void libblis_test_syr2k_experiment( test_params_t* params,
test_op_t* op,
iface_t iface,
num_t datatype,
char* pc_str,
char* sc_str,
unsigned int p_cur,
double* perf,
double* resid )
{
unsigned int n_repeats = params->n_repeats;
unsigned int i;
double time_min = 1e9;
double time;
dim_t m, k;
uplo_t uploc;
trans_t transa, transb;
obj_t kappa;
obj_t alpha, a, b, beta, c;
obj_t c_save;
// Map the dimension specifier to actual dimensions.
m = libblis_test_get_dim_from_prob_size( op->dim_spec[0], p_cur );
k = libblis_test_get_dim_from_prob_size( op->dim_spec[1], p_cur );
// Map parameter characters to BLIS constants.
bli_param_map_char_to_blis_uplo( pc_str[0], &uploc );
bli_param_map_char_to_blis_trans( pc_str[1], &transa );
bli_param_map_char_to_blis_trans( pc_str[2], &transb );
// Create test scalars.
bli_obj_scalar_init_detached( datatype, &kappa );
bli_obj_scalar_init_detached( datatype, &alpha );
bli_obj_scalar_init_detached( datatype, &beta );
// Create test operands (vectors and/or matrices).
libblis_test_mobj_create( params, datatype, transa,
sc_str[0], m, k, &a );
libblis_test_mobj_create( params, datatype, transb,
sc_str[1], m, k, &b );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[2], m, m, &c );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[2], m, m, &c_save );
// Set alpha and beta.
if ( bli_obj_is_real( c ) )
{
bli_setsc( 0.8, 0.0, &alpha );
bli_setsc( -1.0, 0.0, &beta );
}
else
{
// For syr2k, both alpha and beta may be complex since, unlike her2k,
// C is symmetric in both the real and complex cases.
bli_setsc( 0.8, 0.5, &alpha );
bli_setsc( -1.0, 0.5, &beta );
}
// Randomize A and B.
bli_randm( &a );
bli_randm( &b );
// Set the structure and uplo properties of C.
bli_obj_set_struc( BLIS_SYMMETRIC, c );
bli_obj_set_uplo( uploc, c );
// Randomize A, make it densely symmetric, and zero the unstored triangle
// to ensure the implementation is reads only from the stored region.
bli_randm( &c );
bli_mksymm( &c );
bli_mktrim( &c );
// Save C and set its structure and uplo properties.
bli_obj_set_struc( BLIS_SYMMETRIC, c_save );
bli_obj_set_uplo( uploc, c_save );
bli_copym( &c, &c_save );
// Normalize by k.
bli_setsc( 1.0/( double )k, 0.0, &kappa );
bli_scalm( &kappa, &a );
bli_scalm( &kappa, &b );
// Apply the remaining parameters.
bli_obj_set_conjtrans( transa, a );
bli_obj_set_conjtrans( transb, b );
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
bli_copym( &c_save, &c );
time = bli_clock();
libblis_test_syr2k_impl( iface, &alpha, &a, &b, &beta, &c );
time_min = bli_clock_min_diff( time_min, time );
}
// Estimate the performance of the best experiment repeat.
*perf = ( 2.0 * m * m * k ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( c ) ) *perf *= 4.0;
// Perform checks.
libblis_test_syr2k_check( &alpha, &a, &b, &beta, &c, &c_save, resid );
// Zero out performance and residual if output matrix is empty.
libblis_test_check_empty_problem( &c, perf, resid );
// Free the test objects.
bli_obj_free( &a );
bli_obj_free( &b );
bli_obj_free( &c );
bli_obj_free( &c_save );
}
int main( int argc, char** argv )
{
obj_t a, b, c;
obj_t c_save;
obj_t alpha, beta;
dim_t m, n;
dim_t p;
dim_t p_begin, p_end, p_inc;
int m_input, n_input;
num_t dt;
int r, n_repeats;
side_t side;
uplo_t uploa;
f77_char f77_side;
f77_char f77_uploa;
double dtime;
double dtime_save;
double gflops;
bli_init();
//bli_error_checking_level_set( BLIS_NO_ERROR_CHECKING );
n_repeats = 3;
#ifndef PRINT
p_begin = 200;
p_end = 2000;
p_inc = 200;
m_input = -1;
n_input = -1;
#else
p_begin = 16;
p_end = 16;
p_inc = 1;
m_input = 4;
n_input = 4;
#endif
#if 1
//dt = BLIS_FLOAT;
dt = BLIS_DOUBLE;
#else
//dt = BLIS_SCOMPLEX;
dt = BLIS_DCOMPLEX;
#endif
side = BLIS_LEFT;
//side = BLIS_RIGHT;
uploa = BLIS_LOWER;
//uploa = BLIS_UPPER;
bli_param_map_blis_to_netlib_side( side, &f77_side );
bli_param_map_blis_to_netlib_uplo( uploa, &f77_uploa );
for ( p = p_begin; p <= p_end; p += p_inc )
{
if ( m_input < 0 ) m = p * ( dim_t )abs(m_input);
else m = ( dim_t ) m_input;
if ( n_input < 0 ) n = p * ( dim_t )abs(n_input);
else n = ( dim_t ) n_input;
bli_obj_create( dt, 1, 1, 0, 0, &alpha );
bli_obj_create( dt, 1, 1, 0, 0, &beta );
if ( bli_is_left( side ) )
bli_obj_create( dt, m, m, 0, 0, &a );
else
bli_obj_create( dt, n, n, 0, 0, &a );
bli_obj_create( dt, m, n, 0, 0, &b );
bli_obj_create( dt, m, n, 0, 0, &c );
bli_obj_create( dt, m, n, 0, 0, &c_save );
bli_randm( &a );
bli_randm( &b );
bli_randm( &c );
bli_obj_set_struc( BLIS_HERMITIAN, a );
bli_obj_set_uplo( uploa, a );
// Randomize A, make it densely Hermitian, and zero the unstored
// triangle to ensure the implementation reads only from the stored
// region.
bli_randm( &a );
bli_mkherm( &a );
bli_mktrim( &a );
/*
bli_obj_toggle_uplo( a );
bli_obj_inc_diag_off( 1, a );
bli_setm( &BLIS_ZERO, &a );
bli_obj_inc_diag_off( -1, a );
bli_obj_toggle_uplo( a );
bli_obj_set_diag( BLIS_NONUNIT_DIAG, a );
bli_scalm( &BLIS_TWO, &a );
bli_scalm( &BLIS_TWO, &a );
*/
bli_setsc( (2.0/1.0), 1.0, &alpha );
bli_setsc( -(1.0/1.0), 0.0, &beta );
bli_copym( &c, &c_save );
dtime_save = 1.0e9;
for ( r = 0; r < n_repeats; ++r )
{
bli_copym( &c_save, &c );
dtime = bli_clock();
#ifdef PRINT
bli_printm( "a", &a, "%4.1f", "" );
bli_printm( "b", &b, "%4.1f", "" );
bli_printm( "c", &c, "%4.1f", "" );
#endif
#ifdef BLIS
bli_hemm( side,
&alpha,
&a,
&b,
&beta,
&c );
#else
if ( bli_is_float( dt ) )
{
f77_int mm = bli_obj_length( c );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
float* alphap = bli_obj_buffer( alpha );
float* ap = bli_obj_buffer( a );
float* bp = bli_obj_buffer( b );
float* betap = bli_obj_buffer( beta );
float* cp = bli_obj_buffer( c );
ssymm_( &f77_side,
&f77_uploa,
&mm,
&nn,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
else if ( bli_is_double( dt ) )
{
f77_int mm = bli_obj_length( c );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
double* alphap = bli_obj_buffer( alpha );
double* ap = bli_obj_buffer( a );
double* bp = bli_obj_buffer( b );
double* betap = bli_obj_buffer( beta );
double* cp = bli_obj_buffer( c );
dsymm_( &f77_side,
&f77_uploa,
&mm,
&nn,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
else if ( bli_is_scomplex( dt ) )
{
f77_int mm = bli_obj_length( c );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
scomplex* alphap = bli_obj_buffer( alpha );
scomplex* ap = bli_obj_buffer( a );
scomplex* bp = bli_obj_buffer( b );
scomplex* betap = bli_obj_buffer( beta );
scomplex* cp = bli_obj_buffer( c );
chemm_( &f77_side,
&f77_uploa,
&mm,
&nn,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
else if ( bli_is_dcomplex( dt ) )
{
f77_int mm = bli_obj_length( c );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
dcomplex* alphap = bli_obj_buffer( alpha );
dcomplex* ap = bli_obj_buffer( a );
dcomplex* bp = bli_obj_buffer( b );
dcomplex* betap = bli_obj_buffer( beta );
dcomplex* cp = bli_obj_buffer( c );
zhemm_( &f77_side,
&f77_uploa,
&mm,
&nn,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
#endif
#ifdef PRINT
bli_printm( "c after", &c, "%9.5f", "" );
exit(1);
#endif
dtime_save = bli_clock_min_diff( dtime_save, dtime );
}
if ( bli_is_left( side ) )
gflops = ( 2.0 * m * m * n ) / ( dtime_save * 1.0e9 );
else
gflops = ( 2.0 * m * n * n ) / ( dtime_save * 1.0e9 );
if ( bli_is_complex( dt ) ) gflops *= 4.0;
#ifdef BLIS
printf( "data_hemm_blis" );
#else
printf( "data_hemm_%s", BLAS );
#endif
printf( "( %2lu, 1:4 ) = [ %4lu %4lu %10.3e %6.3f ];\n",
( unsigned long )(p - p_begin + 1)/p_inc + 1,
( unsigned long )m,
( unsigned long )n, dtime_save, gflops );
bli_obj_free( &alpha );
bli_obj_free( &beta );
bli_obj_free( &a );
bli_obj_free( &b );
bli_obj_free( &c );
bli_obj_free( &c_save );
}
bli_finalize();
return 0;
}
int main( int argc, char** argv )
{
obj_t a, x;
obj_t a_save;
obj_t alpha;
dim_t m;
dim_t p;
dim_t p_begin, p_end, p_inc;
int m_input;
num_t dt_a, dt_x;
num_t dt_alpha;
int r, n_repeats;
uplo_t uplo;
double dtime;
double dtime_save;
double gflops;
//bli_init();
n_repeats = 3;
#ifndef PRINT
p_begin = 40;
p_end = 2000;
p_inc = 40;
m_input = -1;
#else
p_begin = 16;
p_end = 16;
p_inc = 1;
m_input = 6;
#endif
#if 1
dt_alpha = dt_x = dt_a = BLIS_DOUBLE;
#else
dt_alpha = dt_x = dt_a = BLIS_DCOMPLEX;
#endif
uplo = BLIS_LOWER;
// Begin with initializing the last entry to zero so that
// matlab allocates space for the entire array once up-front.
