FLA_Error FLA_Chol_solve( FLA_Uplo uplo, FLA_Obj A, FLA_Obj B, FLA_Obj X )
{
// Check parameters.
if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
FLA_Chol_solve_check( uplo, A, B, X );
if ( FLA_Obj_is_identical( B, X ) == FALSE )
FLA_Copy_external( B, X );
if ( uplo == FLA_LOWER_TRIANGULAR )
{
FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE,
FLA_NONUNIT_DIAG, FLA_ONE, A, X );
FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE,
FLA_NONUNIT_DIAG, FLA_ONE, A, X );
}
else // if ( uplo == FLA_UPPER_TRIANGULAR )
{
FLA_Trsm_external( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE,
FLA_NONUNIT_DIAG, FLA_ONE, A, X );
FLA_Trsm_external( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE,
FLA_NONUNIT_DIAG, FLA_ONE, A, X );
}
return FLA_SUCCESS;
}
FLA_Error FLA_QR_UT_solve( FLA_Obj A, FLA_Obj T, FLA_Obj B, FLA_Obj X )
{
FLA_Obj W, Y;
FLA_Obj AT, AB;
FLA_Obj YT, YB;
// Check parameters.
if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
FLA_QR_UT_solve_check( A, T, B, X );
FLA_Apply_Q_UT_create_workspace( T, B, &W );
FLA_Obj_create_copy_of( FLA_NO_TRANSPOSE, B, &Y );
FLA_Apply_Q_UT( FLA_LEFT, FLA_CONJ_TRANSPOSE, FLA_FORWARD, FLA_COLUMNWISE,
A, T, W, Y );
FLA_Part_2x1( A, &AT,
&AB, FLA_Obj_width( A ), FLA_TOP );
FLA_Part_2x1( Y, &YT,
&YB, FLA_Obj_width( A ), FLA_TOP );
FLA_Trsm_external( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE,
FLA_NONUNIT_DIAG, FLA_ONE, AT, YT );
FLA_Copy_external( YT, X );
FLA_Obj_free( &Y );
FLA_Obj_free( &W );
return FLA_SUCCESS;
}
FLA_Error FLA_LU_piv_copy_task( FLA_Obj A, FLA_Obj p, FLA_Obj U, fla_lu_t* cntl )
{
FLA_Error r_val;
r_val = FLA_LU_piv_task( A, p, cntl );
FLA_Copy_external( A, U );
return r_val;
}
FLA_Error FLA_UDdate_UT_solve( FLA_Obj R, FLA_Obj bR, FLA_Obj x )
{
// Check parameters.
if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
FLA_UDdate_UT_solve_check( R, bR, x );
// Copy the contents of bR to x so that after the triangular solve, the
// solution resides in x (and bR is preserved).
FLA_Copy_external( bR, x );
// Perform a triangular solve with R the right-hand side.
FLA_Trsm_external( FLA_LEFT, FLA_UPPER_TRIANGULAR,
FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, R, x );
return FLA_SUCCESS;
}
FLA_Error FLA_Copy( FLA_Obj A, FLA_Obj B )
{
FLA_Error r_val;
#ifdef FLA_ENABLE_BLAS1_FRONT_END_CNTL_TREES
// Check parameters.
if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
FLA_Copy_check( A, B );
// Invoke FLA_Copy_internal() with flat control tree that simply calls
// external wrapper.
r_val = FLA_Copy_internal( A, B, fla_copy_cntl_blas );
#else
r_val = FLA_Copy_external( A, B );
#endif
return r_val;
}
FLA_Error FLA_Trmvsx_external( FLA_Uplo uplo, FLA_Trans transa, FLA_Diag diag, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y )
{
FLA_Obj x_copy;
if ( FLA_Check_error_level() == FLA_FULL_ERROR_CHECKING )
FLA_Trmvsx_check( uplo, transa, diag, alpha, A, x, beta, y );
if ( FLA_Obj_has_zero_dim( A ) ) return FLA_SUCCESS;
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, x, &x_copy );
FLA_Copy_external( x, x_copy );
FLA_Trmv_external( uplo, transa, diag, A, x_copy );
FLA_Scal_external( beta, y );
FLA_Axpy_external( alpha, x_copy, y );
FLA_Obj_free( &x_copy );
return FLA_SUCCESS;
}
void time_Sylv_nn(
int variant, int type, int n_repeats, int m, int n, int nb_alg,
FLA_Obj isgn, FLA_Obj A, FLA_Obj B, FLA_Obj C, FLA_Obj C_ref, FLA_Obj scale,
double *dtime, double *diff, double *gflops )
{
int
irep;
double
dtime_old = 1.0e9;
FLA_Obj
C_old;
fla_blocksize_t*
bp;
fla_sylv_t*
cntl_sylv_var;
fla_sylv_t*
cntl_sylv_unb;
fla_gemm_t*
cntl_gemm_blas;
/*
if( type == FLA_ALG_UNBLOCKED && n > 400 )
{
*gflops = 0.0;
*diff = 0.0;
return;
}
*/
bp = FLA_Blocksize_create( nb_alg, nb_alg, nb_alg, nb_alg );
cntl_sylv_unb = FLA_Cntl_sylv_obj_create( FLA_FLAT, FLA_UNB_OPT_VARIANT1, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL );
cntl_gemm_blas = FLA_Cntl_gemm_obj_create( FLA_FLAT, FLA_SUBPROBLEM, NULL, NULL, NULL );
cntl_sylv_var = FLA_Cntl_sylv_obj_create( FLA_FLAT, variant, bp, cntl_sylv_unb, cntl_sylv_unb, cntl_sylv_unb, cntl_gemm_blas, cntl_gemm_blas, cntl_gemm_blas, cntl_gemm_blas, cntl_gemm_blas, cntl_gemm_blas, cntl_gemm_blas, cntl_gemm_blas );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &C_old );
FLA_Copy_external( C, C_old );
for ( irep = 0 ; irep < n_repeats; irep++ ){
FLA_Copy_external( C_old, C );
*dtime = FLA_Clock();
switch( variant ){
case 0:
/* Time reference implementation */
REF_Sylv_nn( isgn, A, B, C, scale );
break;
case 1:{
/* Time variant 1 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var1( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var1( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
case 2:{
/* Time variant 2 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var2( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var2( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
case 3:{
/* Time variant 3 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var3( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var3( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
case 4:{
/* Time variant 4 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var4( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var4( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
case 5:{
/* Time variant 5 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var5( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var5( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
case 6:{
/* Time variant 6 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var6( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var6( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
case 7:{
/* Time variant 7 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var7( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var7( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
case 8:{
/* Time variant 8 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var8( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var8( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
case 9:{
/* Time variant 9 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var9( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var9( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
case 10:{
/* Time variant 10 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var10( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var10( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
case 11:{
/* Time variant 11 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var11( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var11( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
case 12:{
/* Time variant 12 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var12( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var12( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
case 13:{
/* Time variant 13 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var13( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var13( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
case 14:{
/* Time variant 14 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var14( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var14( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
case 15:{
/* Time variant 15 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var15( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var15( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
case 16:{
/* Time variant 16 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var16( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var16( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
case 17:{
/* Time variant 17 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var17( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var17( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
case 18:{
/* Time variant 18 */
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Sylv_nn_opt_var18( isgn, A, B, C, scale );
break;
case FLA_ALG_BLOCKED:
FLA_Sylv_nn_blk_var18( isgn, A, B, C, scale, cntl_sylv_var );
break;
default:
printf("trouble\n");
}
break;
}
}
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
FLA_Cntl_obj_free( cntl_sylv_var );
FLA_Cntl_obj_free( cntl_sylv_unb );
FLA_Cntl_obj_free( cntl_gemm_blas );
