void libblis_test_normfm_check
(
test_params_t* params,
obj_t* beta,
obj_t* x,
obj_t* norm,
double* resid
)
{
num_t dt_real = bli_obj_datatype_proj_to_real( *x );
dim_t m = bli_obj_length( *x );
dim_t n = bli_obj_width( *x );
obj_t m_r, n_r, temp_r;
double junk;
//
// Pre-conditions:
// - x is set to beta.
// Note:
// - beta should have a non-zero imaginary component in the complex
// cases in order to more fully exercise the implementation.
//
// Under these conditions, we assume that the implementation for
//
// norm := normf( x )
//
// is functioning correctly if
//
// norm = sqrt( absqsc( beta ) * m * n )
//
// where m and n are the dimensions of x.
//
bli_obj_scalar_init_detached( dt_real, &temp_r );
bli_obj_scalar_init_detached( dt_real, &m_r );
bli_obj_scalar_init_detached( dt_real, &n_r );
bli_setsc( ( double )m, 0.0, &m_r );
bli_setsc( ( double )n, 0.0, &n_r );
bli_absqsc( beta, &temp_r );
bli_mulsc( &m_r, &temp_r );
bli_mulsc( &n_r, &temp_r );
bli_sqrtsc( &temp_r, &temp_r );
bli_subsc( &temp_r, norm );
bli_getsc( norm, resid, &junk );
}
void libblis_test_addv_check( obj_t* alpha,
obj_t* beta,
obj_t* x,
obj_t* y,
double* resid )
{
num_t dt = bli_obj_datatype( *x );
num_t dt_real = bli_obj_datatype_proj_to_real( *x );
dim_t m = bli_obj_vector_dim( *x );
conj_t conjx = bli_obj_conj_status( *x );
obj_t aplusb;
obj_t alpha_conj;
obj_t norm_r, m_r, temp_r;
double junk;
//
// Pre-conditions:
// - x is set to alpha.
// - y_orig is set to beta.
// Note:
// - alpha and beta should have non-zero imaginary components in the
// complex cases in order to more fully exercise the implementation.
//
// Under these conditions, we assume that the implementation for
//
// y := y_orig + conjx(x)
//
// is functioning correctly if
//
// fnormv(y) - sqrt( absqsc( beta + conjx(alpha) ) * m )
//
// is negligible.
//
bli_obj_scalar_init_detached( dt, &aplusb );
bli_obj_scalar_init_detached( dt_real, &temp_r );
bli_obj_scalar_init_detached( dt_real, &norm_r );
bli_obj_scalar_init_detached( dt_real, &m_r );
bli_obj_scalar_init_detached_copy_of( dt, conjx, alpha, &alpha_conj );
bli_fnormv( y, &norm_r );
bli_copysc( beta, &aplusb );
bli_addsc( &alpha_conj, &aplusb );
bli_setsc( ( double )m, 0.0, &m_r );
bli_absqsc( &aplusb, &temp_r );
bli_mulsc( &m_r, &temp_r );
bli_sqrtsc( &temp_r, &temp_r );
bli_subsc( &temp_r, &norm_r );
bli_getsc( &norm_r, resid, &junk );
}
void libblis_test_scalv_check
(
test_params_t* params,
obj_t* beta,
obj_t* y,
obj_t* y_orig,
double* resid
)
{
num_t dt = bli_obj_dt( y );
num_t dt_real = bli_obj_dt_proj_to_real( y );
dim_t m = bli_obj_vector_dim( y );
obj_t norm_y_r;
obj_t nbeta;
obj_t y2;
double junk;
//
// Pre-conditions:
// - y_orig is randomized.
// Note:
// - beta should have a non-zero imaginary component in the complex
// cases in order to more fully exercise the implementation.
//
// Under these conditions, we assume that the implementation for
//
// y := conjbeta(beta) * y_orig
//
// is functioning correctly if
//
// normf( y + -conjbeta(beta) * y_orig )
//
// is negligible.