for ( p = p_begin; p + p_inc <= p_end; p += p_inc ) ;
#ifdef BLIS
printf( "data_her_blis" );
#else
printf( "data_her_%s", BLAS );
#endif
printf( "( %2lu, 1:2 ) = [ %4lu %7.2f ];\n",
( unsigned long )(p - p_begin + 1)/p_inc + 1,
( unsigned long )0, 0.0 );
for ( p = p_begin; p <= p_end; p += p_inc )
{
if ( m_input < 0 ) m = p * ( dim_t )abs(m_input);
else m = ( dim_t ) m_input;
bli_obj_create( dt_alpha, 1, 1, 0, 0, &alpha );
bli_obj_create( dt_x, m, 1, 0, 0, &x );
bli_obj_create( dt_a, m, m, 0, 0, &a );
bli_obj_create( dt_a, m, m, 0, 0, &a_save );
bli_randm( &x );
bli_randm( &a );
bli_obj_set_struc( BLIS_HERMITIAN, &a );
//bli_obj_set_struc( BLIS_SYMMETRIC, &a );
bli_obj_set_uplo( uplo, &a );
bli_setsc( (2.0/1.0), 0.0, &alpha );
bli_copym( &a, &a_save );
dtime_save = DBL_MAX;
for ( r = 0; r < n_repeats; ++r )
{
bli_copym( &a_save, &a );
dtime = bli_clock();
#ifdef PRINT
bli_printm( "x", &x, "%4.1f", "" );
bli_printm( "a", &a, "%4.1f", "" );
#endif
#ifdef BLIS
//bli_obj_toggle_conj( &x );
//bli_syr( &alpha,
bli_her( &alpha,
&x,
&a );
#else
f77_char uplo = 'L';
f77_int mm = bli_obj_length( &a );
f77_int incx = bli_obj_vector_inc( &x );
f77_int lda = bli_obj_col_stride( &a );
double* alphap = bli_obj_buffer( &alpha );
double* xp = bli_obj_buffer( &x );
double* ap = bli_obj_buffer( &a );
/*
dcomplex* xp = bli_obj_buffer( x );
dcomplex* ap = bli_obj_buffer( &a );
*/
dsyr_( &uplo,
//zher_( &uplo,
&mm,
alphap,
xp, &incx,
ap, &lda );
#endif
#ifdef PRINT
bli_printm( "a after", &a, "%4.1f", "" );
exit(1);
#endif
dtime_save = bli_clock_min_diff( dtime_save, dtime );
}
gflops = ( 1.0 * m * m ) / ( dtime_save * 1.0e9 );
#ifdef BLIS
printf( "data_her_blis" );
#else
printf( "data_her_%s", BLAS );
#endif
printf( "( %2lu, 1:2 ) = [ %4lu %7.2f ];\n",
( unsigned long )(p - p_begin + 1)/p_inc + 1,
( unsigned long )m, gflops );
bli_obj_free( &alpha );
bli_obj_free( &x );
bli_obj_free( &a );
bli_obj_free( &a_save );
}
//bli_finalize();
return 0;
}
int main( int argc, char** argv )
{
obj_t a, b, c;
obj_t c_save;
obj_t alpha, beta;
dim_t m, n;
dim_t p;
dim_t p_begin, p_end, p_inc;
int m_input, n_input;
num_t dt_a, dt_b, dt_c;
num_t dt_alpha, dt_beta;
int r, n_repeats;
side_t side;
uplo_t uplo;
double dtime;
double dtime_save;
double gflops;
bli_init();
n_repeats = 3;
if( argc < 7 )
{
printf("Usage:\n");
printf("test_foo.x m n p_begin p_inc p_end:\n");
exit;
}
int world_size, world_rank, provided;
MPI_Init_thread( NULL, NULL, MPI_THREAD_FUNNELED, &provided );
MPI_Comm_size( MPI_COMM_WORLD, &world_size );
MPI_Comm_rank( MPI_COMM_WORLD, &world_rank );
m_input = strtol( argv[1], NULL, 10 );
n_input = strtol( argv[2], NULL, 10 );
p_begin = strtol( argv[4], NULL, 10 );
p_inc = strtol( argv[5], NULL, 10 );
p_end = strtol( argv[6], NULL, 10 );
#if 1
dt_a = BLIS_DOUBLE;
dt_b = BLIS_DOUBLE;
dt_c = BLIS_DOUBLE;
dt_alpha = BLIS_DOUBLE;
dt_beta = BLIS_DOUBLE;
#else
dt_a = dt_b = dt_c = dt_alpha = dt_beta = BLIS_DCOMPLEX;
#endif
side = BLIS_LEFT;
//side = BLIS_RIGHT;
uplo = BLIS_LOWER;
//uplo = BLIS_UPPER;
for ( p = p_begin + world_rank * p_inc; p <= p_end; p += p_inc * world_size )
{
if ( m_input < 0 ) m = p * ( dim_t )abs(m_input);
else m = ( dim_t ) m_input;
if ( n_input < 0 ) n = p * ( dim_t )abs(n_input);
else n = ( dim_t ) n_input;
bli_obj_create( dt_alpha, 1, 1, 0, 0, &alpha );
bli_obj_create( dt_beta, 1, 1, 0, 0, &beta );
if ( bli_is_left( side ) )
bli_obj_create( dt_a, m, m, 0, 0, &a );
else
bli_obj_create( dt_a, n, n, 0, 0, &a );
bli_obj_create( dt_b, m, n, 0, 0, &b );
bli_obj_create( dt_c, m, n, 0, 0, &c );
bli_obj_create( dt_c, m, n, 0, 0, &c_save );
bli_obj_set_struc( BLIS_TRIANGULAR, a );
bli_obj_set_uplo( uplo, a );
bli_randm( &a );
bli_randm( &c );
bli_randm( &b );
/*
bli_obj_toggle_uplo( a );
bli_obj_inc_diag_off( -1, a );
bli_setm( &BLIS_ZERO, &a );
bli_obj_inc_diag_off( 1, a );
bli_obj_toggle_uplo( a );
bli_obj_set_diag( BLIS_NONUNIT_DIAG, a );
bli_scalm( &BLIS_TWO, &a );
//bli_scalm( &BLIS_TWO, &a );
*/
bli_setsc( (2.0/1.0), 0.0, &alpha );
bli_setsc( (1.0/1.0), 0.0, &beta );
bli_copym( &c, &c_save );
dtime_save = 1.0e9;
for ( r = 0; r < n_repeats; ++r )
{
bli_copym( &c_save, &c );
dtime = bli_clock();
#ifdef PRINT
/*
obj_t ar, ai;
bli_obj_alias_to( a, ar );
bli_obj_alias_to( a, ai );
bli_obj_set_datatype( BLIS_DOUBLE, ar ); ar.rs *= 2; ar.cs *= 2;
bli_obj_set_datatype( BLIS_DOUBLE, ai ); ai.rs *= 2; ai.cs *= 2; ai.buffer = ( double* )ai.buffer + 1;
bli_printm( "ar", &ar, "%4.1f", "" );
bli_printm( "ai", &ai, "%4.1f", "" );
*/
bli_printm( "a", &a, "%4.1f", "" );
bli_printm( "c", &c, "%4.1f", "" );
#endif
#ifdef BLIS
bli_error_checking_level_set( BLIS_NO_ERROR_CHECKING );
bli_trmm( side,
//bli_trmm4m( side,
&alpha,
&a,
&c );
#else
f77_char side = 'L';
f77_char uplo = 'L';
f77_char transa = 'N';
f77_char diag = 'N';
f77_int mm = bli_obj_length( c );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldc = bli_obj_col_stride( c );
double* alphap = bli_obj_buffer( alpha );
double* ap = bli_obj_buffer( a );
double* cp = bli_obj_buffer( c );
dtrmm_( &side,
&uplo,
&transa,
&diag,
&mm,
&nn,
alphap,
ap, &lda,
cp, &ldc );
#endif
#ifdef PRINT
bli_printm( "c after", &c, "%4.1f", "" );
exit(1);
#endif
dtime_save = bli_clock_min_diff( dtime_save, dtime );
}
if ( bli_is_left( side ) )
gflops = ( 1.0 * m * m * n ) / ( dtime_save * 1.0e9 );
else
gflops = ( 1.0 * m * n * n ) / ( dtime_save * 1.0e9 );
if ( bli_is_complex( dt_a ) ) gflops *= 4.0;
#ifdef BLIS
printf( "data_trmm_blis" );
#else
printf( "data_trmm_%s", BLAS );
#endif
printf( "( %2lu, 1:4 ) = [ %4lu %4lu %10.3e %6.3f ];\n",
( unsigned long )(p - p_begin + 1)/p_inc + 1,
( unsigned long )m,
( unsigned long )n, dtime_save, gflops );
bli_obj_free( &alpha );
bli_obj_free( &beta );
bli_obj_free( &a );
bli_obj_free( &b );
bli_obj_free( &c );
bli_obj_free( &c_save );
}
bli_finalize();
return 0;
}
int main( int argc, char** argv )
{
obj_t a, b, c;
obj_t c_save;
obj_t alpha, beta;
dim_t m, n, k;
dim_t p;
dim_t p_begin, p_end, p_inc;
int m_input, n_input, k_input;
num_t dt, dt_real;
char dt_ch;
int r, n_repeats;
trans_t transa;
trans_t transb;
f77_char f77_transa;
f77_char f77_transb;
double dtime;
double dtime_save;
double gflops;
extern blksz_t* gemm_kc;
bli_init();
//bli_error_checking_level_set( BLIS_NO_ERROR_CHECKING );
n_repeats = 3;
dt = DT;
dt_real = bli_datatype_proj_to_real( DT );
p_begin = P_BEGIN;
p_end = P_END;
p_inc = P_INC;
m_input = -1;
n_input = -1;
k_input = -1;
// Extract the kc blocksize for the requested datatype and its
// real analogue.
dim_t kc = bli_blksz_get_def( dt, gemm_kc );
dim_t kc_real = bli_blksz_get_def( dt_real, gemm_kc );
// Assign the k dimension depending on which implementation is
// being tested. Note that the BLIS_NAT case handles the real
// domain cases as well as native complex.
if ( IND == BLIS_NAT ) k_input = kc;
else if ( IND == BLIS_3M1 ) k_input = kc_real / 3;
else if ( IND == BLIS_4M1A ) k_input = kc_real / 2;
else k_input = kc_real;
// Adjust the relative dimensions, if requested.
#if (defined ADJ_MK)
m_input = -2; k_input = -2; n_input = -1;
#elif (defined ADJ_KN)
k_input = -2; n_input = -2; m_input = -1;
#elif (defined ADJ_MN)
m_input = -2; n_input = -2; k_input = -1;
#endif
// Choose the char corresponding to the requested datatype.
if ( bli_is_float( dt ) ) dt_ch = 's';
else if ( bli_is_double( dt ) ) dt_ch = 'd';
else if ( bli_is_scomplex( dt ) ) dt_ch = 'c';
else dt_ch = 'z';
transa = BLIS_NO_TRANSPOSE;
transb = BLIS_NO_TRANSPOSE;
bli_param_map_blis_to_netlib_trans( transa, &f77_transa );
bli_param_map_blis_to_netlib_trans( transb, &f77_transb );
// Begin with initializing the last entry to zero so that
// matlab allocates space for the entire array once up-front.
for ( p = p_begin; p + p_inc <= p_end; p += p_inc ) ;
#ifdef BLIS
printf( "data_%s_%cgemm_%s_blis", THR_STR, dt_ch, STR );
#else
printf( "data_%s_%cgemm_%s", THR_STR, dt_ch, STR );
#endif
printf( "( %2lu, 1:5 ) = [ %4lu %4lu %4lu %10.3e %6.3f ];\n",
( unsigned long )(p - p_begin + 1)/p_inc + 1,
( unsigned long )0,
( unsigned long )0,
( unsigned long )0, 0.0, 0.0 );
for ( p = p_begin; p <= p_end; p += p_inc )
{
if ( m_input < 0 ) m = p / ( dim_t )abs(m_input);
else m = ( dim_t ) m_input;
if ( n_input < 0 ) n = p / ( dim_t )abs(n_input);
else n = ( dim_t ) n_input;
if ( k_input < 0 ) k = p / ( dim_t )abs(k_input);
else k = ( dim_t ) k_input;
bli_obj_create( dt, 1, 1, 0, 0, &alpha );
bli_obj_create( dt, 1, 1, 0, 0, &beta );
bli_obj_create( dt, m, k, 0, 0, &a );
bli_obj_create( dt, k, n, 0, 0, &b );
bli_obj_create( dt, m, n, 0, 0, &c );
//bli_obj_create( dt, m, k, 2, 2*m, &a );
//bli_obj_create( dt, k, n, 2, 2*k, &b );
//bli_obj_create( dt, m, n, 2, 2*m, &c );
bli_obj_create( dt, m, n, 0, 0, &c_save );
bli_randm( &a );
bli_randm( &b );
bli_randm( &c );
bli_obj_set_conjtrans( transa, a );
bli_obj_set_conjtrans( transb, b );
bli_setsc( (2.0/1.0), 0.0, &alpha );
bli_setsc( -(1.0/1.0), 0.0, &beta );
bli_copym( &c, &c_save );
#ifdef BLIS
bli_ind_disable_all_dt( dt );
bli_ind_enable_dt( IND, dt );
#endif
dtime_save = 1.0e9;
for ( r = 0; r < n_repeats; ++r )
{
bli_copym( &c_save, &c );
dtime = bli_clock();
#ifdef PRINT
bli_printm( "a", &a, "%4.1f", "" );
bli_printm( "b", &b, "%4.1f", "" );
bli_printm( "c", &c, "%4.1f", "" );
#endif
#ifdef BLIS
bli_gemm( &alpha,
&a,
&b,
&beta,
&c );
#else
if ( bli_is_float( dt ) )
{
f77_int mm = bli_obj_length( c );
f77_int kk = bli_obj_width_after_trans( a );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
float* alphap = bli_obj_buffer( alpha );
float* ap = bli_obj_buffer( a );
float* bp = bli_obj_buffer( b );
float* betap = bli_obj_buffer( beta );
float* cp = bli_obj_buffer( c );
sgemm_( &f77_transa,
&f77_transb,
&mm,
&nn,
&kk,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
else if ( bli_is_double( dt ) )
{
f77_int mm = bli_obj_length( c );
f77_int kk = bli_obj_width_after_trans( a );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
double* alphap = bli_obj_buffer( alpha );
double* ap = bli_obj_buffer( a );
double* bp = bli_obj_buffer( b );
double* betap = bli_obj_buffer( beta );
double* cp = bli_obj_buffer( c );
dgemm_( &f77_transa,
&f77_transb,
&mm,
&nn,
&kk,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
else if ( bli_is_scomplex( dt ) )
{
f77_int mm = bli_obj_length( c );
f77_int kk = bli_obj_width_after_trans( a );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
scomplex* alphap = bli_obj_buffer( alpha );
scomplex* ap = bli_obj_buffer( a );
scomplex* bp = bli_obj_buffer( b );
scomplex* betap = bli_obj_buffer( beta );
scomplex* cp = bli_obj_buffer( c );
cgemm_( &f77_transa,
&f77_transb,
&mm,
&nn,
&kk,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
else if ( bli_is_dcomplex( dt ) )
{
f77_int mm = bli_obj_length( c );
f77_int kk = bli_obj_width_after_trans( a );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
dcomplex* alphap = bli_obj_buffer( alpha );
dcomplex* ap = bli_obj_buffer( a );
dcomplex* bp = bli_obj_buffer( b );
dcomplex* betap = bli_obj_buffer( beta );
dcomplex* cp = bli_obj_buffer( c );
zgemm_( &f77_transa,
//zgemm3m_( &f77_transa,
&f77_transb,
&mm,
&nn,
&kk,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
#endif
#ifdef PRINT
bli_printm( "c after", &c, "%4.1f", "" );
exit(1);
#endif
dtime_save = bli_clock_min_diff( dtime_save, dtime );
}
gflops = ( 2.0 * m * k * n ) / ( dtime_save * 1.0e9 );
if ( bli_is_complex( dt ) ) gflops *= 4.0;
#ifdef BLIS
printf( "data_%s_%cgemm_%s_blis", THR_STR, dt_ch, STR );
#else
printf( "data_%s_%cgemm_%s", THR_STR, dt_ch, STR );
#endif
printf( "( %2lu, 1:5 ) = [ %4lu %4lu %4lu %10.3e %6.3f ];\n",
( unsigned long )(p - p_begin + 1)/p_inc + 1,
( unsigned long )m,
( unsigned long )k,
( unsigned long )n, dtime_save, gflops );
bli_obj_free( &alpha );
bli_obj_free( &beta );
bli_obj_free( &a );
bli_obj_free( &b );
bli_obj_free( &c );
bli_obj_free( &c_save );
}
bli_finalize();
return 0;
}
void libblis_test_scalm_experiment( test_params_t* params,
test_op_t* op,
iface_t iface,
num_t datatype,
char* pc_str,
char* sc_str,
unsigned int p_cur,
double* perf,
double* resid )
{
unsigned int n_repeats = params->n_repeats;
unsigned int i;
double time_min = 1e9;
double time;
dim_t m, n;
conj_t conjbeta;
obj_t beta, y;
obj_t y_save;
// Map the dimension specifier to actual dimensions.