FLA_Blocksize_free( bp );
if ( variant == 0 ){
FLA_Copy_external( C, C_ref );
*diff = 0.0;
}
else{
*diff = FLA_Max_elemwise_diff( C, C_ref );
}
*gflops = ( m * m * n + n * n * m ) /
dtime_old / 1e9;
if ( FLA_Obj_is_complex( C ) )
*gflops *= 4.0;
*dtime = dtime_old;
FLA_Copy_external( C_old, C );
FLA_Obj_free( &C_old );
}
FLA_Error FLA_Copy_task( FLA_Obj A, FLA_Obj B, fla_copy_t* cntl )
{
return FLA_Copy_external( A, B );
}
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;
if ( param_combo != 4 )
{
*gflops = 0.0;
*diff = 0.0;
return;
}
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &C_old );
FLA_Copy_external( C, C_old );
for ( irep = 0 ; irep < nrepeats; irep++ ){
FLA_Copy_external( C_old, C );
*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, B, FLA_ZERO, C );
break;
case FLA_ALG_FRONT:
FLA_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, B, FLA_ZERO, C );
break;
case FLA_ALG_FRONT:
FLA_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, B, FLA_ZERO, C );
break;
case FLA_ALG_FRONT:
FLA_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, B, FLA_ZERO, C );
break;
case FLA_ALG_FRONT:
FLA_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, B, FLA_ZERO, C );
break;
case FLA_ALG_FRONT:
//FLA_Gemm( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_ONE, A, B, FLA_ZERO, C );
//FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_ONE, A, B, FLA_ZERO, C );
FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_ONE, A, B, FLA_ONE, 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, B, FLA_ZERO, C );
break;
case FLA_ALG_FRONT:
FLA_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, B, FLA_ZERO, C );
break;
case FLA_ALG_FRONT:
FLA_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, B, FLA_ZERO, C );
break;
case FLA_ALG_FRONT:
FLA_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, B, FLA_ZERO, C );
break;
case FLA_ALG_FRONT:
FLA_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 )
{
FLA_Copy_external( C, C_ref );
*diff = 0.0;
}
else
{
*diff = FLA_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;
FLA_Copy_external( C_old, C );
FLA_Obj_free( &C_old );
}
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_QR_UT(
int variant, int type, int nrepeats, int m, int n,
FLA_Obj A, FLA_Obj A_ref, FLA_Obj t, FLA_Obj T, FLA_Obj W, FLA_Obj b, FLA_Obj b_orig,
double *dtime, double *diff, double *gflops )
{
int
irep;
double
dtime_old = 1.0e9;
FLA_Obj
A_save, b_save, norm;
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, A, &A_save );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, b, &b_save );
if ( FLA_Obj_is_single_precision( A ) )
FLA_Obj_create( FLA_FLOAT, 1, 1, 0, 0, &norm );
else
FLA_Obj_create( FLA_DOUBLE, 1, 1, 0, 0, &norm );
FLA_Copy_external( A, A_save );
FLA_Copy_external( b, b_save );
for ( irep = 0 ; irep < nrepeats; irep++ ){
FLA_Copy_external( A_save, A );
*dtime = FLA_Clock();
switch( variant ){
case 0:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_QR_UT( A, t );
break;
case FLA_ALG_FRONT:
FLA_QR_UT( A, T );
break;
default:
printf("trouble\n");
}
break;
}
}
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
if ( type == FLA_ALG_REFERENCE )
{
FLA_Obj AT, AB;
FLA_Obj bT, bB;
FLA_Obj y;
FLA_Obj_create( FLA_Obj_datatype( b ), n, 1, 0, 0, &y );
FLA_Copy_external( b, b_orig );
if ( FLA_Obj_is_real( A ) )
FLA_Apply_Q_blk_external( FLA_LEFT, FLA_TRANSPOSE, FLA_COLUMNWISE, A, t, b );
else
FLA_Apply_Q_blk_external( FLA_LEFT, FLA_CONJ_TRANSPOSE, FLA_COLUMNWISE, A, t, b );
FLA_Part_2x1( A, &AT,
&AB, FLA_Obj_width( A ), FLA_TOP );
FLA_Part_2x1( b, &bT,
&bB, FLA_Obj_width( A ), FLA_TOP );
FLA_Trsm_external( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE,
FLA_NONUNIT_DIAG, FLA_ONE, AT, bT );
FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A_save, bT, FLA_ONE, b_orig );
FLA_Gemv_external( FLA_CONJ_TRANSPOSE, FLA_ONE, A_save, b_orig, FLA_ZERO, y );
FLA_Nrm2_external( y, norm );
FLA_Obj_extract_real_scalar( norm, diff );
FLA_Obj_free( &y );
}
else
{
FLA_Obj x, y;
FLA_Obj_create( FLA_Obj_datatype( b ), n, 1, 0, 0, &y );
FLA_Obj_create( FLA_Obj_datatype( b ), n, 1, 0, 0, &x );
FLA_Copy_external( b, b_orig );
FLA_QR_UT_solve( A, T, b, x );
FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A_save, x, FLA_ONE, b_orig );
FLA_Gemv_external( FLA_CONJ_TRANSPOSE, FLA_ONE, A_save, b_orig, FLA_ZERO, y );
FLA_Nrm2_external( y, norm );
FLA_Obj_extract_real_scalar( norm, diff );
FLA_Obj_free( &x );
FLA_Obj_free( &y );
}
*gflops = ( 2.0 * m * n * n -
( 2.0 / 3.0 ) * n * n * n ) /
dtime_old / 1e9;
if ( FLA_Obj_is_complex( A ) )
*gflops *= 4.0;
*dtime = dtime_old;
FLA_Copy_external( A_save, A );
FLA_Copy_external( b_save, b );
FLA_Obj_free( &A_save );
FLA_Obj_free( &b_save );
FLA_Obj_free( &norm );
}
int main(int argc, char *argv[])
{
int
datatype,
m_input,
m,
p_first, p_last, p_inc,
p,
nb_alg,
variant,
n_repeats,
i, j,
n_variants = N_VARIANTS;
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, b, b_orig, norm;
FLA_Init();
fprintf( stdout, "%c number of repeats:", '%' );
scanf( "%d", &n_repeats );
fprintf( stdout, "%c %d\n", '%', n_repeats );
fprintf( stdout, "%c Enter blocking size:", '%' );
scanf( "%d", &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 (-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;
for ( p = p_first, i = 1; p <= p_last; p += p_inc, i += 1 )
{
m = m_input;
if( m < 0 ) m = p / f2c_abs(m_input);
FLA_Obj_create( datatype, m, m, 0, 0, &A );
FLA_Obj_create( datatype, m, 1, 0, 0, &b );
FLA_Obj_create( datatype, m, 1, 0, 0, &b_orig );
/*
FLA_Obj_create( datatype, m, m, m, 1, &A );
FLA_Obj_create( datatype, m, 1, 1, 1, &b );
FLA_Obj_create( datatype, m, 1, 1, 1, &b_orig );
*/
if ( FLA_Obj_is_single_precision( A ) )
FLA_Obj_create( FLA_FLOAT, 1, 1, 0, 0, &norm );
else
FLA_Obj_create( FLA_DOUBLE, 1, 1, 0, 0, &norm );
FLA_Random_tri_matrix( FLA_UPPER_TRIANGULAR, FLA_NONUNIT_DIAG, A );
FLA_Random_matrix( b );
FLA_Copy_external( b, b_orig );
/*
time_Trinv_un( 0, FLA_ALG_REFERENCE, n_repeats, m, nb_alg,
A, b, b_orig, norm, &dtime, &diff, &gflops );
fprintf( stdout, "data_REF( %d, 1:2 ) = [ %d %6.3lf ]; \n", i, p, gflops );
fflush( stdout );
*/
for ( variant = 1; variant <= n_variants; variant++ ){
fprintf( stdout, "data_var%d( %d, 1:7 ) = [ %d ", variant, i, p );
fflush( stdout );
time_Trinv_un( variant, FLA_ALG_UNBLOCKED, n_repeats, m, nb_alg,
A, b, b_orig, norm, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
time_Trinv_un( variant, FLA_ALG_UNB_OPT, n_repeats, m, nb_alg,
A, b, b_orig, norm, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
time_Trinv_un( variant, FLA_ALG_BLOCKED, n_repeats, m, nb_alg,
A, b, b_orig, norm, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
fprintf( stdout, " ]; \n" );
fflush( stdout );
}
FLA_Obj_free( &A );
FLA_Obj_free( &b );
FLA_Obj_free( &b_orig );
FLA_Obj_free( &norm );
fprintf( stdout, "\n" );
}
/*
fprintf( stdout, "figure;\n" );
fprintf( stdout, "hold on;\n" );
fprintf( stdout, "plot( data_REF( :,1 ), data_REF( :, 2 ), '-' ); \n" );
for ( i = 1; i <= n_variants; i++ ){
fprintf( stdout, "plot( data_var%d( :,1 ), data_var%d( :, 2 ), '%c:%c' ); \n",
variant, variant, colors[ i ], ticks[ i ] );
}
fprintf( stdout, "legend( ... \n" );
fprintf( stdout, "'Reference', ... \n" );
for ( i = 1; i <= n_variants; i++ )
fprintf( stdout, "'FLAME var%d', ... \n", i );
fprintf( stdout, "'Location', 'SouthWest' ); \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 trinv\\_u performance (%s)' );\n",
m_dim_desc );
fprintf( stdout, "print -depsc trinv_l_%s.eps\n", m_dim_tag );
fprintf( stdout, "hold off;\n");
fflush( stdout );
*/
FLA_Finalize( );
}
void time_Her2k_ln(
int variant, int type, int nrepeats, int n, int nb_alg,
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;
fla_blocksize_t*
bp;
fla_gemm_t*
cntl_gemm_blas;
fla_her2k_t*
cntl_her2k_blas;
fla_her2k_t*
cntl_her2k_var;
bp = FLA_Blocksize_create( nb_alg, nb_alg, nb_alg, nb_alg );
cntl_gemm_blas = FLA_Cntl_gemm_obj_create( FLA_FLAT, FLA_SUBPROBLEM, NULL, NULL );
cntl_her2k_blas = FLA_Cntl_her2k_obj_create( FLA_FLAT, FLA_SUBPROBLEM, NULL, NULL, NULL, NULL );
cntl_her2k_var = FLA_Cntl_her2k_obj_create( FLA_FLAT, variant, bp, cntl_her2k_blas, cntl_gemm_blas, cntl_gemm_blas );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &C_old );
FLA_Copy_external( C, C_old );
