &eye );
FLA_Obj_free( &norm );
}
k = 2.00;
if ( FLA_Obj_is_complex( A ) )
{
*gflops = (
( 4.5 * k * m * m ) +
2.0 * ( 3.0 * k * m * m * m ) ) /
dtime_old / 1e9;
}
else
{
*gflops = (
( 4.5 * k * m * m ) +
1.0 * ( 3.0 * k * m * m * m ) ) /
dtime_old / 1e9;
}
*dtime = dtime_old;
FLA_Copy_external( A_save, A );
FLA_Copy_external( G_save, G );
FLA_Copy_external( d_save, d );
FLA_Copy_external( e_save, e );
FLA_Obj_free( &A_save );
FLA_Obj_free( &G_save );
FLA_Obj_free( &d_save );
FLA_Obj_free( &e_save );
}
void time_Hevd_lv_components(
int variant, int type, int n_repeats, int m, int n_iter_max, int k_accum, int b_alg,
FLA_Obj A, FLA_Obj l,
double* dtime, double* diff1, double* diff2, double* gflops,
double* dtime_tred, double* gflops_tred,
double* dtime_tevd, double* gflops_tevd,
double* dtime_appq, double* gflops_appq, int* k_perf )
{
int i;
double k;
double dtime_save = 1.0e9;
double dtime_tred_save = 1.0e9;
double dtime_tevd_save = 1.0e9;
double dtime_appq_save = 1.0e9;
double flops_tred;
double flops_tevd;
double flops_appq;
double mult_tred;
double mult_tevd;
double mult_appq;
FLA_Obj A_save, Z;
if (
( variant == -3 ) ||
( variant == -4 ) ||
( variant == -5 ) ||
//( variant == 0 ) ||
//( variant == -1 ) ||
//( variant == -2 ) ||
//( variant == 1 ) ||
//( variant == 2 ) ||
//( variant == 3 ) ||
//( variant == 4 ) ||
FALSE
)
{
*gflops = 0.0;
*dtime = 0.0;
*diff1 = 0.0;
*diff2 = 0.0;
*dtime_tred = 0.0;
*dtime_tevd = 0.0;
*dtime_appq = 0.0;
*gflops_tred = 0.0;
*gflops_tevd = 0.0;
*gflops_appq = 0.0;
*k_perf = 0;
return;
}
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, A, &A_save );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, A, &Z );
FLA_Copy_external( A, A_save );
for ( i = 0 ; i < n_repeats; i++ ){
FLA_Copy_external( A_save, A );
*dtime = FLA_Clock();
switch( variant ){
case -3:
{
*k_perf = 0;
REF_Hevd_lv( A, l,
dtime_tred, dtime_tevd, dtime_appq );
break;
}
case -4:
{
*k_perf = 0;
REF_Hevdd_lv( A, l,
dtime_tred, dtime_tevd, dtime_appq );
break;
}
case -5:
{
*k_perf = 0;
REF_Hevdr_lv( A, l, Z,
dtime_tred, dtime_tevd, dtime_appq );
break;
}
case 0:
{
*k_perf = 0;
REF_Hevd_lv_components( A, l,
dtime_tred, dtime_tevd, dtime_appq );
break;
}
case -1:
{
*k_perf = 0;
REF_Hevdd_lv_components( A, l,
dtime_tred, dtime_tevd, dtime_appq );
break;
}
case -2:
{
*k_perf = 0;
REF_Hevdr_lv_components( A, l, Z,
dtime_tred, dtime_tevd, dtime_appq );
break;
}
// Time variant 1
case 1:
{
*k_perf = FLA_Hevd_lv_var1_components( n_iter_max, A, l, k_accum, b_alg,
dtime_tred, dtime_tevd, dtime_appq );
break;
}
// Time variant 2
case 2:
{
*k_perf = FLA_Hevd_lv_var2_components( n_iter_max, A, l, k_accum, b_alg,
