FLA_Error FLASH_Gemv( FLA_Trans transa, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y )
{
FLA_Error r_val;
FLA_Bool enable_supermatrix;
// Check parameters.
if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
FLA_Gemv_check( transa, alpha, A, x, beta, y );
// Find the status of SuperMatrix.
enable_supermatrix = FLASH_Queue_get_enabled();
// Temporarily disable SuperMatrix.
FLASH_Queue_disable();
// Execute tasks.
r_val = FLA_Gemv_internal( transa, alpha, A, x, beta, y, flash_gemv_cntl_fm_rp );
// Restore SuperMatrix to its previous status.
if ( enable_supermatrix )
FLASH_Queue_enable();
return r_val;
}
FLA_Error FLASH_Axpy( FLA_Obj alpha, FLA_Obj A, FLA_Obj B )
{
FLA_Error r_val;
FLA_Bool enable_supermatrix;
// Check parameters.
if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
FLA_Axpy_check( alpha, A, B );
// Find the status of SuperMatrix.
enable_supermatrix = FLASH_Queue_get_enabled();
// Temporarily disable SuperMatrix.
FLASH_Queue_disable();
// Execute tasks.
r_val = FLA_Axpy_internal( alpha, A, B, flash_axpy_cntl );
// Restore SuperMatrix to its previous status.
if ( enable_supermatrix )
FLASH_Queue_enable();
return r_val;
}
FLA_Error FLASH_Trsv( FLA_Uplo uplo, FLA_Trans trans, FLA_Diag diag, FLA_Obj A, FLA_Obj x )
{
FLA_Error r_val;
FLA_Bool enable_supermatrix;
// Check parameters.
if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
FLA_Trsv_check( uplo, trans, diag, A, x );
// Find the status of SuperMatrix.
enable_supermatrix = FLASH_Queue_get_enabled();
// Temporarily disable SuperMatrix.
FLASH_Queue_disable();
// Execute tasks.
r_val = FLA_Trsv_internal( uplo, trans, diag, A, x, flash_trsv_cntl );
// Restore SuperMatrix to its previous status.
if ( enable_supermatrix )
FLASH_Queue_enable();
return r_val;
}
int main(int argc, char *argv[])
{
int
datatype,
m_input,
m,
p_first, p_last, p_inc,
p,
n_repeats,
param_combo,
i,
n_param_combos = N_PARAM_COMBOS;
dim_t nb_alg;
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,
diff;
FLA_Obj
A, A_ref;
FLA_Init();
fprintf( stdout, "%c number of repeats: ", '%' );
scanf( "%d", &n_repeats );
fprintf( stdout, "%c %d\n", '%', n_repeats );
fprintf( stdout, "%c enter FLASH blocksize: ", '%' );
scanf( "%u", &nb_alg );
fprintf( stdout, "%c %u\n", '%', nb_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, "\nclear all;\n\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 );
}
//datatype = FLA_FLOAT;
datatype = FLA_DOUBLE;
//datatype = FLA_COMPLEX;
//datatype = FLA_DOUBLE_COMPLEX;
FLASH_Queue_disable();
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);
for ( param_combo = 0; param_combo < n_param_combos; param_combo++ ){
FLASH_Obj_create( datatype, m, m, 1, &nb_alg, &A );
FLASH_Obj_create( datatype, m, m, 1, &nb_alg, &A_ref );
if ( pc_str[param_combo][0] == 'l' )
FLASH_Random_spd_matrix( FLA_LOWER_TRIANGULAR, A );
else
FLASH_Random_spd_matrix( FLA_UPPER_TRIANGULAR, A );
FLASH_Copy( A, A_ref );
fprintf( stdout, "data_chol_%s( %d, 1:5 ) = [ %d ", pc_str[param_combo], i, p );
fflush( stdout );
time_Chol( param_combo, FLA_ALG_REFERENCE, n_repeats, m,
A, A_ref, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