for ( irep = 0 ; irep < nrepeats; irep++ )
{
FLA_Copy_external( C_old, C );
*dtime = FLA_Clock();
switch( variant ){
case 0:
// Time reference implementation
REF_Her2k( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_ONE, A, B, FLA_ONE, C );
break;
case 1:{
// Time variant 1
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Her2k_ln_unb_var1( FLA_ONE, A, B, FLA_ONE, C );
break;
case FLA_ALG_BLOCKED:
FLA_Her2k_ln_blk_var1( FLA_ONE, A, B, FLA_ONE, C, cntl_her2k_var );
break;
default:
printf("trouble\n");
}
break;
}
case 2:{
// Time variant 2
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Her2k_ln_unb_var2( FLA_ONE, A, B, FLA_ONE, C );
break;
case FLA_ALG_BLOCKED:
FLA_Her2k_ln_blk_var2( FLA_ONE, A, B, FLA_ONE, C, cntl_her2k_var );
break;
default:
printf("trouble\n");
}
break;
}
case 3:{
// Time variant 3
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Her2k_ln_unb_var3( FLA_ONE, A, B, FLA_ONE, C );
break;
case FLA_ALG_BLOCKED:
FLA_Her2k_ln_blk_var3( FLA_ONE, A, B, FLA_ONE, C, cntl_her2k_var );
break;
default:
printf("trouble\n");
}
break;
}
case 4:{
// Time variant 4
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Her2k_ln_unb_var4( FLA_ONE, A, B, FLA_ONE, C );
break;
case FLA_ALG_BLOCKED:
FLA_Her2k_ln_blk_var4( FLA_ONE, A, B, FLA_ONE, C, cntl_her2k_var );
break;
default:
printf("trouble\n");
}
break;
}
case 5:{
// Time variant 5
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Her2k_ln_unb_var5( FLA_ONE, A, B, FLA_ONE, C );
break;
case FLA_ALG_BLOCKED:
FLA_Her2k_ln_blk_var5( FLA_ONE, A, B, FLA_ONE, C, cntl_her2k_var );
break;
default:
printf("trouble\n");
}
break;
}
case 6:{
// Time variant 6
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Her2k_ln_unb_var6( FLA_ONE, A, B, FLA_ONE, C );
break;
case FLA_ALG_BLOCKED:
FLA_Her2k_ln_blk_var6( FLA_ONE, A, B, FLA_ONE, C, cntl_her2k_var );
break;
default:
printf("trouble\n");
}
break;
}
case 7:{
// Time variant 7
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Her2k_ln_unb_var7( FLA_ONE, A, B, FLA_ONE, C );
break;
case FLA_ALG_BLOCKED:
FLA_Her2k_ln_blk_var7( FLA_ONE, A, B, FLA_ONE, C, cntl_her2k_var );
break;
default:
printf("trouble\n");
}
break;
}
case 8:{
// Time variant 8
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Her2k_ln_unb_var8( FLA_ONE, A, B, FLA_ONE, C );
break;
case FLA_ALG_BLOCKED:
FLA_Her2k_ln_blk_var8( FLA_ONE, A, B, FLA_ONE, C, cntl_her2k_var );
break;
default:
printf("trouble\n");
}
break;
}
case 9:{
// Time variant 9
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Her2k_ln_unb_var9( FLA_ONE, A, B, FLA_ONE, C );
break;
case FLA_ALG_BLOCKED:
FLA_Her2k_ln_blk_var9( FLA_ONE, A, B, FLA_ONE, C, cntl_her2k_var );
break;
default:
printf("trouble\n");
}
break;
}
case 10:{
// Time variant 10
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Her2k_ln_unb_var10( FLA_ONE, A, B, FLA_ONE, C );
break;
case FLA_ALG_BLOCKED:
FLA_Her2k_ln_blk_var10( FLA_ONE, A, B, FLA_ONE, C, cntl_her2k_var );
break;
default:
printf("trouble\n");
}
break;
}
}
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
FLA_Cntl_obj_free( cntl_her2k_var );
FLA_Cntl_obj_free( cntl_her2k_blas );
FLA_Cntl_obj_free( cntl_gemm_blas );
FLA_Blocksize_free( bp );
if ( variant == 0 )
{
FLA_Copy_external( C, C_ref );
*diff = 0.0;
}
else
{
*diff = FLA_Max_elemwise_diff( C, C_ref );
}
*gflops = 2.0 *
FLA_Obj_length( C ) *
FLA_Obj_width( C ) *
FLA_Obj_width( A ) /
dtime_old /
1e9;
*dtime = dtime_old;
FLA_Copy_external( C_old, C );
FLA_Obj_free( &C_old );
}
int main(int argc, char *argv[])
{
int
datatype,
precision,
m_input, k_input, n_input,
m, k, 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 k_dim_desc[14];
char n_dim_desc[14];
char m_dim_tag[10];
char k_dim_tag[10];
char n_dim_tag[10];
double max_gflops=6.0;
double
dtime,
gflops,
diff;
FLA_Obj
A, Ad, Az, B, Bd, Bz, C, Cd, Cz, C_ref, indexd, indexz;
FLA_Obj alpha0d, alpha0z, alpha1d, alpha1z, normd, normz;
FLA_Obj alphad, alphaz, betad, betaz, rhod, rhoz;
FLA_Obj xd, xz, yd, yz;
FLA_Init( );
fprintf( stdout, "%c number of repeats:", '%' );
scanf( "%d", &n_repeats );
fprintf( stdout, "%c %d\n", '%', n_repeats );
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 k n (-1 means bind to problem size): ", '%' );
scanf( "%d%d%d", &m_input, &k_input, &n_input );
fprintf( stdout, "%c %d %d %d\n", '%', m_input, k_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 ( k_input > 0 ) {
sprintf( k_dim_desc, "k = %d", k_input );
sprintf( k_dim_tag, "k%dc", k_input);
}
else if( k_input < -1 ) {
sprintf( k_dim_desc, "k = p/%d", -k_input );
sprintf( k_dim_tag, "k%dp", -k_input );
}
else if( k_input == -1 ) {
sprintf( k_dim_desc, "k = p" );
sprintf( k_dim_tag, "k%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;
for ( p = p_first, i = 1; p <= p_last; p += p_inc, i += 1 )
{
m = m_input;
k = k_input;
n = n_input;
if( m < 0 ) m = p / f2c_abs(m_input);
if( k < 0 ) k = p / f2c_abs(k_input);
if( n < 0 ) n = p / f2c_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][0] == 'c' ||
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;
}
// If transposing A, switch dimensions.
if ( pc_str[param_combo][0] == 'n' )
FLA_Obj_create( datatype, m, k, 0, 0, &A );
else
FLA_Obj_create( datatype, k, m, 0, 0, &A );
// If transposing B, switch dimensions.
if ( pc_str[param_combo][1] == 'n' )
FLA_Obj_create( datatype, k, n, 0, 0, &B );
else
FLA_Obj_create( datatype, n, k, 0, 0, &B );
FLA_Obj_create( datatype, m, n, 0, 0, &C );
FLA_Obj_create( datatype, m, n, 0, 0, &C_ref );
FLA_Random_matrix( A );
FLA_Random_matrix( B );
FLA_Random_matrix( C );
FLA_Copy_external( C, C_ref );
fprintf( stdout, "data_gemm_%s( %d, 1:5 ) = [ %4d %4d %4d ", pc_str[param_combo], i, m, k, n );
fflush( stdout );
time_Gemm( param_combo, FLA_ALG_REFERENCE, n_repeats, m, k, n,
A, B, C, C_ref, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
/*
time_Gemm( param_combo, FLA_ALG_FRONT, n_repeats, m, k, n,
A, B, C, C_ref, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
*/
fprintf( stdout, " ]; \n" );
fflush( stdout );
FLA_Obj_free( &A );
FLA_Obj_free( &B );
FLA_Obj_free( &C );
FLA_Obj_free( &C_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_gemm_%s( :,1 ), data_gemm_%s( :, 2 ), '%c:%c' ); \n",
pc_str[i], pc_str[i], colors[ i ], ticks[ i ] );
fprintf( stdout, "plot( data_gemm_%s( :,1 ), data_gemm_%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\\_gemm\\_%s', 'fla\\_gemm\\_%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 gemm front-end performance (%s, %s, %s)' );\n",
m_dim_desc, k_dim_desc, n_dim_desc );
fprintf( stdout, "print -depsc gemm_front_%s_%s_%s.eps\n", m_dim_tag, k_dim_tag, n_dim_tag );
fprintf( stdout, "hold off;\n");
fflush( stdout );
*/
FLA_Finalize( );
return 0;
}
void time_Apply_G_rf(
int variant, int type, int n_repeats, int m, int k, int n, int b_alg,
FLA_Obj A, FLA_Obj A_ref, FLA_Obj G, FLA_Obj P,
double *dtime, double *diff, double *gflops )
{
int irep;
double
dtime_old = 1.0e9;
FLA_Obj
A_save, G_save, norm;
if ( FLA_Obj_is_real( A ) )
{
if (
//( variant == 1 && type == FLA_ALG_UNB_OPT ) ||
//( variant == 1 && type == FLA_ALG_UNB_ASM ) ||
//( variant == 1 && type == FLA_ALG_BLOCKED ) ||
//( variant == 2 && type == FLA_ALG_UNB_OPT ) ||
//( variant == 2 && type == FLA_ALG_UNB_ASM ) ||
//( variant == 2 && type == FLA_ALG_BLOCKED ) ||
//( variant == 3 && type == FLA_ALG_UNB_OPT ) ||
//( variant == 3 && type == FLA_ALG_UNB_ASM ) ||
//( variant == 3 && type == FLA_ALG_BLOCKED ) ||
//( variant == 6 && type == FLA_ALG_UNB_OPT ) ||
//( variant == 6 && type == FLA_ALG_UNB_ASM ) ||
//( variant == 6 && type == FLA_ALG_BLOCKED ) ||
//( variant == 9 && type == FLA_ALG_UNB_OPT ) ||
//( variant == 9 && type == FLA_ALG_UNB_ASM ) ||
//( variant == 9 && type == FLA_ALG_BLOCKED ) ||
( variant == 4 ) ||
( variant == 5 ) ||
( variant == 7 ) ||
( variant == 8 ) ||
FALSE
)
{
*gflops = 0.0;
*diff = 0.0;
return;
}
}
else if ( FLA_Obj_is_complex( A ) )
{
if (
//( variant == 1 && type == FLA_ALG_UNB_OPT ) ||
//( variant == 1 && type == FLA_ALG_UNB_ASM ) ||
//( variant == 1 && type == FLA_ALG_BLOCKED ) ||
//( variant == 2 && type == FLA_ALG_UNB_OPT ) ||
//( variant == 2 && type == FLA_ALG_UNB_ASM ) ||
//( variant == 2 && type == FLA_ALG_BLOCKED ) ||
//( variant == 3 && type == FLA_ALG_UNB_OPT ) ||
//( variant == 3 && type == FLA_ALG_UNB_ASM ) ||
//( variant == 3 && type == FLA_ALG_BLOCKED ) ||
