void bli_norm1m_unb_var1( obj_t* x,
obj_t* norm )
{
num_t dt_x = bli_obj_datatype( *x );
doff_t diagoffx = bli_obj_diag_offset( *x );
uplo_t diagx = bli_obj_diag( *x );
uplo_t uplox = bli_obj_uplo( *x );
dim_t m = bli_obj_length( *x );
dim_t n = bli_obj_width( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t rs_x = bli_obj_row_stride( *x );
inc_t cs_x = bli_obj_col_stride( *x );
void* buf_norm = bli_obj_buffer_at_off( *norm );
FUNCPTR_T f;
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_x];
// Invoke the function.
f( diagoffx,
diagx,
uplox,
m,
n,
buf_x, rs_x, cs_x,
buf_norm );
}
void bli_unpackm_blk_var2( obj_t* p,
obj_t* c,
unpackm_t* cntl )
{
num_t dt_cp = bli_obj_datatype( *c );
// Normally we take the parameters from the source argument. But here,
// the packm/unpackm framework is not yet solidified enough for us to
// assume that at this point struc(P) == struc(C), (ie: since
// densification may have marked P's structure as dense when the root
// is upper or lower). So, we take the struc field from C, not P.
struc_t strucc = bli_obj_struc( *c );
doff_t diagoffc = bli_obj_diag_offset( *c );
diag_t diagc = bli_obj_diag( *c );
uplo_t uploc = bli_obj_uplo( *c );
// Again, normally the trans argument is on the source matrix. But we
// know that the packed matrix is not transposed. If there is to be a
// transposition, it is because C was originally transposed when packed.
// Thus, we query C for the trans status, not P. Also, we only query
// the trans status (not the conjugation status), since we probably
// don't want to un-conjugate if the original matrix was conjugated
// when packed.
trans_t transc = bli_obj_onlytrans_status( *c );
dim_t m_c = bli_obj_length( *c );
dim_t n_c = bli_obj_width( *c );
dim_t m_panel = bli_obj_panel_length( *c );
dim_t n_panel = bli_obj_panel_width( *c );
void* buf_p = bli_obj_buffer_at_off( *p );
inc_t rs_p = bli_obj_row_stride( *p );
inc_t cs_p = bli_obj_col_stride( *p );
dim_t pd_p = bli_obj_panel_dim( *p );
inc_t ps_p = bli_obj_panel_stride( *p );
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 );
FUNCPTR_T f;
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_cp];
// Invoke the function.
f( strucc,
diagoffc,
diagc,
uploc,
transc,
m_c,
n_c,
m_panel,
n_panel,
buf_p, rs_p, cs_p,
pd_p, ps_p,
buf_c, rs_c, cs_c );
}
void bli_packm_unb_var1( obj_t* c,
obj_t* p,
packm_thrinfo_t* thread )
{
num_t dt_cp = bli_obj_datatype( *c );
struc_t strucc = bli_obj_struc( *c );
doff_t diagoffc = bli_obj_diag_offset( *c );
diag_t diagc = bli_obj_diag( *c );
uplo_t uploc = bli_obj_uplo( *c );
trans_t transc = bli_obj_conjtrans_status( *c );
dim_t m_p = bli_obj_length( *p );
dim_t n_p = bli_obj_width( *p );
dim_t m_max_p = bli_obj_padded_length( *p );
dim_t n_max_p = bli_obj_padded_width( *p );
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 );
void* buf_p = bli_obj_buffer_at_off( *p );
inc_t rs_p = bli_obj_row_stride( *p );
inc_t cs_p = bli_obj_col_stride( *p );
void* buf_kappa;
FUNCPTR_T f;
// This variant assumes that the computational kernel will always apply
// the alpha scalar of the higher-level operation. Thus, we use BLIS_ONE
// for kappa so that the underlying packm implementation does not scale
// during packing.
buf_kappa = bli_obj_buffer_for_const( dt_cp, BLIS_ONE );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_cp];
if( thread_am_ochief( thread ) ) {
// Invoke the function.
f( strucc,
diagoffc,
diagc,
uploc,
transc,
m_p,
n_p,
m_max_p,
n_max_p,
buf_kappa,
buf_c, rs_c, cs_c,
buf_p, rs_p, cs_p );
}
}
void bli_scalm_unb_var1( obj_t* alpha,
obj_t* x,
cntx_t* cntx )
{
num_t dt_x = bli_obj_datatype( *x );
doff_t diagoffx = bli_obj_diag_offset( *x );
uplo_t diagx = bli_obj_diag( *x );
uplo_t uplox = bli_obj_uplo( *x );
dim_t m = bli_obj_length( *x );
dim_t n = bli_obj_width( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t rs_x = bli_obj_row_stride( *x );
inc_t cs_x = bli_obj_col_stride( *x );
void* buf_alpha;
obj_t x_local;
FUNCPTR_T f;
// Alias x to x_local so we can apply alpha if it is non-unit.