//
bli_obj_create( dt, m, 1, 0, 0, &y2 );
bli_copyv( y_orig, &y2 );
bli_obj_scalar_init_detached( dt, &nbeta );
bli_obj_scalar_init_detached( dt_real, &norm_y_r );
bli_copysc( beta, &nbeta );
bli_mulsc( &BLIS_MINUS_ONE, &nbeta );
bli_scalv( &nbeta, &y2 );
bli_addv( &y2, y );
bli_normfv( y, &norm_y_r );
bli_getsc( &norm_y_r, resid, &junk );
bli_obj_free( &y2 );
}
void libblis_test_fnormv_check( obj_t* beta,
obj_t* x,
obj_t* norm,
double* resid )
{
num_t dt_real = bli_obj_datatype_proj_to_real( *x );
dim_t m = bli_obj_vector_dim( *x );
obj_t m_r, temp_r;
double junk;
//
// Pre-conditions:
// - x is set to beta.
// Note:
// - beta should have a non-zero imaginary component in the complex
// cases in order to more fully exercise the implementation.
//
// Under these conditions, we assume that the implementation for
//
// norm := fnorm( x )
//
// is functioning correctly if
//
// norm = sqrt( absqsc( beta ) * m )
//
// where m is the length of x.
//
bli_obj_init_scalar( dt_real, &temp_r );
bli_obj_init_scalar( dt_real, &m_r );
bli_setsc( ( double )m, 0.0, &m_r );
bli_absqsc( beta, &temp_r );
bli_mulsc( &m_r, &temp_r );
bli_sqrtsc( &temp_r, &temp_r );
bli_subsc( &temp_r, norm );
bli_getsc( norm, resid, &junk );
}
void bli_obj_scalar_apply_scalar( obj_t* alpha,
obj_t* a )
{
obj_t alpha_cast;
obj_t scalar_a;
// Make a copy-cast of alpha of the same datatype as A. This step
// gives us the opportunity to typecast alpha.
bli_obj_scalar_init_detached_copy_of( bli_obj_datatype( *a ),
BLIS_NO_CONJUGATE,
alpha,
&alpha_cast );
// Detach the scalar from A.
bli_obj_scalar_detach( a, &scalar_a );
// Scale the detached scalar by alpha.
bli_mulsc( &alpha_cast, &scalar_a );
// Copy the internal scalar in scalar_a to A.
bli_obj_copy_internal_scalar( scalar_a, *a );
}
void libblis_test_dotxaxpyf_check
(
test_params_t* params,
obj_t* alpha,
obj_t* at,
obj_t* a,
obj_t* w,
obj_t* x,
obj_t* beta,
obj_t* y,
obj_t* z,
obj_t* y_orig,
obj_t* z_orig,
double* resid
)
{
num_t dt = bli_obj_datatype( *y );
num_t dt_real = bli_obj_datatype_proj_to_real( *y );
dim_t m = bli_obj_vector_dim( *z );
dim_t b_n = bli_obj_vector_dim( *y );
dim_t i;
obj_t a1, chi1, psi1, v, q;
obj_t alpha_chi1;
obj_t norm;
double resid1, resid2;
double junk;
//
// Pre-conditions:
// - a is randomized.
// - w is randomized.
// - x is randomized.
// - y is randomized.
// - z is randomized.
// - at is an alias to a.
// Note:
// - alpha and beta should have a non-zero imaginary component in the
// complex cases in order to more fully exercise the implementation.
//
// Under these conditions, we assume that the implementation for
//
// y := beta * y_orig + alpha * conjat(A^T) * conjw(w)
// z := z_orig + alpha * conja(A) * conjx(x)
//
// is functioning correctly if
//
// normf( y - v )
//
// and
//
// normf( z - q )
//
// are negligible, where v and q contain y and z as computed by repeated
// calls to dotxv and axpyv, respectively.
