dim_t bli_gemm_determine_kc_f( dim_t i,
dim_t dim,
obj_t* a,
obj_t* b,
blksz_t* bsize )
{
num_t dt;
dim_t mnr;
dim_t b_alg, b_max;
dim_t b_use;
// We assume that this function is being called from an algorithm that
// is moving "forward" (ie: top to bottom, left to right, top-left
// to bottom-right).
// Extract the execution datatype and use it to query the corresponding
// blocksize and blocksize maximum values from the blksz_t object.
dt = bli_obj_execution_datatype( *a );
b_alg = bli_blksz_get_def( dt, bsize );
b_max = bli_blksz_get_max( dt, bsize );
// Nudge the default and maximum kc blocksizes up to the nearest
// multiple of MR if A is Hermitian or symmetric, or NR if B is
// Hermitian or symmetric. If neither case applies, then we leave
// the blocksizes unchanged.
if ( bli_obj_root_is_herm_or_symm( *a ) )
{
mnr = bli_blksz_get_mr( dt, bsize );
b_alg = bli_align_dim_to_mult( b_alg, mnr );
b_max = bli_align_dim_to_mult( b_max, mnr );
}
else if ( bli_obj_root_is_herm_or_symm( *b ) )
{
mnr = bli_blksz_get_nr( dt, bsize );
b_alg = bli_align_dim_to_mult( b_alg, mnr );
b_max = bli_align_dim_to_mult( b_max, mnr );
}
b_use = bli_determine_blocksize_f_sub( i, dim, b_alg, b_max );
return b_use;
}
dim_t bli_determine_blocksize_b( dim_t i,
dim_t dim,
obj_t* obj,
bszid_t bszid,
cntx_t* cntx )
{
num_t dt;
blksz_t* bsize;
dim_t b_alg, b_max;
dim_t b_use;
// Extract the execution datatype and use it to query the corresponding
// blocksize and blocksize maximum values from the blksz_t object.
dt = bli_obj_execution_datatype( *obj );
bsize = bli_cntx_get_blksz( bszid, cntx );
b_alg = bli_blksz_get_def( dt, bsize );
b_max = bli_blksz_get_max( dt, bsize );
b_use = bli_determine_blocksize_b_sub( i, dim, b_alg, b_max );
return b_use;
}
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_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_trsm_blk_var1f( obj_t* a,
obj_t* b,
obj_t* c,
trsm_t* cntl,
trsm_thrinfo_t* thread )
{
obj_t b_pack_s;
obj_t a1_pack_s;
obj_t a1, c1;
obj_t* b_pack = NULL;
obj_t* a1_pack = NULL;
dim_t i;
dim_t b_alg;
dim_t m_trans;
dim_t offA;
// Initialize object for packing B.
if( thread_am_ochief( thread ) ) {
bli_obj_init_pack( &b_pack_s );
bli_packm_init( b, &b_pack_s,
cntl_sub_packm_b( cntl ) );
}
b_pack = thread_obroadcast( thread, &b_pack_s );
// Initialize object for packing B.
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &a1_pack_s );
}
a1_pack = thread_ibroadcast( thread, &a1_pack_s );
// Pack B1 (if instructed).
bli_packm_int( b, b_pack,
cntl_sub_packm_b( cntl ),
trsm_thread_sub_opackm( thread ) );
// Set the default length of and offset to the non-zero part of A.
m_trans = bli_obj_length_after_trans( *a );
offA = 0;
// If A is lower triangular, we have to adjust where the non-zero part of
// A begins.
if ( bli_obj_is_lower( *a ) )
offA = bli_abs( bli_obj_diag_offset_after_trans( *a ) );
dim_t start, end;
num_t dt = bli_obj_execution_datatype( *a );
bli_get_range_t2b( thread, offA, m_trans,
//bli_lcm( bli_info_get_default_nr( BLIS_TRSM, dt ), bli_info_get_default_mr( BLIS_TRSM, dt ) ),
bli_info_get_default_mc( BLIS_TRSM, dt ),
&start, &end );
// Partition along the remaining portion of the m dimension.
for ( i = start; i < end; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_f( i, end, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and C1.
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Initialize object for packing A1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
cntl_sub_packm_a( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
cntl_sub_packm_a( cntl ),
trsm_thread_sub_ipackm( thread ) );
// Perform trsm subproblem.
bli_trsm_int( &BLIS_ONE,
a1_pack,
b_pack,
&BLIS_ONE,
&c1,
cntl_sub_trsm( cntl ),
trsm_thread_sub_trsm( thread ) );
thread_ibarrier( thread );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
thread_obarrier( thread );
if( thread_am_ochief( thread ) )
bli_packm_release( b_pack, cntl_sub_packm_b( cntl ) );
if( thread_am_ichief( thread ) )
bli_packm_release( a1_pack, cntl_sub_packm_a( cntl ) );
}
void bli_trsm_blk_var1b( obj_t* a,
obj_t* b,
obj_t* c,
trsm_t* cntl,
trsm_thrinfo_t* thread )
{
obj_t b_pack_s;
obj_t a1_pack_s;
obj_t a1, c1;
obj_t* b_pack = NULL;
obj_t* a1_pack = NULL;
dim_t i;
dim_t b_alg;
// Prune any zero region that exists along the partitioning dimension.
bli_trsm_prune_unref_mparts_m( a, b, c );
// Initialize object for packing B.
if( thread_am_ochief( thread ) ) {
bli_obj_init_pack( &b_pack_s );
bli_packm_init( b, &b_pack_s,
cntl_sub_packm_b( cntl ) );
}
b_pack = thread_obroadcast( thread, &b_pack_s );
// Initialize object for packing B.