//( variant == 6 && type == FLA_ALG_UNB_OPT ) ||
//( variant == 6 && type == FLA_ALG_UNB_ASM ) ||
//( variant == 6 && type == FLA_ALG_BLOCKED ) ||
//( variant == 9 && type == FLA_ALG_UNB_OPT ) ||
//( variant == 9 && type == FLA_ALG_UNB_ASM ) ||
//( variant == 9 && type == FLA_ALG_BLOCKED ) ||
( variant == 4 ) ||
( variant == 5 ) ||
( variant == 7 ) ||
( variant == 8 ) ||
FALSE
)
{
*gflops = 0.0;
*diff = 0.0;
return;
}
}
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, A, &A_save );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, G, &G_save );
FLA_Obj_create( FLA_Obj_datatype_proj_to_real( A ), 1, 1, 0, 0, &norm );
//dim_t b_flash_m = b_alg;
//dim_t b_flash_n = n;
//FLASH_Obj_create_hier_copy_of_flat_ext( A, 1, &b_flash_m, &b_flash_n, &AH );
//printf ( "flash dims: %d x %d\n", FLA_Obj_length( AH ), FLA_Obj_width( AH ) );
FLA_Copy_external( A, A_save );
FLA_Copy_external( G, G_save );
for ( irep = 0 ; irep < n_repeats; irep++ ){
FLA_Copy_external( A_save, A );
FLA_Copy_external( G_save, G );
//FLASH_Obj_hierarchify( A_save, AH );
*dtime = FLA_Clock();
switch( variant ){
case 0:
break;
// Time variant 1
case 1:
{
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Apply_G_rf_opt_var1( G, A );
break;
case FLA_ALG_UNB_ASM:
FLA_Apply_G_rf_asm_var1( G, A );
break;
case FLA_ALG_BLOCKED:
FLA_Apply_G_rf_blk_var1( G, A, b_alg );
break;
}
break;
}
// Time variant 2
case 2:
{
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Apply_G_rf_opt_var2( G, A );
break;
case FLA_ALG_UNB_ASM:
FLA_Apply_G_rf_asm_var2( G, A );
break;
case FLA_ALG_BLOCKED:
FLA_Apply_G_rf_blk_var2( G, A, b_alg );
break;
}
break;
}
// Time variant 3
case 3:
{
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Apply_G_rf_opt_var3( G, A );
break;
case FLA_ALG_UNB_ASM:
FLA_Apply_G_rf_asm_var3( G, A );
break;
case FLA_ALG_BLOCKED:
FLA_Apply_G_rf_blk_var3( G, A, b_alg );
break;
}
break;
}
// Time variant 6
case 6:
{
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Apply_G_rf_opt_var6( G, A );
break;
case FLA_ALG_UNB_ASM:
FLA_Apply_G_rf_asm_var6( G, A );
break;
case FLA_ALG_BLOCKED:
FLA_Apply_G_rf_blk_var6( G, A, b_alg );
break;
}
break;
}
// Time variant 9
case 9:
{
switch( type ){
case FLA_ALG_UNB_OPT:
FLA_Apply_G_rf_opt_var9( G, A );
break;
case FLA_ALG_UNB_ASM:
FLA_Apply_G_rf_asm_var9( G, A );
break;
case FLA_ALG_BLOCKED:
FLA_Apply_G_rf_blk_var9( G, A, b_alg );
break;
}
break;
}
}
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
if ( variant == 1 && type == FLA_ALG_UNB_OPT )
{
//FLA_Obj_show( "A_ref", A, "%9.2e + %9.2e ", "" );
//FLA_Obj_show( "A", A, "%9.2e ", "" );
FLA_Copy( A, A_ref );
*diff = 0.0;
}
else
{
//FLA_Obj_show( "A", A, "%9.2e + %9.2e ", "" );
//if ( variant == 7 && type == FLA_ALG_UNB_ASM )
//FLA_Obj_show( "A", A, "%9.2e", "" );
//if ( variant == 9 ) FLASH_Obj_flatten( AH, A );
FLA_Axpy( FLA_MINUS_ONE, A_ref, A );
FLA_Norm_frob( A, norm );
FLA_Obj_extract_real_scalar( norm, diff );
//*diff = FLA_Max_elemwise_diff( A_ref, A );
}
*gflops = 6.0 * k * m * ( n - 1 ) /
dtime_old / 1e9;
if ( FLA_Obj_is_complex( A ) )
*gflops *= 2.0;
*dtime = dtime_old;
FLA_Copy_external( A_save, A );
FLA_Copy_external( G_save, G );
//FLASH_Obj_free( &AH );
FLA_Obj_free( &A_save );
FLA_Obj_free( &G_save );
FLA_Obj_free( &norm );
}
void time_Syrk_ln(
int variant, int type, int nrepeats, int n, int nb_alg,
FLA_Obj A, FLA_Obj B, FLA_Obj C, FLA_Obj Cref,
double *dtime, double *diff, double *gflops )
{
int
irep,
info, lwork;
double
dtime_old,
d_minus_one = -1.0, d_one = 1.0;
FLA_Obj
Cold;
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &Cold );
FLA_Copy_external( C, Cold );
for ( irep = 0 ; irep < nrepeats; irep++ ){
FLA_Copy_external( Cold, C );
*dtime = FLA_Clock();
switch( variant ){
case 0:
// Time reference implementation
REF_Syrk_ln( FLA_ONE, A, FLA_ONE, C );
break;
default:
printf("trouble\n");
break;
}
if ( irep == 0 )
dtime_old = FLA_Clock() - *dtime;
else{
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
}
if ( variant == 0 ){
FLA_Copy_external( C, Cref );
*diff = 0.0;
}
else{
*diff = FLA_Max_elemwise_diff( C, Cref );
}
*gflops = 1.0 *
FLA_Obj_length( A ) *
FLA_Obj_length( A ) *
FLA_Obj_width( A ) /
dtime_old /
1e9;
*dtime = dtime_old;
FLA_Copy_external( Cold, C );
FLA_Obj_free( &Cold );
}
void time_Gemm_pp_nn(
int variant, int type, int nrepeats, int n, int nb_alg,
FLA_Obj A, FLA_Obj B, FLA_Obj C, FLA_Obj Cref,
double *dtime, double *diff, double *mflops )
{
int
irep,
info, lwork;
double
dtime_old,
d_minus_one = -1.0, d_one = 1.0;
FLA_Obj
Cold;
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &Cold );
FLA_Copy_external( C, Cold );
for ( irep = 0 ; irep < nrepeats; irep++ ){
FLA_Copy_external( Cold, C );
*dtime = FLA_Clock();
switch( variant ){
case 0:
// Time reference implementation
REF_Gemm( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
ONE, A, B, FLA_ONE, C );
break;
case 1:{
// Time variant 1
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Gemm_pp_nn_var1( FLA_ONE, A, B, C, nb_alg );
break;
case FLA_ALG_BLOCKED:
REF_Gemm( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
ONE, A, B, FLA_ONE, C );
break;
default:
printf("trouble\n");
}
break;
}
}
if ( irep == 0 )
dtime_old = FLA_Clock() - *dtime;
else{
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
}
if ( variant == 0 ){
FLA_Copy_external( C, Cref );
*diff = 0.0;
}
else{
*diff = FLA_Max_elemwise_diff( C, Cref );
}
*mflops = 2.0 *
FLA_Obj_length( C ) *
FLA_Obj_width( C ) *
FLA_Obj_width( A ) /
dtime_old /
1000000;
*dtime = dtime_old;
FLA_Copy_external( Cold, C );
FLA_Obj_free( &Cold );
}
void time_Copyt(
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;
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &C_old );
FLA_Copy_external( C, C_old );
for ( irep = 0 ; irep < nrepeats; irep++ ){
FLA_Copy_external( C_old, C );
*dtime = FLA_Clock();
switch( param_combo ){
// Time parameter combination 0
case 0:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Copyt( FLA_NO_TRANSPOSE, A, C );
break;
case FLA_ALG_FRONT:
FLA_Copyt( FLA_NO_TRANSPOSE, A, C );
break;
default:
printf("trouble\n");
}
break;
}
case 1:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Copyt( FLA_TRANSPOSE, A, C );
break;
case FLA_ALG_FRONT:
FLA_Copyt( FLA_TRANSPOSE, A, C );
break;
default:
printf("trouble\n");
}
break;
}
case 2:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Copyt( FLA_CONJ_NO_TRANSPOSE, A, C );
break;
case FLA_ALG_FRONT:
FLA_Copyt( FLA_CONJ_NO_TRANSPOSE, A, C );
break;
default:
printf("trouble\n");
}
break;
}
case 3:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Copyt( FLA_CONJ_TRANSPOSE, A, C );
break;
case FLA_ALG_FRONT:
FLA_Copyt( FLA_CONJ_TRANSPOSE, A, C );
break;
default:
printf("trouble\n");
}
break;
}
}
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
if ( type == FLA_ALG_REFERENCE )
{
FLA_Copy_external( C, C_ref );
*diff = 0.0;
}
else
{
*diff = FLA_Max_elemwise_diff( C, C_ref );
}
*gflops = 2.0 * m * n /
dtime_old /
1.0e9;
if ( FLA_Obj_is_complex( A ) )
*gflops *= 4.0;
*dtime = dtime_old;
FLA_Copy_external( C_old, C );
FLA_Obj_free( &C_old );
}
FLA_Error REF_Svdd_uv_components( FLA_Obj A, FLA_Obj s, FLA_Obj U, FLA_Obj V,
double* dtime_bred, double* dtime_bsvd, double* dtime_appq,
double* dtime_qrfa, double* dtime_gemm )
/*
{
*dtime_bred = 1;
*dtime_bsvd = 1;
*dtime_appq = 1;
*dtime_qrfa = 1;
*dtime_gemm = 1;
return FLA_Svdd_external( FLA_SVD_VECTORS_ALL, A, s, U, V );
}
*/
{
FLA_Datatype dt_A;
FLA_Datatype dt_A_real;
dim_t m_A, n_A;
dim_t min_m_n;
FLA_Obj tq, tu, tv, d, e, Ur, Vr, W;
FLA_Obj eT, epsilonB;
FLA_Uplo uplo = FLA_UPPER_TRIANGULAR;
double crossover_ratio = 16.0 / 10.0;
double dtime_temp;
dt_A = FLA_Obj_datatype( A );
dt_A_real = FLA_Obj_datatype_proj_to_real( A );
m_A = FLA_Obj_length( A );
n_A = FLA_Obj_width( A );
min_m_n = FLA_Obj_min_dim( A );
FLA_Obj_create( dt_A, min_m_n, 1, 0, 0, &tq );
FLA_Obj_create( dt_A, min_m_n, 1, 0, 0, &tu );
FLA_Obj_create( dt_A, min_m_n, 1, 0, 0, &tv );
FLA_Obj_create( dt_A_real, min_m_n, 1, 0, 0, &d );
FLA_Obj_create( dt_A_real, min_m_n, 1, 0, 0, &e );
FLA_Obj_create( dt_A_real, n_A, n_A, 0, 0, &Ur );
FLA_Obj_create( dt_A_real, n_A, n_A, 0, 0, &Vr );
FLA_Part_2x1( e, &eT,
&epsilonB, 1, FLA_BOTTOM );
if ( m_A >= n_A )
{
if ( m_A < crossover_ratio * n_A )
{
dtime_temp = FLA_Clock();
{
// Reduce to bidiagonal form.