dtime_tred, dtime_tevd, dtime_appq );
break;
}
}
*dtime = FLA_Clock() - *dtime;
if ( *dtime < dtime_save )
{
dtime_save = *dtime;
dtime_tred_save = *dtime_tred;
dtime_tevd_save = *dtime_tevd;
dtime_appq_save = *dtime_appq;
}
}
*dtime = dtime_save;
*dtime_tred = dtime_tred_save;
*dtime_tevd = dtime_tevd_save;
*dtime_appq = dtime_appq_save;
//if ( variant == -3 || variant == 0 )
//printf( "\ndtime is %9.3e\n", *dtime );
{
FLA_Obj V, A_rev_evd, norm, eye;
if ( variant == -2 || variant == -5 ) FLA_Copy( Z, A );
FLA_Obj_create_copy_of( FLA_NO_TRANSPOSE, A, &V );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, A, &A_rev_evd );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, A, &eye );
FLA_Obj_create( FLA_Obj_datatype_proj_to_real( A ), 1, 1, 0, 0, &norm );
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_NO_CONJUGATE, l, A );
FLA_Gemm( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
FLA_ONE, A, V, FLA_ZERO, A_rev_evd );
FLA_Triangularize( FLA_LOWER_TRIANGULAR, FLA_NONUNIT_DIAG, A_rev_evd );
//FLA_Obj_show( "A_rev_evd", A_rev_evd, "%9.2e + %9.2e ", "" );
FLA_Axpy( FLA_MINUS_ONE, A_save, A_rev_evd );
FLA_Norm_frob( A_rev_evd, norm );
FLA_Obj_extract_real_scalar( norm, diff1 );
FLA_Set_to_identity( eye );
FLA_Copy( V, A_rev_evd );
FLA_Gemm( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
FLA_ONE, V, A_rev_evd, FLA_MINUS_ONE, eye );
FLA_Norm_frob( eye, norm );
FLA_Obj_extract_real_scalar( norm, diff2 );
FLA_Obj_free( &V );
FLA_Obj_free( &A_rev_evd );
FLA_Obj_free( &eye );
FLA_Obj_free( &norm );
}
k = 2.00;
flops_tred = ( ( 4.0 / 3.0 ) * m * m * m );
flops_tevd = ( 4.5 * k * m * m +
3.0 * k * m * m * m );
if ( variant == -1 || variant == -2 ||
variant == -4 || variant == -5 )
flops_appq = ( 2.0 * m * m * m );
else
flops_appq = ( 4.0 / 3.0 * m * m * m );
/*
if ( FLA_Obj_is_complex( A ) )
{
*gflops = ( 4.0 * flops_tred +
2.0 * flops_tevd +
4.0 * flops_appq ) / *dtime / 1e9;
*gflops_tred = ( 4.0 * flops_tred ) / *dtime_tred / 1e9;
*gflops_tevd = ( 2.0 * flops_tevd ) / *dtime_tevd / 1e9;
*gflops_appq = ( 4.0 * flops_appq ) / *dtime_appq / 1e9;
}
else
{
*gflops = ( 1.0 * flops_tred +
1.0 * flops_tevd +
1.0 * flops_appq ) / *dtime / 1e9;
*gflops_tred = ( 1.0 * flops_tred ) / *dtime_tred / 1e9;
*gflops_tevd = ( 1.0 * flops_tevd ) / *dtime_tevd / 1e9;
*gflops_appq = ( 1.0 * flops_appq ) / *dtime_appq / 1e9;
}
*/
if ( FLA_Obj_is_complex( A ) )
{
mult_tred = 4.0;
mult_tevd = 2.0;
mult_appq = 4.0;
}
else
{
mult_tred = 1.0;
mult_tevd = 1.0;
mult_appq = 1.0;
}
*gflops = ( mult_tred * flops_tred +
mult_tevd * flops_tevd +
mult_appq * flops_appq ) / *dtime / 1e9;
*gflops_tred = ( mult_tred * flops_tred ) / *dtime_tred / 1e9;