//
bli_obj_scalar_init_detached( dt_real, &norm );
bli_obj_scalar_init_detached( dt, &alpha_chi1 );
bli_obj_create( dt, b_n, 1, 0, 0, &v );
bli_obj_create( dt, m, 1, 0, 0, &q );
bli_copyv( y_orig, &v );
bli_copyv( z_orig, &q );
// v := beta * v + alpha * conjat(at) * conjw(w)
for ( i = 0; i < b_n; ++i )
{
bli_acquire_mpart_l2r( BLIS_SUBPART1, i, 1, at, &a1 );
bli_acquire_vpart_f2b( BLIS_SUBPART1, i, 1, &v, &psi1 );
bli_dotxv( alpha, &a1, w, beta, &psi1 );
}
// q := q + alpha * conja(a) * conjx(x)
for ( i = 0; i < b_n; ++i )
{
bli_acquire_mpart_l2r( BLIS_SUBPART1, i, 1, a, &a1 );
bli_acquire_vpart_f2b( BLIS_SUBPART1, i, 1, x, &chi1 );
bli_copysc( &chi1, &alpha_chi1 );
bli_mulsc( alpha, &alpha_chi1 );
bli_axpyv( &alpha_chi1, &a1, &q );
}
bli_subv( y, &v );
bli_normfv( &v, &norm );
bli_getsc( &norm, &resid1, &junk );
bli_subv( z, &q );
bli_normfv( &q, &norm );
bli_getsc( &norm, &resid2, &junk );
*resid = bli_fmaxabs( resid1, resid2 );
bli_obj_free( &v );
bli_obj_free( &q );
}
void libblis_test_syr2_check( obj_t* alpha,
obj_t* x,
obj_t* y,
obj_t* a,
obj_t* a_orig,
double* resid )
{
num_t dt = bli_obj_datatype( *a );
num_t dt_real = bli_obj_datatype_proj_to_real( *a );
dim_t m_a = bli_obj_length( *a );
obj_t xt, yt;
obj_t t, v, w1, w2;
obj_t tau, rho, norm;
double junk;
//
// Pre-conditions:
// - x is randomized.
// - y is randomized.
// - a is randomized and symmetric.
// Note:
// - alpha should have a non-zero imaginary component in the
// complex cases in order to more fully exercise the implementation.
//
// Under these conditions, we assume that the implementation for
//
// A := A_orig + alpha * conjx(x) * conjy(y)^T + alpha * conjy(y) * conjx(x)^T
//
// is functioning correctly if
//
// normf( v - w )
//
// is negligible, where
//
// v = A * t
// w = ( A_orig + alpha * conjx(x) * conjy(y)^T + alpha * conjy(y) * conjx(x)^T ) * t
// = A_orig * t + alpha * conjx(x) * conjy(y)^T * t + alpha * conjy(y) * conjx(x)^T * t
// = A_orig * t + alpha * conjx(x) * conjy(y)^T * t + alpha * conjy(y) * rho
// = A_orig * t + alpha * conjx(x) * conjy(y)^T * t + w1
// = A_orig * t + alpha * conjx(x) * rho + w1
// = A_orig * t + w2 + w1
//
bli_mksymm( a );
bli_mksymm( a_orig );
bli_obj_set_struc( BLIS_GENERAL, *a );
bli_obj_set_struc( BLIS_GENERAL, *a_orig );
bli_obj_set_uplo( BLIS_DENSE, *a );
bli_obj_set_uplo( BLIS_DENSE, *a_orig );
bli_obj_scalar_init_detached( dt, &tau );
bli_obj_scalar_init_detached( dt, &rho );