time_Chol( param_combo, FLA_ALG_FRONT, n_repeats, m,
A, A_ref, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
fprintf( stdout, " ]; \n" );
fflush( stdout );
FLASH_Obj_free( &A );
FLASH_Obj_free( &A_ref );
}
fprintf( stdout, "\n" );
}
fprintf( stdout, "figure;\n" );
fprintf( stdout, "hold on;\n" );
for ( i = 0; i < n_param_combos; i++ ) {
fprintf( stdout, "plot( data_chol_%s( :,1 ), data_chol_%s( :, 2 ), '%c:%c' ); \n",
pc_str[i], pc_str[i], colors[ i ], ticks[ i ] );
fprintf( stdout, "plot( data_chol_%s( :,1 ), data_chol_%s( :, 4 ), '%c-.%c' ); \n",
pc_str[i], pc_str[i], colors[ i ], ticks[ i ] );
}
fprintf( stdout, "legend( ... \n" );
for ( i = 0; i < n_param_combos; i++ )
fprintf( stdout, "'ref\\_chol\\_%s', 'fla\\_chol\\_%s', ... \n", pc_str[i], pc_str[i] );
fprintf( stdout, "'Location', 'SouthEast' ); \n" );
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 chol front-end performance (%s)' );\n", m_dim_desc );
fprintf( stdout, "print -depsc chol_front_%s.eps\n", m_dim_tag );
fprintf( stdout, "hold off;\n");
fflush( stdout );
FLA_Finalize( );
return 0;
}
int main(int argc, char *argv[])
{
int
datatype,
precision,
nb_alg, bm, bn,
m_input, n_input,
m, n,
p_first, p_last, p_inc,
p,
n_repeats,
param_combo,
i,
n_param_combos = N_PARAM_COMBOS;
char *colors = "brkgmcbrkgmcbrkgmc";
char *ticks = "o+*xso+*xso+*xso+*xs";
char m_dim_desc[14];
char n_dim_desc[14];
char m_dim_tag[10];
char n_dim_tag[10];
double max_gflops=6.0;
double
dtime,
gflops,
diff;
FLA_Obj
A, A_save, A_flat, B, B_ref, T, T_flat, W, t;
FLA_Init( );
fprintf( stdout, "%c number of repeats: ", '%' );
scanf( "%d", &n_repeats );
fprintf( stdout, "%c %d\n", '%', n_repeats );
fprintf( stdout, "%c enter FLASH blocksize: ", '%' );
scanf( "%d", &nb_alg );
fprintf( stdout, "%c %d\n", '%', nb_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 n (-1 means bind to problem size): ", '%' );
scanf( "%d%d", &m_input, &n_input );
fprintf( stdout, "%c %d %d\n", '%', m_input, n_input );
fprintf( stdout, "\nclear all;\n\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 );
}
if ( n_input > 0 ) {
sprintf( n_dim_desc, "n = %d", n_input );
sprintf( n_dim_tag, "n%dc", n_input);
}
else if( n_input < -1 ) {
sprintf( n_dim_desc, "n = p/%d", -n_input );
sprintf( n_dim_tag, "n%dp", -n_input );
}
else if( n_input == -1 ) {
sprintf( n_dim_desc, "n = p" );
sprintf( n_dim_tag, "n%dp", 1 );
}
//precision = FLA_SINGLE_PRECISION;
precision = FLA_DOUBLE_PRECISION;
FLASH_Queue_disable();
for ( p = p_first, i = 1; p <= p_last; p += p_inc, i += 1 )
{
m = m_input;
n = n_input;
if( m < 0 ) m = p / abs(m_input);
if( n < 0 ) n = p / abs(n_input);
for ( param_combo = 0; param_combo < n_param_combos; param_combo++ ){
// Determine datatype based on trans argument.
if ( pc_str[param_combo][1] == 'c' )
{
if ( precision == FLA_SINGLE_PRECISION )
datatype = FLA_COMPLEX;
else
datatype = FLA_DOUBLE_COMPLEX;
}
else
{
if ( precision == FLA_SINGLE_PRECISION )
datatype = FLA_FLOAT;
else
datatype = FLA_DOUBLE;
}
bm = nb_alg / 4;
bn = nb_alg;
// If multiplying Q on the left, A is m x m; ...on the right, A is n x n.
