FLA_Error FLASH_Random_spd_matrix( FLA_Uplo uplo, FLA_Obj H )
{
FLA_Obj F;
FLA_Error e_val;
if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
{
e_val = FLA_Check_valid_uplo( uplo );
FLA_Check_error_code( e_val );
}
// Exit early if one dimension is zero.
if ( FLA_Obj_has_zero_dim( H ) ) return FLA_SUCCESS;
// Create a temporary flat copy of the hierarchical object.
FLASH_Obj_create_flat_copy_of_hier( H, &F );
// Randomize the flat matrix object to be SPD.
FLA_Random_spd_matrix( uplo, F );
// Copy the flat object's contents back to the hierarchical object.
FLASH_Obj_hierarchify( F, H );
// Free the temporary flat object.
FLASH_Obj_free( &F );
return FLA_SUCCESS;
}
FLA_Error FLASH_Random_matrix( FLA_Obj H )
{
FLA_Obj F;
// Exit early if one dimension is zero.
if ( FLA_Obj_has_zero_dim( H ) ) return FLA_SUCCESS;
// Create a temporary flat copy of the hierarchical object.
FLASH_Obj_create_flat_copy_of_hier( H, &F );
// Randomize the flat matrix object.
FLA_Random_matrix( F );
// Copy the flat object's contents back to the hierarchical object.
FLASH_Obj_hierarchify( F, H );
// Free the temporary flat object.
FLA_Obj_free( &F );
return FLA_SUCCESS;
}
FLA_Error FLASH_Triangularize( FLA_Uplo uplo, FLA_Diag diag, FLA_Obj A )
{
FLA_Error r_val;
FLA_Obj A_flat;
// Exit early if one dimension is zero.
if ( FLA_Obj_has_zero_dim( A ) ) return FLA_SUCCESS;
// Create a temporary flat copy of the hierarchical object.
FLASH_Obj_create_flat_copy_of_hier( A, &A_flat );
// Triangularize the flat matrix object as specified by uplo and diag.
r_val = FLA_Triangularize( uplo, diag, A_flat );
// Copy the flat object's contents back to the hierarchical object.
FLASH_Obj_hierarchify( A_flat, A );
// Free the temporary flat object.
FLA_Obj_free( &A_flat );
return r_val;
}
void time_Copy(
int param_combo, int type, int nrepeats, int m, int n,
FLA_Obj A, FLA_Obj C, FLA_Obj C_ref,
double *dtime, double *diff, double *gflops )
{
int
irep;
double
dtime_old = 1.0e9;
FLA_Obj
C_old, A_flat, C_flat;
FLASH_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &C_old );
FLASH_Obj_create_flat_conf_to_hier( FLA_NO_TRANSPOSE, A, &A_flat );
FLASH_Obj_create_flat_conf_to_hier( FLA_NO_TRANSPOSE, C, &C_flat );
FLASH_Copy( C, C_old );
for ( irep = 0 ; irep < nrepeats; irep++ )
{
FLASH_Copy( C_old, C );
FLASH_Obj_flatten( A, A_flat );
FLASH_Obj_flatten( C, C_flat );
*dtime = FLA_Clock();
switch( param_combo ){
// Time parameter combination 0
case 0:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Copy( A_flat, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Copy( A, C );
break;
default:
printf("trouble\n");
}
break;
}
}
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
if ( type == FLA_ALG_REFERENCE )
{
FLASH_Obj_hierarchify( C_flat, C_ref );
*diff = 0.0;
}
else
{
*diff = FLASH_Max_elemwise_diff( C, C_ref );
}
*gflops = 2.0 * m * n /
dtime_old /
1.0e9;
*dtime = dtime_old;
FLASH_Copy( C_old, C );
FLASH_Obj_free( &C_old );
FLASH_Obj_free( &A_flat );
FLASH_Obj_free( &C_flat );
}
void time_Apply_Q(
int param_combo, int type, int nrepeats, int m, int n,
FLA_Obj A, FLA_Obj B, FLA_Obj B_ref, FLA_Obj t, FLA_Obj T, FLA_Obj W,
double *dtime, double *diff, double *gflops )
{
int
irep;
double
dtime_old = 1.0e9;
FLA_Obj
B_save, A_flat, B_flat;
FLASH_Obj_create_conf_to( FLA_NO_TRANSPOSE, B, &B_save );
FLASH_Obj_create_flat_conf_to_hier( FLA_NO_TRANSPOSE, A, &A_flat );
FLASH_Obj_create_flat_conf_to_hier( FLA_NO_TRANSPOSE, B, &B_flat );
FLASH_Copy( B, B_save );
for ( irep = 0 ; irep < nrepeats; irep++ )
{
FLASH_Copy( B_save, B );
FLASH_Obj_flatten( A, A_flat );
FLASH_Obj_flatten( B, B_flat );
*dtime = FLA_Clock();
switch( param_combo ){
// Time parameter combination 0
case 0:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Apply_Q( FLA_LEFT, FLA_TRANSPOSE, FLA_COLUMNWISE, A_flat, t, B_flat );
break;
case FLA_ALG_FRONT:
//printf("\n");
FLASH_Apply_Q_UT( FLA_LEFT, FLA_CONJ_TRANSPOSE, FLA_FORWARD, FLA_COLUMNWISE, A, T, W, B );
break;
default:
printf("trouble\n");
}
break;
}
}
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
if ( type == FLA_ALG_REFERENCE )
{
FLA_Trsm_external( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE,
FLA_NONUNIT_DIAG, FLA_ONE, A_flat, B_flat );
FLASH_Obj_hierarchify( B_flat, B_ref );
*diff = 0.0;
}
else
{
FLASH_Trsm( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, A, B );
*diff = FLASH_Max_elemwise_diff( B, B_ref );
}
*gflops = 2.0 *
FLASH_Obj_scalar_length( A ) *
FLASH_Obj_scalar_width( A ) *
FLASH_Obj_scalar_width( B ) /
dtime_old /
1.0e9;
if ( FLA_Obj_is_complex( A ) )
*gflops *= 4.0;
*dtime = dtime_old;
FLASH_Copy( B_save, B );
FLASH_Obj_free( &B_save );
FLASH_Obj_free( &A_flat );
FLASH_Obj_free( &B_flat );
}
void libfla_test_symm_experiment( test_params_t params,
unsigned int var,
char* sc_str,
FLA_Datatype datatype,
unsigned int p_cur,
unsigned int pci,
unsigned int n_repeats,
signed int impl,
double* perf,
double* residual )
{
dim_t b_flash = params.b_flash;
dim_t b_alg_flat = params.b_alg_flat;
double time_min = 1e9;
double time;
unsigned int i;
unsigned int m;
signed int m_input = -1;
unsigned int n;
signed int n_input = -1;
FLA_Side side;
FLA_Uplo uplo;
FLA_Obj A, B, C, x, y, z, w, norm;
FLA_Obj alpha, beta;
FLA_Obj C_save;
FLA_Obj A_test, B_test, C_test;
// Determine the dimensions.