FLA_Bidiag_blk_external( A, tu, tv );
FLA_Bidiag_UT_extract_diagonals( A, d, eT );
}
*dtime_bred = FLA_Clock() - dtime_temp;
dtime_temp = FLA_Clock();
{
// Divide-and-conquor algorithm.
FLA_Bsvdd_external( uplo, d, e, Ur, Vr );
}
*dtime_bsvd = FLA_Clock() - dtime_temp;
dtime_temp = FLA_Clock();
{
// Form U.
FLA_Copy_external( Ur, U );
FLA_Bidiag_apply_U_external( FLA_LEFT, FLA_NO_TRANSPOSE, A, tu, U );
// Form V.
FLA_Copy_external( Vr, V );
FLA_Bidiag_apply_V_external( FLA_RIGHT, FLA_CONJ_TRANSPOSE, A, tv, V );
}
*dtime_appq = FLA_Clock() - dtime_temp;
*dtime_qrfa = 0.0;
*dtime_gemm = 0.0;
}
else
{
FLA_Obj AT,
AB;
FLA_Obj UL, UR;
FLA_Part_2x1( A, &AT,
&AB, n_A, FLA_TOP );
FLA_Part_1x2( U, &UL, &UR, n_A, FLA_LEFT );
// Create a temporary n-by-n matrix R.
FLA_Obj_create( dt_A, n_A, n_A, 0, 0, &W );
dtime_temp = FLA_Clock();
{
// Perform a QR factorization.
FLA_QR_blk_external( A, tq );
FLA_Copyr_external( FLA_LOWER_TRIANGULAR, A, UL );
FLA_Setr( FLA_LOWER_TRIANGULAR, FLA_ZERO, A );
}
*dtime_qrfa = FLA_Clock() - dtime_temp;
dtime_temp = FLA_Clock();
{
// Form Q.
FLA_QR_form_Q_external( U, tq );
}
*dtime_appq = FLA_Clock() - dtime_temp;
dtime_temp = FLA_Clock();
{
// Reduce R to bidiagonal form.
FLA_Bidiag_blk_external( AT, tu, tv );
FLA_Bidiag_UT_extract_diagonals( A, d, eT );
}
*dtime_bred = FLA_Clock() - dtime_temp;
dtime_temp = FLA_Clock();
{
// Divide-and-conquor algorithm.
FLA_Bsvdd_external( uplo, d, e, Ur, Vr );
}
*dtime_bsvd = FLA_Clock() - dtime_temp;
dtime_temp = FLA_Clock();
{
// Form U in W.
FLA_Copy_external( Ur, W );
FLA_Bidiag_apply_U_external( FLA_LEFT, FLA_NO_TRANSPOSE, AT, tu, W );
// Form V.
FLA_Copy_external( Vr, V );
FLA_Bidiag_apply_V_external( FLA_RIGHT, FLA_CONJ_TRANSPOSE, AT, tv, V );
}
*dtime_appq += FLA_Clock() - dtime_temp;
dtime_temp = FLA_Clock();
{
// Multiply R into U, storing the result in A and then copying
// back to U.
FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_ONE, UL, W, FLA_ZERO, A );
FLA_Copy( A, UL );
}
*dtime_gemm = FLA_Clock() - dtime_temp;
// Free R.
FLA_Obj_free( &W );
}
}
else
{
FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
}
// Copy singular values to output vector.
FLA_Copy( d, s );
// Sort singular values and vectors.
FLA_Sort_svd( FLA_BACKWARD, s, U, V );
FLA_Obj_free( &tq );
FLA_Obj_free( &tu );
FLA_Obj_free( &tv );
FLA_Obj_free( &d );
FLA_Obj_free( &e );
FLA_Obj_free( &Ur );
FLA_Obj_free( &Vr );
return FLA_SUCCESS;
}
int main(int argc, char *argv[])
{
int
m_input,
m,
p_first, p_last, p_inc,
p,
n_repeats,
param_combo,
i,
n_param_combos = N_PARAM_COMBOS;
FLA_Datatype datatype;
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, b, b_orig, norm;
FLA_Init();
fprintf( stdout, "%c number of repeats: ", '%' );
scanf( "%d", &n_repeats );
fprintf( stdout, "%c %d\n", '%', n_repeats );
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;
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++ ){
FLA_Obj_create( datatype, m, m, 0, 0, &A );
FLA_Obj_create( datatype, m, 1, 0, 0, &b );
FLA_Obj_create( datatype, m, 1, 0, 0, &b_orig );
if ( FLA_Obj_is_single_precision( A ) )
FLA_Obj_create( FLA_FLOAT, 1, 1, 0, 0, &norm );
else
FLA_Obj_create( FLA_DOUBLE, 1, 1, 0, 0, &norm );
if ( pc_str[param_combo][0] == 'l' )
FLA_Random_spd_matrix( FLA_LOWER_TRIANGULAR, A );
else
FLA_Random_spd_matrix( FLA_UPPER_TRIANGULAR, A );
FLA_Copy_external( b, b_orig );
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, b, b_orig, norm, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
time_Chol( param_combo, FLA_ALG_FRONT, n_repeats, m,
A, b, b_orig, norm, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
fprintf( stdout, " ]; \n" );
fflush( stdout );
FLA_Obj_free( &A );
FLA_Obj_free( &b );
FLA_Obj_free( &b_orig );
FLA_Obj_free( &norm );
}
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,
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;
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, b, b_orig, 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 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;
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, 0, 0, &C );
FLA_Obj_create( datatype, m, 1, 0, 0, &b );
FLA_Obj_create( datatype, m, 1, 0, 0, &b_orig );
if ( FLA_Obj_is_single_precision( C ) )
FLA_Obj_create( FLA_FLOAT, 1, 1, 0, 0, &b_norm );
else
FLA_Obj_create( FLA_DOUBLE, 1, 1, 0, 0, &b_norm );
FLA_Random_matrix( C );
FLA_Random_matrix( b );
FLA_Copy_external( b, b_orig );
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,
// C, b, b_orig, 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,
C, b, b_orig, 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( &b );
FLA_Obj_free( &b_orig );
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,
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, B, C, C_ref;
FLA_Init( );
fprintf( stdout, "%c number of repeats:", '%' );
scanf( "%d", &n_repeats );
fprintf( stdout, "%c %d\n", '%', n_repeats );
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;
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++ ){
// If multiplying A 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, m, m, 0, 0, &A );
else
FLA_Obj_create( datatype, n, n, 0, 0, &A );
FLA_Obj_create( datatype, m, n, 0, 0, &B );
FLA_Obj_create( datatype, m, n, 0, 0, &C );
FLA_Obj_create( datatype, m, n, 0, 0, &C_ref );
FLA_Random_matrix( A );
FLA_Random_matrix( B );
FLA_Random_matrix( C );
FLA_Copy_external( C, C_ref );
fprintf( stdout, "data_symm_%s( %d, 1:5 ) = [ %d ", pc_str[param_combo], i, p );
fflush( stdout );
time_Symm( param_combo, FLA_ALG_REFERENCE, n_repeats, m, n,
A, B, C, C_ref, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
time_Symm( param_combo, FLA_ALG_FRONT, n_repeats, m, n,
A, B, C, C_ref, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
fprintf( stdout, " ]; \n" );
fflush( stdout );
FLA_Obj_free( &A );
FLA_Obj_free( &B );
FLA_Obj_free( &C );
FLA_Obj_free( &C_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_symm_%s( :,1 ), data_symm_%s( :, 2 ), '%c:%c' ); \n",
pc_str[i], pc_str[i], colors[ i ], ticks[ i ] );
fprintf( stdout, "plot( data_symm_%s( :,1 ), data_symm_%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\\_symm\\_%s', 'fla\\_symm\\_%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 symm front-end performance (%s, %s)' );\n",
m_dim_desc, n_dim_desc );
fprintf( stdout, "print -depsc symm_front_%s_%s.eps\n", m_dim_tag, n_dim_tag );
fprintf( stdout, "hold off;\n");