if ( pc_str[param_combo][0] == 'l' )
{
FLA_Obj_create( datatype, nb_alg, nb_alg, &A_flat );
FLASH_Obj_create( datatype, nb_alg, nb_alg, 1, &nb_alg, &A );
FLASH_Obj_create( datatype, nb_alg, nb_alg, 1, &nb_alg, &A_save );
FLA_Obj_create( datatype, bm, bn, &T_flat );
FLASH_Obj_create_ext( datatype, bm, bn, 1, &bm, &bn, &T );
FLASH_Obj_create_ext( datatype, bm, n, 1, &bm, &bn, &W );
}
else
{
FLASH_Obj_create( datatype, n, n, 1, &nb_alg, &A );
}
FLASH_Obj_create( datatype, nb_alg, n, 1, &nb_alg, &B );
FLASH_Obj_create( datatype, nb_alg, n, 1, &nb_alg, &B_ref );
FLA_Obj_create( datatype, nb_alg, 1, &t );
FLASH_Random_matrix( A );
FLASH_Random_matrix( B );
fprintf( stdout, "data_applyq_%s( %d, 1:5 ) = [ %d ", pc_str[param_combo], i, p );
fflush( stdout );
FLASH_Copy( A, A_save );
FLASH_Obj_flatten( A, A_flat );
FLA_QR_blk_external( A_flat, t );
FLASH_Obj_hierarchify( A_flat, A );
time_Apply_Q( param_combo, FLA_ALG_REFERENCE, n_repeats, m, n,
A, B, B_ref, t, T, W, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
FLASH_Copy( A_save, A );
FLASH_Obj_flatten( A, A_flat );
FLA_QR_UT( A_flat, t, T_flat );
FLASH_Obj_hierarchify( A_flat, A );
FLASH_Obj_hierarchify( T_flat, T );
time_Apply_Q( param_combo, FLA_ALG_FRONT, n_repeats, m, n,
A, B, B_ref, t, T, W, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
fprintf( stdout, " ]; \n" );
fflush( stdout );
FLASH_Obj_free( &A );
FLA_Obj_free( &A_flat );
FLASH_Obj_free( &B );
FLASH_Obj_free( &B_ref );
FLA_Obj_free( &t );
FLASH_Obj_free( &T );
FLA_Obj_free( &T_flat );
FLASH_Obj_free( &W );
}
fprintf( stdout, "\n" );
}
fprintf( stdout, "figure;\n" );
fprintf( stdout, "hold on;\n" );
for ( i = 0; i < n_param_combos; i++ ) {
fprintf( stdout, "plot( data_applyq_%s( :,1 ), data_applyq_%s( :, 2 ), '%c:%c' ); \n",
pc_str[i], pc_str[i], colors[ i ], ticks[ i ] );
fprintf( stdout, "plot( data_applyq_%s( :,1 ), data_applyq_%s( :, 4 ), '%c-.%c' ); \n",
pc_str[i], pc_str[i], colors[ i ], ticks[ i ] );
}
fprintf( stdout, "legend( ... \n" );
for ( i = 0; i < n_param_combos; i++ )
fprintf( stdout, "'ref\\_applyq\\_%s', 'fla\\_applyq\\_%s', ... \n", pc_str[i], pc_str[i] );
fprintf( stdout, "'Location', 'SouthEast' ); \n" );
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 applyq front-end performance (%s, %s)' );\n",
m_dim_desc, n_dim_desc );
fprintf( stdout, "print -depsc applyq_front_%s_%s.eps\n", m_dim_tag, n_dim_tag );
fprintf( stdout, "hold off;\n");
fflush( stdout );
FLA_Finalize( );
return 0;
}
int main(int argc, char *argv[])
{
int
datatype,
m_input, n_input,
m, n, min_m_n,
p_first, p_last, p_inc,
pp,
pivot_combo,
n_repeats,
i,
n_pivot_combos = N_PIVOT_COMBOS;
dim_t
nb_alg,
nb_flash;
char *colors = "brkgmcbrkg";
char *ticks = "o+*xso+*xs";