if ( m_input < 0 ) m = p_cur / abs(m_input);
else m = p_cur;
if ( n_input < 0 ) n = p_cur / abs(n_input);
else n = p_cur;
// Translate parameter characters to libflame constants.
FLA_Param_map_char_to_flame_side( &pc_str[pci][0], &side );
FLA_Param_map_char_to_flame_uplo( &pc_str[pci][1], &uplo );
// Create the matrices for the current operation.
if ( side == FLA_LEFT )
{
libfla_test_obj_create( datatype, FLA_NO_TRANSPOSE, sc_str[0], m, m, &A );
// Create vectors for use in test.
FLA_Obj_create( datatype, n, 1, 0, 0, &x );
FLA_Obj_create( datatype, m, 1, 0, 0, &y );
FLA_Obj_create( datatype, m, 1, 0, 0, &z );
FLA_Obj_create( datatype, m, 1, 0, 0, &w );
}
else
{
libfla_test_obj_create( datatype, FLA_NO_TRANSPOSE, sc_str[0], n, n, &A );
// Create vectors for use in test.
FLA_Obj_create( datatype, n, 1, 0, 0, &x );
FLA_Obj_create( datatype, m, 1, 0, 0, &y );
FLA_Obj_create( datatype, m, 1, 0, 0, &z );
FLA_Obj_create( datatype, n, 1, 0, 0, &w );
}
libfla_test_obj_create( datatype, FLA_NO_TRANSPOSE, sc_str[1], m, n, &B );
libfla_test_obj_create( datatype, FLA_NO_TRANSPOSE, sc_str[2], m, n, &C );
// Create a norm scalar.
FLA_Obj_create( FLA_Obj_datatype_proj_to_real( A ), 1, 1, 0, 0, &norm );
// Initialize the test matrices.
FLA_Random_symm_matrix( uplo, A );
FLA_Random_matrix( B );
FLA_Random_matrix( C );
// Initialize the test vectors.
FLA_Random_matrix( x );
FLA_Set( FLA_ZERO, y );
FLA_Set( FLA_ZERO, z );
FLA_Set( FLA_ZERO, w );
// Set constants.
alpha = FLA_TWO;
beta = FLA_MINUS_ONE;
// Save the original object contents in a temporary object.
FLA_Obj_create_copy_of( FLA_NO_TRANSPOSE, C, &C_save );
// Use hierarchical matrices if we're testing the FLASH front-end.
if ( impl == FLA_TEST_HIER_FRONT_END )
{
FLASH_Obj_create_hier_copy_of_flat( A, 1, &b_flash, &A_test );
FLASH_Obj_create_hier_copy_of_flat( B, 1, &b_flash, &B_test );
FLASH_Obj_create_hier_copy_of_flat( C, 1, &b_flash, &C_test );
}
else
{
A_test = A;
B_test = B;
C_test = C;
}
// Create a control tree for the individual variants.
if ( impl == FLA_TEST_FLAT_UNB_VAR ||
impl == FLA_TEST_FLAT_OPT_VAR ||
impl == FLA_TEST_FLAT_BLK_VAR ||
impl == FLA_TEST_FLAT_UNB_EXT ||
impl == FLA_TEST_FLAT_BLK_EXT )
libfla_test_symm_cntl_create( var, b_alg_flat );
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
if ( impl == FLA_TEST_HIER_FRONT_END )
FLASH_Obj_hierarchify( C_save, C_test );
else
FLA_Copy_external( C_save, C_test );
time = FLA_Clock();
libfla_test_symm_impl( impl, side, uplo, alpha, A_test, B_test, beta, C_test );
time = FLA_Clock() - time;
time_min = min( time_min, time );
}
// Copy the solution to flat matrix X.
if ( impl == FLA_TEST_HIER_FRONT_END )
{
FLASH_Obj_flatten( C_test, C );
}
else
{
// No action needed since C_test and C refer to the same object.
