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 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 );
}
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 time_LU(
int pivot_combo, int type, int nrepeats, int m, int n, dim_t nb_alg, dim_t nb_flash,
FLA_Obj A, FLA_Obj p, FLA_Obj x, FLA_Obj b, FLA_Obj norm,
double *dtime, double *diff, double *gflops )
{
int
irep;
double
dtime_old = 1.0e9;
FLA_Obj AH_save, b_save;
FLA_Obj AH, pH, bH, LH;
FLASH_LU_incpiv_create_hier_matrices( A, 1, &nb_flash, nb_alg,
&AH, &pH, &LH );
FLASH_Obj_create_hier_copy_of_flat( b, 1, &nb_flash, &bH );
FLASH_Obj_create_copy_of( FLA_NO_TRANSPOSE, AH, &AH_save );
FLA_Obj_create_copy_of( FLA_NO_TRANSPOSE, b, &b_save );
for ( irep = 0 ; irep < nrepeats; irep++ )
{
FLASH_Copy( AH_save, AH );
*dtime = FLA_Clock();
switch( pivot_combo ){
case 0:
{
switch( type )
{
case FLA_ALG_FRONT_OPT0:
FLASH_LU_incpiv_noopt( AH, pH, LH );
break;
case FLA_ALG_FRONT_OPT1:
FLASH_LU_incpiv_opt1( AH, pH, LH );
break;
default:
printf("trouble\n");
}
break;
}
}
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
{
FLASH_FS_incpiv( AH, pH, LH, bH );
FLASH_Trsv( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
AH, bH );
FLASH_Obj_flatten( bH, x );
FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_ONE,
A, x, FLA_MINUS_ONE, b );
FLA_Nrm2_external( b, norm );
FLA_Obj_extract_real_scalar( norm, diff );
}
*gflops = 2.0 / 3.0 * m * m * n /
dtime_old / 1e9;
if ( FLA_Obj_is_complex( A ) ) *gflops *= 4.0;
*dtime = dtime_old;
FLA_Copy( b_save, b );
FLASH_Obj_free( &AH );
FLASH_Obj_free( &pH );
FLASH_Obj_free( &bH );
FLASH_Obj_free( &LH );
FLA_Obj_free( &b_save );
FLASH_Obj_free( &AH_save );
}
void time_Syr2k(
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, 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_Syr2k( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_ONE, A_flat, B_flat, FLA_ZERO, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Syr2k( FLA_LOWER_TRIANGULAR, FLA_NO_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_Syr2k( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_ONE, A_flat, B_flat, FLA_ZERO, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Syr2k( FLA_LOWER_TRIANGULAR, FLA_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_Syr2k( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_ONE, A_flat, B_flat, FLA_ZERO, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Syr2k( FLA_UPPER_TRIANGULAR, FLA_NO_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_Syr2k( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_ONE, A_flat, B_flat, FLA_ZERO, C_flat );
break;
case FLA_ALG_FRONT:
FLASH_Syr2k( FLA_UPPER_TRIANGULAR, 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 * 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( &B_flat );
FLASH_Obj_free( &C_flat );
}
void libfla_test_lu_piv_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;
signed int m_input = -1;
signed int n_input = -1;
FLA_Obj A, p, x, b, norm;
FLA_Obj A_save;
FLA_Obj A_test, p_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;
// Create the matrices for the current operation.
libfla_test_obj_create( datatype, FLA_NO_TRANSPOSE, sc_str[0], m, n, &A );
FLA_Obj_create( FLA_INT, min( m, n ), 1, 0, 0, &p );
// 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, m, 1, 0, 0, &x );
FLA_Obj_create( datatype, m, 1, 0, 0, &b );
// 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_Obj_create_hier_copy_of_flat( A, 1, &b_flash, &A_test );
FLASH_Obj_create_hier_copy_of_flat( p, 1, &b_flash, &p_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;
p_test = p;
}
// 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_lu_piv_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_lu_piv_impl( impl, A_test, p_test );
time = FLA_Clock() - time;
time_min = min( time_min, time );
}
#ifndef AMD_ONLY_PERFORMANCE
// Perform a linear solve with the result.
if ( impl == FLA_TEST_HIER_FRONT_END )
{
FLASH_LU_piv_solve( A_test, p_test, b_test, x_test );
FLASH_Obj_flatten( x_test, x );
}
else
{
FLA_LU_piv_solve( A_test, p_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( &p_test );
FLASH_Obj_free( &b_test );
FLASH_Obj_free( &x_test );
}
#endif
// 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_lu_piv_cntl_free();
// Compute the performance of the best experiment repeat.