*gflops_tevd = ( mult_tevd * flops_tevd ) / *dtime_tevd / 1e9;
*gflops_appq = ( mult_appq * flops_appq ) / *dtime_appq / 1e9;
FLA_Copy_external( A_save, A );
FLA_Obj_free( &A_save );
FLA_Obj_free( &Z );
}
int main(int argc, char *argv[])
{
int
m_input,
m,
p_first, p_last, p_inc,
p,
b_alg,
variant,
n_repeats,
i,
datatype,
n_variants = 1;
char *colors = "brkgmcbrkg";
char *ticks = "o+*xso+*xs";
char m_dim_desc[14];
char m_dim_tag[10];
double max_gflops=6.0;
double safemin;
double
dtime,
gflops,
diff;
FLA_Obj
A, l, Q, T, W;
FLA_Init();
fprintf( stdout, "%c number of repeats:", '%' );
scanf( "%d", &n_repeats );
fprintf( stdout, "%c %d\n", '%', n_repeats );
fprintf( stdout, "%c Enter blocking size:", '%' );
scanf( "%d", &b_alg );
fprintf( stdout, "%c %d\n", '%', b_alg );
fprintf( stdout, "%c enter problem size first, last, inc:", '%' );
scanf( "%d%d%d", &p_first, &p_last, &p_inc );
fprintf( stdout, "%c %d %d %d\n", '%', p_first, p_last, p_inc );
fprintf( stdout, "%c enter m (-1 means bind to problem size): ", '%' );
scanf( "%d", &m_input );
fprintf( stdout, "%c %d\n", '%', m_input );
fprintf( stdout, "\n" );
if ( m_input > 0 ) {
sprintf( m_dim_desc, "m = %d", m_input );
sprintf( m_dim_tag, "m%dc", m_input);
}
else if( m_input < -1 ) {
sprintf( m_dim_desc, "m = p/%d", -m_input );
sprintf( m_dim_tag, "m%dp", -m_input );
}
else if( m_input == -1 ) {
sprintf( m_dim_desc, "m = p" );
sprintf( m_dim_tag, "m%dp", 1 );
}
/*
char ch = 's';
safemin = dlamch_( &ch );
printf( "safemin = %23.15e\n", safemin );
ch = 'e';
double eps = dlamch_( &ch );
printf( "eps dla = %23.15e\n", eps );
printf( "eps fla = %23.15e\n", FLA_EPSILON_D );
*/
for ( p = p_first, i = 1; p <= p_last; p += p_inc, i += 1 )
{
m = m_input;
if( m < 0 ) m = p / f2c_abs(m_input);
//datatype = FLA_FLOAT;
//datatype = FLA_DOUBLE;
//datatype = FLA_COMPLEX;
datatype = FLA_DOUBLE_COMPLEX;
FLA_Obj_create( datatype, m, m, 0, 0, &A );
FLA_Obj_create( datatype, m, m, 0, 0, &Q );
FLA_Obj_create( datatype, 32, m, 0, 0, &T );
FLA_Obj_create( datatype, 32, m, 0, 0, &W );
FLA_Obj_create( FLA_Obj_datatype_proj_to_real( A ), m, 1, 0, 0, &l );
//FLA_Random_herm_matrix( FLA_LOWER_TRIANGULAR, A );
//FLA_Random_spd_matrix( FLA_LOWER_TRIANGULAR, A );
FLA_Random_matrix( A );
FLA_Obj_set_to_identity( Q );
FLA_QR_UT( A, T );
FLA_Apply_Q_UT( FLA_LEFT, FLA_CONJ_TRANSPOSE, FLA_FORWARD, FLA_COLUMNWISE, A, T, W, Q );
fill_eigenvalues( l );
//FLA_Obj_show( "eig", l, "%9.2e ", "" );
FLA_Apply_diag_matrix( FLA_LEFT, FLA_NO_CONJUGATE, l, Q );
FLA_Apply_Q_UT( FLA_LEFT, FLA_NO_TRANSPOSE, FLA_FORWARD, FLA_COLUMNWISE, A, T, W, Q );
FLA_Triangularize( FLA_LOWER_TRIANGULAR, FLA_NONUNIT_DIAG, Q );