}
// Free the hierarchical matrices if we're testing the FLASH front-end.
if ( impl == FLA_TEST_HIER_FRONT_END )
{
FLASH_Obj_free( &A_test );
FLASH_Obj_free( &B_test );
FLASH_Obj_free( &C_test );
}
// Free the control trees if we're testing the variants.
if ( impl == FLA_TEST_FLAT_UNB_VAR ||
impl == FLA_TEST_FLAT_OPT_VAR ||
impl == FLA_TEST_FLAT_BLK_VAR ||
impl == FLA_TEST_FLAT_UNB_EXT ||
impl == FLA_TEST_FLAT_BLK_EXT )
libfla_test_symm_cntl_free();
// Compute the performance of the best experiment repeat.
if ( side == FLA_LEFT )
*perf = ( 1 * m * m * n ) / time_min / FLOPS_PER_UNIT_PERF;
else
*perf = ( 1 * m * n * n ) / time_min / FLOPS_PER_UNIT_PERF;
if ( FLA_Obj_is_complex( A ) ) *perf *= 4.0;
// Compute:
// y = C * x
// and compare to
// z = ( beta * C_orig + alpha * A * B ) x (side = left)
// z = ( beta * C_orig + alpha * B * A ) x (side = right)
FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_ONE, C, x, FLA_ZERO, y );
if ( side == FLA_LEFT )
{
FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_ONE, B, x, FLA_ZERO, w );
FLA_Symv_external( uplo, alpha, A, w, FLA_ZERO, z );
}
else
{
FLA_Symv_external( uplo, FLA_ONE, A, x, FLA_ZERO, w );
FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, B, w, FLA_ZERO, z );
}
FLA_Gemv_external( FLA_NO_TRANSPOSE, beta, C_save, x, FLA_ONE, z );
// Compute || y - z ||.
//FLA_Axpy_external( FLA_MINUS_ONE, y, z );
//FLA_Nrm2_external( z, norm );
//FLA_Obj_extract_real_scalar( norm, residual );
*residual = FLA_Max_elemwise_diff( y, z );
// Free the supporting flat objects.
FLA_Obj_free( &C_save );
// Free the flat test matrices.
FLA_Obj_free( &A );
FLA_Obj_free( &B );
FLA_Obj_free( &C );
FLA_Obj_free( &x );
FLA_Obj_free( &y );
FLA_Obj_free( &z );
FLA_Obj_free( &w );
FLA_Obj_free( &norm );
}
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;
}
void time_Trmm(
int param_combo, int type, int nrepeats, int m, int n,
FLA_Obj A, FLA_Obj C, FLA_Obj C_ref,
double *dtime, double *diff, double *gflops )
{
int
irep;
double
dtime_old = 1.0e9;
FLA_Obj
C_old, A_flat, C_flat;
FLASH_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &C_old );
FLASH_Obj_create_flat_conf_to_hier( FLA_NO_TRANSPOSE, A, &A_flat );
FLASH_Obj_create_flat_conf_to_hier( FLA_NO_TRANSPOSE, C, &C_flat );
FLASH_Copy( C, C_old );
for ( irep = 0 ; irep < nrepeats; irep++ )
{
FLASH_Copy( C_old, C );
FLASH_Obj_flatten( A, A_flat );
FLASH_Obj_flatten( C, C_flat );
*dtime = FLA_Clock();
switch( param_combo ){
// Time parameter combination 0
case 0:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Trmm( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A_flat, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Trmm( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 1
case 1:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Trmm( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A_flat, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Trmm( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 2
case 2:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Trmm( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A_flat, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Trmm( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 3
case 3:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Trmm( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A_flat, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Trmm( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 4
case 4:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Trmm( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A_flat, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Trmm( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 5
case 5:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Trmm( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A_flat, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Trmm( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 6
case 6:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Trmm( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A_flat, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Trmm( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 7
case 7:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Trmm( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A_flat, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Trmm( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 8
case 8:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Trmm( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A_flat, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Trmm( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 9
case 9:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Trmm( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A_flat, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Trmm( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 10
case 10:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Trmm( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A_flat, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Trmm( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 11
case 11:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Trmm( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A_flat, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Trmm( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_TWO, A, C );
break;
default:
printf("trouble\n");
}
break;
}
}
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
if ( type == FLA_ALG_REFERENCE )
{
FLASH_Obj_hierarchify( C_flat, C_ref );
*diff = 0.0;
}
else
{
*diff = FLASH_Max_elemwise_diff( C, C_ref );
}
*gflops = 1.0 *
FLASH_Obj_scalar_length( C ) *
FLASH_Obj_scalar_width( C ) *
FLASH_Obj_scalar_width( A ) /
dtime_old /
1.0e9;
if ( param_combo == 0 ||
param_combo == 3 ||
param_combo == 6 ||
param_combo == 9 )
*gflops *= 4.0;
*dtime = dtime_old;
FLASH_Copy( C_old, C );
FLASH_Obj_free( &C_old );
FLASH_Obj_free( &A_flat );
FLASH_Obj_free( &C_flat );
}
void libfla_test_qrut_experiment( test_params_t params,
unsigned int var,
char* sc_str,
FLA_Datatype datatype,
unsigned int p_cur,
unsigned int pci,
unsigned int n_repeats,
signed int impl,
double* perf,
double* residual )
{
dim_t b_flash = params.b_flash;
dim_t b_alg_flat = params.b_alg_flat;
double time_min = 1e9;
double time;
unsigned int i;
unsigned int m, n;
unsigned int min_m_n;
signed int m_input = -2;
signed int n_input = -1;
FLA_Obj A, T, x, b, y, norm;
FLA_Obj A_save;
FLA_Obj A_test, T_test, x_test, b_test;
// Determine the dimensions.
if ( m_input < 0 ) m = p_cur * abs(m_input);
else m = p_cur;
if ( n_input < 0 ) n = p_cur * abs(n_input);
else n = p_cur;
// Compute the minimum dimension.
min_m_n = min( m, n );
// Create the matrices for the current operation.
libfla_test_obj_create( datatype, FLA_NO_TRANSPOSE, sc_str[0], m, n, &A );
if ( impl == FLA_TEST_FLAT_FRONT_END ||
( impl == FLA_TEST_FLAT_BLK_VAR && var == 1 ) )
libfla_test_obj_create( datatype, FLA_NO_TRANSPOSE, sc_str[1], b_alg_flat, min_m_n, &T );
else if ( var == 2 )
libfla_test_obj_create( datatype, FLA_NO_TRANSPOSE, sc_str[1], min_m_n, min_m_n, &T );
else
libfla_test_obj_create( datatype, FLA_NO_TRANSPOSE, sc_str[1], 1, min_m_n, &T );
// Initialize the test matrices.