bli_obj_scalar_init_detached( dt_real, &norm );
bli_obj_create( dt, m_a, 1, 0, 0, &t );
bli_obj_create( dt, m_a, 1, 0, 0, &v );
bli_obj_create( dt, m_a, 1, 0, 0, &w1 );
bli_obj_create( dt, m_a, 1, 0, 0, &w2 );
bli_obj_alias_to( *x, xt );
bli_obj_alias_to( *y, yt );
bli_setsc( 1.0/( double )m_a, -1.0/( double )m_a, &tau );
bli_setv( &tau, &t );
bli_gemv( &BLIS_ONE, a, &t, &BLIS_ZERO, &v );
bli_dotv( &xt, &t, &rho );
bli_mulsc( alpha, &rho );
bli_scal2v( &rho, y, &w1 );
bli_dotv( &yt, &t, &rho );
bli_mulsc( alpha, &rho );
bli_scal2v( &rho, x, &w2 );
bli_addv( &w2, &w1 );
bli_gemv( &BLIS_ONE, a_orig, &t, &BLIS_ONE, &w1 );
bli_subv( &w1, &v );
bli_normfv( &v, &norm );
bli_getsc( &norm, resid, &junk );
bli_obj_free( &t );
bli_obj_free( &v );
bli_obj_free( &w1 );
bli_obj_free( &w2 );
}
void bli_herk_l_ker_var2( obj_t* a,
obj_t* b,
obj_t* c,
gemm_t* cntl,
herk_thrinfo_t* thread )
{
num_t dt_exec = bli_obj_execution_datatype( *c );
doff_t diagoffc = bli_obj_diag_offset( *c );
pack_t schema_a = bli_obj_pack_schema( *a );
pack_t schema_b = bli_obj_pack_schema( *b );
dim_t m = bli_obj_length( *c );
dim_t n = bli_obj_width( *c );
dim_t k = bli_obj_width( *a );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t cs_a = bli_obj_col_stride( *a );
inc_t pd_a = bli_obj_panel_dim( *a );
inc_t ps_a = bli_obj_panel_stride( *a );
void* buf_b = bli_obj_buffer_at_off( *b );
inc_t rs_b = bli_obj_row_stride( *b );
inc_t pd_b = bli_obj_panel_dim( *b );
inc_t ps_b = bli_obj_panel_stride( *b );
void* buf_c = bli_obj_buffer_at_off( *c );
inc_t rs_c = bli_obj_row_stride( *c );
inc_t cs_c = bli_obj_col_stride( *c );
obj_t scalar_a;
obj_t scalar_b;
void* buf_alpha;
void* buf_beta;
FUNCPTR_T f;
func_t* gemm_ukrs;
void* gemm_ukr;
// Detach and multiply the scalars attached to A and B.
bli_obj_scalar_detach( a, &scalar_a );
bli_obj_scalar_detach( b, &scalar_b );
bli_mulsc( &scalar_a, &scalar_b );
// Grab the addresses of the internal scalar buffers for the scalar
// merged above and the scalar attached to C.
buf_alpha = bli_obj_internal_scalar_buffer( scalar_b );
buf_beta = bli_obj_internal_scalar_buffer( *c );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_exec];
// Extract from the control tree node the func_t object containing
// the gemm micro-kernel function addresses, and then query the
// function address corresponding to the current datatype.
gemm_ukrs = cntl_gemm_ukrs( cntl );
gemm_ukr = bli_func_obj_query( dt_exec, gemm_ukrs );
// Invoke the function.