m = libblis_test_get_dim_from_prob_size( op->dim_spec[0], p_cur );
n = libblis_test_get_dim_from_prob_size( op->dim_spec[1], p_cur );
// Map parameter characters to BLIS constants.
bli_param_map_char_to_blis_conj( pc_str[0], &conjbeta );
// Create test scalars.
bli_obj_scalar_init_detached( datatype, &beta );
// Create test operands (vectors and/or matrices).
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[0], m, n, &y );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[0], m, n, &y_save );
// Set beta to 0 + i.
//bli_setsc( 0.0, 1.0, &beta );
if ( bli_obj_is_real( y ) )
bli_setsc( -2.0, 0.0, &beta );
else
bli_setsc( 0.0, -2.0, &beta );
// Randomize and save y.
bli_randm( &y );
bli_copym( &y, &y_save );
// Apply the parameters.
bli_obj_set_conj( conjbeta, beta );
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
bli_copym( &y_save, &y );
time = bli_clock();
libblis_test_scalm_impl( iface, &beta, &y );
time_min = bli_clock_min_diff( time_min, time );
}
// Estimate the performance of the best experiment repeat.
*perf = ( 1.0 * m * n ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( y ) ) *perf *= 6.0;
// Perform checks.
libblis_test_scalm_check( &beta, &y, &y_save, resid );
// Zero out performance and residual if output matrix is empty.
libblis_test_check_empty_problem( &y, perf, resid );
// Free the test objects.
bli_obj_free( &y );
bli_obj_free( &y_save );
}
int main( int argc, char** argv )
{
obj_t a, b, c;
obj_t c_save;
obj_t alpha, beta;
dim_t m, n, k;
dim_t p;
dim_t p_begin, p_end, p_inc;
int m_input, n_input, k_input;
num_t dt;
int r, n_repeats;
trans_t transa;
trans_t transb;
f77_char f77_transa;
f77_char f77_transb;
double dtime;
double dtime_save;
double gflops;
bli_init();
//bli_error_checking_level_set( BLIS_NO_ERROR_CHECKING );
n_repeats = 3;
#ifndef PRINT
p_begin = 200;
p_end = 2000;
p_inc = 200;
m_input = -1;
n_input = -1;
k_input = -1;
#else
p_begin = 16;
p_end = 16;
p_inc = 1;
m_input = 5;
k_input = 6;
n_input = 4;
#endif
#if 1
//dt = BLIS_FLOAT;
dt = BLIS_DOUBLE;
#else
//dt = BLIS_SCOMPLEX;
dt = BLIS_DCOMPLEX;
#endif
transa = BLIS_NO_TRANSPOSE;
transb = BLIS_NO_TRANSPOSE;
bli_param_map_blis_to_netlib_trans( transa, &f77_transa );
bli_param_map_blis_to_netlib_trans( transb, &f77_transb );
for ( p = p_begin; p <= p_end; p += p_inc )
{
if ( m_input < 0 ) m = p * ( dim_t )abs(m_input);
else m = ( dim_t ) m_input;
if ( n_input < 0 ) n = p * ( dim_t )abs(n_input);
else n = ( dim_t ) n_input;
if ( k_input < 0 ) k = p * ( dim_t )abs(k_input);
else k = ( dim_t ) k_input;
bli_obj_create( dt, 1, 1, 0, 0, &alpha );
bli_obj_create( dt, 1, 1, 0, 0, &beta );
bli_obj_create( dt, m, k, 0, 0, &a );
bli_obj_create( dt, k, n, 0, 0, &b );
bli_obj_create( dt, m, n, 0, 0, &c );
bli_obj_create( dt, m, n, 0, 0, &c_save );
bli_randm( &a );
bli_randm( &b );
bli_randm( &c );
bli_obj_set_conjtrans( transa, a );
bli_obj_set_conjtrans( transb, b );
bli_setsc( (0.9/1.0), 0.2, &alpha );
bli_setsc( -(1.1/1.0), 0.3, &beta );
bli_copym( &c, &c_save );
dtime_save = DBL_MAX;
for ( r = 0; r < n_repeats; ++r )
{
bli_copym( &c_save, &c );
dtime = bli_clock();
#ifdef PRINT
bli_printm( "a", &a, "%4.1f", "" );
bli_printm( "b", &b, "%4.1f", "" );
bli_printm( "c", &c, "%4.1f", "" );
#endif
#ifdef BLIS
bli_gemm( &alpha,
&a,
&b,
&beta,
&c );
#else
if ( bli_is_float( dt ) )
{
f77_int mm = bli_obj_length( c );
f77_int kk = bli_obj_width_after_trans( a );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
float* alphap = bli_obj_buffer( alpha );
float* ap = bli_obj_buffer( a );
float* bp = bli_obj_buffer( b );
float* betap = bli_obj_buffer( beta );
float* cp = bli_obj_buffer( c );
sgemm_( &f77_transa,
&f77_transb,
&mm,
&nn,
&kk,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
else if ( bli_is_double( dt ) )
{
f77_int mm = bli_obj_length( c );
f77_int kk = bli_obj_width_after_trans( a );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
double* alphap = bli_obj_buffer( alpha );
double* ap = bli_obj_buffer( a );
double* bp = bli_obj_buffer( b );
double* betap = bli_obj_buffer( beta );
double* cp = bli_obj_buffer( c );
dgemm_( &f77_transa,
&f77_transb,
&mm,
&nn,
&kk,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
else if ( bli_is_scomplex( dt ) )
{
f77_int mm = bli_obj_length( c );
f77_int kk = bli_obj_width_after_trans( a );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
scomplex* alphap = bli_obj_buffer( alpha );
scomplex* ap = bli_obj_buffer( a );
scomplex* bp = bli_obj_buffer( b );
scomplex* betap = bli_obj_buffer( beta );
scomplex* cp = bli_obj_buffer( c );
cgemm_( &f77_transa,
&f77_transb,
&mm,
&nn,
&kk,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
else if ( bli_is_dcomplex( dt ) )
{
f77_int mm = bli_obj_length( c );
f77_int kk = bli_obj_width_after_trans( a );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
dcomplex* alphap = bli_obj_buffer( alpha );
dcomplex* ap = bli_obj_buffer( a );
dcomplex* bp = bli_obj_buffer( b );
dcomplex* betap = bli_obj_buffer( beta );
dcomplex* cp = bli_obj_buffer( c );
zgemm_( &f77_transa,
&f77_transb,
&mm,
&nn,
&kk,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
#endif
#ifdef PRINT
bli_printm( "c after", &c, "%4.1f", "" );
exit(1);
#endif
dtime_save = bli_clock_min_diff( dtime_save, dtime );
}
gflops = ( 2.0 * m * k * n ) / ( dtime_save * 1.0e9 );
if ( bli_is_complex( dt ) ) gflops *= 4.0;
#ifdef BLIS
printf( "data_gemm_blis" );
#else
printf( "data_gemm_%s", BLAS );
#endif
printf( "( %2lu, 1:5 ) = [ %4lu %4lu %4lu %10.3e %6.3f ];\n",
( unsigned long )(p - p_begin + 1)/p_inc + 1,
( unsigned long )m,
( unsigned long )k,
( unsigned long )n, dtime_save, gflops );
bli_obj_free( &alpha );
bli_obj_free( &beta );
bli_obj_free( &a );
bli_obj_free( &b );
bli_obj_free( &c );
bli_obj_free( &c_save );
}
bli_finalize();
return 0;
}
int main( int argc, char** argv )
{
obj_t a, c;
obj_t c_save;
obj_t alpha;
dim_t m, n;
dim_t p;
dim_t p_begin, p_end, p_inc;
int m_input, n_input;
ind_t ind;
num_t dt;
char dt_ch;
int r, n_repeats;
side_t side;
uplo_t uploa;
trans_t transa;
diag_t diaga;
f77_char f77_side;
f77_char f77_uploa;
f77_char f77_transa;
f77_char f77_diaga;
double dtime;
double dtime_save;
double gflops;
//bli_init();
//bli_error_checking_level_set( BLIS_NO_ERROR_CHECKING );
n_repeats = 3;
dt = DT;
ind = IND;
p_begin = P_BEGIN;
p_end = P_END;
p_inc = P_INC;
m_input = -1;
n_input = -1;
// Supress compiler warnings about unused variable 'ind'.