FLA_Random_matrix( A );
// Save the original object contents in a temporary object.
FLA_Obj_create_copy_of( FLA_NO_TRANSPOSE, A, &A_save );
// Create vectors to form a linear system.
FLA_Obj_create( datatype, n, 1, 0, 0, &x );
FLA_Obj_create( datatype, m, 1, 0, 0, &b );
FLA_Obj_create( datatype, n, 1, 0, 0, &y );
// Create a real scalar object to hold the norm of A.
FLA_Obj_create( FLA_Obj_datatype_proj_to_real( A ), 1, 1, 0, 0, &norm );
// Create a random right-hand side vector.
FLA_Random_matrix( b );
// Use hierarchical matrices if we're testing the FLASH front-end.
if ( impl == FLA_TEST_HIER_FRONT_END )
{
FLASH_QR_UT_create_hier_matrices( A, 1, &b_flash, &A_test, &T_test );
FLASH_Obj_create_hier_copy_of_flat( b, 1, &b_flash, &b_test );
FLASH_Obj_create_hier_copy_of_flat( x, 1, &b_flash, &x_test );
}
else
{
A_test = A;
T_test = T;
}
// Create a control tree for the individual variants.
if ( impl == FLA_TEST_FLAT_UNB_VAR ||
impl == FLA_TEST_FLAT_OPT_VAR ||
impl == FLA_TEST_FLAT_BLK_VAR )
libfla_test_qrut_cntl_create( var, b_alg_flat );
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
if ( impl == FLA_TEST_HIER_FRONT_END )
FLASH_Obj_hierarchify( A_save, A_test );
else
FLA_Copy_external( A_save, A_test );
time = FLA_Clock();
libfla_test_qrut_impl( impl, A_test, T_test );
time = FLA_Clock() - time;
time_min = min( time_min, time );
}
// Perform a linear solve with the result.
if ( impl == FLA_TEST_HIER_FRONT_END )
{
FLASH_QR_UT_solve( A_test, T_test, b_test, x_test );
FLASH_Obj_flatten( x_test, x );
}
else
{
FLA_QR_UT_solve( A_test, T_test, b, x );
}
// Free the hierarchical matrices if we're testing the FLASH front-end.
if ( impl == FLA_TEST_HIER_FRONT_END )
{
FLASH_Obj_free( &A_test );
FLASH_Obj_free( &T_test );
FLASH_Obj_free( &b_test );
FLASH_Obj_free( &x_test );
}
// Free the control trees if we're testing the variants.
if ( impl == FLA_TEST_FLAT_UNB_VAR ||
impl == FLA_TEST_FLAT_OPT_VAR ||
impl == FLA_TEST_FLAT_BLK_VAR )
libfla_test_qrut_cntl_free();
// Compute the performance of the best experiment repeat.
*perf = ( 2.0 * m * n * n -
( 2.0 / 3.0 ) * n * n * n ) / time_min / FLOPS_PER_UNIT_PERF;
if ( FLA_Obj_is_complex( A ) ) *perf *= 4.0;
// Compute the residual.
FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_ONE, A_save, x, FLA_MINUS_ONE, b );
FLA_Gemv_external( FLA_CONJ_TRANSPOSE, FLA_ONE, A_save, b, FLA_ZERO, y );
FLA_Nrm2_external( y, norm );
FLA_Obj_extract_real_scalar( norm, residual );
// Free the supporting flat objects.
FLA_Obj_free( &x );
FLA_Obj_free( &b );
FLA_Obj_free( &y );
FLA_Obj_free( &norm );
FLA_Obj_free( &A_save );
// Free the flat test matrices.
