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_obj_scalar_reset( obj_t* a )
{
num_t dt = bli_obj_datatype( *a );
void* scalar_a = bli_obj_internal_scalar_buffer( *a );
void* one = bli_obj_buffer_for_const( dt, BLIS_ONE );
if ( bli_is_float( dt ) ) *(( float* )scalar_a) = *(( float* )one);
else if ( bli_is_double( dt ) ) *(( double* )scalar_a) = *(( double* )one);
else if ( bli_is_scomplex( dt ) ) *(( scomplex* )scalar_a) = *(( scomplex* )one);
else if ( bli_is_dcomplex( dt ) ) *(( dcomplex* )scalar_a) = *(( dcomplex* )one);
// Alternate implementation:
//bli_obj_scalar_attach( BLIS_NO_CONJUGATE, &BLIS_ONE, a );
}
void bli_obj_free( obj_t* obj )
{
if ( bli_error_checking_is_enabled() )
bli_obj_free_check( obj );
// Don't dereference obj if it is NULL.
if ( obj != NULL )
{
// Idiot safety: Don't try to free the buffer field if the object
// is a detached scalar (ie: if the buffer pointer refers to the
// address of the internal scalar buffer).
if ( bli_obj_buffer( *obj ) != bli_obj_internal_scalar_buffer( *obj ) )
bli_free_user( bli_obj_buffer( *obj ) );
}
}
void bli_obj_scalar_init_detached( num_t dt,
obj_t* beta )
{
void* p;
// Initialize beta without a buffer and then attach its internal buffer.
bli_obj_create_without_buffer( dt, 1, 1, beta );
// Query the address of the object's internal scalar buffer.
p = bli_obj_internal_scalar_buffer( *beta );
// Update the object.
bli_obj_set_buffer( p, *beta );
bli_obj_set_strides( 1, 1, *beta );
bli_obj_set_imag_stride( 1, *beta );
}
void bli_obj_create_without_buffer( num_t dt,
dim_t m,
dim_t n,
obj_t* obj )
{
siz_t elem_size;
void* s;
if ( bli_error_checking_is_enabled() )
bli_obj_create_without_buffer_check( dt, m, n, obj );
// Query the size of one element of the object's pre-set datatype.
elem_size = bli_datatype_size( dt );
// Set any default properties that are appropriate.
bli_obj_set_defaults( *obj );
// Set the object root to itself, since obj is not presumed to be a view
// into a larger matrix. This is typically the only time this field is
// ever set; henceforth, subpartitions and aliases to this object will
// get copies of this field, and thus always have access to its
// "greatest-grand" parent (ie: the original parent, or "root", object).
// However, there ARE a few places where it is convenient to reset the
// root field explicitly via bli_obj_set_as_root(). (We do not list
// those places here. Just grep for bli_obj_set_as_root within the
// top-level 'frame' directory to see them.
bli_obj_set_as_root( *obj );
// Set individual fields.
bli_obj_set_buffer( NULL, *obj );
bli_obj_set_datatype( dt, *obj );
bli_obj_set_elem_size( elem_size, *obj );
bli_obj_set_target_datatype( dt, *obj );
bli_obj_set_execution_datatype( dt, *obj );
bli_obj_set_dims( m, n, *obj );
bli_obj_set_offs( 0, 0, *obj );
bli_obj_set_diag_offset( 0, *obj );
// Set the internal scalar to 1.0.
s = bli_obj_internal_scalar_buffer( *obj );
if ( bli_is_float( dt ) ) { bli_sset1s( *(( float* )s) ); }
else if ( bli_is_double( dt ) ) { bli_dset1s( *(( double* )s) ); }
else if ( bli_is_scomplex( dt ) ) { bli_cset1s( *(( scomplex* )s) ); }
else if ( bli_is_dcomplex( dt ) ) { bli_zset1s( *(( dcomplex* )s) ); }
}
bool_t bli_obj_scalar_has_nonzero_imag( obj_t* a )
{
bool_t r_val = FALSE;
num_t dt = bli_obj_datatype( *a );
void* scalar_a = bli_obj_internal_scalar_buffer( *a );
if ( bli_is_real( dt ) )
{
r_val = FALSE;
}
else if ( bli_is_scomplex( dt ) )
{
r_val = ( bli_cimag( *(( scomplex* )scalar_a) ) != 0.0F );
}
else if ( bli_is_dcomplex( dt ) )
{
r_val = ( bli_zimag( *(( dcomplex* )scalar_a) ) != 0.0 );
}
return r_val;
}
void bli_trsm_rl_ker_var2( obj_t* a,
obj_t* b,
obj_t* c,
trsm_t* cntl,
trsm_thrinfo_t* thread )
{
num_t dt_exec = bli_obj_execution_datatype( *c );
doff_t diagoffb = bli_obj_diag_offset( *b );
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 );
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 );
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 );
void* buf_alpha1;
void* buf_alpha2;
FUNCPTR_T f;
func_t* gemmtrsm_ukrs;
func_t* gemm_ukrs;
void* gemmtrsm_ukr;
void* gemm_ukr;
// Grab the address of the internal scalar buffer for the scalar
// attached to A. This will be the alpha scalar used in the gemmtrsm
// subproblems (ie: the scalar that would be applied to the packed
// copy of A prior to it being updated by the trsm subproblem). This
// scalar may be unit, if for example it was applied during packing.