( void )ind;
#if 0
cntx_t* cntx;
ind_t ind_mod = ind;
// A hack to use 3m1 as 1mpb (with 1m as 1mbp).
if ( ind == BLIS_3M1 ) ind_mod = BLIS_1M;
// Initialize a context for the current induced method and datatype.
cntx = bli_gks_query_ind_cntx( ind_mod, dt );
// Set k to the kc blocksize for the current datatype.
k_input = bli_cntx_get_blksz_def_dt( dt, BLIS_KC, cntx );
#elif 1
//k_input = 256;
#endif
// Choose the char corresponding to the requested datatype.
if ( bli_is_float( dt ) ) dt_ch = 's';
else if ( bli_is_double( dt ) ) dt_ch = 'd';
else if ( bli_is_scomplex( dt ) ) dt_ch = 'c';
else dt_ch = 'z';
#if 0
side = BLIS_LEFT;
#else
side = BLIS_RIGHT;
#endif
#if 0
uploa = BLIS_LOWER;
#else
uploa = BLIS_UPPER;
#endif
transa = BLIS_NO_TRANSPOSE;
diaga = BLIS_NONUNIT_DIAG;
bli_param_map_blis_to_netlib_side( side, &f77_side );
bli_param_map_blis_to_netlib_uplo( uploa, &f77_uploa );
bli_param_map_blis_to_netlib_trans( transa, &f77_transa );
bli_param_map_blis_to_netlib_diag( diaga, &f77_diaga );
// Begin with initializing the last entry to zero so that
// matlab allocates space for the entire array once up-front.
for ( p = p_begin; p + p_inc <= p_end; p += p_inc ) ;
#ifdef BLIS
printf( "data_%s_%ctrmm_%s_blis", THR_STR, dt_ch, STR );
#else
printf( "data_%s_%ctrmm_%s", THR_STR, dt_ch, STR );
#endif
printf( "( %2lu, 1:3 ) = [ %4lu %4lu %7.2f ];\n",
( unsigned long )(p - p_begin + 1)/p_inc + 1,
( unsigned long )0,
( unsigned long )0, 0.0 );
for ( p = p_begin; p <= p_end; p += p_inc )
{
if ( m_input < 0 ) m = p / ( dim_t )abs(m_input);
else m = ( dim_t ) m_input;
if ( n_input < 0 ) n = p / ( dim_t )abs(n_input);
else n = ( dim_t ) n_input;
bli_obj_create( dt, 1, 1, 0, 0, &alpha );
if ( bli_does_trans( side ) )
bli_obj_create( dt, m, m, 0, 0, &a );
else
bli_obj_create( dt, n, n, 0, 0, &a );
bli_obj_create( dt, m, n, 0, 0, &c );
bli_obj_create( dt, m, n, 0, 0, &c_save );
bli_randm( &a );
bli_randm( &c );
bli_obj_set_struc( BLIS_TRIANGULAR, &a );
bli_obj_set_uplo( uploa, &a );
bli_obj_set_conjtrans( transa, &a );
bli_obj_set_diag( diaga, &a );
bli_randm( &a );
bli_mktrim( &a );
bli_setsc( (2.0/1.0), 0.0, &alpha );
bli_copym( &c, &c_save );
#if 0 //def BLIS
bli_ind_disable_all_dt( dt );
bli_ind_enable_dt( ind, dt );
#endif
dtime_save = DBL_MAX;
for ( r = 0; r < n_repeats; ++r )
{
bli_copym( &c_save, &c );
dtime = bli_clock();
#ifdef PRINT
bli_printm( "a", &a, "%4.1f", "" );
bli_printm( "c", &c, "%4.1f", "" );
#endif
#ifdef BLIS
bli_trmm( side,
&alpha,
&a,
&c );
#else
if ( bli_is_float( dt ) )
{
f77_int mm = bli_obj_length( &c );
f77_int kk = bli_obj_width( &c );
f77_int lda = bli_obj_col_stride( &a );
f77_int ldc = bli_obj_col_stride( &c );
float* alphap = bli_obj_buffer( &alpha );
float* ap = bli_obj_buffer( &a );
float* cp = bli_obj_buffer( &c );
strmm_( &f77_side,
&f77_uploa,
&f77_transa,
&f77_diaga,
&mm,
&kk,
alphap,
ap, &lda,
cp, &ldc );
}
else if ( bli_is_double( dt ) )
{
f77_int mm = bli_obj_length( &c );
f77_int kk = bli_obj_width( &c );
f77_int lda = bli_obj_col_stride( &a );
f77_int ldc = bli_obj_col_stride( &c );
double* alphap = bli_obj_buffer( &alpha );
double* ap = bli_obj_buffer( &a );
double* cp = bli_obj_buffer( &c );
dtrmm_( &f77_side,
&f77_uploa,
&f77_transa,
&f77_diaga,
&mm,
&kk,
alphap,
ap, &lda,
cp, &ldc );
}
else if ( bli_is_scomplex( dt ) )
{
f77_int mm = bli_obj_length( &c );
f77_int kk = bli_obj_width( &c );
f77_int lda = bli_obj_col_stride( &a );
f77_int ldc = bli_obj_col_stride( &c );
scomplex* alphap = bli_obj_buffer( &alpha );
scomplex* ap = bli_obj_buffer( &a );
scomplex* cp = bli_obj_buffer( &c );
ctrmm_( &f77_side,
&f77_uploa,
&f77_transa,
&f77_diaga,
&mm,
&kk,
alphap,
ap, &lda,
cp, &ldc );
}
else if ( bli_is_dcomplex( dt ) )
{
f77_int mm = bli_obj_length( &c );
f77_int kk = bli_obj_width( &c );
f77_int lda = bli_obj_col_stride( &a );
f77_int ldc = bli_obj_col_stride( &c );
dcomplex* alphap = bli_obj_buffer( &alpha );
dcomplex* ap = bli_obj_buffer( &a );
dcomplex* cp = bli_obj_buffer( &c );
ztrmm_( &f77_side,
&f77_uploa,
&f77_transa,
&f77_diaga,
&mm,
&kk,
alphap,
ap, &lda,
cp, &ldc );
}
#endif
#ifdef PRINT
bli_printm( "c after", &c, "%4.1f", "" );
exit(1);
#endif
dtime_save = bli_clock_min_diff( dtime_save, dtime );
}
if ( bli_is_left( side ) )
gflops = ( 1.0 * m * m * n ) / ( dtime_save * 1.0e9 );
else
gflops = ( 1.0 * m * n * n ) / ( dtime_save * 1.0e9 );
if ( bli_is_complex( dt ) ) gflops *= 4.0;
#ifdef BLIS
printf( "data_%s_%ctrmm_%s_blis", THR_STR, dt_ch, STR );
#else
printf( "data_%s_%ctrmm_%s", THR_STR, dt_ch, STR );
#endif
printf( "( %2lu, 1:3 ) = [ %4lu %4lu %7.2f ];\n",
( unsigned long )(p - p_begin + 1)/p_inc + 1,
( unsigned long )m,
( unsigned long )n, gflops );
bli_obj_free( &alpha );
bli_obj_free( &a );
bli_obj_free( &c );
bli_obj_free( &c_save );
}
//bli_finalize();
return 0;
}
int main( int argc, char** argv )
{
obj_t a, b, c;
obj_t c_save;
obj_t alpha, beta;
dim_t m, n;
dim_t p;
dim_t p_begin, p_end, p_inc;
int m_input, n_input;
num_t dt_a, dt_b, dt_c;
num_t dt_alpha, dt_beta;
int r, n_repeats;
side_t side;
uplo_t uplo;
double dtime;
double dtime_save;
double gflops;
bli_init();
n_repeats = 3;
#ifndef PRINT
p_begin = 1000;
p_end = 1000;
p_inc = 40;
m_input = -1;
n_input = -1;
#else
p_begin = 16;
p_end = 16;
p_inc = 1;
m_input = 8 ;
n_input = 4 ;
#endif
#if 0
dt_a = BLIS_DOUBLE;
dt_b = BLIS_DOUBLE;
dt_c = BLIS_DOUBLE;
dt_alpha = BLIS_DOUBLE;
dt_beta = BLIS_DOUBLE;
#else
dt_a = dt_b = dt_c = dt_alpha = dt_beta = BLIS_FLOAT;
//dt_a = dt_b = dt_c = dt_alpha = dt_beta = BLIS_SCOMPLEX;
#endif
side = BLIS_LEFT;
//side = BLIS_RIGHT;
uplo = BLIS_LOWER;
//uplo = BLIS_UPPER;
for ( p = p_begin; p <= p_end; p += p_inc )
{
if ( m_input < 0 ) m = p * ( dim_t )abs(m_input);
else m = ( dim_t ) m_input;
if ( n_input < 0 ) n = p * ( dim_t )abs(n_input);
else n = ( dim_t ) n_input;
bli_obj_create( dt_alpha, 1, 1, 0, 0, &alpha );
bli_obj_create( dt_beta, 1, 1, 0, 0, &beta );
if ( bli_is_left( side ) )
bli_obj_create( dt_a, m, m, 0, 0, &a );
else
bli_obj_create( dt_a, n, n, 0, 0, &a );
bli_obj_create( dt_b, m, n, 0, 0, &b );
bli_obj_create( dt_c, m, n, 0, 0, &c );
bli_obj_create( dt_c, m, n, 0, 0, &c_save );
bli_obj_set_struc( BLIS_TRIANGULAR, a );
bli_obj_set_uplo( uplo, a );
//bli_obj_set_diag( BLIS_UNIT_DIAG, a );
bli_randm( &a );
bli_randm( &c );
bli_randm( &b );
/*
{
obj_t a2;
bli_obj_alias_to( a, a2 );
bli_obj_toggle_uplo( a2 );
bli_obj_inc_diag_off( 1, a2 );
bli_setm( &BLIS_ZERO, &a2 );
bli_obj_inc_diag_off( -2, a2 );
bli_obj_toggle_uplo( a2 );
bli_obj_set_diag( BLIS_NONUNIT_DIAG, a2 );
bli_scalm( &BLIS_TWO, &a2 );
//bli_scalm( &BLIS_TWO, &a );
}
*/
bli_setsc( (2.0/1.0), 0.0, &alpha );
bli_setsc( -(1.0/1.0), 0.0, &beta );
bli_copym( &c, &c_save );
dtime_save = 1.0e9;
for ( r = 0; r < n_repeats; ++r )
{
bli_copym( &c_save, &c );
dtime = bli_clock();
#ifdef PRINT
/*
obj_t ar, ai;
bli_obj_alias_to( a, ar );
bli_obj_alias_to( a, ai );
bli_obj_set_datatype( BLIS_DOUBLE, ar ); ar.rs *= 2; ar.cs *= 2;
bli_obj_set_datatype( BLIS_DOUBLE, ai ); ai.rs *= 2; ai.cs *= 2; ai.buffer = ( double* )ai.buffer + 1;
bli_printm( "ar", &ar, "%4.1f", "" );
bli_printm( "ai", &ai, "%4.1f", "" );
*/
bli_invertd( &a );
bli_printm( "a", &a, "%4.1f", "" );
bli_invertd( &a );
bli_printm( "c", &c, "%4.1f", "" );
#endif
#ifdef BLIS
//bli_error_checking_level_set( BLIS_NO_ERROR_CHECKING );
bli_trsm( side,
//bli_trsm4m( side,
//bli_trsm3m( side,
&alpha,
&a,
&c );
#else
if ( bli_is_real( dt_a ) )
{
f77_char side = 'L';
f77_char uplo = 'L';
f77_char transa = 'N';
f77_char diag = 'N';
f77_int mm = bli_obj_length( c );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldc = bli_obj_col_stride( c );
float * alphap = bli_obj_buffer( alpha );
float * ap = bli_obj_buffer( a );
float * cp = bli_obj_buffer( c );
strsm_( &side,
&uplo,
&transa,
&diag,
&mm,
&nn,
alphap,
ap, &lda,
cp, &ldc );
}
else // if ( bli_is_complex( dt_a ) )
{
f77_char side = 'L';
f77_char uplo = 'L';
f77_char transa = 'N';
f77_char diag = 'N';
f77_int mm = bli_obj_length( c );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldc = bli_obj_col_stride( c );
scomplex* alphap = bli_obj_buffer( alpha );
scomplex* ap = bli_obj_buffer( a );
scomplex* cp = bli_obj_buffer( c );
ctrsm_( &side,
//ztrsm_( &side,
&uplo,
&transa,
&diag,
&mm,
&nn,
alphap,
ap, &lda,
cp, &ldc );
}
#endif
#ifdef PRINT
bli_printm( "c after", &c, "%4.1f", "" );
exit(1);
#endif
dtime_save = bli_clock_min_diff( dtime_save, dtime );
}
if ( bli_is_left( side ) )
gflops = ( 1.0 * m * m * n ) / ( dtime_save * 1.0e9 );
else
gflops = ( 1.0 * m * n * n ) / ( dtime_save * 1.0e9 );
if ( bli_is_complex( dt_a ) ) gflops *= 4.0;
#ifdef BLIS
printf( "data_trsm_blis" );
#else
printf( "data_trsm_%s", BLAS );
#endif
printf( "( %2lu, 1:4 ) = [ %4lu %4lu %10.3e %6.3f ];\n",
( unsigned long )(p - p_begin + 1)/p_inc + 1,
( unsigned long )m,
( unsigned long )n, dtime_save, gflops );
bli_obj_free( &alpha );
bli_obj_free( &beta );
bli_obj_free( &a );
bli_obj_free( &b );
bli_obj_free( &c );
bli_obj_free( &c_save );
}
bli_finalize();
return 0;
}
int main( int argc, char** argv )
{
//bli_init();
#if 0
obj_t a, b, c;
obj_t aa, bb, cc;
dim_t m, n, k;
num_t dt;
uplo_t uploa, uplob, uploc;
{
dt = BLIS_DOUBLE;
m = 6;
k = 6;
n = 6;
bli_obj_create( dt, m, k, 0, 0, &a );
bli_obj_create( dt, k, n, 0, 0, &b );