FLA_Obj_free( &A );
FLA_Obj_free( &T );
}
void time_Syrk(
int param_combo, int type, int nrepeats, int m, int k,
FLA_Obj A, FLA_Obj B, FLA_Obj C, FLA_Obj C_ref,
double *dtime, double *diff, double *gflops )
{
int
irep;
double
dtime_old = 1.0e9;
FLA_Obj
C_old, A_flat, C_flat;
FLASH_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &C_old );
FLASH_Obj_create_flat_conf_to_hier( FLA_NO_TRANSPOSE, A, &A_flat );
FLASH_Obj_create_flat_conf_to_hier( FLA_NO_TRANSPOSE, C, &C_flat );
FLASH_Copy( C, C_old );
for ( irep = 0 ; irep < nrepeats; irep++ )
{
FLASH_Copy( C_old, C );
FLASH_Obj_flatten( A, A_flat );
FLASH_Obj_flatten( C, C_flat );
*dtime = FLA_Clock();
switch( param_combo ){
// Time parameter combination 0
case 0:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Syrk( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_ONE, A_flat, FLA_ZERO, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Syrk( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_ONE, A, FLA_ZERO, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 1
case 1:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Syrk( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_ONE, A_flat, FLA_ZERO, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Syrk( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_ONE, A, FLA_ZERO, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 2
case 2:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Syrk( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_ONE, A_flat, FLA_ZERO, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Syrk( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_ONE, A, FLA_ZERO, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 3
case 3:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Syrk( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_ONE, A_flat, FLA_ZERO, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Syrk( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_ONE, A, FLA_ZERO, C );
break;
default:
printf("trouble\n");
}
break;
}
}
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
if ( type == FLA_ALG_REFERENCE )
{
FLASH_Obj_hierarchify( C_flat, C_ref );
*diff = 0.0;
}
else
{
*diff = FLASH_Max_elemwise_diff( C, C_ref );
}
*gflops = 1.0 * m * m * k /
dtime_old /
1.0e9;
if ( FLA_Obj_is_complex( C ) )
*gflops *= 4.0;
*dtime = dtime_old;
FLASH_Copy( C_old, C );
FLASH_Obj_free( &C_old );
FLASH_Obj_free( &A_flat );
FLASH_Obj_free( &C_flat );
}
void libfla_test_apqut_experiment( test_params_t params,
unsigned int var,
char* sc_str,
FLA_Datatype datatype,
unsigned int p_cur,
unsigned int pci,
unsigned int n_repeats,
signed int impl,
double* perf,
double* residual )
{
dim_t b_flash = params.b_flash;
dim_t b_alg_flat = params.b_alg_flat;
double time_min = 1e9;
double time;
unsigned int i;
unsigned int m, n;
unsigned int min_m_n;
signed int m_input;
signed int n_input;
FLA_Side side;
FLA_Trans trans;
FLA_Direct direct;
FLA_Store storev;
FLA_Obj A, T, W, B, eye, norm;
FLA_Obj B_save;
FLA_Obj A_test, T_test, W_test, B_test;
// Translate parameter characters to libflame constants.
FLA_Param_map_char_to_flame_side( &pc_str[pci][0], &side );
FLA_Param_map_char_to_flame_trans( &pc_str[pci][1], &trans );
FLA_Param_map_char_to_flame_direct( &pc_str[pci][2], &direct );
FLA_Param_map_char_to_flame_storev( &pc_str[pci][3], &storev );
// We want to make sure the Apply_Q_UT routines work with rectangular
// matrices. So we use m > n when testing with column-wise storage (via
// QR factorization) and m < n when testing with row-wise storage (via
// LQ factorization).
if ( storev == FLA_COLUMNWISE )
{
m_input = -1;
n_input = -1;
//m_input = -1;
//n_input = -1;
}
else // if ( storev == FLA_ROWWISE )
{
m_input = -1;
n_input = -1;
//m_input = -1;
//n_input = -1;
}
// Determine the dimensions.
if ( m_input < 0 ) m = p_cur * abs(m_input);
else m = p_cur;
if ( n_input < 0 ) n = p_cur * abs(n_input);
else n = p_cur;
// Compute the minimum dimension.
min_m_n = min( m, n );
// Create the matrices for the current operation.
libfla_test_obj_create( datatype, FLA_NO_TRANSPOSE, sc_str[0], m, n, &A );
libfla_test_obj_create( datatype, FLA_NO_TRANSPOSE, sc_str[1], b_alg_flat, min_m_n, &T );
if ( storev == FLA_COLUMNWISE )
libfla_test_obj_create( datatype, FLA_NO_TRANSPOSE, sc_str[2], m, m, &B );
else
libfla_test_obj_create( datatype, FLA_NO_TRANSPOSE, sc_str[2], n, n, &B );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, B, &eye );
FLA_Apply_Q_UT_create_workspace( T, B, &W );
// Create a real scalar object to hold the norm of A.
FLA_Obj_create( FLA_Obj_datatype_proj_to_real( A ), 1, 1, 0, 0, &norm );
// Initialize the test matrices.
FLA_Random_matrix( A );
FLA_Set_to_identity( B );
FLA_Set_to_identity( eye );
// Save the original object contents in a temporary object.
FLA_Obj_create_copy_of( FLA_NO_TRANSPOSE, B, &B_save );
// Use hierarchical matrices if we're testing the FLASH front-end.
if ( impl == FLA_TEST_HIER_FRONT_END )
{
if ( storev == FLA_COLUMNWISE )
FLASH_QR_UT_create_hier_matrices( A, 1, &b_flash, &A_test, &T_test );
else // if ( storev == FLA_ROWWISE )
FLASH_LQ_UT_create_hier_matrices( A, 1, &b_flash, &A_test, &T_test );
FLASH_Obj_create_hier_copy_of_flat( B, 1, &b_flash, &B_test );
FLASH_Apply_Q_UT_create_workspace( T_test, B_test, &W_test );
}
else // if ( impl == FLA_TEST_FLAT_FRONT_END )
{
A_test = A;
T_test = T;
W_test = W;
B_test = B;
}
// Compute a Householder factorization.
if ( impl == FLA_TEST_HIER_FRONT_END )
{
if ( storev == FLA_COLUMNWISE ) FLASH_QR_UT( A_test, T_test );
else FLASH_LQ_UT( A_test, T_test );
}
else // if ( impl == FLA_TEST_FLAT_FRONT_END )
{
if ( storev == FLA_COLUMNWISE ) FLA_QR_UT( A_test, T_test );
else FLA_LQ_UT( A_test, T_test );
}
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
if ( impl == FLA_TEST_HIER_FRONT_END )
FLASH_Obj_hierarchify( B_save, B_test );
else
FLA_Copy_external( B_save, B_test );
time = FLA_Clock();
libfla_test_apqut_impl( impl, side, trans, direct, storev,
A_test, T_test, W_test, B_test );
time = FLA_Clock() - time;
time_min = min( time_min, time );
}
// Multiply by its conjugate-transpose to get what should be (near) identity
// and then subtract from actual identity to get what should be (near) zero.
if ( impl == FLA_TEST_HIER_FRONT_END )
{
FLASH_Obj_flatten( B_test, B );
FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
FLA_ONE, B, B, FLA_MINUS_ONE, eye );
}
else // if ( impl == FLA_TEST_FLAT_FRONT_END )
{
FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
FLA_ONE, B, B, FLA_MINUS_ONE, eye );
}
// Free the hierarchical matrices if we're testing the FLASH front-end.
if ( impl == FLA_TEST_HIER_FRONT_END )
{
FLASH_Obj_free( &A_test );
FLASH_Obj_free( &T_test );
FLASH_Obj_free( &W_test );
FLASH_Obj_free( &B_test );
}
// Compute the norm of eye, which contains I - Q * Q'.