f( diagoffc,
schema_a,
schema_b,
m,
n,
k,
buf_alpha,
buf_a, cs_a, pd_a, ps_a,
buf_b, rs_b, pd_b, ps_b,
buf_beta,
buf_c, rs_c, cs_c,
gemm_ukr,
thread );
}
void bli_herk_u_ker_var2
(
obj_t* a,
obj_t* b,
obj_t* c,
cntx_t* cntx,
cntl_t* cntl,
thrinfo_t* thread
)
{
num_t dt_exec = bli_obj_exec_dt( c );
doff_t diagoffc = bli_obj_diag_offset( c );
pack_t schema_a = bli_obj_pack_schema( a );
pack_t schema_b = bli_obj_pack_schema( b );
dim_t m = bli_obj_length( c );
dim_t n = bli_obj_width( c );
dim_t k = bli_obj_width( a );
void* buf_a = bli_obj_buffer_at_off( a );
inc_t cs_a = bli_obj_col_stride( a );
inc_t is_a = bli_obj_imag_stride( a );
dim_t pd_a = bli_obj_panel_dim( a );
inc_t ps_a = bli_obj_panel_stride( a );
void* buf_b = bli_obj_buffer_at_off( b );
inc_t rs_b = bli_obj_row_stride( b );
inc_t is_b = bli_obj_imag_stride( b );
dim_t pd_b = bli_obj_panel_dim( b );
inc_t ps_b = bli_obj_panel_stride( b );
void* buf_c = bli_obj_buffer_at_off( c );
inc_t rs_c = bli_obj_row_stride( c );
inc_t cs_c = bli_obj_col_stride( c );
obj_t scalar_a;
obj_t scalar_b;
void* buf_alpha;
void* buf_beta;
FUNCPTR_T f;
// Detach and multiply the scalars attached to A and B.
bli_obj_scalar_detach( a, &scalar_a );
bli_obj_scalar_detach( b, &scalar_b );
bli_mulsc( &scalar_a, &scalar_b );
// Grab the addresses of the internal scalar buffers for the scalar
// merged above and the scalar attached to C.
buf_alpha = bli_obj_internal_scalar_buffer( &scalar_b );
buf_beta = bli_obj_internal_scalar_buffer( c );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_exec];
// Invoke the function.
f( diagoffc,
schema_a,
schema_b,
m,
n,
k,
buf_alpha,
buf_a, cs_a, is_a,
pd_a, ps_a,
buf_b, rs_b, is_b,
pd_b, ps_b,
buf_beta,
buf_c, rs_c, cs_c,
cntx,
thread );
}
void libblis_test_axpyf_check( obj_t* alpha,
obj_t* a,
obj_t* x,
obj_t* y,
obj_t* y_orig,
double* resid )
{
num_t dt = bli_obj_datatype( *y );
num_t dt_real = bli_obj_datatype_proj_to_real( *y );
dim_t m = bli_obj_vector_dim( *y );
dim_t b_n = bli_obj_width( *a );
dim_t i;
obj_t a1, chi1, v;
obj_t alpha_chi1;
obj_t norm;
double junk;
//
// Pre-conditions:
// - a is randomized.
// - x is randomized.
// - y is randomized.
// Note:
// - alpha should have a non-zero imaginary component in the complex
// cases in order to more fully exercise the implementation.
//
// Under these conditions, we assume that the implementation for
//
// y := y_orig + alpha * conja(A) * conjx(x)
//
// is functioning correctly if
//
// normf( y - v )
//
// is negligible, where v contains y as computed by repeated calls to
// axpyv.
//
bli_obj_scalar_init_detached( dt_real, &norm );
bli_obj_scalar_init_detached( dt, &alpha_chi1 );
bli_obj_create( dt, m, 1, 0, 0, &v );
bli_copyv( y_orig, &v );
for ( i = 0; i < b_n; ++i )
{
bli_acquire_mpart_l2r( BLIS_SUBPART1, i, 1, a, &a1 );
bli_acquire_vpart_f2b( BLIS_SUBPART1, i, 1, x, &chi1 );
bli_copysc( &chi1, &alpha_chi1 );
bli_mulsc( alpha, &alpha_chi1 );
bli_axpyv( &alpha_chi1, &a1, &v );
}
bli_subv( y, &v );
bli_normfv( &v, &norm );
bli_getsc( &norm, resid, &junk );
bli_obj_free( &v );
}
void libblis_test_her_check( obj_t* alpha,
obj_t* x,
obj_t* a,
obj_t* a_orig,
double* resid )
{
num_t dt = bli_obj_datatype( *a );
num_t dt_real = bli_obj_datatype_proj_to_real( *a );
dim_t m_a = bli_obj_length( *a );
obj_t xh, t, v, w;
obj_t tau, rho, norm;
double junk;
//
// Pre-conditions:
// - x is randomized.