bli_obj_alias_to( *x, x_local );
// If alpha is non-unit, apply it to the scalar attached to x.
if ( !bli_obj_equals( alpha, &BLIS_ONE ) )
{
bli_obj_scalar_apply_scalar( alpha, &x_local );
}
// Grab the address of the internal scalar buffer for the scalar
// attached to x.
buf_alpha_x = bli_obj_internal_scalar_buffer( *x );
// Index into the type combination array to extract the correct
// function pointer.
// NOTE: We use dt_x for both alpha and x because alpha was obtained
// from the attached scalar of x, which is guaranteed to be of the
// same datatype as x.
f = ftypes[dt_x][dt_x];
// Invoke the function.
// NOTE: We unconditionally pass in BLIS_NO_CONJUGATE for alpha
// because it would have already been conjugated by the front-end.
f( BLIS_NO_CONJUGATE,
diagoffx,
diagx,
uplox,
m,
n,
buf_alpha,
buf_x, rs_x, cs_x );
}
void bli_axpym_unb_var1( obj_t* alpha,
obj_t* x,
obj_t* y,
cntx_t* cntx )
{
num_t dt_x = bli_obj_datatype( *x );
num_t dt_y = bli_obj_datatype( *y );
doff_t diagoffx = bli_obj_diag_offset( *x );
diag_t diagx = bli_obj_diag( *x );
uplo_t uplox = bli_obj_uplo( *x );
trans_t transx = bli_obj_conjtrans_status( *x );
dim_t m = bli_obj_length( *y );
dim_t n = bli_obj_width( *y );
inc_t rs_x = bli_obj_row_stride( *x );
inc_t cs_x = bli_obj_col_stride( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t rs_y = bli_obj_row_stride( *y );
inc_t cs_y = bli_obj_col_stride( *y );
void* buf_y = bli_obj_buffer_at_off( *y );
num_t dt_alpha;
void* buf_alpha;
FUNCPTR_T f;
// If alpha is a scalar constant, use dt_x to extract the address of the
// corresponding constant value; otherwise, use the datatype encoded
// within the alpha object and extract the buffer at the alpha offset.
bli_set_scalar_dt_buffer( alpha, dt_x, dt_alpha, buf_alpha );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_alpha][dt_x][dt_y];
// Invoke the function.
f( diagoffx,
diagx,
uplox,
transx,
m,
n,
buf_alpha,
buf_x, rs_x, cs_x,
buf_y, rs_y, cs_y );
}
void bli_trmv_unf_var2( obj_t* alpha,
obj_t* a,
obj_t* x,
cntx_t* cntx,
trmv_t* cntl )
{
num_t dt_a = bli_obj_datatype( *a );
num_t dt_x = bli_obj_datatype( *x );
uplo_t uplo = bli_obj_uplo( *a );
trans_t trans = bli_obj_conjtrans_status( *a );
diag_t diag = bli_obj_diag( *a );
dim_t m = bli_obj_length( *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 );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t incx = bli_obj_vector_inc( *x );
num_t dt_alpha;
void* buf_alpha;
FUNCPTR_T f;
// The datatype of alpha MUST be the type union of a and x. This is to
// prevent any unnecessary loss of information during computation.
dt_alpha = bli_datatype_union( dt_a, dt_x );
buf_alpha = bli_obj_buffer_for_1x1( dt_alpha, *alpha );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_a][dt_x];
// Invoke the function.
f( uplo,
trans,
diag,
m,
buf_alpha,
buf_a, rs_a, cs_a,
buf_x, incx );
}
void bli_addm_unb_var1( obj_t* x,
obj_t* y,
cntx_t* cntx )
{
num_t dt_x = bli_obj_datatype( *x );
num_t dt_y = bli_obj_datatype( *y );
doff_t diagoffx = bli_obj_diag_offset( *x );
diag_t diagx = bli_obj_diag( *x );
uplo_t uplox = bli_obj_uplo( *x );
trans_t transx = bli_obj_conjtrans_status( *x );
dim_t m = bli_obj_length( *y );
dim_t n = bli_obj_width( *y );
inc_t rs_x = bli_obj_row_stride( *x );
inc_t cs_x = bli_obj_col_stride( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t rs_y = bli_obj_row_stride( *y );
inc_t cs_y = bli_obj_col_stride( *y );
void* buf_y = bli_obj_buffer_at_off( *y );
FUNCPTR_T f;
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_x][dt_y];
// Invoke the function.