FLA_Norm1( eye, norm );
FLA_Obj_extract_real_scalar( norm, residual );
// Compute the performance of the best experiment repeat.
*perf = ( 4.0 * m * min_m_n * n -
2.0 * min_m_n * min_m_n * n ) / time_min / FLOPS_PER_UNIT_PERF;
if ( FLA_Obj_is_complex( A ) ) *perf *= 4.0;
// Free the supporting flat objects.
FLA_Obj_free( &B_save );
// Free the flat test matrices.
FLA_Obj_free( &A );
FLA_Obj_free( &T );
FLA_Obj_free( &W );
FLA_Obj_free( &B );
FLA_Obj_free( &eye );
FLA_Obj_free( &norm );
}
void time_SPDinv(
int param_combo, int type, int nrepeats, int m,
FLA_Obj C, FLA_Obj C_ref,
double *dtime, double *diff, double *gflops )
{
int
irep;
double
dtime_old = 1.0e9;
FLA_Obj
C_old, C_flat;
FLASH_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &C_old );
FLASH_Obj_create_flat_conf_to_hier( FLA_NO_TRANSPOSE, C, &C_flat );
FLASH_Copy( C, C_old );
for ( irep = 0 ; irep < nrepeats; irep++ )
{
FLASH_Copy( C_old, C );
FLASH_Obj_flatten( C, C_flat );
*dtime = FLA_Clock();
switch( param_combo ){
case 0:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_SPDinv( FLA_LOWER_TRIANGULAR, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_SPDinv( FLA_LOWER_TRIANGULAR, C );
break;
default:
printf("trouble\n");
}
break;
}
case 1:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_SPDinv( FLA_UPPER_TRIANGULAR, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_SPDinv( FLA_UPPER_TRIANGULAR, C );
break;
default:
printf("trouble\n");
}
break;
}
}
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
if ( type == FLA_ALG_REFERENCE ){
FLASH_Obj_hierarchify( C_flat, C_ref );
*diff = 0.0;
}
else{
*diff = FLASH_Max_elemwise_diff( C, C_ref );
}
*gflops = 1.0 *
FLASH_Obj_scalar_length( C ) *
FLASH_Obj_scalar_length( C ) *
FLASH_Obj_scalar_length( C ) /
dtime_old / 1e9;
*dtime = dtime_old;
FLASH_Copy( C_old, C );
FLASH_Obj_free( &C_old );
FLASH_Obj_free( &C_flat );
}
void time_Trinv(
int param_combo, int type, int nrepeats, int m,
FLA_Diag diag, FLA_Obj A, FLA_Obj A_ref,
double *dtime, double *diff, double *gflops )
{
int
irep;
double
dtime_old = 1.0e9;
FLA_Obj
A_old, A_flat;
FLASH_Obj_create_conf_to( FLA_NO_TRANSPOSE, A, &A_old );
FLASH_Obj_create_flat_conf_to_hier( FLA_NO_TRANSPOSE, A, &A_flat );
FLASH_Copy( A, A_old );
for ( irep = 0 ; irep < nrepeats; irep++ )
{
FLASH_Copy( A_old, A );
FLASH_Obj_flatten( A, A_flat );
*dtime = FLA_Clock();
switch( param_combo ){
case 0:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Trinv( FLA_LOWER_TRIANGULAR, diag, A_flat );
break;
case FLA_ALG_FRONT:
FLASH_Trinv( FLA_LOWER_TRIANGULAR, diag, A );
break;
default:
printf("trouble\n");
}
break;
}
case 1:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Trinv( FLA_UPPER_TRIANGULAR, diag, A_flat );
break;
case FLA_ALG_FRONT:
FLASH_Trinv( FLA_UPPER_TRIANGULAR, diag, A );
break;
default:
printf("trouble\n");
}
break;
}
}
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
if ( type == FLA_ALG_REFERENCE ){
FLASH_Obj_hierarchify( A_flat, A_ref );
*diff = 0.0;
}
else{
*diff = FLASH_Max_elemwise_diff( A, A_ref );
}
*gflops = 1.0 / 4.0 * m * m * m /
dtime_old / 1e9;
*dtime = dtime_old;
FLASH_Copy( A_old, A );
FLASH_Obj_free( &A_old );
FLASH_Obj_free( &A_flat );
}
void libfla_test_eig_gest_experiment( test_params_t params,
unsigned int var,
char* sc_str,
FLA_Datatype datatype,
unsigned int p_cur,
unsigned int pci,
unsigned int n_repeats,
signed int impl,
double* perf,
double* residual )
{
dim_t b_flash = params.b_flash;
dim_t b_alg_flat = params.b_alg_flat;
double time_min = 1e9;
double time;
unsigned int i;
unsigned int m;
signed int m_input = -1;
FLA_Uplo inv;
FLA_Uplo uplo;
FLA_Obj A, B, Y, norm;
FLA_Obj A_save, B_save;
FLA_Obj A_test, B_test, Y_test;
// Determine the dimensions.
if ( m_input < 0 ) m = p_cur / abs(m_input);
else m = p_cur;
// Translate parameter characters to libflame constants.