bli_obj_create( dt, m, n, 0, 0, &c );
uploa = BLIS_UPPER;
uploa = BLIS_LOWER;
bli_obj_set_struc( BLIS_TRIANGULAR, &a );
bli_obj_set_uplo( uploa, &a );
bli_obj_set_diag_offset( -2, &a );
uplob = BLIS_UPPER;
uplob = BLIS_LOWER;
bli_obj_set_struc( BLIS_TRIANGULAR, &b );
bli_obj_set_uplo( uplob, &b );
bli_obj_set_diag_offset( -2, &b );
uploc = BLIS_UPPER;
//uploc = BLIS_LOWER;
//uploc = BLIS_ZEROS;
//uploc = BLIS_DENSE;
bli_obj_set_struc( BLIS_HERMITIAN, &c );
//bli_obj_set_struc( BLIS_TRIANGULAR, &c );
bli_obj_set_uplo( uploc, &c );
bli_obj_set_diag_offset( 1, &c );
bli_obj_alias_to( &a, &aa ); (void)aa;
bli_obj_alias_to( &b, &bb ); (void)bb;
bli_obj_alias_to( &c, &cc ); (void)cc;
bli_randm( &a );
bli_randm( &b );
bli_randm( &c );
//bli_mkherm( &a );
//bli_mktrim( &a );
bli_prune_unref_mparts( &cc, BLIS_M,
&aa, BLIS_N );
bli_printm( "c orig", &c, "%4.1f", "" );
bli_printm( "c alias", &cc, "%4.1f", "" );
bli_printm( "a orig", &a, "%4.1f", "" );
bli_printm( "a alias", &aa, "%4.1f", "" );
//bli_obj_print( "a struct", &a );
}
#endif
dim_t p_begin, p_max, p_inc;
gint_t m_input, n_input;
char uploa_ch;
doff_t diagoffa;
dim_t bf;
dim_t n_way;
char part_dim_ch;
bool_t go_fwd;
char out_ch;
obj_t a;
blksz_t bfs;
thrinfo_t thrinfo;
dim_t m, n;
uplo_t uploa;
bool_t part_m_dim, part_n_dim;
bool_t go_bwd;
dim_t p;
num_t dt;
dim_t start, end;
dim_t width;
siz_t area;
gint_t t_begin, t_stop, t_inc;
dim_t t;
if ( argc == 13 )
{
sscanf( argv[1], "%u", &p_begin );
sscanf( argv[2], "%u", &p_max );
sscanf( argv[3], "%u", &p_inc );
sscanf( argv[4], "%d", &m_input );
sscanf( argv[5], "%d", &n_input );
sscanf( argv[6], "%c", &uploa_ch );
sscanf( argv[7], "%d", &diagoffa );
sscanf( argv[8], "%u", &bf );
sscanf( argv[9], "%u", &n_way );
sscanf( argv[10], "%c", &part_dim_ch );
sscanf( argv[11], "%u", &go_fwd );
sscanf( argv[12], "%c", &out_ch );
}
else
{
printf( "\n" );
printf( " %s\n", argv[0] );
printf( "\n" );
printf( " Simulate the dimension ranges assigned to threads when\n" );
printf( " partitioning a matrix for parallelism in BLIS.\n" );
printf( "\n" );
printf( " Usage:\n" );
printf( "\n" );
printf( " %s p_beg p_max p_inc m n uplo doff bf n_way part_dim go_fwd out\n", argv[0] );
printf( "\n" );
printf( " p_beg: the first problem size p to test.\n" );
printf( " p_max: the maximum problem size p to test.\n" );
printf( " p_inc: the increase in problem size p between tests.\n" );
printf( " m: the m dimension:\n" );
printf( " n: the n dimension:\n" );
printf( " if m,n = -1: bind m,n to problem size p.\n" );
printf( " if m,n = 0: bind m,n to p_max.\n" );
printf( " if m,n > 0: hold m,n = c constant for all p.\n" );
printf( " uplo: the uplo field of the matrix being partitioned:\n" );
printf( " 'l': lower-stored (BLIS_LOWER)\n" );
printf( " 'u': upper-stored (BLIS_UPPER)\n" );
printf( " 'd': densely-stored (BLIS_DENSE)\n" );
printf( " doff: the diagonal offset of the matrix being partitioned.\n" );
printf( " bf: the simulated blocking factor. all thread ranges must\n" );
printf( " be a multiple of bf, except for the range that contains\n" );
printf( " the edge case (if one exists). the blocking factor\n" );
printf( " would typically correspond to a register blocksize.\n" );
printf( " n_way: the number of ways of parallelism for which we are\n" );
printf( " partitioning (i.e.: the number of threads, or thread\n" );
printf( " groups).\n" );
printf( " part_dim: the dimension to partition:\n" );
printf( " 'm': partition the m dimension.\n" );
printf( " 'n': partition the n dimension.\n" );
printf( " go_fwd: the direction to partition:\n" );
printf( " '1': forward, e.g. left-to-right (part_dim = 'm') or\n" );
printf( " top-to-bottom (part_dim = 'n')\n" );
printf( " '0': backward, e.g. right-to-left (part_dim = 'm') or\n" );
printf( " bottom-to-top (part_dim = 'n')\n" );
printf( " NOTE: reversing the direction does not change the\n" );
printf( " subpartitions' widths, but it does change which end of\n" );
printf( " the index range receives the edge case, if it exists.\n" );
printf( " out: the type of output per thread-column:\n" );
printf( " 'w': the width (and area) of the thread's subpartition\n" );
printf( " 'r': the actual ranges of the thread's subpartition\n" );
printf( " where the start and end points of each range are\n" );
printf( " inclusive and exclusive, respectively.\n" );
printf( "\n" );
exit(1);
}
if ( m_input == 0 ) m_input = p_max;
if ( n_input == 0 ) n_input = p_max;
if ( part_dim_ch == 'm' ) { part_m_dim = TRUE; part_n_dim = FALSE; }
else { part_m_dim = FALSE; part_n_dim = TRUE; }
go_bwd = !go_fwd;
if ( uploa_ch == 'l' ) uploa = BLIS_LOWER;
else if ( uploa_ch == 'u' ) uploa = BLIS_UPPER;
else uploa = BLIS_DENSE;
if ( part_n_dim )
{
if ( bli_is_upper( uploa ) ) { t_begin = n_way-1; t_stop = -1; t_inc = -1; }
else /* if lower or dense */ { t_begin = 0; t_stop = n_way; t_inc = 1; }
}
else // if ( part_m_dim )
{
if ( bli_is_lower( uploa ) ) { t_begin = n_way-1; t_stop = -1; t_inc = -1; }
else /* if upper or dense */ { t_begin = 0; t_stop = n_way; t_inc = 1; }
}
printf( "\n" );
printf( " part: %3s doff: %3d bf: %3d output: %s\n",
( part_n_dim ? ( go_fwd ? "l2r" : "r2l" )
: ( go_fwd ? "t2b" : "b2t" ) ),
( int )diagoffa, ( int )bf,
( out_ch == 'w' ? "width(area)" : "ranges" ) );
printf( " uplo: %3c nt: %3u\n", uploa_ch, ( unsigned )n_way );
printf( "\n" );
printf( " " );
for ( t = t_begin; t != t_stop; t += t_inc )
{
if ( part_n_dim )
{
if ( t == t_begin ) printf( "left... " );
else if ( t == t_stop-t_inc ) printf( " ...right" );
else printf( " " );
}
else // if ( part_m_dim )
{
if ( t == t_begin ) printf( "top... " );
else if ( t == t_stop-t_inc ) printf( " ...bottom" );
else printf( " " );
}
}
printf( "\n" );
printf( "%4c x %4c ", 'm', 'n' );
for ( t = t_begin; t != t_stop; t += t_inc )
{
printf( "%9s %u ", "thread", ( unsigned )t );
}
printf( "\n" );
printf( "-------------" );
for ( t = t_begin; t != t_stop; t += t_inc )
{
printf( "-------------" );
}
printf( "\n" );
for ( p = p_begin; p <= p_max; p += p_inc )
{
if ( m_input < 0 ) m = ( dim_t )p;
else m = ( dim_t )m_input;
if ( n_input < 0 ) n = ( dim_t )p;
else n = ( dim_t )n_input;
dt = BLIS_DOUBLE;
bli_obj_create( dt, m, n, 0, 0, &a );
bli_obj_set_struc( BLIS_TRIANGULAR, &a );
bli_obj_set_uplo( uploa, &a );
bli_obj_set_diag_offset( diagoffa, &a );
bli_randm( &a );
bli_blksz_init_easy( &bfs, bf, bf, bf, bf );
printf( "%4u x %4u ", ( unsigned )m, ( unsigned )n );
for ( t = t_begin; t != t_stop; t += t_inc )
{
thrinfo.n_way = n_way;
thrinfo.work_id = t;
if ( part_n_dim && go_fwd )
area = bli_thread_get_range_weighted_l2r( &thrinfo, &a, &bfs, &start, &end );
else if ( part_n_dim && go_bwd )
area = bli_thread_get_range_weighted_r2l( &thrinfo, &a, &bfs, &start, &end );
else if ( part_m_dim && go_fwd )
area = bli_thread_get_range_weighted_t2b( &thrinfo, &a, &bfs, &start, &end );
else // ( part_m_dim && go_bwd )
area = bli_thread_get_range_weighted_b2t( &thrinfo, &a, &bfs, &start, &end );
width = end - start;
if ( out_ch == 'w' ) printf( "%4u(%6u) ", ( unsigned )width,
( unsigned )area );
else printf( "[%4u,%4u) ", ( unsigned )start,
( unsigned )end );
}
printf( "\n" );
bli_obj_free( &a );
}
//bli_finalize();
return 0;
}
int main( int argc, char** argv )
{
obj_t a, b, c;
obj_t x, y;
obj_t alpha, beta;
dim_t m;
num_t dt_a, dt_b, dt_c;
num_t dt_alpha, dt_beta;
int ii;
#ifdef NBLIS
bli_init();
#endif
m = 4000;
dt_a = BLIS_DOUBLE;
dt_b = BLIS_DOUBLE;
dt_c = BLIS_DOUBLE;
dt_alpha = BLIS_DOUBLE;
dt_beta = BLIS_DOUBLE;
{
#ifdef NBLIS
bli_obj_create( dt_alpha, 1, 1, 0, 0, &alpha );
bli_obj_create( dt_beta, 1, 1, 0, 0, &beta );
bli_obj_create( dt_a, m, 1, 0, 0, &x );
bli_obj_create( dt_a, m, 1, 0, 0, &y );
bli_obj_create( dt_a, m, m, 0, 0, &a );
bli_obj_create( dt_b, m, m, 0, 0, &b );
bli_obj_create( dt_c, m, m, 0, 0, &c );
bli_randm( &a );
bli_randm( &b );
bli_randm( &c );
bli_setsc( (2.0/1.0), 0.0, &alpha );
bli_setsc( -(1.0/1.0), 0.0, &beta );
#endif
#ifdef NBLAS
x.buffer = malloc( m * 1 * sizeof( double ) );
y.buffer = malloc( m * 1 * sizeof( double ) );
alpha.buffer = malloc( 1 * sizeof( double ) );
beta.buffer = malloc( 1 * sizeof( double ) );
a.buffer = malloc( m * m * sizeof( double ) );
a.m = m;
a.n = m;
a.cs = m;
b.buffer = malloc( m * m * sizeof( double ) );
b.m = m;
b.n = m;
b.cs = m;
c.buffer = malloc( m * m * sizeof( double ) );
c.m = m;
c.n = m;
c.cs = m;
*((double*)alpha.buffer) = 2.0;
*((double*)beta.buffer) = -1.0;
#endif
#ifdef NBLIS
#if NBLIS >= 1
for ( ii = 0; ii < 2000000000; ++ii )
{
bli_gemm( &BLIS_ONE,
&a,
&b,
&BLIS_ONE,
&c );
}
#endif
#if NBLIS >= 2
{
bli_hemm( BLIS_LEFT,
&BLIS_ONE,
&a,
&b,
&BLIS_ONE,
&c );
}
#endif
#if NBLIS >= 3
{
bli_herk( &BLIS_ONE,
&a,
&BLIS_ONE,
&c );
}
#endif
#if NBLIS >= 4
{
bli_her2k( &BLIS_ONE,
&a,
&b,
&BLIS_ONE,
&c );
}
#endif
#if NBLIS >= 5
{
bli_trmm( BLIS_LEFT,
&BLIS_ONE,
&a,
&c );
}
#endif
#if NBLIS >= 6
{
bli_trsm( BLIS_LEFT,
&BLIS_ONE,
&a,
&c );
}
#endif
#endif
#ifdef NBLAS
#if NBLAS >= 1
for ( ii = 0; ii < 2000000000; ++ii )
{
f77_char transa = 'N';
f77_char transb = 'N';
f77_int mm = bli_obj_length( c );
f77_int kk = bli_obj_width_after_trans( a );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
double* alphap = bli_obj_buffer( alpha );
double* ap = bli_obj_buffer( a );
double* bp = bli_obj_buffer( b );
double* betap = bli_obj_buffer( beta );