fflush( stdout );
*/
FLA_Finalize( );
return 0;
}
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
m_input, k_input, n_input,
m, n, k,
p_first, p_last, p_inc,
p,
nb_alg,
n_repeats,
variant,
n_threads,
n_thread_experiments,
i, j;
int n_threads_exp[64];
char *colors = "brkgmckkk";
char *ticks = "o+*xso+*x";
char m_dim_desc[14];
char k_dim_desc[14];
char n_dim_desc[14];
char m_dim_tag[5];
char k_dim_tag[5];
char n_dim_tag[5];
char nth_str[32];
double max_gflops=6.0;
double
dtime,
gflops,
diff,
d_n;
FLA_Obj
A, B, C, C_ref;
/* Initialize FLAME */
FLA_Init( );
fprintf( stdout, "%c number of repeats:", '%' );
scanf( "%d", &n_repeats );
fprintf( stdout, "%c %d\n", '%', n_repeats );
fprintf( stdout, "%c Enter blocking size:", '%' );
scanf( "%d", &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 k n (-1 means bind to problem size: ", '%' );
scanf( "%d%d%d", &m_input, &k_input, &n_input );
fprintf( stdout, "%c %d %d %d\n", '%', m_input, k_input, n_input );
fprintf( stdout, "%c enter number of thread experiments: ", '%' );
scanf( "%d", &n_thread_experiments );
fprintf( stdout, "%c %d\n", '%', n_thread_experiments );
fprintf( stdout, "%c enter number of threads for each experiment (separated by spaces): ", '%' );
for( i = 0; i < n_thread_experiments; ++i )
scanf( "%d", &n_threads_exp[i] );
fprintf( stdout, "%c", '%' );
for( i = 0; i < n_thread_experiments; ++i )
fprintf( stdout, " %d", n_threads_exp[i] );
/* Delete all existing data structures */
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 ( k_input > 0 ) {
sprintf( k_dim_desc, "k = %d", k_input );
sprintf( k_dim_tag, "k%dc", k_input);
}
else if( k_input < -1 ) {
sprintf( k_dim_desc, "k = p/%d", -k_input );
sprintf( k_dim_tag, "k%dp", -k_input );
}
else if( k_input == -1 ) {
sprintf( k_dim_desc, "k = p" );
sprintf( k_dim_tag, "k%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 );
}
m = p_last;
k = p_last;
n = p_last;
sprintf( nth_str, "OMP_NUM_THREADS=%d", n_threads_exp[ n_thread_experiments-1 ] );
putenv( nth_str );
blas_cpu_number = n_threads_exp[ n_thread_experiments-1 ];
blas_thread_init();
for ( p = p_first, i = 1; p <= p_last; p += p_inc, i += 1 )
{
m = m_input;
k = k_input;
n = n_input;
if( m < 0 ) m = p / abs(m_input);
if( k < 0 ) k = p / abs(k_input);
if( n < 0 ) n = p / abs(n_input);
FLA_Obj_create( FLA_DOUBLE, m, k, &A );
FLA_Obj_create( FLA_DOUBLE, k, n, &B );
FLA_Obj_create( FLA_DOUBLE, m, n, &C );
FLA_Obj_create( FLA_DOUBLE, m, n, &C_ref );
/* Generate random matrices A, C */
if( p > 4000 ){
FLA_Random_matrix( A );
FLA_Random_matrix( B );
FLA_Random_matrix( C );
FLA_Copy_external( C, C_ref );
}
blas_cpu_number = 1;
//time_Gemm_nn( 0, FLA_ALG_REFERENCE, n_repeats, p, nb_alg,
// A, B, C, C_ref, &dtime, &diff, &gflops );
//fprintf( stdout, "data_REF( %d, 1:2 ) = [ %d %6.3lf ]; \n", i, p, gflops );
//fflush( stdout );
for ( j = 0; j < n_thread_experiments; j++ ){
n_threads = n_threads_exp[j];
blas_cpu_number = n_threads;
fprintf( stdout, "data_nth%d( %d, 1:3 ) = [ %d ", n_threads, i, p );
fflush( stdout );
time_Gemm_nn( 0, FLA_ALG_REFERENCE, n_repeats, p, nb_alg,
A, B, C, C_ref, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
fprintf( stdout, " ]; \n" );
fflush( stdout );
}
fprintf( stdout, "\n" );
FLA_Obj_free( &A );
FLA_Obj_free( &B );
FLA_Obj_free( &C );
FLA_Obj_free( &C_ref );
}
/* Print the MATLAB commands to plot the data */
/* Delete all existing figures */
fprintf( stdout, "figure;\n" );
/* Indicate that you want to add to the existing plot */
fprintf( stdout, "hold on;\n" );
/* Plot the data for the other numbers of threads */
for ( i = 0; i < n_thread_experiments; i++ ){
fprintf( stdout, "plot( data_nth%d( :,1 ), data_nth%d( :, 2 ), '%c:%c' ); \n",
n_threads_exp[ i ], n_threads_exp[ i ], colors[ i ], ticks[ i ] );
}
fprintf( stdout, "legend( ... \n" );
for ( i = 0; i < n_thread_experiments-1; i++ )
fprintf( stdout, "'%d threads', ... \n", n_threads_exp[ i ] );
fprintf( stdout, "'%d threads', 'Location', 'Best' ); \n", n_threads_exp[ n_thread_experiments-1 ] );
fprintf( stdout, "xlabel( 'problem size p' );\n" );
fprintf( stdout, "ylabel( 'GFLOPS/sec.' );\n" );
fprintf( stdout, "axis( [ 0 %d 0 %.2f ] ); \n", p_last, n_threads_exp[n_thread_experiments-1] * max_gflops );
fprintf( stdout, "title( 'Goto BLAS dgemm performance (%s, %s, %s)' );\n",
m_dim_desc, k_dim_desc, n_dim_desc );
fprintf( stdout, "print -depsc gemm_nn_goto_p_%s_%s_%s.eps\n", m_dim_tag, k_dim_tag, n_dim_tag );
fprintf( stdout, "hold off;\n");
fflush( stdout );
FLA_Finalize( );
}
FLA_Error FLA_Eig_gest_il_unb_var5( FLA_Obj A, FLA_Obj Y, FLA_Obj B )
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Obj BTL, BTR, B00, b01, B02,
BBL, BBR, b10t, beta11, b12t,
B20, b21, B22;
//FLA_Obj yT, y01,
// yB, psi11,
// y21;
//FLA_Obj y21_l, y21_r;
FLA_Obj psi11, y12t,
y21, Y22;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
FLA_Part_2x2( B, &BTL, &BTR,
&BBL, &BBR, 0, 0, FLA_TL );
//FLA_Part_2x1( Y, &yT,
// &yB, 0, FLA_TOP );
FLA_Part_2x2( Y, &psi11, &y12t,
&y21, &Y22, 1, 1, FLA_TL );
while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
/* ************* */ /* ************************** */
&a10t, /**/ &alpha11, &a12t,
ABL, /**/ ABR, &A20, /**/ &a21, &A22,
1, 1, FLA_BR );
FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &b01, &B02,
/* ************* */ /* ************************* */
&b10t, /**/ &beta11, &b12t,
BBL, /**/ BBR, &B20, /**/ &b21, &B22,
1, 1, FLA_BR );
//FLA_Repart_2x1_to_3x1( yT, &y01,
// /* ** */ /* ***** */
// &psi11,
// yB, &y21, 1, FLA_BOTTOM );
/*------------------------------------------------------------*/
//FLA_Part_1x2( y21, &y21_l, &y21_r, 1, FLA_LEFT );
// alpha11 = inv(beta11) * alpha11 * inv(conj(beta11));
// = inv(beta11) * alpha11 * inv(beta11);
FLA_Inv_scal_external( beta11, alpha11 );
FLA_Inv_scal_external( beta11, alpha11 );
//// y21 = b21 * alpha11;
//FLA_Copy_external( b21, y21_l );
//FLA_Scal_external( alpha11, y21_l );
// psi11 = - 1/2 * alpha11;
FLA_Copy_external( alpha11, psi11 );
FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 );
// a21 = a21 * inv(conj(beta11));
// = a21 * inv(beta11);
FLA_Inv_scal_external( beta11, a21 );
//// a21 = a21 - 1/2 * y21;
//FLA_Axpy_external( FLA_MINUS_ONE_HALF, y21_l, a21 );
// a21 = a21 - 1/2 * alpha11 * b21;
FLA_Axpy_external( psi11, b21, a21 );
// A22 = A22 - a21 * b21' - b21 * a21';
FLA_Her2c_external( FLA_LOWER_TRIANGULAR, FLA_NO_CONJUGATE,
FLA_MINUS_ONE, a21, b21, A22 );