FLA_Copy( Q, A );
time_Hevd_ln( 0, FLA_ALG_REFERENCE, n_repeats, m, b_alg,
A, l, &dtime, &diff, &gflops );
fprintf( stdout, "data_REFs( %d, 1:2 ) = [ %d %6.3lf %6.2le ]; \n", i, p, gflops, diff );
fflush( stdout );
time_Hevd_ln( -1, FLA_ALG_REFERENCE, n_repeats, m, b_alg,
A, l, &dtime, &diff, &gflops );
fprintf( stdout, "data_REFd( %d, 1:2 ) = [ %d %6.3lf %6.2le ]; \n", i, p, gflops, diff );
fflush( stdout );
for ( variant = 1; variant <= n_variants; variant++ ){
fprintf( stdout, "data_var%d( %d, 1:9 ) = [ %d ", variant, i, p );
fflush( stdout );
time_Hevd_ln( variant, FLA_ALG_UNBLOCKED, n_repeats, m, b_alg,
A, l, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
//time_Hevd_ln( variant, FLA_ALG_UNB_OPT, n_repeats, m, b_alg,
// A, l, &dtime, &diff, &gflops );
//fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
//fflush( stdout );
fprintf( stdout, "];\n" );
fflush( stdout );
}
fprintf( stdout, "\n" );
FLA_Obj_free( &A );
FLA_Obj_free( &T );
FLA_Obj_free( &W );
FLA_Obj_free( &Q );
FLA_Obj_free( &l );
}
/*
fprintf( stdout, "figure;\n" );
fprintf( stdout, "plot( data_REF( :,1 ), data_REF( :, 2 ), '-' ); \n" );
fprintf( stdout, "hold on;\n" );
for ( i = 1; i <= n_variants; i++ ) {
fprintf( stdout, "plot( data_var%d( :,1 ), data_var%d( :, 2 ), '%c:%c' ); \n",
i, i, colors[ i-1 ], ticks[ i-1 ] );
fprintf( stdout, "plot( data_var%d( :,1 ), data_var%d( :, 4 ), '%c-.%c' ); \n",
i, i, colors[ i-1 ], ticks[ i-1 ] );
}
fprintf( stdout, "legend( ... \n" );
fprintf( stdout, "'Reference', ... \n" );
for ( i = 1; i < n_variants; i++ )
fprintf( stdout, "'unb\\_var%d', 'blk\\_var%d', ... \n", i, i );
fprintf( stdout, "'unb\\_var%d', 'blk\\_var%d' ); \n", i, i );
fprintf( stdout, "xlabel( 'problem size p' );\n" );
fprintf( stdout, "ylabel( 'GFLOPS/sec.' );\n" );
fprintf( stdout, "axis( [ 0 %d 0 %.2f ] ); \n", p_last, max_gflops );
fprintf( stdout, "title( 'FLAME Hevd_ln performance (%s, %s)' );\n",
m_dim_desc, n_dim_desc );
fprintf( stdout, "print -depsc tridiag_%s_%s.eps\n", m_dim_tag, n_dim_tag );
fprintf( stdout, "hold off;\n");
fflush( stdout );
*/
FLA_Finalize( );
return 0;
}
FLA_Error FLA_Svd_ext_u_unb_var1( FLA_Svd_type jobu, FLA_Svd_type jobv,
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, C; // C is dummy.
dim_t m_A, n_A, min_m_n;
dim_t n_GH;
double crossover_ratio = 17.0 / 9.0;
FLA_Bool u_is_formed = FALSE,
v_is_formed = FALSE;
int apply_scale;
n_GH = k_accum;
m_A = FLA_Obj_length( A );
n_A = FLA_Obj_width( A );
min_m_n = min( m_A, n_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.
if ( FLA_Obj_is_complex( A ) )
{
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 );
// Scale matrix A if necessary.