double* cp = bli_obj_buffer( c );
dgemm_( &transa,
&transb,
&mm,
&nn,
&kk,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
#endif
#if NBLAS >= 2
{
f77_char side = 'L';
f77_char uplo = 'L';
f77_int mm = bli_obj_length( c );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
double* alphap = bli_obj_buffer( alpha );
double* ap = bli_obj_buffer( a );
double* bp = bli_obj_buffer( b );
double* betap = bli_obj_buffer( beta );
double* cp = bli_obj_buffer( c );
dsymm_( &side,
&uplo,
&mm,
&nn,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
#endif
#if NBLAS >= 3
{
f77_char uplo = 'L';
f77_char trans = 'N';
f77_int mm = bli_obj_length( c );
f77_int kk = bli_obj_width( a );
f77_int lda = bli_obj_col_stride( a );
f77_int ldc = bli_obj_col_stride( c );
double* alphap = bli_obj_buffer( alpha );
double* ap = bli_obj_buffer( a );
double* betap = bli_obj_buffer( beta );
double* cp = bli_obj_buffer( c );
dsyrk_( &uplo,
&trans,
&mm,
&kk,
alphap,
ap, &lda,
betap,
cp, &ldc );
}
#endif
#if NBLAS >= 4
{
f77_char uplo = 'L';
f77_char trans = 'N';
f77_int mm = bli_obj_length( c );
f77_int kk = bli_obj_width( a );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
double* alphap = bli_obj_buffer( alpha );
double* ap = bli_obj_buffer( a );
double* bp = bli_obj_buffer( b );
double* betap = bli_obj_buffer( beta );
double* cp = bli_obj_buffer( c );
dsyr2k_( &uplo,
&trans,
&mm,
&kk,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
#endif
#if NBLAS >= 5
{
f77_char side = 'L';
f77_char uplo = 'L';
f77_char trans = 'N';
f77_char diag = 'N';
f77_int mm = bli_obj_length( c );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldc = bli_obj_col_stride( c );
double* alphap = bli_obj_buffer( alpha );
double* ap = bli_obj_buffer( a );
double* cp = bli_obj_buffer( c );
dtrmm_( &side,
&uplo,
&trans,
&diag,
&mm,
&nn,
alphap,
ap, &lda,
cp, &ldc );
}
#endif
#if NBLAS >= 6
{
f77_char side = 'L';
f77_char uplo = 'L';
f77_char trans = 'N';
f77_char diag = 'N';
f77_int mm = bli_obj_length( c );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldc = bli_obj_col_stride( c );
double* alphap = bli_obj_buffer( alpha );
double* ap = bli_obj_buffer( a );
double* cp = bli_obj_buffer( c );
dtrsm_( &side,
&uplo,
&trans,
&diag,
&mm,
&nn,
alphap,
ap, &lda,
cp, &ldc );
}
#endif
#if NBLAS >= 7
{
f77_char transa = 'N';
f77_char transb = 'N';
f77_int mm = bli_obj_length( c );
f77_int kk = bli_obj_width_after_trans( a );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
dcomplex* alphap = bli_obj_buffer( alpha );
dcomplex* ap = bli_obj_buffer( a );
dcomplex* bp = bli_obj_buffer( b );
dcomplex* betap = bli_obj_buffer( beta );
dcomplex* cp = bli_obj_buffer( c );
zgemm_( &transa,
&transb,
&mm,
&nn,
&kk,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
#endif
#if NBLAS >= 8
{
f77_char side = 'L';
f77_char uplo = 'L';
f77_int mm = bli_obj_length( c );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
dcomplex* alphap = bli_obj_buffer( alpha );
dcomplex* ap = bli_obj_buffer( a );
dcomplex* bp = bli_obj_buffer( b );
dcomplex* betap = bli_obj_buffer( beta );
dcomplex* cp = bli_obj_buffer( c );
zhemm_( &side,
&uplo,
&mm,
&nn,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
#endif
#if NBLAS >= 9
{
f77_char uplo = 'L';
f77_char trans = 'N';
f77_int mm = bli_obj_length( c );
f77_int kk = bli_obj_width( a );
f77_int lda = bli_obj_col_stride( a );
f77_int ldc = bli_obj_col_stride( c );
double* alphap = bli_obj_buffer( alpha );
dcomplex* ap = bli_obj_buffer( a );
double* betap = bli_obj_buffer( beta );
dcomplex* cp = bli_obj_buffer( c );
zherk_( &uplo,
&trans,
&mm,
&kk,
alphap,
ap, &lda,
betap,
cp, &ldc );
}
#endif
#if NBLAS >= 10
{
f77_char uplo = 'L';
f77_char trans = 'N';
f77_int mm = bli_obj_length( c );
f77_int kk = bli_obj_width( a );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
dcomplex* alphap = bli_obj_buffer( alpha );
dcomplex* ap = bli_obj_buffer( a );
dcomplex* bp = bli_obj_buffer( b );
double* betap = bli_obj_buffer( beta );
dcomplex* cp = bli_obj_buffer( c );
zher2k_( &uplo,
&trans,
&mm,
&kk,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
}
#endif
#if NBLAS >= 11
{
f77_char side = 'L';
f77_char uplo = 'L';
f77_char trans = 'N';
f77_char diag = 'N';
f77_int mm = bli_obj_length( c );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldc = bli_obj_col_stride( c );
dcomplex* alphap = bli_obj_buffer( alpha );
dcomplex* ap = bli_obj_buffer( a );
dcomplex* cp = bli_obj_buffer( c );
ztrmm_( &side,
&uplo,
&trans,
&diag,
&mm,
&nn,
alphap,
ap, &lda,
cp, &ldc );
}
#endif
#if NBLAS >= 12
{
f77_char side = 'L';
f77_char uplo = 'L';
f77_char trans = 'N';
f77_char diag = 'N';
f77_int mm = bli_obj_length( c );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldc = bli_obj_col_stride( c );
dcomplex* alphap = bli_obj_buffer( alpha );
dcomplex* ap = bli_obj_buffer( a );
dcomplex* cp = bli_obj_buffer( c );
ztrsm_( &side,
&uplo,
&trans,
&diag,
&mm,
&nn,
alphap,
ap, &lda,
cp, &ldc );
}
#endif
#endif
#ifdef NBLIS
bli_obj_free( &x );
bli_obj_free( &y );
bli_obj_free( &alpha );
bli_obj_free( &beta );
bli_obj_free( &a );
bli_obj_free( &b );
bli_obj_free( &c );
#endif
#ifdef NBLAS
free( x.buffer );
free( y.buffer );
free( alpha.buffer );
free( beta.buffer );
free( a.buffer );
free( b.buffer );
free( c.buffer );
#endif
}
#ifdef NBLIS
bli_finalize();
#endif
return 0;
}
int main( int argc, char** argv )
{
obj_t a, c;
obj_t c_save;
obj_t alpha, beta;
dim_t m, k;
dim_t p;
dim_t p_begin, p_end, p_inc;
int m_input, k_input;
num_t dt_a, dt_c;
num_t dt_alpha, dt_beta;
int r, n_repeats;
uplo_t uplo;
double dtime;
double dtime_save;
double gflops;
bli_init();
n_repeats = 3;
if( argc < 7 )
{
printf("Usage:\n");
printf("test_foo.x m n k p_begin p_inc p_end:\n");
exit;
}
int world_size, world_rank, provided;
MPI_Init_thread( NULL, NULL, MPI_THREAD_FUNNELED, &provided );
MPI_Comm_size( MPI_COMM_WORLD, &world_size );
MPI_Comm_rank( MPI_COMM_WORLD, &world_rank );
m_input = strtol( argv[1], NULL, 10 );
k_input = strtol( argv[3], NULL, 10 );
p_begin = strtol( argv[4], NULL, 10 );
p_inc = strtol( argv[5], NULL, 10 );
p_end = strtol( argv[6], NULL, 10 );
dt_a = BLIS_DOUBLE;
dt_c = BLIS_DOUBLE;
dt_alpha = BLIS_DOUBLE;
dt_beta = BLIS_DOUBLE;
uplo = BLIS_LOWER;
for ( p = p_begin + world_rank * p_inc; p <= p_end; p += p_inc * world_size )
{
if ( m_input < 0 ) m = p * ( dim_t )abs(m_input);
else m = ( dim_t ) m_input;
if ( k_input < 0 ) k = p * ( dim_t )abs(k_input);
else k = ( dim_t ) k_input;
bli_obj_create( dt_alpha, 1, 1, 0, 0, &alpha );
bli_obj_create( dt_beta, 1, 1, 0, 0, &beta );
bli_obj_create( dt_a, m, k, 0, 0, &a );
bli_obj_create( dt_c, m, m, 0, 0, &c );
bli_obj_create( dt_c, m, m, 0, 0, &c_save );
bli_randm( &a );
bli_randm( &c );
bli_obj_set_struc( BLIS_HERMITIAN, &c );
bli_obj_set_uplo( uplo, &c );
bli_setsc( (2.0/1.0), 0.0, &alpha );
bli_setsc( (1.0/1.0), 0.0, &beta );
bli_copym( &c, &c_save );
dtime_save = 1.0e9;
for ( r = 0; r < n_repeats; ++r )
{
bli_copym( &c_save, &c );
dtime = bli_clock();
#ifdef PRINT
bli_printm( "a", &a, "%4.1f", "" );
bli_printm( "c", &c, "%4.1f", "" );
#endif
#ifdef BLIS
//bli_error_checking_level_set( BLIS_NO_ERROR_CHECKING );
bli_herk( &alpha,
&a,
&beta,
&c );
#else
f77_char uploa = 'L';
f77_char transa = 'N';
f77_int mm = bli_obj_length( &c );
f77_int kk = bli_obj_width_after_trans( &a );
f77_int lda = bli_obj_col_stride( &a );
f77_int ldc = bli_obj_col_stride( &c );
double* alphap = bli_obj_buffer( &alpha );
double* ap = bli_obj_buffer( &a );
double* betap = bli_obj_buffer( &beta );
double* cp = bli_obj_buffer( &c );
dsyrk_( &uploa,
&transa,
&mm,
&kk,
alphap,
ap, &lda,
betap,
cp, &ldc );
#endif
#ifdef PRINT
bli_printm( "c after", &c, "%4.1f", "" );
exit(1);
#endif
dtime_save = bli_clock_min_diff( dtime_save, dtime );
}
gflops = ( 1.0 * m * k * m ) / ( dtime_save * 1.0e9 );
#ifdef BLIS
printf( "data_herk_blis" );
#else
printf( "data_herk_%s", BLAS );
#endif
printf( "( %2lu, 1:4 ) = [ %4lu %4lu %10.3e %6.3f ];\n",
( unsigned long )(p - p_begin + 1)/p_inc + 1,
( unsigned long )m,
( unsigned long )k, dtime_save, gflops );
bli_obj_free( &alpha );
bli_obj_free( &beta );
bli_obj_free( &a );
bli_obj_free( &c );
bli_obj_free( &c_save );
}
bli_finalize();
return 0;
}
int main( int argc, char** argv )
{
obj_t a, b, c;
obj_t c_save;
obj_t alpha, beta;
dim_t m, n, k;
dim_t p;
dim_t p_begin, p_end, p_inc;
int m_input, n_input, k_input;
num_t dt_a, dt_b, dt_c;
num_t dt_alpha, dt_beta;
int r, n_repeats;
double dtime;
double dtime_save;
double gflops;
bli_init();
n_repeats = 3;
#ifndef PRINT
p_begin = 40;
p_end = 2000;
p_inc = 40;
m_input = -1;
n_input = -1;
k_input = -1;
#else
p_begin = 16;
p_end = 16;
p_inc = 1;
m_input = 8;
k_input = 16;
n_input = 16;
#endif
dt_a = BLIS_DOUBLE;
dt_b = BLIS_DOUBLE;
dt_c = BLIS_DOUBLE;
dt_alpha = BLIS_DOUBLE;
dt_beta = BLIS_DOUBLE;
for ( p = p_begin; p <= p_end; p += p_inc )
{
if ( m_input < 0 ) m = p * ( dim_t )abs(m_input);
else m = ( dim_t ) m_input;
if ( n_input < 0 ) n = p * ( dim_t )abs(n_input);
else n = ( dim_t ) n_input;
if ( k_input < 0 ) k = p * ( dim_t )abs(k_input);
else k = ( dim_t ) k_input;
bli_obj_create( dt_alpha, 1, 1, 0, 0, &alpha );
bli_obj_create( dt_beta, 1, 1, 0, 0, &beta );
bli_obj_create( dt_a, m, k, 0, 0, &a );
bli_obj_create( dt_b, k, n, 0, 0, &b );
bli_obj_create( dt_c, m, n, 0, 0, &c );