// - a is randomized and Hermitian.
// Note:
// - alpha must be real-valued.
//
// Under these conditions, we assume that the implementation for
//
// A := A_orig + alpha * conjx(x) * conjx(x)^H
//
// is functioning correctly if
//
// normf( v - w )
//
// is negligible, where
//
// v = A * t
// w = ( A_orig + alpha * conjx(x) * conjx(x)^H ) * t
// = A_orig * t + alpha * conjx(x) * conjx(x)^H * t
// = A_orig * t + alpha * conjx(x) * rho
// = A_orig * t + w
//
bli_mkherm( a );
bli_mkherm( a_orig );
bli_obj_set_struc( BLIS_GENERAL, *a );
bli_obj_set_struc( BLIS_GENERAL, *a_orig );
bli_obj_set_uplo( BLIS_DENSE, *a );
bli_obj_set_uplo( BLIS_DENSE, *a_orig );
bli_obj_scalar_init_detached( dt, &tau );
bli_obj_scalar_init_detached( dt, &rho );
bli_obj_scalar_init_detached( dt_real, &norm );
bli_obj_create( dt, m_a, 1, 0, 0, &t );
bli_obj_create( dt, m_a, 1, 0, 0, &v );
bli_obj_create( dt, m_a, 1, 0, 0, &w );
bli_obj_alias_with_conj( BLIS_CONJUGATE, *x, xh );
bli_setsc( 1.0/( double )m_a, -1.0/( double )m_a, &tau );
bli_setv( &tau, &t );
bli_gemv( &BLIS_ONE, a, &t, &BLIS_ZERO, &v );
bli_dotv( &xh, &t, &rho );
bli_mulsc( alpha, &rho );
bli_scal2v( &rho, x, &w );
bli_gemv( &BLIS_ONE, a_orig, &t, &BLIS_ONE, &w );
bli_subv( &w, &v );
bli_normfv( &v, &norm );
bli_getsc( &norm, resid, &junk );
bli_obj_free( &t );
bli_obj_free( &v );
bli_obj_free( &w );
}
void bli_trmm_rl_ker_var2( obj_t* a,
obj_t* b,
obj_t* c,
trmm_t* cntl )
{
num_t dt_exec = bli_obj_execution_datatype( *c );
doff_t diagoffb = bli_obj_diag_offset( *b );
dim_t m = bli_obj_length( *c );
dim_t n = bli_obj_width( *c );
dim_t k = bli_obj_width( *a );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t rs_a = bli_obj_row_stride( *a );
inc_t cs_a = bli_obj_col_stride( *a );
inc_t ps_a = bli_obj_panel_stride( *a );
void* buf_b = bli_obj_buffer_at_off( *b );
inc_t rs_b = bli_obj_row_stride( *b );
inc_t cs_b = bli_obj_col_stride( *b );
inc_t ps_b = bli_obj_panel_stride( *b );
void* buf_c = bli_obj_buffer_at_off( *c );
inc_t rs_c = bli_obj_row_stride( *c );
inc_t cs_c = bli_obj_col_stride( *c );
obj_t scalar_a;
obj_t scalar_b;
void* buf_alpha;
void* buf_beta;
FUNCPTR_T f;
// Detach and multiply the scalars attached to A and B.
bli_obj_scalar_detach( a, &scalar_a );
bli_obj_scalar_detach( b, &scalar_b );
bli_mulsc( &scalar_a, &scalar_b );
// Grab the addresses of the internal scalar buffers for the scalar
// merged above and the scalar attached to C.
buf_alpha = bli_obj_internal_scalar_buffer( scalar_b );
buf_beta = bli_obj_internal_scalar_buffer( *c );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_exec];
// Invoke the function.