FLA_Max_abs_value( A, scale );
apply_scale =
( FLA_Obj_gt( scale, FLA_OVERFLOW_SQUARE_THRES ) == TRUE ) -
( FLA_Obj_lt( scale, FLA_UNDERFLOW_SQUARE_THRES ) == TRUE );
if ( apply_scale )
FLA_Scal( apply_scale > 0 ? FLA_SAFE_MIN : FLA_SAFE_INV_MIN, 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 );
if ( FLA_Obj_is_complex( A ) )
FLA_Bidiag_UT_realify( A, rL, rR );
FLA_Bidiag_UT_extract_real_diagonals( A, d, e );
// Form U and V.
if ( u_is_formed == FALSE )
{
switch ( jobu )
{
case FLA_SVD_VECTORS_MIN_OVERWRITE:
if ( jobv != FLA_SVD_VECTORS_NONE )
FLA_Bidiag_UT_form_V_ext( FLA_UPPER_TRIANGULAR, A, S, FLA_NO_TRANSPOSE, V );
v_is_formed = TRUE; // For this case, V should be formed here.
U = A;
case FLA_SVD_VECTORS_ALL:
case FLA_SVD_VECTORS_MIN_COPY:
FLA_Bidiag_UT_form_U_ext( FLA_UPPER_TRIANGULAR, A, T, FLA_NO_TRANSPOSE, U );
u_is_formed = TRUE;
break;
case FLA_SVD_VECTORS_NONE:
// Do nothing
break;
}
}
if ( v_is_formed == FALSE )
{
if ( jobv == FLA_SVD_VECTORS_MIN_OVERWRITE )
{
FLA_Bidiag_UT_form_V_ext( FLA_UPPER_TRIANGULAR, A, S, FLA_CONJ_TRANSPOSE, A );
v_is_formed = TRUE; /* and */
V = A; // This V is actually V^H.
// V^H -> V
FLA_Obj_flip_base( &V );
FLA_Obj_flip_view( &V );
if ( FLA_Obj_is_complex( A ) )
FLA_Conjugate( V );
}
else if ( jobv != FLA_SVD_VECTORS_NONE )
{
FLA_Bidiag_UT_form_V_ext( FLA_UPPER_TRIANGULAR, A, S, FLA_NO_TRANSPOSE, V );
v_is_formed = TRUE;
}
}
// For complex matrices, apply realification transformation.
if ( FLA_Obj_is_complex( A ) && jobu != FLA_SVD_VECTORS_NONE )
{
FLA_Obj UL, UR;
FLA_Part_1x2( U, &UL, &UR, min_m_n, FLA_LEFT );
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_CONJUGATE, rL, UL );
}
if ( FLA_Obj_is_complex( A ) && jobv != FLA_SVD_VECTORS_NONE )
{
FLA_Obj VL, VR;
FLA_Part_1x2( V, &VL, &VR, min_m_n, FLA_LEFT );
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_NO_CONJUGATE, rR, VL );
}
// Perform a singular value decomposition on the upper bidiagonal matrix.
r_val = FLA_Bsvd_ext_opt_var1( n_iter_max,
d, e, G, H,
jobu, U, jobv, V,
FALSE, C, // C is not referenced
b_alg );
}
else // if ( crossover_ratio * n_A <= m_A )
{
FLA_Obj TQ, R;
FLA_Obj AT,
AB;
// Perform a QR factorization on A.