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &a1_pack_s );
}
a1_pack = thread_ibroadcast( thread, &a1_pack_s );
// Pack B1 (if instructed).
bli_packm_int( b, b_pack,
cntl_sub_packm_b( cntl ),
trsm_thread_sub_opackm( thread ) );
dim_t my_start, my_end;
num_t dt = bli_obj_execution_datatype( *a );
dim_t bf = ( bli_obj_root_is_triangular( *a ) ?
bli_info_get_default_mr( BLIS_TRSM, dt ) :
bli_info_get_default_nr( BLIS_TRSM, dt ) );
bli_get_range_b2t( thread, a, bf,
&my_start, &my_end );
// Partition along the remaining portion of the m dimension.
for ( i = my_start; i < my_end; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_b( i, my_end, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and C1.
bli_acquire_mpart_b2t( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_b2t( BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Initialize object for packing A1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
cntl_sub_packm_a( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
cntl_sub_packm_a( cntl ),
trsm_thread_sub_ipackm( thread ) );
// Perform trsm subproblem.
bli_trsm_int( &BLIS_ONE,
a1_pack,
b_pack,
&BLIS_ONE,
&c1,
cntl_sub_trsm( cntl ),
trsm_thread_sub_trsm( thread ) );
thread_ibarrier( thread );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
thread_obarrier( thread );
if( thread_am_ochief( thread ) )
bli_packm_release( b_pack, cntl_sub_packm_b( cntl ) );
if( thread_am_ichief( thread ) )
bli_packm_release( a1_pack, cntl_sub_packm_a( cntl ) );
}
void bli_trsm_blk_var2b( obj_t* a,
obj_t* b,
obj_t* c,
trsm_t* cntl,
trsm_thrinfo_t* thread )
{
obj_t a_pack_s;
obj_t b1_pack_s, c1_pack_s;
obj_t b1, c1;
obj_t* a_pack = NULL;
obj_t* b1_pack = NULL;
obj_t* c1_pack = NULL;
dim_t i;
dim_t b_alg;
dim_t n_trans;
// Initialize pack objects for A that are passed into packm_init().
if( thread_am_ochief( thread ) ) {
bli_obj_init_pack( &a_pack_s );
// Initialize object for packing A.
bli_packm_init( a, &a_pack_s,
cntl_sub_packm_a( cntl ) );
// Scale C by beta (if instructed).
bli_scalm_int( &BLIS_ONE,
c,
cntl_sub_scalm( cntl ) );
}
a_pack = thread_obroadcast( thread, &a_pack_s );
// Initialize pack objects for B and C that are passed into packm_init().
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &b1_pack_s );
bli_obj_init_pack( &c1_pack_s );
}
b1_pack = thread_ibroadcast( thread, &b1_pack_s );
c1_pack = thread_ibroadcast( thread, &c1_pack_s );
// Pack A (if instructed).
bli_packm_int( a, a_pack,
cntl_sub_packm_a( cntl ),
trmm_thread_sub_opackm( thread ) );
// Query dimension in partitioning direction.
n_trans = bli_obj_width_after_trans( *b );
dim_t start, end;
num_t dt = bli_obj_execution_datatype( *a );
bli_get_range_r2l( thread, 0, n_trans,
//bli_lcm( bli_info_get_default_nr( BLIS_TRSM, dt ),
// bli_info_get_default_mr( BLIS_TRSM, dt ) ),
bli_lcm( bli_blksz_get_nr( dt, cntl_blocksize( cntl ) ),
bli_blksz_get_mr( dt, cntl_blocksize( cntl ) ) ),
&start, &end );
// Partition along the n dimension.
for ( i = start; i < end; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_b( i, end, b,
cntl_blocksize( cntl ) );
// Acquire partitions for B1 and C1.
bli_acquire_mpart_r2l( BLIS_SUBPART1,
i, b_alg, b, &b1 );
bli_acquire_mpart_r2l( BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Initialize objects for packing A1 and B1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &b1, b1_pack,
cntl_sub_packm_b( cntl ) );
bli_packm_init( &c1, c1_pack,
cntl_sub_packm_c( cntl ) );
}
thread_ibarrier( thread );
// Pack B1 (if instructed).
bli_packm_int( &b1, b1_pack,
cntl_sub_packm_b( cntl ),
trsm_thread_sub_ipackm( thread ) );
// Pack C1 (if instructed).
bli_packm_int( &c1, c1_pack,
cntl_sub_packm_c( cntl ),
trsm_thread_sub_ipackm( thread ) );
// Perform trsm subproblem.
bli_trsm_int( &BLIS_ONE,
a_pack,
b1_pack,
&BLIS_ONE,
c1_pack,
cntl_sub_trsm( cntl ),
trsm_thread_sub_trsm( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ),
trsm_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
thread_obarrier( thread );
if( thread_am_ochief( thread ) )
bli_packm_release( a_pack, cntl_sub_packm_a( cntl ) );
if( thread_am_ichief( thread ) ) {
bli_packm_release( b1_pack, cntl_sub_packm_b( cntl ) );
bli_packm_release( c1_pack, cntl_sub_packm_c( cntl ) );
}
}
void bli_trmm_ru_ker_var2( obj_t* a,
obj_t* b,
obj_t* c,
cntx_t* cntx,
gemm_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 );
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,
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,
cntx,
thread );
}
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 );
}