f( diagoffb,
m,
n,
k,
buf_alpha,
buf_a, rs_a, cs_a, ps_a,
buf_b, rs_b, cs_b, ps_b,
buf_beta,
buf_c, rs_c, cs_c );
}
void libblis_test_ger_check
(
test_params_t* params,
obj_t* alpha,
obj_t* x,
obj_t* y,
obj_t* a,
obj_t* a_orig,
double* resid
)
{
num_t dt = bli_obj_dt( a );
num_t dt_real = bli_obj_dt_proj_to_real( a );
dim_t m_a = bli_obj_length( a );
dim_t n_a = bli_obj_width( a );
obj_t t, v, w;
obj_t rho, norm;
double junk;
//
// Pre-conditions:
// - x is randomized.
// - y is randomized.
// - a is identity.
// Note:
// - alpha should have a non-zero imaginary component in the
// complex cases in order to more fully exercise the implementation.
//
// Under these conditions, we assume that the implementation for
//
// A := A_orig + alpha * conjx(x) * conjy(y)
//
// is functioning correctly if
//
// normf( v - w )
//
// is negligible, where
//
// v = A * t
// w = ( A_orig + alpha * conjx(x) * conjy(y)^T ) * t
// = A_orig * t + alpha * conjx(x) * conjy(y)^T * t
// = A_orig * t + alpha * conjx(x) * rho
// = A_orig * t + w
//
bli_obj_scalar_init_detached( dt, &rho );
bli_obj_scalar_init_detached( dt_real, &norm );
bli_obj_create( dt, n_a, 1, 0, 0, &t );
bli_obj_create( dt, m_a, 1, 0, 0, &v );
bli_obj_create( dt, m_a, 1, 0, 0, &w );
libblis_test_vobj_randomize( params, TRUE, &t );
bli_gemv( &BLIS_ONE, a, &t, &BLIS_ZERO, &v );
bli_dotv( y, &t, &rho );
bli_mulsc( alpha, &rho );
bli_scal2v( &rho, x, &w );
bli_gemv( &BLIS_ONE, a_orig, &t, &BLIS_ONE, &w );
bli_subv( &w, &v );
bli_normfv( &v, &norm );
bli_getsc( &norm, resid, &junk );
bli_obj_free( &t );
bli_obj_free( &v );
bli_obj_free( &w );
}
void bli_trmm_ru_ker_var2( obj_t* a,
obj_t* b,
obj_t* c,
trmm_t* cntl,
trmm_thrinfo_t* thread )
{
num_t dt_exec = bli_obj_execution_datatype( *c );
doff_t diagoffb = bli_obj_diag_offset( *b );
dim_t m = bli_obj_length( *c );
dim_t n = bli_obj_width( *c );
dim_t k = bli_obj_width( *a );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t cs_a = bli_obj_col_stride( *a );
inc_t pd_a = bli_obj_panel_dim( *a );
inc_t ps_a = bli_obj_panel_stride( *a );
void* buf_b = bli_obj_buffer_at_off( *b );
inc_t rs_b = bli_obj_row_stride( *b );
inc_t pd_b = bli_obj_panel_dim( *b );
inc_t ps_b = bli_obj_panel_stride( *b );
void* buf_c = bli_obj_buffer_at_off( *c );
inc_t rs_c = bli_obj_row_stride( *c );
inc_t cs_c = bli_obj_col_stride( *c );
obj_t scalar_a;
obj_t scalar_b;
void* buf_alpha;
void* buf_beta;
FUNCPTR_T f;
func_t* gemm_ukrs;
void* gemm_ukr;
// Detach and multiply the scalars attached to A and B.
bli_obj_scalar_detach( a, &scalar_a );
bli_obj_scalar_detach( b, &scalar_b );
bli_mulsc( &scalar_a, &scalar_b );
// Grab the addresses of the internal scalar buffers for the scalar
// merged above and the scalar attached to C.