FLA_QR_UT_create_T( A, &TQ );
FLA_QR_UT( A, 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 );
// Form U; if necessary overwrite on A.
if ( u_is_formed == FALSE )
{
switch ( jobu )
{
case FLA_SVD_VECTORS_MIN_OVERWRITE:
U = A;
case FLA_SVD_VECTORS_ALL:
case FLA_SVD_VECTORS_MIN_COPY:
FLA_QR_UT_form_Q( A, TQ, U );
u_is_formed = TRUE;
break;
case FLA_SVD_VECTORS_NONE:
// Do nothing
break;
}
}
FLA_Obj_free( &TQ );
// 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 );
if ( FLA_Obj_is_complex( R ) )
FLA_Bidiag_UT_realify( R, rL, rR );
FLA_Bidiag_UT_extract_real_diagonals( R, d, e );
if ( v_is_formed == FALSE )
{
if ( jobv == FLA_SVD_VECTORS_MIN_OVERWRITE )
{
FLA_Bidiag_UT_form_V_ext( FLA_UPPER_TRIANGULAR, R, S, FLA_CONJ_TRANSPOSE, AT );
v_is_formed = TRUE; /* and */
V = AT; // This V is actually V^H.
// V^H -> V
FLA_Obj_flip_base( &V );
FLA_Obj_flip_view( &V );
if ( FLA_Obj_is_complex( A ) )
FLA_Conjugate( V );
}
else if ( jobv != FLA_SVD_VECTORS_NONE )
{
FLA_Bidiag_UT_form_V_ext( FLA_UPPER_TRIANGULAR, R, S, FLA_NO_TRANSPOSE, V );
v_is_formed = TRUE;
}
}
// Apply householder vectors U in R.
FLA_Bidiag_UT_form_U_ext( FLA_UPPER_TRIANGULAR, R, T, FLA_NO_TRANSPOSE, R );
// Apply the realifying scalars in rL and rR to U and V, respectively.
if ( FLA_Obj_is_complex( A ) && jobu != FLA_SVD_VECTORS_NONE )
{
FLA_Obj RL, RR;
FLA_Part_1x2( R, &RL, &RR, min_m_n, FLA_LEFT );
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_CONJUGATE, rL, RL );
}
if ( FLA_Obj_is_complex( A ) && jobv != FLA_SVD_VECTORS_NONE )
{
FLA_Obj VL, VR;
FLA_Part_1x2( V, &VL, &VR, min_m_n, FLA_LEFT );
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_NO_CONJUGATE, rR, VL );
}
// Perform a singular value decomposition on the bidiagonal matrix.
r_val = FLA_Bsvd_ext_opt_var1( n_iter_max,
d, e, G, H,
jobu, R, jobv, V,
FALSE, C,
b_alg );
// Multiply R into U, storing the result in A and then copying back
// to U.
if ( jobu != FLA_SVD_VECTORS_NONE )
{
FLA_Obj UL, UR;
FLA_Part_1x2( U, &UL, &UR, min_m_n, FLA_LEFT );
if ( jobu == FLA_SVD_VECTORS_MIN_OVERWRITE ||
jobv == FLA_SVD_VECTORS_MIN_OVERWRITE )
{
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, UL, &C );
FLA_Gemm( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_ONE, UL, R, FLA_ZERO, C );
FLA_Copy( C, UL );
FLA_Obj_free( &C );
}
else
{
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 );
// No sort is required as it is applied on FLA_Bsvd.
if ( apply_scale )
FLA_Scal( apply_scale < 0 ? FLA_SAFE_MIN : FLA_SAFE_INV_MIN, s );
// When V is overwritten, flip it again.
if ( jobv == FLA_SVD_VECTORS_MIN_OVERWRITE )
{
// Always apply conjugation first wrt dimensions used; then, flip base.