char m_dim_desc[14];
char n_dim_desc[14];
char m_dim_tag[10];
char n_dim_tag[10];
double max_gflops=6.0;
double
dtime,
gflops,
diff;
FLA_Obj C, p, b, b_ref, b_norm;
FLA_Init();
fprintf( stdout, "%c number of repeats: ", '%' );
scanf( "%d", &n_repeats );
fprintf( stdout, "%c %d\n", '%', n_repeats );
fprintf( stdout, "%c enter algorithmic blocksize: ", '%' );
scanf( "%u", &nb_alg );
fprintf( stdout, "%c %u\n", '%', nb_alg );
fprintf( stdout, "%c enter FLASH blocksize: ", '%' );
scanf( "%u", &nb_flash );
fprintf( stdout, "%c %u\n", '%', nb_flash );
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 n (-1 means bind to problem size): ", '%' );
scanf( "%d %d", &m_input, &n_input );
fprintf( stdout, "%c %d %d\n", '%', m_input, n_input );
fprintf( stdout, "\nclear all;\n\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 );
}
if ( n_input > 0 ) {
sprintf( n_dim_desc, "n = %d", n_input );
sprintf( n_dim_tag, "n%dc", n_input);
}
else if( n_input < -1 ) {
sprintf( n_dim_desc, "n = p/%d", -n_input );
sprintf( n_dim_tag, "n%dp", -n_input );
}
else if( n_input == -1 ) {
sprintf( n_dim_desc, "n = p" );
sprintf( n_dim_tag, "n%dp", 1 );
}
//datatype = FLA_FLOAT;
datatype = FLA_DOUBLE;
//datatype = FLA_COMPLEX;
//datatype = FLA_DOUBLE_COMPLEX;
FLASH_Queue_disable();
for ( pp = p_first, i = 1; pp <= p_last; pp += p_inc, i += 1 )
{
m = m_input;
n = n_input;
if( m < 0 ) m = pp / abs(m_input);
if( n < 0 ) n = pp / abs(n_input);
min_m_n = min( m, n );
for ( pivot_combo = 0; pivot_combo < n_pivot_combos; pivot_combo++ ){
FLA_Obj_create( datatype, m, n, &C );
FLA_Obj_create( FLA_INT, min_m_n, 1, &p );
FLA_Obj_create( datatype, m, 1, &b );
FLA_Obj_create( datatype, m, 1, &b_ref );
FLA_Obj_create( datatype, 1, 1, &b_norm );
FLA_Random_matrix( C );
FLA_Random_matrix( b );
FLA_Copy_external( b, b_ref );
fprintf( stdout, "data_lu_%s( %d, 1:5 ) = [ %d ", pc_str[pivot_combo], i, pp );
fflush( stdout );
time_LU( pivot_combo, FLA_ALG_REFERENCE, n_repeats, m, n, nb_alg, nb_flash,
C, p, b, b_ref, b_norm, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
time_LU( pivot_combo, FLA_ALG_FRONT, n_repeats, m, n, nb_alg, nb_flash,
C, p, b, b_ref, b_norm, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
fprintf( stdout, " ]; \n" );
fflush( stdout );
FLA_Obj_free( &C );
FLA_Obj_free( &p );
FLA_Obj_free( &b );
FLA_Obj_free( &b_ref );
FLA_Obj_free( &b_norm );
}
fprintf( stdout, "\n" );
}
fprintf( stdout, "figure;\n" );
fprintf( stdout, "hold on;\n" );
for ( i = 0; i < n_pivot_combos; i++ ) {
fprintf( stdout, "plot( data_lu_%s( :,1 ), data_lu_%s( :, 2 ), '%c:%c' ); \n",
pc_str[i], pc_str[i], colors[ i ], ticks[ i ] );
fprintf( stdout, "plot( data_lu_%s( :,1 ), data_lu_%s( :, 4 ), '%c-.%c' ); \n",
pc_str[i], pc_str[i], colors[ i ], ticks[ i ] );