buf_alpha1 = bli_obj_internal_scalar_buffer( *a );
// Grab the address of the internal scalar buffer for the scalar
// attached to C. This will be the "beta" scalar used in the gemm-only
// subproblems that correspond to micro-panels that do not intersect
// the diagonal. We need this separate scalar because it's possible
// that the alpha attached to B was reset, if it was applied during
// packing.
buf_alpha2 = 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 objects containing
// the gemmtrsm and gemm micro-kernel function addresses, and then
// query the function addresses corresponding to the current datatype.
gemmtrsm_ukrs = cntl_gemmtrsm_u_ukrs( cntl );
gemm_ukrs = cntl_gemm_ukrs( cntl );
gemmtrsm_ukr = bli_func_obj_query( dt_exec, gemmtrsm_ukrs );
gemm_ukr = bli_func_obj_query( dt_exec, gemm_ukrs );
// Invoke the function.
f( diagoffb,
schema_a,
schema_b,
m,
n,
k,
buf_alpha1,
buf_a, cs_a, pd_a, ps_a,
buf_b, rs_b, pd_b, ps_b,
buf_alpha2,
buf_c, rs_c, cs_c,
gemmtrsm_ukr,
gemm_ukr,
thread );
}
void bli_trsm_ru_ker_var2( obj_t* a,
obj_t* b,
obj_t* c,
trsm_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 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 );
void* buf_alpha1;
void* buf_alpha2;
FUNCPTR_T f;
func_t* gemmtrsm_ukrs;
func_t* gemm_ukrs;
void* gemmtrsm_ukr;
void* gemm_ukr;
// Grab the address of the internal scalar buffer for the scalar
// attached to A. This will be the alpha scalar used in the gemmtrsm
// subproblems (ie: the scalar that would be applied to the packed
// copy of A prior to it being updated by the trsm subproblem). This
// scalar may be unit, if for example it was applied during packing.
buf_alpha1 = bli_obj_internal_scalar_buffer( *a );
// Grab the address of the internal scalar buffer for the scalar
// attached to C. This will be the "beta" scalar used in the gemm-only
// subproblems that correspond to micro-panels that do not intersect
// the diagonal. We need this separate scalar because it's possible
// that the alpha attached to B was reset, if it was applied during
// packing.
buf_alpha2 = 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 objects containing
// the gemmtrsm and gemm micro-kernel function addresses, and then
// query the function addresses corresponding to the current datatype.
gemmtrsm_ukrs = cntl_gemmtrsm_l_ukrs( cntl );
gemm_ukrs = cntl_gemm_ukrs( cntl );
gemmtrsm_ukr = bli_func_obj_query( dt_exec, gemmtrsm_ukrs );
gemm_ukr = bli_func_obj_query( dt_exec, gemm_ukrs );
// Invoke the function.
f( diagoffb,
m,
n,
k,
buf_alpha1,
buf_a, cs_a, pd_a, ps_a,
buf_b, rs_b, pd_b, ps_b,
buf_alpha2,
buf_c, rs_c, cs_c,
gemmtrsm_ukr,
gemm_ukr );
}
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 bli_trsm_rl_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 );
doff_t diagoffb = bli_obj_diag_offset( *b );
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 );
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 );
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 );
void* buf_alpha1;
void* buf_alpha2;
FUNCPTR_T f;
// Grab the address of the internal scalar buffer for the scalar
// attached to A (the non-triangular matrix). This will be the alpha
// scalar used in the gemmtrsm subproblems (ie: the scalar that would
// be applied to the packed copy of A prior to it being updated by
// the trsm subproblem). This scalar may be unit, if for example it
// was applied during packing.
buf_alpha1 = bli_obj_internal_scalar_buffer( *a );
// Grab the address of the internal scalar buffer for the scalar
// attached to C. This will be the "beta" scalar used in the gemm-only
// subproblems that correspond to micro-panels that do not intersect
// the diagonal. We need this separate scalar because it's possible
// that the alpha attached to B was reset, if it was applied during
// packing.
buf_alpha2 = 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,
schema_a,
schema_b,
m,
n,
k,
buf_alpha1,
buf_a, cs_a, pd_a, ps_a,
buf_b, rs_b, pd_b, ps_b,
buf_alpha2,
buf_c, rs_c, cs_c,
cntx,
thread );
}
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 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 );
}