void libblis_test_gemm_impl
(
iface_t iface,
obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c
)
{
switch ( iface )
{
case BLIS_TEST_SEQ_FRONT_END:
#if 0
//bli_printm( "alpha", alpha, "%5.2f", "" );
//bli_printm( "beta", beta, "%5.2f", "" );
bli_printm( "a", a, "%5.2f", "" );
bli_printm( "b", b, "%5.2f", "" );
bli_printm( "c", c, "%5.2f", "" );
#endif
//if ( bli_obj_length( b ) == 16 &&
// bli_obj_stor3_from_strides( c, a, b ) == BLIS_CRR )
//bli_printm( "c before", c, "%6.3f", "" );
bli_gemm( alpha, a, b, beta, c );
#if 0
if ( bli_obj_length( c ) == 12 &&
bli_obj_stor3_from_strides( c, a, b ) == BLIS_RRR )
bli_printm( "c after", c, "%6.3f", "" );
#endif
break;
default:
libblis_test_printf_error( "Invalid interface type.\n" );
}
}
void libblis_test_gemm_impl( iface_t iface,
obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c )
{
switch ( iface )
{
case BLIS_TEST_SEQ_FRONT_END:
bli_gemm( alpha, a, b, beta, c );
//bli_gemm4m( alpha, a, b, beta, c );
//bli_gemm3m( alpha, a, b, beta, c );
break;
default:
libblis_test_printf_error( "Invalid interface type.\n" );
}
}
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;
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 );
bli_init();
n_repeats = 3;
#ifndef PRINT
p_begin = 16;
p_end = 2048;
p_inc = 16;
m_input = 10240;
n_input = 10240;
k_input = -1;
#else
p_begin = 24;
p_end = 24;
p_inc = 1;
m_input = -1;
k_input = -1;
n_input = -1;
#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 + 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;
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( (1.0/1.0), 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
char transa = 'N';
char transb = 'N';
int mm = bli_obj_length( c );
int kk = bli_obj_width_after_trans( a );
int nn = bli_obj_width( c );
int lda = bli_obj_col_stride( a );
int ldb = bli_obj_col_stride( b );
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 );
//if(world_rank == 0){
#ifdef BLIS
printf( "data_gemm_blis" );
#else
printf( "data_gemm_%s", BLAS );
#endif
printf( "( %2ld, 1:5 ) = [ %4lu %4lu %4lu %10.3e %6.3f %d ];\n",
(p - p_begin + 1)/p_inc + 1, m, k, n, dtime_save, gflops, world_rank );
//}
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();
MPI_Finalize();
return 0;
}
void libblis_test_gemm_md_check
(
test_params_t* params,
obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
obj_t* c_orig,
double* resid
)
{
num_t dt_real = bli_obj_dt_proj_to_real( c );
num_t dt_comp = bli_obj_dt_proj_to_complex( c );
num_t dt;
dim_t m = bli_obj_length( c );
dim_t n = bli_obj_width( c );
dim_t k = bli_obj_width_after_trans( a );
obj_t norm;
obj_t t, v, w, z;
double junk;
// Compute our reference checksum in the real domain if all operands
// are real, and in the complex domain otherwise. Also implicit in this
// is that we use the storage precision of C to determine the precision
// in which we perform the reference checksum.
if ( bli_obj_is_real( a ) &&
bli_obj_is_real( b ) &&
bli_obj_is_real( c ) ) dt = dt_real;
else dt = dt_comp;
// This function works in a manner similar to that of the function
// libblis_test_gemm_check(), except that we project a, b, and c into
// the complex domain (regardless of their storage datatype), and then
// proceed with the checking accordingly.
obj_t a2, b2, c2, c0;
bli_obj_scalar_init_detached( dt_real, &norm );
bli_obj_create( dt, n, 1, 0, 0, &t );
bli_obj_create( dt, m, 1, 0, 0, &v );
bli_obj_create( dt, k, 1, 0, 0, &w );
bli_obj_create( dt, m, 1, 0, 0, &z );
libblis_test_vobj_randomize( params, TRUE, &t );
// We need to zero out the imaginary part of t in order for our
// checks to work in all cases. Otherwise, the imaginary parts
// could affect intermediate products, depending on the order that
// they are executed.
bli_setiv( &BLIS_ZERO, &t );
// Create complex equivalents of a, b, c_orig, and c.
bli_obj_create( dt, m, k, 0, 0, &a2 );
bli_obj_create( dt, k, n, 0, 0, &b2 );
bli_obj_create( dt, m, n, 0, 0, &c2 );
bli_obj_create( dt, m, n, 0, 0, &c0 );
// Cast a, b, c_orig, and c into the datatype of our temporary objects.
bli_castm( a, &a2 );
bli_castm( b, &b2 );
bli_castm( c_orig, &c2 );
bli_castm( c, &c0 );
bli_gemv( &BLIS_ONE, &c0, &t, &BLIS_ZERO, &v );
#if 0
if ( bli_obj_is_scomplex( c ) &&
bli_obj_is_float( a ) &&
bli_obj_is_float( b ) )
{
bli_printm( "test_gemm.c: a", a, "%7.3f", "" );
bli_printm( "test_gemm.c: b", b, "%7.3f", "" );
bli_printm( "test_gemm.c: c orig", c_orig, "%7.3f", "" );
bli_printm( "test_gemm.c: c computed", c, "%7.3f", "" );
}
#endif
#if 0
bli_gemm( alpha, &a2, &b2, beta, &c2 );
bli_gemv( &BLIS_ONE, &c2, &t, &BLIS_ZERO, &z );
if ( bli_obj_is_real( c ) ) bli_setiv( &BLIS_ZERO, &z );
#else
bli_gemv( &BLIS_ONE, &b2, &t, &BLIS_ZERO, &w );
bli_gemv( alpha, &a2, &w, &BLIS_ZERO, &z );
bli_gemv( beta, &c2, &t, &BLIS_ONE, &z );
if ( bli_obj_is_real( c ) ) bli_setiv( &BLIS_ZERO, &z );
#endif
bli_subv( &z, &v );
bli_normfv( &v, &norm );
bli_getsc( &norm, resid, &junk );
bli_obj_free( &t );
bli_obj_free( &v );
bli_obj_free( &w );
bli_obj_free( &z );
bli_obj_free( &a2 );
bli_obj_free( &b2 );
bli_obj_free( &c2 );
bli_obj_free( &c0 );
}
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;
}
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, 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;
}
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;
}