}
fprintf( stdout, "legend( ... \n" );
for ( i = 0; i < n_pivot_combos; i++ )
fprintf( stdout, "'ref\\_lu\\_%s', 'fla\\_lu\\_%s', ... \n", pc_str[i], pc_str[i] );
fprintf( stdout, "'Location', 'SouthEast' ); \n" );
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 LU front-end performance (%s, %s)' );\n",
m_dim_desc, n_dim_desc );
fprintf( stdout, "print -depsc lu_front_%s_%s.eps\n", m_dim_tag, n_dim_tag );
fprintf( stdout, "hold off;\n");
fflush( stdout );
FLA_Finalize( );
return 0;
}
int main(int argc, char *argv[])
{
int
datatype,
nb_alg,
m_input, n_input,
m, n,
p_first, p_last, p_inc,
p,
n_repeats,
param_combo,
i, j,
n_param_combos = N_PARAM_COMBOS;
int sign;
char *colors = "brkgmcbrkg";
char *ticks = "o+*xso+*xs";
char m_dim_desc[14];
char n_dim_desc[14];
char m_dim_tag[10];
char n_dim_tag[10];
double max_gflops=6.0;
double
dtime,
gflops,
diff;
FLA_Obj
A, B, C, C_ref, scale, isgn, norm;
FLA_Init();
fprintf( stdout, "%c number of repeats: ", '%' );
scanf( "%d", &n_repeats );
fprintf( stdout, "%c %d\n", '%', n_repeats );
fprintf( stdout, "%c enter FLASH blocksize: ", '%' );
scanf( "%d", &nb_alg );
fprintf( stdout, "%c %d\n", '%', nb_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 sign (-1 or 1):", '%' );
scanf( "%d", &sign );
fprintf( stdout, "%c %d\n", '%', sign );
fprintf( stdout, "%c enter m n (-1 means bind to problem size): ", '%' );
scanf( "%d %d", &m_input, &n_input );
fprintf( stdout, "%c %d %d\n", '%', m_input, n_input );
fprintf( stdout, "\nclear all;\n\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 );
}
if ( n_input > 0 ) {
sprintf( n_dim_desc, "n = %d", n_input );
sprintf( n_dim_tag, "n%dc", n_input);
}
else if( n_input < -1 ) {
sprintf( n_dim_desc, "n = p/%d", -n_input );
sprintf( n_dim_tag, "n%dp", -n_input );
}
else if( n_input == -1 ) {
sprintf( n_dim_desc, "n = p" );
sprintf( n_dim_tag, "n%dp", 1 );
}
if ( 0 < sign )
isgn = FLA_ONE;
else
isgn = FLA_MINUS_ONE;
//datatype = FLA_FLOAT;
datatype = FLA_DOUBLE;
//datatype = FLA_COMPLEX;
//datatype = FLA_DOUBLE_COMPLEX;
if ( datatype == FLA_DOUBLE || datatype == FLA_DOUBLE_COMPLEX )
{
FLA_Obj_create( FLA_DOUBLE, 1, 1, &scale );
FLA_Obj_create( FLA_DOUBLE, 1, 1, &norm );
}
else if ( datatype == FLA_FLOAT || datatype == FLA_COMPLEX )
{
FLA_Obj_create( FLA_FLOAT, 1, 1, &scale );
FLA_Obj_create( FLA_FLOAT, 1, 1, &norm );
}
FLASH_Queue_disable();
for ( p = p_first, i = 1; p <= p_last; p += p_inc, i += 1 )
{
m = m_input;
n = n_input;
if( m < 0 ) m = p / abs(m_input);
if( n < 0 ) n = p / abs(n_input);
for ( param_combo = 0; param_combo < n_param_combos; param_combo++ ){
FLASH_Obj_create( datatype, m, m, 1, &nb_alg, &A );
FLASH_Obj_create( datatype, n, n, 1, &nb_alg, &B );