bli_obj_create( dt_c, m, n, 0, 0, &c_save );
bli_randm( &a );
bli_randm( &b );
bli_randm( &c );
bli_setsc( (2.0/1.0), 0.0, &alpha );
bli_setsc( -(1.0/1.0), 0.0, &beta );
bli_copym( &c, &c_save );
dtime_save = 1.0e9;
for ( r = 0; r < n_repeats; ++r )
{
bli_copym( &c_save, &c );
dtime = bli_clock();
#ifdef PRINT
bli_printm( "a", &a, "%4.1f", "" );
bli_printm( "b", &b, "%4.1f", "" );
bli_printm( "c", &c, "%4.1f", "" );
#endif
#ifdef BLIS
//bli_error_checking_level_set( BLIS_NO_ERROR_CHECKING );
bli_gemm( &alpha,
&a,
&b,
&beta,
&c );
#else
f77_char transa = 'N';
f77_char transb = 'N';
f77_int mm = bli_obj_length( c );
f77_int kk = bli_obj_width_after_trans( a );
f77_int nn = bli_obj_width( c );
f77_int lda = bli_obj_col_stride( a );
f77_int ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
double* alphap = bli_obj_buffer( alpha );
double* ap = bli_obj_buffer( a );
double* bp = bli_obj_buffer( b );
double* betap = bli_obj_buffer( beta );
double* cp = bli_obj_buffer( c );
dgemm_( &transa,
&transb,
&mm,
&nn,
&kk,
alphap,
ap, &lda,
bp, &ldb,
betap,
cp, &ldc );
#endif
#ifdef PRINT
bli_printm( "c after", &c, "%4.1f", "" );
exit(1);
#endif
dtime_save = bli_clock_min_diff( dtime_save, dtime );
}
gflops = ( 2.0 * m * k * n ) / ( dtime_save * 1.0e9 );
#ifdef BLIS
printf( "data_gemm_blis" );
#else
printf( "data_gemm_%s", BLAS );
#endif
printf( "( %2lu, 1:5 ) = [ %4lu %4lu %4lu %10.3e %6.3f ];\n",
( unsigned long )(p - p_begin + 1)/p_inc + 1,
( unsigned long )m,
( unsigned long )k,
( unsigned long )n, dtime_save, gflops );
bli_obj_free( &alpha );
bli_obj_free( &beta );
bli_obj_free( &a );
bli_obj_free( &b );
bli_obj_free( &c );
bli_obj_free( &c_save );
}
bli_finalize();
return 0;
}
void runBenchmark(enum BenchmarkType benchmarkType, int32_t size, struct BenchmarkResult* out) {
const int32_t experiments = 10;
struct BenchmarkResult result = {
.time = __builtin_nan(""),
.flops = __builtin_nan(""),
.throughput = __builtin_nan("")
};
switch (benchmarkType) {
case BenchmarkTypeNaiveDGEMM:
{
double* a = malloc(size * size * sizeof(double));
double* b = malloc(size * size * sizeof(double));
double* c = malloc(size * size * sizeof(double));
for (int32_t i = 0; i < size * size; i++) {
a[i] = ((double) rand()) / ((double) RAND_MAX);
b[i] = ((double) rand()) / ((double) RAND_MAX);
}
memset(c, 0, size * size * sizeof(double));
for (int32_t experiment = 0; experiment < experiments; ++experiment) {
const double timeStart = time_sec();
dgemm_naive(size, a, b, c);
result.time = fmin(result.time, time_sec() - timeStart);
}
result.flops = 2.0 * size * size * size / result.time;
free(a);
free(b);
free(c);
break;
}
case BenchmarkTypeBlockedDGEMM:
{
double* a = malloc(size * size * sizeof(double));
double* b = malloc(size * size * sizeof(double));
double* c = malloc(size * size * sizeof(double));
for (int32_t i = 0; i < size * size; i++) {
a[i] = ((double) rand()) / ((double) RAND_MAX);
b[i] = ((double) rand()) / ((double) RAND_MAX);
}
memset(c, 0, size * size * sizeof(double));
for (int32_t experiment = 0; experiment < experiments; ++experiment) {
const double timeStart = time_sec();
dgemm_blocked(size, a, b, c);
result.time = fmin(result.time, time_sec() - timeStart);
}
result.flops = 2.0 * size * size * size / result.time;
free(a);
free(b);
free(c);
break;
}
case BenchmarkTypeBlisDGEMM:
{
obj_t alpha, beta;
bli_obj_scalar_init_detached(BLIS_DOUBLE, &alpha);
bli_obj_scalar_init_detached(BLIS_DOUBLE, &beta);
bli_setsc( 1.0, 0.0, &alpha);
bli_setsc( 0.0, 0.0, &beta);
obj_t a, b, c;
bli_obj_create(BLIS_DOUBLE, size, size, 0, 0, &a);
bli_obj_create(BLIS_DOUBLE, size, size, 0, 0, &b);
bli_obj_create(BLIS_DOUBLE, size, size, 0, 0, &c);
bli_randm(&a);
bli_randm(&b);
bli_randm(&c);
for (int32_t i = 0; i < experiments; ++i) {
const double timeStart = time_sec();
bli_gemm(&alpha, &a, &b, &beta, &c);
result.time = fmin(result.time, time_sec() - timeStart);
}
result.flops = 2.0 * size * size * size / result.time;
bli_obj_free(&a);
bli_obj_free(&b);
bli_obj_free(&c);
break;
}
case BenchmarkTypePointerChasing:
{
struct xor_shift xor_shift = xor_shift_init(UINT32_C(1), __builtin_ctz(size));
uint32_t last_index = 1;
void** data = (void**) malloc(size * sizeof(void*));
data[0] = &data[1];
for (size_t i = 0; i < size; i++) {
const uint32_t index = xor_shift_next(&xor_shift);
data[last_index] = &data[index];
last_index = index;
}
const int32_t iterations = 16777216 / size;
for (int32_t experiment = 0; experiment < experiments; ++experiment) {
const double timeStart = time_sec();
for (int32_t iteration = 0; iteration < iterations; iteration++) {
chase_pointers(data);
}
result.time = fmin(result.time, (time_sec() - timeStart) / ((double) iterations) / ((double) size));
}
free((void*) data);
break;
}
default:
__builtin_unreachable();
}
*out = result;
}
void libblis_test_gemm_experiment( test_params_t* params,
test_op_t* op,
iface_t iface,
num_t datatype,
char* pc_str,
char* sc_str,
unsigned int p_cur,
double* perf,
double* resid )
{
unsigned int n_repeats = params->n_repeats;
unsigned int i;
double time_min = 1e9;
double time;
dim_t m, n, k;
trans_t transa;
trans_t transb;
obj_t kappa;
obj_t alpha, a, b, beta, c;
obj_t c_save;
// Map the dimension specifier to actual dimensions.
m = libblis_test_get_dim_from_prob_size( op->dim_spec[0], p_cur );
n = libblis_test_get_dim_from_prob_size( op->dim_spec[1], p_cur );
k = libblis_test_get_dim_from_prob_size( op->dim_spec[2], p_cur );
// Map parameter characters to BLIS constants.
bli_param_map_char_to_blis_trans( pc_str[0], &transa );
bli_param_map_char_to_blis_trans( pc_str[1], &transb );
// Create test scalars.
bli_obj_scalar_init_detached( datatype, &kappa );
bli_obj_scalar_init_detached( datatype, &alpha );
bli_obj_scalar_init_detached( datatype, &beta );
// Create test operands (vectors and/or matrices).
libblis_test_mobj_create( params, datatype, transa,
sc_str[0], m, k, &a );
libblis_test_mobj_create( params, datatype, transb,
sc_str[1], k, n, &b );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[2], m, n, &c );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[2], m, n, &c_save );
// Set alpha and beta.
if ( bli_obj_is_real( c ) )
{
bli_setsc( 1.2, 0.0, &alpha );
bli_setsc( -1.0, 0.0, &beta );
}
else
{
bli_setsc( 1.2, 0.8, &alpha );
bli_setsc( -1.0, 1.0, &beta );
}
// Randomize A, B, and C, and save C.
bli_randm( &a );
bli_randm( &b );
bli_randm( &c );
bli_copym( &c, &c_save );
// Normalize by k.
bli_setsc( 1.0/( double )k, 0.0, &kappa );
bli_scalm( &kappa, &a );
bli_scalm( &kappa, &b );
// Apply the parameters.
bli_obj_set_conjtrans( transa, a );
bli_obj_set_conjtrans( transb, b );
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
bli_copym( &c_save, &c );
time = bli_clock();
libblis_test_gemm_impl( iface, &alpha, &a, &b, &beta, &c );
time_min = bli_clock_min_diff( time_min, time );
}
// Estimate the performance of the best experiment repeat.
*perf = ( 2.0 * m * n * k ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( c ) ) *perf *= 4.0;
// Perform checks.
libblis_test_gemm_check( &alpha, &a, &b, &beta, &c, &c_save, resid );
// Zero out performance and residual if output matrix is empty.
libblis_test_check_empty_problem( &c, perf, resid );
// Free the test objects.
bli_obj_free( &a );
bli_obj_free( &b );
bli_obj_free( &c );
bli_obj_free( &c_save );
}
void libblis_test_dotxf_experiment( test_params_t* params,
test_op_t* op,
iface_t iface,
num_t datatype,
char* pc_str,
char* sc_str,
unsigned int p_cur,
double* perf,
double* resid )
{
unsigned int n_repeats = params->n_repeats;
unsigned int i;
double time_min = 1e9;
double time;
dim_t m, b_n;
conj_t conjat, conjx;
obj_t alpha, a, x, beta, y;
obj_t y_save;
// Map the dimension specifier to an actual dimension.
m = libblis_test_get_dim_from_prob_size( op->dim_spec[0], p_cur );
// Query the operation's fusing factor for the current datatype.
b_n = bli_dotxf_fusefac( datatype );
// Store the fusing factor so that the driver can retrieve the value
// later when printing results.
op->dim_aux[0] = b_n;
// Map parameter characters to BLIS constants.
bli_param_map_char_to_blis_conj( pc_str[0], &conjat );
bli_param_map_char_to_blis_conj( pc_str[1], &conjx );
// Create test scalars.
bli_obj_scalar_init_detached( datatype, &alpha );
bli_obj_scalar_init_detached( datatype, &beta );
// Create test operands (vectors and/or matrices).
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[0], m, b_n, &a );
libblis_test_vobj_create( params, datatype, sc_str[1], m, &x );
libblis_test_vobj_create( params, datatype, sc_str[2], b_n, &y );
libblis_test_vobj_create( params, datatype, sc_str[2], b_n, &y_save );
// Set alpha.
if ( bli_obj_is_real( y ) )
{
bli_setsc( 1.2, 0.0, &alpha );
bli_setsc( -1.0, 0.0, &beta );
}
else
{
bli_setsc( 1.2, 0.1, &alpha );
bli_setsc( -1.0, -0.1, &beta );
}
// Randomize A, x, and y, and save y.
bli_randm( &a );
bli_randv( &x );
bli_randv( &y );
bli_copyv( &y, &y_save );
// Apply the parameters.
bli_obj_set_conj( conjat, a );
bli_obj_set_conj( conjx, x );
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
bli_copyv( &y_save, &y );
time = bli_clock();
libblis_test_dotxf_impl( iface, &alpha, &a, &x, &beta, &y );
time_min = bli_clock_min_diff( time_min, time );
}
// Estimate the performance of the best experiment repeat.
