ind_t bli_ind_oper_find_avail( opid_t oper, num_t dt )
{
ind_t im;
// If the datatype is real, return native execution.
if ( !bli_is_complex( dt ) ) return BLIS_NAT;
// If the operation is not level-3, return native execution.
if ( !bli_opid_is_level3( oper ) ) return BLIS_NAT;
// Iterate over all induced methods and search for the first one
// that is available (ie: both implemented and enabled) for the
// current operation and datatype.
for ( im = 0; im < BLIS_NUM_IND_METHODS; ++im )
{
void* func = bli_ind_oper_get_func( oper, im );
bool_t stat = bli_ind_oper_get_enable( oper, im, dt );
if ( func != NULL &&
stat == TRUE ) return im;
}
// This return statement should never execute since the native index
// should be found even if all induced methods are unavailable. We
// include it simply to avoid a compiler warning.
return BLIS_NAT;
}
void bli_ind_oper_enable_only( opid_t oper, ind_t method, num_t dt )
{
if ( !bli_is_complex( dt ) ) return;
if ( !bli_opid_is_level3( oper ) ) return;
bli_ind_oper_set_enable_all( oper, dt, FALSE );
bli_ind_oper_set_enable( oper, method, dt, TRUE );
}
num_t bli_ind_map_cdt_to_index( num_t dt )
{
// A non-complex datatype should never be passed in.
if ( !bli_is_complex( dt ) ) bli_abort();
// Map the complex datatype to a zero-based index.
if ( bli_is_scomplex( dt ) ) return 0;
else /* if ( bli_is_dcomplex( dt ) ) */ return 1;
}
void bli_ind_set_enable_dt( ind_t method, num_t dt, bool_t status )
{
opid_t iop;
if ( !bli_is_complex( dt ) ) return;
// Iterate over all level-3 operation ids.
for ( iop = 0; iop < BLIS_NUM_LEVEL3_OPS; ++iop )
{
bli_ind_oper_set_enable( iop, method, dt, status );
}
}
err_t bli_check_datatype_real_proj_of( num_t dt_c, num_t dt_r )
{
err_t e_val = BLIS_SUCCESS;
if ( ( dt_c == BLIS_CONSTANT && bli_is_complex( dt_r ) ) ||
( dt_c == BLIS_FLOAT && dt_r != BLIS_FLOAT ) ||
( dt_c == BLIS_DOUBLE && dt_r != BLIS_DOUBLE ) ||
( dt_c == BLIS_SCOMPLEX && dt_r != BLIS_FLOAT ) ||
( dt_c == BLIS_DCOMPLEX && dt_r != BLIS_DOUBLE ) )
e_val = BLIS_EXPECTED_REAL_PROJ_OF;
return e_val;
}
bool_t bli_ind_oper_is_avail( opid_t oper, ind_t method, num_t dt )
{
void* func;
bool_t stat;
// If the datatype is real, it is never available.
if ( !bli_is_complex( dt ) ) return FALSE;
func = bli_ind_oper_get_func( oper, method );
stat = bli_ind_oper_get_enable( oper, method, dt );
return ( func != NULL && stat == TRUE );
}
void bli_ind_oper_set_enable( opid_t oper, ind_t method, num_t dt, bool_t status )
{
num_t idt;
if ( !bli_is_complex( dt ) ) return;
if ( !bli_opid_is_level3( oper ) ) return;
// Disallow changing status of native execution.
if ( method == BLIS_NAT ) return;
idt = bli_ind_map_cdt_to_index( dt );
bli_ind_oper_st[ method ][ oper ][ idt ] = status;
}
void bli_ind_oper_set_enable_all( opid_t oper, num_t dt, bool_t status )
{
ind_t im;
if ( !bli_is_complex( dt ) ) return;
if ( !bli_opid_is_level3( oper ) ) return;
for ( im = 0; im < BLIS_NUM_IND_METHODS; ++im )
{
// Native execution should always stay enabled.