FLA_Param_map_char_to_flame_inv( &pc_str[pci][0], &inv );
FLA_Param_map_char_to_flame_uplo( &pc_str[pci][1], &uplo );
if ( inv == FLA_NO_INVERSE &&
( ( impl == FLA_TEST_FLAT_UNB_VAR && var == 3 ) ||
( impl == FLA_TEST_FLAT_OPT_VAR && var == 3 ) ||
( impl == FLA_TEST_FLAT_BLK_VAR && var == 3 ) )
)
{
*perf = 0.0;
*residual = 0.0;
return;
}
// Create the matrices for the current operation.
libfla_test_obj_create( datatype, FLA_NO_TRANSPOSE, sc_str[0], m, m, &A );
libfla_test_obj_create( datatype, FLA_NO_TRANSPOSE, sc_str[1], m, m, &Y );
libfla_test_obj_create( datatype, FLA_NO_TRANSPOSE, sc_str[2], m, m, &B );
// Initialize the test matrices.
FLA_Random_spd_matrix( uplo, A );
FLA_Scalr( uplo, FLA_TWO, A );
FLA_Hermitianize( uplo, A );
FLA_Random_spd_matrix( uplo, B );
FLA_Scalr( uplo, FLA_TWO, B );
FLA_Chol( uplo, B );
// Save the original object contents in a temporary object.
FLA_Obj_create_copy_of( FLA_NO_TRANSPOSE, A, &A_save );
FLA_Obj_create_copy_of( FLA_NO_TRANSPOSE, B, &B_save );
// Create a real scalar object to hold the norm of A.
FLA_Obj_create( FLA_Obj_datatype_proj_to_real( A ), 1, 1, 0, 0, &norm );
// Use hierarchical matrices if we're testing the FLASH front-end.
if ( impl == FLA_TEST_HIER_FRONT_END )
{
FLASH_Obj_create_hier_copy_of_flat( A, 1, &b_flash, &A_test );
FLASH_Obj_create_hier_copy_of_flat( Y, 1, &b_flash, &Y_test );
FLASH_Obj_create_hier_copy_of_flat( B, 1, &b_flash, &B_test );
}
else
{
A_test = A;
Y_test = Y;
B_test = B;
}
// Create a control tree for the individual variants.
if ( impl == FLA_TEST_FLAT_UNB_VAR ||
impl == FLA_TEST_FLAT_OPT_VAR ||
impl == FLA_TEST_FLAT_BLK_VAR )
libfla_test_eig_gest_cntl_create( var, b_alg_flat );
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
if ( impl == FLA_TEST_HIER_FRONT_END )
{
FLASH_Obj_hierarchify( A_save, A_test );
FLASH_Obj_hierarchify( B_save, B_test );
}
else
{
FLA_Copy_external( A_save, A_test );
FLA_Copy_external( B_save, B_test );
}
time = FLA_Clock();
libfla_test_eig_gest_impl( impl, inv, uplo, A_test, Y_test, B_test );
time = FLA_Clock() - time;
time_min = min( time_min, time );
}
// Check our solution.
if ( impl == FLA_TEST_HIER_FRONT_END )
{
FLA_Trans trans_left, trans_right;
FLASH_Hermitianize( uplo, A_test );
if ( ( inv == FLA_NO_INVERSE && uplo == FLA_LOWER_TRIANGULAR ) ||
( inv == FLA_INVERSE && uplo == FLA_UPPER_TRIANGULAR ) )
{
trans_left = FLA_CONJ_TRANSPOSE;
trans_right = FLA_NO_TRANSPOSE;
}
else
{
trans_left = FLA_NO_TRANSPOSE;
trans_right = FLA_CONJ_TRANSPOSE;
}
if ( inv == FLA_NO_INVERSE )
{
FLASH_Trsm( FLA_LEFT, uplo, trans_left, FLA_NONUNIT_DIAG,
FLA_ONE, B_test, A_test );
FLASH_Trsm( FLA_RIGHT, uplo, trans_right, FLA_NONUNIT_DIAG,
FLA_ONE, B_test, A_test );
}
else // if ( inv == FLA_INVERSE )
{
FLASH_Trmm( FLA_LEFT, uplo, trans_left, FLA_NONUNIT_DIAG,
FLA_ONE, B_test, A_test );
FLASH_Trmm( FLA_RIGHT, uplo, trans_right, FLA_NONUNIT_DIAG,
FLA_ONE, B_test, A_test );
}
FLASH_Obj_flatten( A_test, A );
}
else
{
FLA_Trans trans_left, trans_right;
FLA_Hermitianize( uplo, A_test );
if ( ( inv == FLA_NO_INVERSE && uplo == FLA_LOWER_TRIANGULAR ) ||
( inv == FLA_INVERSE && uplo == FLA_UPPER_TRIANGULAR ) )
{
trans_left = FLA_CONJ_TRANSPOSE;
trans_right = FLA_NO_TRANSPOSE;
}
else
{
trans_left = FLA_NO_TRANSPOSE;
trans_right = FLA_CONJ_TRANSPOSE;
}
if ( inv == FLA_NO_INVERSE )
{
FLA_Trsm( FLA_LEFT, uplo, trans_left, FLA_NONUNIT_DIAG,
FLA_ONE, B_test, A_test );
FLA_Trsm( FLA_RIGHT, uplo, trans_right, FLA_NONUNIT_DIAG,
FLA_ONE, B_test, A_test );
}
else // if ( inv == FLA_INVERSE )
{
FLA_Trmm( FLA_LEFT, uplo, trans_left, FLA_NONUNIT_DIAG,
FLA_ONE, B_test, A_test );
FLA_Trmm( FLA_RIGHT, uplo, trans_right, FLA_NONUNIT_DIAG,
FLA_ONE, B_test, A_test );
}
}
// Free the hierarchical matrices if we're testing the FLASH front-end.
if ( impl == FLA_TEST_HIER_FRONT_END )
{
FLASH_Obj_free( &A_test );
FLASH_Obj_free( &Y_test );
FLASH_Obj_free( &B_test );
}
// Free the control trees if we're testing the variants.
if ( impl == FLA_TEST_FLAT_UNB_VAR ||
impl == FLA_TEST_FLAT_OPT_VAR ||
impl == FLA_TEST_FLAT_BLK_VAR )
libfla_test_eig_gest_cntl_free();
// Compute the performance of the best experiment repeat.
