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_blk_var2( obj_t* c,
obj_t* p )
{
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 );
dim_t pd_p = bli_obj_panel_dim( *p );
inc_t ps_p = bli_obj_panel_stride( *p );
void* buf_kappa;
FUNCPTR_T f;
// This variant assumes 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
// 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];
// 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,
pd_p, ps_p );
}
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_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 );
}
void bli_packm_blk_var1_md
(
obj_t* c,
obj_t* p,
cntx_t* cntx,
cntl_t* cntl,
thrinfo_t* t
)
{
num_t dt_c = bli_obj_dt( c );
num_t dt_p = bli_obj_dt( p );
trans_t transc = bli_obj_conjtrans_status( c );
pack_t schema = bli_obj_pack_schema( 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;
void* buf_kappa;
FUNCPTR_T f;
// Treatment of kappa (ie: packing during scaling) depends on
// whether we are executing an induced method.
if ( bli_is_nat_packed( schema ) )
{
// This branch is 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_p, &BLIS_ONE );
}
else // if ( bli_is_ind_packed( schema ) )
{
obj_t* kappa_p;
// 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 ( 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;
}
// Acquire the buffer to the kappa chosen above.
buf_kappa = bli_obj_buffer_for_1x1( dt_p, kappa_p );
}
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_c][dt_p];
// Invoke the function.
f(
transc,
schema,
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,
cntx,
t );
}
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 );
}