buf_alpha = bli_obj_internal_scalar_buffer( scalar_b );
buf_beta = bli_obj_internal_scalar_buffer( *c );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_exec];
// Adjust cs_a and rs_b if A and B were packed for 4m or 3m. This
// is needed because cs_a and rs_b are used to index into the
// micro-panels of A and B, respectively, and since the pointer
// types in the macro-kernel (scomplex or dcomplex) will result
// in pointer arithmetic that moves twice as far as it should,
// given the datatypes actually stored (float or double), we must
// halve the strides to compensate.
if ( bli_obj_is_panel_packed_4m( *a ) ||
bli_obj_is_panel_packed_3m( *a ) ) { cs_a /= 2; rs_b /= 2; }
// Extract from the control tree node the func_t object containing
// the gemm micro-kernel function addresses, and then query the
// function address corresponding to the current datatype.
gemm_ukrs = cntl_gemm_ukrs( cntl );
gemm_ukr = bli_func_obj_query( dt_exec, gemm_ukrs );
// Invoke the function.
f( diagoffb,
m,
n,
k,
buf_alpha,
buf_a, cs_a, pd_a, ps_a,
buf_b, rs_b, pd_b, ps_b,
buf_beta,
buf_c, rs_c, cs_c,
gemm_ukr,
thread );
}
void bli_gemm_ker_var2
(
obj_t* a,
obj_t* b,
obj_t* c,
cntx_t* cntx,
cntl_t* cntl,
thrinfo_t* thread
)
{
num_t dt_exec = bli_obj_execution_datatype( *c );
pack_t schema_a = bli_obj_pack_schema( *a );
pack_t schema_b = bli_obj_pack_schema( *b );
dim_t m = bli_obj_length( *c );
dim_t n = bli_obj_width( *c );
dim_t k = bli_obj_width( *a );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t cs_a = bli_obj_col_stride( *a );
inc_t is_a = bli_obj_imag_stride( *a );
dim_t pd_a = bli_obj_panel_dim( *a );
inc_t ps_a = bli_obj_panel_stride( *a );
void* buf_b = bli_obj_buffer_at_off( *b );
inc_t rs_b = bli_obj_row_stride( *b );
inc_t is_b = bli_obj_imag_stride( *b );
dim_t pd_b = bli_obj_panel_dim( *b );
inc_t ps_b = bli_obj_panel_stride( *b );
void* buf_c = bli_obj_buffer_at_off( *c );
inc_t rs_c = bli_obj_row_stride( *c );
inc_t cs_c = bli_obj_col_stride( *c );
obj_t scalar_a;
obj_t scalar_b;
void* buf_alpha;
void* buf_beta;
FUNCPTR_T f;
// Detach and multiply the scalars attached to A and B.
bli_obj_scalar_detach( a, &scalar_a );
bli_obj_scalar_detach( b, &scalar_b );
bli_mulsc( &scalar_a, &scalar_b );
// Grab the addresses of the internal scalar buffers for the scalar
// merged above and the scalar attached to C.
buf_alpha = bli_obj_internal_scalar_buffer( scalar_b );
buf_beta = bli_obj_internal_scalar_buffer( *c );
// If 1m is being employed on a column- or row-stored matrix with a
// real-valued beta, we can use the real domain macro-kernel, which
// eliminates a little overhead associated with the 1m virtual
// micro-kernel.
#if 1
if ( bli_is_1m_packed( schema_a ) )
{
bli_l3_ind_recast_1m_params
(
dt_exec,
schema_a,
c,
m, n, k,
pd_a, ps_a,
pd_b, ps_b,
rs_c, cs_c
);
}
#endif
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_exec];
// Invoke the function.
f( schema_a,
schema_b,
m,
n,
k,
buf_alpha,
buf_a, cs_a, is_a,
pd_a, ps_a,
buf_b, rs_b, is_b,
pd_b, ps_b,
buf_beta,
buf_c, rs_c, cs_c,
cntx,
thread );
}