FLASH_Obj_create( datatype, m, n, 1, &nb_alg, &C );
FLASH_Obj_create( datatype, m, n, 1, &nb_alg, &C_ref );
FLASH_Random_matrix( A );
FLASH_Random_matrix( B );
FLASH_Random_matrix( C );
FLASH_Norm1( A, norm );
FLASH_Obj_shift_diagonal( FLA_NO_CONJUGATE, norm, A );
FLASH_Norm1( B, norm );
if ( FLA_Obj_is( isgn, FLA_MINUS_ONE ) )
FLA_Negate( norm );
FLASH_Obj_shift_diagonal( FLA_NO_CONJUGATE, norm, B );
FLASH_Copy( C, C_ref );
fprintf( stdout, "data_sylv_%s( %d, 1:5 ) = [ %d ", pc_str[param_combo], i, p );
fflush( stdout );
time_Sylv( param_combo, FLA_ALG_REFERENCE, n_repeats, m, n,
isgn, A, B, C, C_ref, scale, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
time_Sylv( param_combo, FLA_ALG_FRONT, n_repeats, m, n,
isgn, A, B, C, C_ref, scale, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
fprintf( stdout, " ]; \n" );
fflush( stdout );
FLASH_Obj_free( &A );
FLASH_Obj_free( &B );
FLASH_Obj_free( &C );
FLASH_Obj_free( &C_ref );
}
fprintf( stdout, "\n" );
}
FLA_Obj_free( &scale );
FLA_Obj_free( &norm );
fprintf( stdout, "figure;\n" );
fprintf( stdout, "hold on;\n" );
for ( i = 0; i < n_param_combos; i++ ) {
fprintf( stdout, "plot( data_sylv_%s( :,1 ), data_sylv_%s( :, 2 ), '%c:%c' ); \n",
pc_str[i], pc_str[i], colors[ i ], ticks[ i ] );
fprintf( stdout, "plot( data_sylv_%s( :,1 ), data_sylv_%s( :, 4 ), '%c-.%c' ); \n",
pc_str[i], pc_str[i], colors[ i ], ticks[ i ] );
}
fprintf( stdout, "legend( ... \n" );
for ( i = 0; i < n_param_combos; i++ )
fprintf( stdout, "'ref\\_sylv\\_%s', 'fla\\_sylv\\_%s', ... \n", pc_str[i], pc_str[i] );
fprintf( stdout, "'Location', 'SouthEast' ); \n" );
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 sylv front-end performance (%s)' );\n", m_dim_desc );
fprintf( stdout, "print -depsc sylv_front_%s.eps\n", m_dim_tag );
fprintf( stdout, "hold off;\n");
fflush( stdout );
FLA_Finalize( );
}
int main(int argc, char *argv[])
{
int
datatype,
m_input, n_input,
m, n,
p_first, p_last, p_inc,
p,
n_repeats,
param_combo,
i,
n_param_combos = N_PARAM_COMBOS;
dim_t
nb_flash,
nb_alg;
char *colors = "brkgmcbrkgmcbrkgmc";
char *ticks = "o+*xso+*xso+*xso+*xs";
char m_dim_desc[14];
char m_dim_tag[10];
double max_gflops=6.0;
double
dtime,
gflops,
diff;
FLA_Obj A_flat, T_flat, W_flat, B_flat;
FLA_Obj A, TW, W1, B, B_ref;
FLA_Init( );
fprintf( stdout, "%c number of repeats: ", '%' );
scanf( "%d", &n_repeats );
fprintf( stdout, "%c %d\n", '%', n_repeats );
fprintf( stdout, "%c enter algorithmic blocksize: ", '%' );
scanf( "%u", &nb_alg );
fprintf( stdout, "%c %u\n", '%', nb_alg );
fprintf( stdout, "%c enter FLASH blocksize: ", '%' );
scanf( "%u", &nb_flash );
fprintf( stdout, "%c %u\n", '%', nb_flash );
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 n (-1 means bind to problem size): ", '%' );