*perf = 2.0 / 3.0 * m * m * m / time_min / FLOPS_PER_UNIT_PERF;
if ( FLA_Obj_is_complex( A ) ) *perf *= 4.0;
#ifndef AMD_ONLY_PERFORMANCE
// Compute the residual.
FLA_Gemv_external( FLA_NO_TRANSPOSE,
FLA_ONE, A_save, x, FLA_MINUS_ONE, b );
FLA_Nrm2_external( b, norm );
FLA_Obj_extract_real_scalar( norm, residual );
#else
*residual = 0.0;
#endif
// Free the supporting flat objects.
FLA_Obj_free( &p );
FLA_Obj_free( &x );
FLA_Obj_free( &b );
FLA_Obj_free( &norm );
FLA_Obj_free( &A_save );
// Free the flat test matrices.
FLA_Obj_free( &A );
}
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 = 1.0 * 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 );
}
int main(int argc, char *argv[])
{
int
datatype,
n_blocks_m,
n_threads,
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 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_flat_ref, A_flat, B, B_flat, D, D_flat, t, T, T_flat;
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 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, "%c enter the number of SuperMatrix threads: ", '%' );
scanf( "%d", &n_threads );
fprintf( stdout, "%c %d\n", '%', n_threads );
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_set_num_threads( n_threads );
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);
nb_flash = n;
for ( param_combo = 0; param_combo < n_param_combos; param_combo++ )
{
FLA_Obj_create( datatype, m, nb_flash, &A_flat );
FLA_Obj_create( datatype, m, nb_flash, &A_flat_ref );
FLA_Obj_create( datatype, m, nb_flash, &T_flat );
FLA_Obj_create( datatype, nb_flash, 1, &t );
FLASH_Obj_create( datatype, m, nb_flash, 1, &nb_flash, &A );
n_blocks_m = FLA_Obj_length( A );
FLASH_Obj_create_ext( datatype, nb_alg * n_blocks_m, nb_flash, 1, &nb_alg, &nb_flash, &T );
FLA_Set( FLA_ZERO, T_flat );
FLASH_Set( FLA_ZERO, T );
FLASH_Random_matrix( A );
FLASH_Obj_flatten( A, A_flat );
FLA_Part_2x1( A, &B,
&D, 1, FLA_TOP );
FLA_Part_2x1( A_flat, &B_flat,
&D_flat, FLA_Obj_width( A_flat ), FLA_TOP );
FLA_Triangularize( FLA_UPPER_TRIANGULAR, FLA_NONUNIT_DIAG, *(FLASH_OBJ_PTR_AT(B)) );
FLA_Triangularize( FLA_UPPER_TRIANGULAR, FLA_NONUNIT_DIAG, B_flat );
fprintf( stdout, "data_qr2ut_%s( %d, 1:5 ) = [ %d ", pc_str[param_combo], i, p );
fflush( stdout );
time_QR2_UT( param_combo, FLA_ALG_REFERENCE, n_repeats, m, n,
A, A_flat_ref, B, B_flat, D, D_flat, A_flat, t, T, T_flat, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
time_QR2_UT( param_combo, FLA_ALG_FRONT, n_repeats, m, n,
A, A_flat_ref, B, B_flat, D, D_flat, A_flat, t, T, T_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( &A_flat_ref );
FLA_Obj_free( &T_flat );
FLA_Obj_free( &t );
FLASH_Obj_free( &A );
FLASH_Obj_free( &T );
}
fprintf( stdout, "\n" );
}
fprintf( stdout, "figure;\n" );
fprintf( stdout, "hold on;\n" );
for ( i = 0; i < n_param_combos; i++ ) {
fprintf( stdout, "plot( data_qr2ut_%s( :,1 ), data_qr2ut_%s( :, 2 ), '%c:%c' ); \n",
pc_str[i], pc_str[i], colors[ i ], ticks[ i ] );
fprintf( stdout, "plot( data_qr2ut_%s( :,1 ), data_qr2ut_%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\\_qr2ut\\_%s', 'fla\\_qr2ut\\_%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 qr2ut front-end performance (%s, %s)' );\n",
m_dim_desc, n_dim_desc );
fprintf( stdout, "print -depsc qr2ut_front_%s_%s.eps\n", m_dim_tag, n_dim_tag );
fprintf( stdout, "hold off;\n");
fflush( stdout );
FLA_Finalize( );
return 0;
}