if ( FLA_Obj_is_complex( V ) )
FLA_Conjugate( V );
FLA_Obj_flip_base( &V );
}
FLA_Obj_free( &scale );
FLA_Obj_free( &T );
FLA_Obj_free( &S );
if ( FLA_Obj_is_complex( A ) )
{
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_Datatype realtype = REALTYPE;
FLA_Obj
A, TU, TV,
A_copy, A_recovered,
U, V, Vb, B, Be, d, e,
DU, DV;
FLA_Obj
ATL, ATR,
ABL, ABR, Ae;
FLA_Uplo uplo;
dim_t m, n, min_m_n;
FLA_Error init_result;
double residual_A = 0.0;
if ( argc == 3 ) {
m = atoi(argv[1]);
n = atoi(argv[2]);
min_m_n = min(m,n);
} else {
fprintf(stderr, " \n");
fprintf(stderr, "Usage: %s m n\n", argv[0]);
fprintf(stderr, " m : matrix length\n");
fprintf(stderr, " n : matrix width\n");
fprintf(stderr, " \n");
return -1;
}
if ( m == 0 || n == 0 )
return 0;
FLA_Init_safe( &init_result );
// FLAME Bidiag setup
FLA_Obj_create( datatype, m, n, 0, 0, &A );
FLA_Bidiag_UT_create_T( A, &TU, &TV );
// Rand A and create A_copy.
FLA_Random_matrix( A );
{
scomplex *buff_A = FLA_Obj_buffer_at_view( A );
buff_A[0].real = 4.4011e-01; buff_A[0].imag = -4.0150e-09; buff_A[2].real = -2.2385e-01; buff_A[2].imag = -1.5546e-01; buff_A[4].real = -6.3461e-02; buff_A[4].imag = 2.7892e-01; buff_A[6].real = -1.3197e-01; buff_A[6].imag = 5.0888e-01;
buff_A[1].real = 3.3352e-01; buff_A[1].imag = -6.6346e-02; buff_A[3].real = -1.9307e-01; buff_A[3].imag = -8.4066e-02; buff_A[5].real = -6.0446e-03; buff_A[5].imag = 2.2094e-01; buff_A[7].real = -2.3299e-02; buff_A[7].imag = 4.0553e-01;
}
//FLA_Set_to_identity( A );
//FLA_Scal( FLA_MINUS_ONE, A );
if ( m >= n ) {
uplo = FLA_UPPER_TRIANGULAR;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, min_m_n - 1, 1, FLA_TL );
Ae = ATR;
} else {
uplo = FLA_LOWER_TRIANGULAR;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 1, min_m_n - 1, FLA_TL );
Ae = ABL;
}
FLA_Obj_create_copy_of( FLA_NO_TRANSPOSE, A, &A_copy );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, A, &A_recovered );
// Bidiag test
{
FLA_Obj norm;
FLA_Bool apply_scale;
FLA_Obj_create( realtype, 1,1, 0,0, &norm );
FLA_Max_abs_value( A, norm );
apply_scale = FLA_Obj_gt( norm, FLA_OVERFLOW_SQUARE_THRES );
if ( apply_scale ) FLA_Scal( FLA_SAFE_MIN, A );
FLA_Bidiag_UT( A, TU, TV );
if ( apply_scale ) FLA_Bidiag_UT_scale_diagonals( FLA_SAFE_INV_MIN, A );
FLA_Obj_free( &norm );
}
// Orthonomal basis U, V.