*perf = 1.0 * m * m * m / time_min / FLOPS_PER_UNIT_PERF;
if ( FLA_Obj_is_complex( A ) ) *perf *= 4.0;
// Compute the residual.
FLA_Axpy_external( FLA_MINUS_ONE, A_save, A );
FLA_Norm1( A, norm );
FLA_Obj_extract_real_scalar( norm, residual );
// Free the supporting flat objects.
FLA_Obj_free( &norm );
FLA_Obj_free( &A_save );
FLA_Obj_free( &B_save );
// Free the flat test matrices.
FLA_Obj_free( &A );
FLA_Obj_free( &Y );
FLA_Obj_free( &B );
}
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;
}
void time_Gemm(
int param_combo, int type, int nrepeats, int m, int k, int n,
FLA_Obj A, FLA_Obj B, FLA_Obj C, FLA_Obj C_ref,
double *dtime, double *diff, double *gflops )
{
int
irep;
double
dtime_old = 1.0e9;
FLA_Obj
C_old, A_flat, B_flat, C_flat;
FLASH_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &C_old );
FLASH_Obj_create_flat_conf_to_hier( FLA_NO_TRANSPOSE, A, &A_flat );
FLASH_Obj_create_flat_conf_to_hier( FLA_NO_TRANSPOSE, B, &B_flat );
FLASH_Obj_create_flat_conf_to_hier( FLA_NO_TRANSPOSE, C, &C_flat );
FLASH_Copy( C, C_old );
for ( irep = 0 ; irep < nrepeats; irep++ )
{
FLASH_Copy( C_old, C );
FLASH_Obj_flatten( A, A_flat );
FLASH_Obj_flatten( B, B_flat );
FLASH_Obj_flatten( C, C_flat );
*dtime = FLA_Clock();
switch( param_combo ){
// Time parameter combination 0
case 0:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Gemm( FLA_CONJ_TRANSPOSE, FLA_CONJ_TRANSPOSE, FLA_ONE, A_flat, B_flat, FLA_ZERO, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Gemm( FLA_CONJ_TRANSPOSE, FLA_CONJ_TRANSPOSE, FLA_ONE, A, B, FLA_ZERO, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 1
case 1:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Gemm( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_ONE, A_flat, B_flat, FLA_ZERO, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Gemm( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_ONE, A, B, FLA_ZERO, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 2
case 2:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Gemm( FLA_CONJ_TRANSPOSE, FLA_TRANSPOSE, FLA_ONE, A_flat, B_flat, FLA_ZERO, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Gemm( FLA_CONJ_TRANSPOSE, FLA_TRANSPOSE, FLA_ONE, A, B, FLA_ZERO, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 3
case 3:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Gemm( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, FLA_ONE, A_flat, B_flat, FLA_ZERO, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Gemm( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, FLA_ONE, A, B, FLA_ZERO, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 4
case 4:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Gemm( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_ONE, A_flat, B_flat, FLA_ZERO, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Gemm( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_ONE, A, B, FLA_ZERO, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 5
case 5:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Gemm( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, FLA_ONE, A_flat, B_flat, FLA_ZERO, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Gemm( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, FLA_ONE, A, B, FLA_ZERO, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 6
case 6:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Gemm( FLA_TRANSPOSE, FLA_CONJ_TRANSPOSE, FLA_ONE, A_flat, B_flat, FLA_ZERO, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Gemm( FLA_TRANSPOSE, FLA_CONJ_TRANSPOSE, FLA_ONE, A, B, FLA_ZERO, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 7
case 7:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Gemm( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_ONE, A_flat, B_flat, FLA_ZERO, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Gemm( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_ONE, A, B, FLA_ZERO, C );
break;
default:
printf("trouble\n");
}
break;
}
// Time parameter combination 8
case 8:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Gemm( FLA_TRANSPOSE, FLA_TRANSPOSE, FLA_ONE, A_flat, B_flat, FLA_ZERO, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Gemm( FLA_TRANSPOSE, FLA_TRANSPOSE, FLA_ONE, A, B, FLA_ZERO, C );
break;
default:
printf("trouble\n");
}
break;
}
}
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
if ( type == FLA_ALG_REFERENCE )
{
FLASH_Obj_hierarchify( C_flat, C_ref );
*diff = 0.0;
}
else
{
*diff = FLASH_Max_elemwise_diff( C, C_ref );
}
*gflops = 2.0 * m * k * n /
dtime_old /
1.0e9;
if ( param_combo == 0 ||
param_combo == 1 ||
param_combo == 2 ||
param_combo == 3 ||
param_combo == 6 )
*gflops *= 4.0;
*dtime = dtime_old;
FLASH_Copy( C_old, C );
FLASH_Obj_free( &C_old );
FLASH_Obj_free( &A_flat );
FLASH_Obj_free( &B_flat );
FLASH_Obj_free( &C_flat );
}