if ( im != BLIS_NAT )
bli_ind_oper_set_enable( oper, im, dt, status );
}
}
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;
int r, n_repeats;
uplo_t uploc;
trans_t transa;
f77_char f77_uploc;
f77_char f77_transa;
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;
k_input = -1;
#else
p_begin = 16;
p_end = 16;
p_inc = 1;
m_input = 3;
k_input = 1;
#endif
#if 1
//dt = BLIS_FLOAT;
dt = BLIS_DOUBLE;
#else
//dt = BLIS_SCOMPLEX;
dt = BLIS_DCOMPLEX;
#endif
uploc = BLIS_LOWER;
//uploc = BLIS_UPPER;
transa = BLIS_NO_TRANSPOSE;
bli_param_map_blis_to_netlib_uplo( uploc, &f77_uploc );
bli_param_map_blis_to_netlib_trans( transa, &f77_transa );
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 ( 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 );
if ( bli_does_trans( transa ) )
bli_obj_create( dt, k, m, 0, 0, &a );
else
bli_obj_create( dt, m, k, 0, 0, &a );
bli_obj_create( dt, m, m, 0, 0, &c );
bli_obj_create( dt, m, m, 0, 0, &c_save );
bli_randm( &a );
bli_randm( &c );
bli_obj_set_struc( BLIS_HERMITIAN, c );
bli_obj_set_uplo( uploc, c );
bli_obj_set_conjtrans( transa, 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
bli_printm( "a", &a, "%4.1f", "" );
bli_printm( "c", &c, "%4.1f", "" );
#endif
#ifdef BLIS
bli_herk( &alpha,
&a,
&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 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* betap = bli_obj_buffer( beta );
float* cp = bli_obj_buffer( c );
ssyrk_( &f77_uploc,
&f77_transa,
&mm,
&kk,
alphap,
ap, &lda,
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 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_( &f77_uploc,
&f77_transa,
&mm,
&kk,
alphap,
ap, &lda,
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 lda = bli_obj_col_stride( a );
f77_int ldc = bli_obj_col_stride( c );
float* alphap = bli_obj_buffer( alpha );
scomplex* ap = bli_obj_buffer( a );
float* betap = bli_obj_buffer( beta );
scomplex* cp = bli_obj_buffer( c );
cherk_( &f77_uploc,
&f77_transa,
&mm,
&kk,
alphap,
ap, &lda,
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 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_( &f77_uploc,
&f77_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 );
if ( bli_is_complex( dt ) ) gflops *= 4.0;
#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;
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 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 );
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_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 + 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_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_offset( 1, &a2 );
bli_setm( &BLIS_ZERO, &a2 );
bli_obj_inc_diag_offset( -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_dt( BLIS_DOUBLE, &ar ); ar.rs *= 2; ar.cs *= 2;
bli_obj_set_dt( 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 )
{
obj_t a, c;
obj_t c_save;
obj_t alpha;
dim_t m, n;
dim_t p;
dim_t p_begin, p_max, 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_max = P_MAX;
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_max; p += p_inc ) ;
printf( "data_%s_%ctrsm_%s", THR_STR, dt_ch, STR );
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_max; 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_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, &c );
//bli_obj_create( dt, m, n, n, 1, &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 );
// Load the diagonal of A to make it more likely to be invertible.
bli_shiftd( &BLIS_TWO, &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_trsm( 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 = ( float* )bli_obj_buffer( &alpha );
float* ap = ( float* )bli_obj_buffer( &a );
float* cp = ( float* )bli_obj_buffer( &c );
strsm_( &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 = ( double* )bli_obj_buffer( &alpha );
double* ap = ( double* )bli_obj_buffer( &a );
double* cp = ( double* )bli_obj_buffer( &c );
dtrsm_( &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 );
#ifdef EIGEN
float* alphap = ( float* )bli_obj_buffer( &alpha );
float* ap = ( float* )bli_obj_buffer( &a );
float* cp = ( float* )bli_obj_buffer( &c );
#else
scomplex* alphap = ( scomplex* )bli_obj_buffer( &alpha );
scomplex* ap = ( scomplex* )bli_obj_buffer( &a );
scomplex* cp = ( scomplex* )bli_obj_buffer( &c );
#endif
ctrsm_( &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 );
#ifdef EIGEN
double* alphap = ( double* )bli_obj_buffer( &alpha );
double* ap = ( double* )bli_obj_buffer( &a );
double* cp = ( double* )bli_obj_buffer( &c );
#else
dcomplex* alphap = ( dcomplex* )bli_obj_buffer( &alpha );
dcomplex* ap = ( dcomplex* )bli_obj_buffer( &a );
dcomplex* cp = ( dcomplex* )bli_obj_buffer( &c );
#endif
ztrsm_( &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;
printf( "data_%s_%ctrsm_%s", THR_STR, dt_ch, STR );
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;
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, 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 bli_ind_disable_dt( ind_t method, num_t dt )
{
if ( !bli_is_complex( dt ) ) return;
bli_ind_set_enable_dt( method, dt, FALSE );
}
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;
}