*perf = ( 2.0 * m * b_n ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( y ) ) *perf *= 4.0;
// Perform checks.
libblis_test_dotxf_check( &alpha, &a, &x, &beta, &y, &y_save, resid );
// Zero out performance and residual if output vector is empty.
libblis_test_check_empty_problem( &y, perf, resid );
// Free the test objects.
bli_obj_free( &a );
bli_obj_free( &x );
bli_obj_free( &y );
bli_obj_free( &y_save );
}
void libblis_test_trmm3_experiment( test_params_t* params,
test_op_t* op,
mt_impl_t impl,
num_t datatype,
char* pc_str,
char* sc_str,
unsigned int p_cur,
double* perf,
double* resid )
{
unsigned int n_repeats = params->n_repeats;
unsigned int i;
double time_min = 1e9;
double time;
dim_t m, n;
dim_t mn_side;
side_t side;
uplo_t uploa;
trans_t transa;
diag_t diaga;
trans_t transb;
obj_t kappa;
obj_t alpha, a, b, beta, c;
obj_t c_save;
// Map the dimension specifier to actual dimensions.
m = libblis_test_get_dim_from_prob_size( op->dim_spec[0], p_cur );
n = libblis_test_get_dim_from_prob_size( op->dim_spec[1], p_cur );
// Map parameter characters to BLIS constants.
bli_param_map_char_to_blis_side( pc_str[0], &side );
bli_param_map_char_to_blis_uplo( pc_str[1], &uploa );
bli_param_map_char_to_blis_trans( pc_str[2], &transa );
bli_param_map_char_to_blis_diag( pc_str[3], &diaga );
bli_param_map_char_to_blis_trans( pc_str[4], &transb );
// Create test scalars.
bli_obj_scalar_init_detached( datatype, &kappa );
bli_obj_scalar_init_detached( datatype, &alpha );
bli_obj_scalar_init_detached( datatype, &beta );
// Create test operands (vectors and/or matrices).
bli_set_dim_with_side( side, m, n, mn_side );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[0], mn_side, mn_side, &a );
libblis_test_mobj_create( params, datatype, transb,
sc_str[1], m, n, &b );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[2], m, n, &c );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[2], m, n, &c_save );
// Set alpha and beta.
if ( bli_obj_is_real( c ) )
{
bli_setsc( 0.8, 0.0, &alpha );
bli_setsc( -1.0, 0.0, &beta );
}
else
{
bli_setsc( 0.8, 0.6, &alpha );
bli_setsc( -1.0, 0.5, &beta );
}
// Set the structure and uplo properties of A.
bli_obj_set_struc( BLIS_TRIANGULAR, a );
bli_obj_set_uplo( uploa, a );
// Randomize A, make it densely triangular.
bli_randm( &a );
bli_mktrim( &a );
// Randomize B and C, and save C.
bli_randm( &b );
bli_randm( &c );
bli_copym( &c, &c_save );
// Normalize by m.
bli_setsc( 1.0/( double )m, 0.0, &kappa );
bli_scalm( &kappa, &b );
// Apply the remaining parameters.
bli_obj_set_conjtrans( transa, a );
bli_obj_set_diag( diaga, a );
bli_obj_set_conjtrans( transb, b );
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
bli_copym( &c_save, &c );
time = bli_clock();
libblis_test_trmm3_impl( impl, side, &alpha, &a, &b, &beta, &c );
time_min = bli_clock_min_diff( time_min, time );
}
// Estimate the performance of the best experiment repeat.
*perf = ( 1.0 * mn_side * m * n ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( c ) ) *perf *= 4.0;
// Perform checks.
libblis_test_trmm3_check( side, &alpha, &a, &b, &beta, &c, &c_save, resid );
// Zero out performance and residual if output matrix is empty.
libblis_test_check_empty_problem( &c, perf, resid );
// Free the test objects.
bli_obj_free( &a );
bli_obj_free( &b );
bli_obj_free( &c );
bli_obj_free( &c_save );
}
void libblis_test_trsv_experiment( test_params_t* params,
test_op_t* op,
mt_impl_t impl,
num_t datatype,
char* pc_str,
char* sc_str,
unsigned int p_cur,
double* perf,
double* resid )
{
unsigned int n_repeats = params->n_repeats;
unsigned int i;
double time_min = 1e9;
double time;
dim_t m;
uplo_t uploa;
trans_t transa;
diag_t diaga;
obj_t kappa;
obj_t alpha, a, x;
obj_t x_save;
// Map the dimension specifier to an actual dimension.
m = libblis_test_get_dim_from_prob_size( op->dim_spec[0], p_cur );
// Map parameter characters to BLIS constants.
bli_param_map_char_to_blis_uplo( pc_str[0], &uploa );
bli_param_map_char_to_blis_trans( pc_str[1], &transa );
bli_param_map_char_to_blis_diag( pc_str[2], &diaga );
// Create test scalars.
bli_obj_scalar_init_detached( datatype, &alpha );
bli_obj_scalar_init_detached( datatype, &kappa );
// Create test operands (vectors and/or matrices).
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[0], m, m, &a );
libblis_test_vobj_create( params, datatype,
sc_str[1], m, &x );
libblis_test_vobj_create( params, datatype,
sc_str[1], m, &x_save );
// Set alpha.
if ( bli_obj_is_real( x ) )
bli_setsc( 2.0, 0.0, &alpha );
else
bli_setsc( 2.0, -1.0, &alpha );
// Set the structure and uplo properties of A.
bli_obj_set_struc( BLIS_TRIANGULAR, a );
bli_obj_set_uplo( uploa, a );
// Randomize A, make it densely triangular.
bli_randm( &a );
bli_mktrim( &a );
// Randomize x and save.
bli_randv( &x );
bli_copyv( &x, &x_save );
// Normalize vectors by m.
bli_setsc( 1.0/( double )m, 0.0, &kappa );
bli_scalv( &kappa, &x );
bli_scalv( &kappa, &x_save );
// Apply the remaining parameters.
bli_obj_set_conjtrans( transa, a );
bli_obj_set_diag( diaga, a );
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
bli_copym( &x_save, &x );
time = bli_clock();
libblis_test_trsv_impl( impl, &alpha, &a, &x );
time_min = bli_clock_min_diff( time_min, time );
}
// Estimate the performance of the best experiment repeat.
*perf = ( 1.0 * m * m ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( x ) ) *perf *= 4.0;
// Perform checks.
libblis_test_trsv_check( &alpha, &a, &x, &x_save, resid );
// Zero out performance and residual if output vector is empty.
libblis_test_check_empty_problem( &x, perf, resid );
// Free the test objects.
bli_obj_free( &a );
bli_obj_free( &x );
bli_obj_free( &x_save );
}
void libblis_test_her_experiment( test_params_t* params,
test_op_t* op,
iface_t iface,
num_t datatype,
char* pc_str,
char* sc_str,
unsigned int p_cur,
double* perf,
double* resid )
{
unsigned int n_repeats = params->n_repeats;
unsigned int i;
double time_min = 1e9;
double time;
dim_t m;
uplo_t uploa;
conj_t conjx;
obj_t alpha, x, a;
obj_t a_save;
// Map the dimension specifier to an actual dimension.
m = libblis_test_get_dim_from_prob_size( op->dim_spec[0], p_cur );
// Map parameter characters to BLIS constants.
bli_param_map_char_to_blis_uplo( pc_str[0], &uploa );
bli_param_map_char_to_blis_conj( pc_str[1], &conjx );
// Create test scalars.
bli_obj_scalar_init_detached( datatype, &alpha );
// Create test operands (vectors and/or matrices).
libblis_test_vobj_create( params, datatype,
sc_str[0], m, &x );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[1], m, m, &a );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[1], m, m, &a_save );
// Set alpha.
//bli_copysc( &BLIS_MINUS_ONE, &alpha );
bli_setsc( -1.0, 0.0, &alpha );
// Randomize x.
bli_randv( &x );
// Set the structure and uplo properties of A.
bli_obj_set_struc( BLIS_HERMITIAN, a );
bli_obj_set_uplo( uploa, a );
// Randomize A, make it densely Hermitian, and zero the unstored triangle
// to ensure the implementation is reads only from the stored region.
bli_randm( &a );
bli_mkherm( &a );
bli_mktrim( &a );
// Save A and set its structure and uplo properties.
bli_obj_set_struc( BLIS_HERMITIAN, a_save );
bli_obj_set_uplo( uploa, a_save );
bli_copym( &a, &a_save );
// Apply the remaining parameters.
bli_obj_set_conj( conjx, x );
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
bli_copym( &a_save, &a );
time = bli_clock();
libblis_test_her_impl( iface, &alpha, &x, &a );
time_min = bli_clock_min_diff( time_min, time );
}
// Estimate the performance of the best experiment repeat.
*perf = ( 1.0 * m * m ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( a ) ) *perf *= 4.0;
// Perform checks.
libblis_test_her_check( &alpha, &x, &a, &a_save, resid );
// Zero out performance and residual if output matrix is empty.
libblis_test_check_empty_problem( &a, perf, resid );
// Free the test objects.
bli_obj_free( &x );
bli_obj_free( &a );
bli_obj_free( &a_save );
}
int main( int argc, char** argv )
{
obj_t a, x;
obj_t x_save;
obj_t alpha;
dim_t m;
dim_t p;
dim_t p_begin, p_end, p_inc;
int m_input;
num_t dt_a, dt_x;
num_t dt_alpha;
int r, n_repeats;
uplo_t uplo;
double dtime;
double dtime_save;
double gflops;
//bli_init();
n_repeats = 3;
#ifndef PRINT
p_begin = 40;
p_end = 2000;
p_inc = 40;
m_input = -1;
#else
p_begin = 16;
p_end = 16;
p_inc = 1;
m_input = 15;
n_input = 15;
#endif
dt_alpha = dt_a = dt_x = BLIS_DOUBLE;
uplo = BLIS_LOWER;
// Begin with initializing the last entry to zero so that
// matlab allocates space for the entire array once up-front.
for ( p = p_begin; p + p_inc <= p_end; p += p_inc ) ;
#ifdef BLIS
printf( "data_trsv_blis" );
#else
printf( "data_trv_%s", BLAS );
#endif
printf( "( %2lu, 1:2 ) = [ %4lu %7.2f ];\n",
( unsigned long )(p - p_begin + 1)/p_inc + 1,
( unsigned long )0, 0.0 );
for ( p = p_begin; p <= p_end; p += p_inc )
{
if ( m_input < 0 ) m = p * ( dim_t )abs(m_input);
else m = ( dim_t ) m_input;
bli_obj_create( dt_alpha, 1, 1, 0, 0, &alpha );
bli_obj_create( dt_a, m, m, 0, 0, &a );
bli_obj_create( dt_x, m, 1, 0, 0, &x );
bli_obj_create( dt_x, m, 1, 0, 0, &x_save );
bli_randm( &a );
bli_randm( &x );
bli_obj_set_struc( BLIS_TRIANGULAR, &a );
bli_obj_set_uplo( uplo, &a );
bli_obj_set_onlytrans( BLIS_NO_TRANSPOSE, &a );
bli_obj_set_diag( BLIS_NONUNIT_DIAG, &a );
// Randomize A and zero the unstored triangle to ensure the
// implementation reads only from the stored region.
bli_randm( &a );
bli_mktrim( &a );
// Load the diagonal of A to make it more likely to be invertible.
bli_shiftd( &BLIS_TWO, &a );
bli_setsc( (1.0/1.0), 0.0, &alpha );
bli_copym( &x, &x_save );
dtime_save = DBL_MAX;
for ( r = 0; r < n_repeats; ++r )
{
bli_copym( &x_save, &x );
dtime = bli_clock();
#ifdef PRINT
bli_printm( "a", &a, "%4.1f", "" );
bli_printm( "x", &x, "%4.1f", "" );
#endif
#ifdef BLIS
bli_trsv( &BLIS_ONE,
&a,
&x );
#else
f77_char uploa = 'L';
f77_char transa = 'N';
f77_char diaga = 'N';
f77_int mm = bli_obj_length( &a );
f77_int lda = bli_obj_col_stride( &a );
f77_int incx = bli_obj_vector_inc( &x );
double* ap = bli_obj_buffer( &a );
double* xp = bli_obj_buffer( &x );
dtrsv_( &uploa,
&transa,
&diaga,
&mm,
ap, &lda,
xp, &incx );
#endif
#ifdef PRINT
bli_printm( "x after", &x, "%4.1f", "" );
exit(1);
#endif
dtime_save = bli_clock_min_diff( dtime_save, dtime );
}
gflops = ( 1.0 * m * m ) / ( dtime_save * 1.0e9 );
#ifdef BLIS
printf( "data_trsv_blis" );
#else
printf( "data_trsv_%s", BLAS );
#endif
printf( "( %2lu, 1:2 ) = [ %4lu %7.2f ];\n",
( unsigned long )(p - p_begin + 1)/p_inc + 1,
( unsigned long )m, gflops );
bli_obj_free( &alpha );
bli_obj_free( &a );
bli_obj_free( &x );
bli_obj_free( &x_save );
}
//bli_finalize();
return 0;
}