f( diagoffx,
diagx,
uplox,
transx,
m,
n,
buf_x, rs_x, cs_x,
buf_y, rs_y, cs_y );
}
void bli_packm_blk_var1( obj_t* c,
obj_t* p,
packm_thrinfo_t* t )
{
num_t dt_cp = bli_obj_datatype( *c );
struc_t strucc = bli_obj_struc( *c );
doff_t diagoffc = bli_obj_diag_offset( *c );
diag_t diagc = bli_obj_diag( *c );
uplo_t uploc = bli_obj_uplo( *c );
trans_t transc = bli_obj_conjtrans_status( *c );
pack_t schema = bli_obj_pack_schema( *p );
bool_t invdiag = bli_obj_has_inverted_diag( *p );
bool_t revifup = bli_obj_is_pack_rev_if_upper( *p );
bool_t reviflo = bli_obj_is_pack_rev_if_lower( *p );
dim_t m_p = bli_obj_length( *p );
dim_t n_p = bli_obj_width( *p );
dim_t m_max_p = bli_obj_padded_length( *p );
dim_t n_max_p = bli_obj_padded_width( *p );
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 );
void* buf_p = bli_obj_buffer_at_off( *p );
inc_t rs_p = bli_obj_row_stride( *p );
inc_t cs_p = bli_obj_col_stride( *p );
inc_t is_p = bli_obj_imag_stride( *p );
dim_t pd_p = bli_obj_panel_dim( *p );
inc_t ps_p = bli_obj_panel_stride( *p );
obj_t kappa;
/*---initialize pointer to stop gcc complaining 2-9-16 GH --- */
obj_t* kappa_p = {0};
void* buf_kappa;
func_t* packm_kers;
void* packm_ker;
FUNCPTR_T f;
// Treatment of kappa (ie: packing during scaling) depends on
// whether we are executing an induced method.
if ( bli_is_ind_packed( schema ) )
{
// The value for kappa we use will depend on whether the scalar
// attached to A has a nonzero imaginary component. If it does,
// then we will apply the scalar during packing to facilitate
// implementing induced complex domain algorithms in terms of
// real domain micro-kernels. (In the aforementioned situation,
// applying a real scalar is easy, but applying a complex one is
// harder, so we avoid the need altogether with the code below.)
if( thread_am_ochief( t ) )
{
if ( bli_obj_scalar_has_nonzero_imag( p ) )
{
// Detach the scalar.
bli_obj_scalar_detach( p, &kappa );
// Reset the attached scalar (to 1.0).
bli_obj_scalar_reset( p );
kappa_p = κ
}
else
{
// If the internal scalar of A has only a real component, then
// we will apply it later (in the micro-kernel), and so we will
// use BLIS_ONE to indicate no scaling during packing.
kappa_p = &BLIS_ONE;
}
}
kappa_p = thread_obroadcast( t, kappa_p );
// Acquire the buffer to the kappa chosen above.
buf_kappa = bli_obj_buffer_for_1x1( dt_cp, *kappa_p );
}
else // if ( bli_is_nat_packed( schema ) )
{
// This branch if for native execution, where we assume that
// the micro-kernel will always apply the alpha scalar of the
// higher-level operation. Thus, we use BLIS_ONE for kappa so
// that the underlying packm implementation does not perform
// any scaling during packing.
buf_kappa = bli_obj_buffer_for_const( dt_cp, BLIS_ONE );
}
// Choose the correct func_t object based on the pack_t schema.
if ( bli_is_4mi_packed( schema ) ) packm_kers = packm_struc_cxk_4mi_kers;
else if ( bli_is_3mi_packed( schema ) ||
bli_is_3ms_packed( schema ) ) packm_kers = packm_struc_cxk_3mis_kers;
else if ( bli_is_ro_packed( schema ) ||
bli_is_io_packed( schema ) ||
bli_is_rpi_packed( schema ) ) packm_kers = packm_struc_cxk_rih_kers;
else packm_kers = packm_struc_cxk_kers;
// Query the datatype-specific function pointer from the func_t object.
packm_ker = bli_func_obj_query( dt_cp, packm_kers );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_cp];
// Invoke the function.
f( strucc,
diagoffc,
diagc,
uploc,
transc,
schema,
invdiag,
revifup,
reviflo,
m_p,
n_p,
m_max_p,
n_max_p,
buf_kappa,
buf_c, rs_c, cs_c,
buf_p, rs_p, cs_p,
is_p,
pd_p, ps_p,
packm_ker,
t );
}