//// a21 = a21 - 1/2 * y21;
//FLA_Axpy_external( FLA_MINUS_ONE_HALF, y21_l, a21 );
// a21 = a21 - 1/2 * alpha11 * b21;
FLA_Axpy_external( psi11, b21, a21 );
// a21 = inv( tril( B22 ) ) * a21;
FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
B22, a21 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
a10t, alpha11, /**/ a12t,
/* ************** */ /* ************************ */
&ABL, /**/ &ABR, A20, a21, /**/ A22,
FLA_TL );
FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, b01, /**/ B02,
b10t, beta11, /**/ b12t,
/* ************** */ /* *********************** */
&BBL, /**/ &BBR, B20, b21, /**/ B22,
FLA_TL );
//FLA_Cont_with_3x1_to_2x1( &yT, y01,
// psi11,
// /* ** */ /* ***** */
// &yB, y21, FLA_TOP );
}
return FLA_SUCCESS;
}
void time_Gemm_hh(
int variant, int type, int nrepeats, int n, int nb_alg,
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;
fla_blocksize_t*
bp;
fla_gemm_t*
cntl_gemm_blas;
fla_gemm_t*
cntl_gemm_var;
bp = FLA_Blocksize_create( nb_alg, nb_alg, nb_alg, nb_alg );
cntl_gemm_blas = FLA_Cntl_gemm_obj_create( FLA_FLAT, FLA_SUBPROBLEM, NULL, NULL );
cntl_gemm_var = FLA_Cntl_gemm_obj_create( FLA_FLAT, variant, bp, cntl_gemm_blas );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &C_old );
FLA_Copy_external( C, C_old );
for ( irep = 0 ; irep < nrepeats; irep++ ){
FLA_Copy_external( C_old, C );
*dtime = FLA_Clock();
switch( variant ){
// Time reference implementation
case 0:
REF_Gemm( FLA_CONJ_TRANSPOSE, FLA_CONJ_TRANSPOSE,
FLA_ONE, A, B, FLA_ONE, C );
break;
// Time variant 1
case 1:{
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Gemm_hh_unb_var1( FLA_ONE, A, B, FLA_ONE, C );
break;
case FLA_ALG_BLOCKED:
FLA_Gemm_hh_blk_var1( FLA_ONE, A, B, FLA_ONE, C, cntl_gemm_var );
break;
default:
printf("trouble\n");
}
break;
}
// Time variant 2
case 2:{
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Gemm_hh_unb_var2( FLA_ONE, A, B, FLA_ONE, C );
break;
case FLA_ALG_BLOCKED:
FLA_Gemm_hh_blk_var2( FLA_ONE, A, B, FLA_ONE, C, cntl_gemm_var );
break;
default:
printf("trouble\n");
}
break;
}
// Time variant 3
case 3:{
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Gemm_hh_unb_var3( FLA_ONE, A, B, FLA_ONE, C );
break;
case FLA_ALG_BLOCKED:
FLA_Gemm_hh_blk_var3( FLA_ONE, A, B, FLA_ONE, C, cntl_gemm_var );
break;
default:
printf("trouble\n");
}
break;
}
// Time variant 4
case 4:{
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Gemm_hh_unb_var4( FLA_ONE, A, B, FLA_ONE, C );
break;
case FLA_ALG_BLOCKED:
FLA_Gemm_hh_blk_var4( FLA_ONE, A, B, FLA_ONE, C, cntl_gemm_var );
break;
default:
printf("trouble\n");
}
break;
}
// Time variant 5
case 5:{
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Gemm_hh_unb_var5( FLA_ONE, A, B, FLA_ONE, C );
break;
case FLA_ALG_BLOCKED:
FLA_Gemm_hh_blk_var5( FLA_ONE, A, B, FLA_ONE, C, cntl_gemm_var );
break;
default:
printf("trouble\n");
}
break;
}
// Time variant 6
case 6:{
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Gemm_hh_unb_var6( FLA_ONE, A, B, FLA_ONE, C );
break;
case FLA_ALG_BLOCKED:
FLA_Gemm_hh_blk_var6( FLA_ONE, A, B, FLA_ONE, C, cntl_gemm_var );
break;
default:
printf("trouble\n");
}
break;
}
}
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
FLA_Cntl_obj_free( cntl_gemm_var );
FLA_Cntl_obj_free( cntl_gemm_blas );
FLA_Blocksize_free( bp );
if ( variant == 0 )
{
FLA_Copy_external( C, C_ref );
*diff = 0.0;
}
else
{
*diff = FLA_Max_elemwise_diff( C, C_ref );
}
*gflops = 2.0 *
FLA_Obj_length( C ) *
FLA_Obj_width( C ) *
FLA_Obj_width( A ) /
dtime_old /
1.0e9;
if ( FLA_Obj_is_complex( C ) )
*gflops *= 4.0;
*dtime = dtime_old;
FLA_Copy_external( C_old, C );
FLA_Obj_free( &C_old );
}
FLA_Error FLA_Hess_UT_blk_var4( FLA_Obj A, FLA_Obj T )
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj UT, U0,
UB, U1,
U2;
FLA_Obj YT, Y0,
YB, Y1,
Y2;
FLA_Obj ZT, Z0,
ZB, Z1,
Z2;
FLA_Obj TL, TR, T0, T1, T2;
FLA_Obj U, Y, Z;
FLA_Obj ABR_l;
FLA_Obj UB_l, U2_l;
FLA_Obj YB_l, Y2_l;
FLA_Obj ZB_l, Z2_l;
FLA_Obj WT_l;
FLA_Obj T1_tl;
FLA_Obj none, none2, none3;
FLA_Obj UB_tl,
UB_bl;
FLA_Datatype datatype_A;
dim_t m_A;
dim_t b_alg, b, bb;
b_alg = FLA_Obj_length( T );
datatype_A = FLA_Obj_datatype( A );
m_A = FLA_Obj_length( A );
FLA_Obj_create( datatype_A, m_A, b_alg, 0, 0, &U );
FLA_Obj_create( datatype_A, m_A, b_alg, 0, 0, &Y );
FLA_Obj_create( datatype_A, m_A, b_alg, 0, 0, &Z );
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
FLA_Part_2x1( U, &UT,
&UB, 0, FLA_TOP );
FLA_Part_2x1( Y, &YT,
&YB, 0, FLA_TOP );
FLA_Part_2x1( Z, &ZT,
&ZB, 0, FLA_TOP );
FLA_Part_1x2( T, &TL, &TR, 0, FLA_LEFT );
while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) )
{
b = min( FLA_Obj_length( ABR ), b_alg );
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
/* ************* */ /* ******************** */
&A10, /**/ &A11, &A12,
ABL, /**/ ABR, &A20, /**/ &A21, &A22,
b, b, FLA_BR );
FLA_Repart_2x1_to_3x1( UT, &U0,
/* ** */ /* ** */
&U1,
UB, &U2, b, FLA_BOTTOM );
FLA_Repart_2x1_to_3x1( YT, &Y0,
/* ** */ /* ** */
&Y1,
YB, &Y2, b, FLA_BOTTOM );
FLA_Repart_2x1_to_3x1( ZT, &Z0,
/* ** */ /* ** */
&Z1,
ZB, &Z2, b, FLA_BOTTOM );
FLA_Repart_1x2_to_1x3( TL, /**/ TR, &T0, /**/ &T1, &T2,
b, FLA_RIGHT );
/*------------------------------------------------------------*/
FLA_Part_2x2( T1, &T1_tl, &none,
&none2, &none3, b, b, FLA_TL );
bb = min( FLA_Obj_length( ABR ) - 1, b_alg );
FLA_Part_1x2( ABR, &ABR_l, &none, bb, FLA_LEFT );
FLA_Part_1x2( UB, &UB_l, &none, bb, FLA_LEFT );
FLA_Part_1x2( YB, &YB_l, &none, bb, FLA_LEFT );
FLA_Part_1x2( ZB, &ZB_l, &none, bb, FLA_LEFT );
FLA_Part_2x1( UB_l, &none,
&U2_l, b, FLA_TOP );
FLA_Part_2x1( YB_l, &none,
&Y2_l, b, FLA_TOP );
FLA_Part_2x1( ZB_l, &none,
&Z2_l, b, FLA_TOP );
// [ ABR, YB, ZB, T1 ] = FLA_Hess_UT_step_unb_var4( ABR, YB, ZB, T1, b );
//FLA_Hess_UT_step_unb_var4( ABR, YB, ZB, T1_tl );
//FLA_Hess_UT_step_ofu_var4( ABR, YB, ZB, T1_tl );
FLA_Hess_UT_step_opt_var4( ABR, YB, ZB, T1_tl );
// Build UB from ABR, with explicit unit subdiagonal and zeros.
FLA_Copy_external( ABR_l, UB_l );
FLA_Part_2x1( UB_l, &UB_tl,
&UB_bl, 1, FLA_TOP );
FLA_Triangularize( FLA_LOWER_TRIANGULAR, FLA_UNIT_DIAG, UB_bl );
FLA_Set( FLA_ZERO, UB_tl );
// ATR = ATR - ATR * UB * inv( triu( T ) ) * UB' );
if ( FLA_Obj_length( ATR ) > 0 )
{
// NOTE: We use ZT as temporary workspace.
FLA_Part_1x2( ZT, &WT_l, &none, bb, FLA_LEFT );
FLA_Part_2x2( T1, &T1_tl, &none,
&none2, &none3, bb, bb, FLA_TL );
// WT_l = ATR * UB_l * inv( triu( T ) ).
FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_ONE, ATR, UB_l, FLA_ZERO, WT_l );
FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_ONE, T1_tl, WT_l );
// ATR = ATR - WT_l * UB_l'
FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
FLA_MINUS_ONE, WT_l, UB_l, FLA_ONE, ATR );
}
// A22 = A22 - U2 * Y2' - Z2 * U2';
FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
FLA_MINUS_ONE, U2_l, Y2_l, FLA_ONE, A22 );
FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
FLA_MINUS_ONE, Z2_l, U2_l, FLA_ONE, A22 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
A10, A11, /**/ A12,
/* ************** */ /* ****************** */
&ABL, /**/ &ABR, A20, A21, /**/ A22,
FLA_TL );
FLA_Cont_with_3x1_to_2x1( &UT, U0,
U1,
/* ** */ /* ** */
&UB, U2, FLA_TOP );
FLA_Cont_with_3x1_to_2x1( &YT, Y0,
Y1,
/* ** */ /* ** */
&YB, Y2, FLA_TOP );
FLA_Cont_with_3x1_to_2x1( &ZT, Z0,
Z1,
/* ** */ /* ** */
&ZB, Z2, FLA_TOP );
FLA_Cont_with_1x3_to_1x2( &TL, /**/ &TR, T0, T1, /**/ T2,
FLA_LEFT );
}
FLA_Obj_free( &U );
FLA_Obj_free( &Y );
FLA_Obj_free( &Z );
return FLA_SUCCESS;
}
FLA_Error FLA_Eig_gest_nl_unb_var4( FLA_Obj A, FLA_Obj Y, FLA_Obj B )
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Obj BTL, BTR, B00, b01, B02,
BBL, BBR, b10t, beta11, b12t,
B20, b21, B22;
//FLA_Obj yL, yR, y10t, psi11, y12t;
//FLA_Obj y10t_t,
// y10t_b;
FLA_Obj psi11, y12t,
y21, Y22;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
FLA_Part_2x2( B, &BTL, &BTR,
&BBL, &BBR, 0, 0, FLA_TL );
//FLA_Part_1x2( Y, &yL, &yR, 0, FLA_LEFT );
FLA_Part_2x2( Y, &psi11, &y12t,
&y21, &Y22, 1, 1, FLA_TL );
while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
/* ************* */ /* ************************** */
&a10t, /**/ &alpha11, &a12t,
ABL, /**/ ABR, &A20, /**/ &a21, &A22,
1, 1, FLA_BR );
FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &b01, &B02,
/* ************* */ /* ************************* */
&b10t, /**/ &beta11, &b12t,
BBL, /**/ BBR, &B20, /**/ &b21, &B22,
1, 1, FLA_BR );
//FLA_Repart_1x2_to_1x3( yL, /**/ yR, &y10t, /**/ &psi11, &y12t,
// 1, FLA_RIGHT );
/*------------------------------------------------------------*/
//FLA_Part_2x1( y10t, &y10t_t,
// &y10t_b, 1, FLA_TOP );
//// y10t = alpha11 * b10t;
//FLA_Copy_external( b10t, y10t_t );
//FLA_Scal_external( alpha11, y10t_t );
// psi11 = 1/2 * alpha11;
FLA_Copy_external( alpha11, psi11 );
FLA_Scal_external( FLA_ONE_HALF, psi11 );
//// a10t = a10t + 1/2 * y10t;
//FLA_Axpy_external( FLA_ONE_HALF, y10t_t, a10t );
// a10t = a10t + 1/2 * alpha11 * b10t;
FLA_Axpy_external( psi11, b10t, a10t );
// A00 = A00 + a10t' * b10t + b10t' * a10t;
FLA_Her2c_external( FLA_LOWER_TRIANGULAR, FLA_CONJUGATE,
FLA_ONE, a10t, b10t, A00 );
//// a10t = a10t + 1/2 * y10t;
//FLA_Axpy_external( FLA_ONE_HALF, y10t_t, a10t );
// a10t = a10t + 1/2 * alpha11 * b10t;
FLA_Axpy_external( psi11, b10t, a10t );
// a10t = conj(beta11) * a10t;
// = beta11 * a10t;
FLA_Scal_external( beta11, a10t );
// alpha11 = conj(beta11) * alpha11 * beta11;
// = beta11 * alpha11 * beta11;
FLA_Scal_external( beta11, alpha11 );
FLA_Scal_external( beta11, alpha11 );
// A20 = A20 + a21 * b10t;
FLA_Ger_external( FLA_ONE, a21, b10t, A20 );
// a21 = a21 * beta11;
FLA_Scal_external( beta11, a21 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
a10t, alpha11, /**/ a12t,
/* ************** */ /* ************************ */
&ABL, /**/ &ABR, A20, a21, /**/ A22,
FLA_TL );
FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, b01, /**/ B02,
b10t, beta11, /**/ b12t,
/* ************** */ /* *********************** */
&BBL, /**/ &BBR, B20, b21, /**/ B22,
FLA_TL );
//FLA_Cont_with_1x3_to_1x2( &yL, /**/ &yR, y10t, psi11, /**/ y12t,
// FLA_LEFT );
}
return FLA_SUCCESS;
}
int main(int argc, char *argv[])
{
int
m_input, n_input,
m, n,
p_first, p_last, p_inc,
p,
nb_alg,
n_repeats,
variant,
i, j,
datatype,
n_variants = N_VARIANTS;
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;
/* Initialize FLAME */
FLA_Init( );
fprintf( stdout, "%c number of repeats: ", '%' );
scanf( "%d", &n_repeats );
fprintf( stdout, "%c %d\n", '%', n_repeats );
fprintf( stdout, "%c Enter blocking size: ", '%' );
scanf( "%d", &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 );
/* Delete all existing data structures */
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 );
}
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);
//datatype = FLA_COMPLEX;
datatype = FLA_DOUBLE_COMPLEX;
/* Allocate space for the matrices */
FLA_Obj_create( datatype, m, m, &A );
FLA_Obj_create( datatype, m, n, &C );
FLA_Obj_create( datatype, m, n, &C_ref );
/* Generate random matrices A, C */
FLA_Random_tri_matrix( FLA_LOWER_TRIANGULAR, FLA_UNIT_DIAG, A );
FLA_Random_matrix( C );
FLA_Copy_external( C, C_ref );
/* Time the reference implementation */
time_Trmm_luh( 0, FLA_ALG_REFERENCE, n_repeats, p, nb_alg,
A, B, C, C_ref, &dtime, &diff, &gflops );
fprintf( stdout, "data_REF( %d, 1:2 ) = [ %d %6.3lf ]; \n", i, p, gflops );
fflush( stdout );
for ( variant = 1; variant <= n_variants; variant++ ){
//fprintf( stdout, "data_var%d( %d, 1:7 ) = [ %d ", variant, i, p );
fprintf( stdout, "data_var%d( %d, 1:5 ) = [ %d ", variant, i, p );
fflush( stdout );
time_Trmm_luh( variant, FLA_ALG_UNBLOCKED, n_repeats, p, nb_alg,
A, B, C, C_ref, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
time_Trmm_luh( variant, FLA_ALG_BLOCKED, n_repeats, p, nb_alg,
A, B, C, C_ref, &dtime, &diff, &gflops );
fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
fflush( stdout );
//time_Trmm_luh( variant, FLA_ALG_OPTIMIZED, n_repeats, p, nb_alg,
// A, B, C, C_ref, &dtime, &diff, &gflops );
//fprintf( stdout, "%6.3lf %6.2le ", gflops, diff );
//fflush( stdout );
fprintf( stdout, " ]; \n" );
fflush( stdout );
}
fprintf( stdout, "\n" );
FLA_Obj_free( &A );
FLA_Obj_free( &C );
FLA_Obj_free( &C_ref );
}
/* Print the MATLAB commands to plot the data */
/* Delete all existing figures */
fprintf( stdout, "figure;\n" );
/* Plot the performance of the reference implementation */
fprintf( stdout, "plot( data_REF( :,1 ), data_REF( :, 2 ), '-' ); \n" );
/* Indicate that you want to add to the existing plot */
fprintf( stdout, "hold on;\n" );
/* Plot the data for the other numbers of threads */
for ( i = 1; i <= n_variants; i++ ) {
fprintf( stdout, "plot( data_var%d( :,1 ), data_var%d( :, 2 ), '%c:%c' ); \n",
i, i, colors[ i-1 ], ticks[ i-1 ] );
fprintf( stdout, "plot( data_var%d( :,1 ), data_var%d( :, 4 ), '%c-.%c' ); \n",
i, i, colors[ i-1 ], ticks[ i-1 ] );
//fprintf( stdout, "plot( data_var%d( :,1 ), data_var%d( :, 6 ), '%c--%c' ); \n",
// i, i, colors[ i-1 ], ticks[ i-1 ] );
}
fprintf( stdout, "legend( ... \n" );
fprintf( stdout, "'Reference', ... \n" );
for ( i = 1; i < n_variants; i++ )
//fprintf( stdout, "'unb\\_var%d', 'blk\\_var%d', 'opt\\_var%d', ... \n", i, i, i );
fprintf( stdout, "'unb\\_var%d', 'blk\\_var%d', ... \n", i, i );
i = n_variants;
fprintf( stdout, "'unb\\_var%d', 'blk\\_var%d' ); \n", i, i );
fprintf( stdout, "xlabel( 'problem size p' );\n" );
fprintf( stdout, "ylabel( 'GFLOPS/sec.' );\n" );
fprintf( stdout, "axis( [ 0 %d 0 %.2f ] ); \n", p_last, max_gflops );
fprintf( stdout, "title( 'FLAME trmm\\_luc performance (%s, %s)' );\n",
m_dim_desc, n_dim_desc );
fprintf( stdout, "print -depsc trmm_luc_%s_%s.eps\n", m_dim_tag, n_dim_tag );
fprintf( stdout, "hold off;\n");
fflush( stdout );
FLA_Finalize( );
}
void time_Her2k(
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;
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &C_old );
FLA_Copy_external( C, C_old );
for ( irep = 0 ; irep < nrepeats; irep++ ){
FLA_Copy_external( C_old, C );
*dtime = FLA_Clock();
switch( param_combo ){
// Time parameter combination 0
case 0:{
switch( type ){
case FLA_ALG_REFERENCE:
REF_Her2k( FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_ONE, A, B, FLA_ZERO, C );
break;
case FLA_ALG_FRONT:
FLA_Her2k( FLA_LOWER_TRIANGULAR, 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_Her2k( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_ONE, A, B, FLA_ZERO, C );
break;
case FLA_ALG_FRONT:
FLA_Her2k( FLA_LOWER_TRIANGULAR, 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_Her2k( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_ONE, A, B, FLA_ZERO, C );
break;
case FLA_ALG_FRONT:
FLA_Her2k( FLA_UPPER_TRIANGULAR, FLA_CONJ_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_Her2k( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_ONE, A, B, FLA_ZERO, C );
break;
case FLA_ALG_FRONT:
FLA_Her2k( FLA_UPPER_TRIANGULAR, FLA_NO_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 )
{
FLA_Copy_external( C, C_ref );
*diff = 0.0;
}
else
{
*diff = FLA_Max_elemwise_diff( C, C_ref );
}
*gflops = 4.0 * 2.0 * m * m * k /
dtime_old /
1.0e9;
*dtime = dtime_old;
FLA_Copy_external( C_old, C );
FLA_Obj_free( &C_old );
}