FLA_Obj_create( datatype, m, min_m_n, 0, 0, &U ); FLA_Set( FLA_ZERO, U );
FLA_Obj_create( datatype, min_m_n, n, 0, 0, &V ); FLA_Set( FLA_ZERO, V );
FLA_Bidiag_UT_form_U_ext( uplo, A, TU, FLA_NO_TRANSPOSE, U );
FLA_Bidiag_UT_form_V_ext( uplo, A, TV, FLA_CONJ_TRANSPOSE, V );
if ( FLA_Obj_is_complex( A ) ){
FLA_Obj rL, rR;
FLA_Obj_create( datatype, min_m_n, 1, 0, 0, &rL );
FLA_Obj_create( datatype, min_m_n, 1, 0, 0, &rR );
FLA_Obj_fshow( stdout, " - Factor no realified - ", A, "% 6.4e", "------");
FLA_Bidiag_UT_realify( A, rL, rR );
FLA_Obj_fshow( stdout, " - Factor realified - ", A, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - rL - ", rL, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - rR - ", rR, "% 6.4e", "------");
FLA_Apply_diag_matrix( FLA_RIGHT, FLA_CONJUGATE, rL, U );
FLA_Apply_diag_matrix( FLA_LEFT, FLA_CONJUGATE, rR, V );
FLA_Obj_free( &rL );
FLA_Obj_free( &rR );
}
// U^H U
FLA_Obj_create( datatype, min_m_n, min_m_n, 0, 0, &DU );
FLA_Gemm_external( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_ONE, U, U, FLA_ZERO, DU );
// V^H V
FLA_Obj_create( datatype, min_m_n, min_m_n, 0, 0, &DV );
FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
FLA_ONE, V, V, FLA_ZERO, DV );
// Recover the matrix
FLA_Obj_create( datatype, min_m_n, min_m_n, 0, 0, &B );
FLA_Set( FLA_ZERO, B );
// Set B
FLA_Obj_create( datatype, min_m_n, 1, 0, 0, &d );
FLA_Set_diagonal_vector( A, d );
FLA_Set_diagonal_matrix( d, B );
FLA_Obj_free( &d );
if ( min_m_n > 1 ) {
FLA_Obj_create( datatype, min_m_n - 1 , 1, 0, 0, &e );
FLA_Set_diagonal_vector( Ae, e );
if ( uplo == FLA_UPPER_TRIANGULAR ) {
FLA_Part_2x2( B, &ATL, &ATR,
&ABL, &ABR, min_m_n - 1, 1, FLA_TL );
Be = ATR;
} else {
FLA_Part_2x2( B, &ATL, &ATR,
&ABL, &ABR, 1, min_m_n - 1, FLA_TL );
Be = ABL;
}
FLA_Set_diagonal_matrix( e, Be );
FLA_Obj_free( &e );
}
// Vb := B (V^H)
FLA_Obj_create_copy_of( FLA_NO_TRANSPOSE, V, &Vb );
FLA_Trmm_external( FLA_LEFT, uplo, FLA_NO_TRANSPOSE,
FLA_NONUNIT_DIAG, FLA_ONE, B, Vb );
// A := U Vb
FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_ONE, U, Vb, FLA_ZERO, A_recovered );
residual_A = FLA_Max_elemwise_diff( A_copy, A_recovered );
if (1) {
FLA_Obj_fshow( stdout, " - Given - ", A_copy, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - Factor - ", A, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - TU - ", TU, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - TV - ", TV, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - B - ", B, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - U - ", U, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - V - ", V, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - Vb - ", Vb, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - U'U - ", DU, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - VV' - ", DV, "% 6.4e", "------");
FLA_Obj_fshow( stdout, " - Recovered A - ", A_recovered, "% 6.4e", "------");
fprintf( stdout, "lapack2flame: %lu x %lu: ", m, n);
fprintf( stdout, "recovery A = %12.10e\n\n", residual_A ) ;
}
FLA_Obj_free( &A );
FLA_Obj_free( &TU );
FLA_Obj_free( &TV );
FLA_Obj_free( &B );
FLA_Obj_free( &U );
FLA_Obj_free( &V );
FLA_Obj_free( &Vb );
FLA_Obj_free( &DU );
FLA_Obj_free( &DV );
FLA_Obj_free( &A_copy );
FLA_Obj_free( &A_recovered );
FLA_Finalize_safe( init_result );
}