scanf( "%d%d", &m_input, &n_input );
fprintf( stdout, "%c %d %d\n", '%', m_input, n_input );
fprintf( stdout, "\nclear all;\n\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 );
}
//datatype = FLA_FLOAT;
//datatype = FLA_DOUBLE;
//datatype = FLA_COMPLEX;
datatype = FLA_DOUBLE_COMPLEX;
FLASH_Queue_disable();
for ( p = p_first, i = 1; p <= p_last; p += p_inc, i += 1 )
{
m = m_input;
n = n_input;
if( m < 0 ) m = p / abs(m_input);
if( n < 0 ) n = p / abs(n_input);
for ( param_combo = 0; param_combo < n_param_combos; param_combo++ )
{
FLA_Obj_create( datatype, m, m, &A_flat );
FLA_Obj_create( datatype, nb_alg, m, &T_flat );
FLA_Obj_create( datatype, nb_alg, n, &W_flat );
FLA_Obj_create( datatype, m, n, &B_flat );
FLA_Random_matrix( A_flat );
FLA_Random_matrix( B_flat );
FLASH_Obj_create_ext( datatype, m, n, 1, &nb_flash, &nb_flash, &B );
FLASH_Obj_create_ext( datatype, m, n, 1, &nb_flash, &nb_flash, &B_ref );
FLASH_QR_UT_inc_create_hier_matrices( A_flat, 1, &nb_flash, nb_alg, &A, &TW );
FLASH_Apply_Q_UT_inc_create_workspace( TW, B, &W1 );
FLASH_Obj_hierarchify( B_flat, B );
fprintf( stdout, "data_qrutinc_%s( %d, 1:5 ) = [ %d ", pc_str[param_combo], i, p );
fflush( stdout );
time_Apply_Q_UT_inc( param_combo, FLA_ALG_REFERENCE, n_repeats, m, n,
A, TW, W1, B, B_ref, A_flat, T_flat, W_flat, B_flat,
&dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
time_Apply_Q_UT_inc( param_combo, FLA_ALG_FRONT, n_repeats, m, n,
A, TW, W1, B, B_ref, A_flat, T_flat, W_flat, B_flat,
&dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
fprintf( stdout, " ]; \n" );
fflush( stdout );
FLA_Obj_free( &A_flat );
FLA_Obj_free( &T_flat );
FLA_Obj_free( &W_flat );
FLA_Obj_free( &B_flat );
FLASH_Obj_free( &A );
FLASH_Obj_free( &TW );
FLASH_Obj_free( &W1 );
FLASH_Obj_free( &B );
FLASH_Obj_free( &B_ref );
}
fprintf( stdout, "\n" );
}
fprintf( stdout, "figure;\n" );
fprintf( stdout, "hold on;\n" );
for ( i = 0; i < n_param_combos; i++ ) {
fprintf( stdout, "plot( data_qrutinc_%s( :,1 ), data_qrutinc_%s( :, 2 ), '%c:%c' ); \n",
pc_str[i], pc_str[i], colors[ i ], ticks[ i ] );
fprintf( stdout, "plot( data_qrutinc_%s( :,1 ), data_qrutinc_%s( :, 4 ), '%c-.%c' ); \n",
pc_str[i], pc_str[i], colors[ i ], ticks[ i ] );
}
fprintf( stdout, "legend( ... \n" );
for ( i = 0; i < n_param_combos; i++ )
fprintf( stdout, "'ref\\_qrutinc\\_%s', 'fla\\_qrutinc\\_%s', ... \n", pc_str[i], pc_str[i] );
fprintf( stdout, "'Location', 'SouthEast' ); \n" );
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 qrutinc front-end performance (%s)' );\n",
m_dim_desc );
fprintf( stdout, "print -depsc qrutinc_front_%s.eps\n", m_dim_tag );
fprintf( stdout, "hold off;\n");
fflush( stdout );
FLA_Finalize( );
return 0;
}