err_t bli_check_triangular_object( obj_t* a )
{
err_t e_val = BLIS_SUCCESS;
if ( !bli_obj_is_triangular( *a ) )
e_val = BLIS_EXPECTED_TRIANGULAR_OBJECT;
return e_val;
}
void bli_l3_cntl_create_if
(
obj_t* a,
obj_t* b,
obj_t* c,
cntx_t* cntx,
cntl_t* cntl_orig,
cntl_t** cntl_use
)
{
// If the control tree pointer is NULL, we construct a default
// tree as a function of the operation family.
if ( cntl_orig == NULL )
{
opid_t family = bli_cntx_get_family( cntx );
if ( family == BLIS_GEMM ||
family == BLIS_HERK ||
family == BLIS_TRMM )
{
*cntl_use = bli_gemm_cntl_create( family );
}
else // if ( family == BLIS_TRSM )
{
side_t side;
if ( bli_obj_is_triangular( *a ) ) side = BLIS_LEFT;
else side = BLIS_RIGHT;
*cntl_use = bli_trsm_cntl_create( side );
}
}
else
{
// If the user provided a control tree, create a copy and use it
// instead (so that it can be used to cache things like pack mem_t
// entries).
*cntl_use = bli_cntl_copy( cntl_orig );
}
}
void bli_trmm_lu_blk_var1( obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
trmm_t* cntl )
{
obj_t a1, a1_pack;
obj_t b_pack;
obj_t c1, c1_pack;
dim_t i;
dim_t b_alg;
dim_t mT_trans;
// Initialize all pack objects that are passed into packm_init().
bli_obj_init_pack( &a1_pack );
bli_obj_init_pack( &b_pack );
bli_obj_init_pack( &c1_pack );
// If A is [upper] triangular, use the diagonal offset of A to determine
// the length of the non-zero region.
if ( bli_obj_is_triangular( *a ) )
mT_trans = bli_abs( bli_obj_diag_offset_after_trans( *a ) ) +
bli_obj_width_after_trans( *a );
else // if ( bli_obj_is_general( *a )
mT_trans = bli_obj_length_after_trans( *a );
// Scale C by beta (if instructed).
bli_scalm_int( beta,
c,
cntl_sub_scalm( cntl ) );
// Initialize object for packing B.
bli_packm_init( b, &b_pack,
cntl_sub_packm_b( cntl ) );
// Pack B and scale by alpha (if instructed).
bli_packm_int( alpha,
b, &b_pack,
cntl_sub_packm_b( cntl ) );
// Partition along the m dimension.
for ( i = 0; i < mT_trans; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_f( i, mT_trans, 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 objects for packing A1 and C1.
bli_packm_init( &a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
bli_packm_init( &c1, &c1_pack,
cntl_sub_packm_c( cntl ) );
// Pack A1 and scale by alpha (if instructed).
bli_packm_int( alpha,
&a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
// Pack C1 and scale by beta (if instructed).
bli_packm_int( beta,
&c1, &c1_pack,
cntl_sub_packm_c( cntl ) );
// Perform trmm subproblem.
bli_trmm_int( BLIS_LEFT,
alpha,
&a1_pack,
&b_pack,
beta,
&c1_pack,
cntl_sub_trmm( cntl ) );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( &c1_pack, &c1,
cntl_sub_unpackm_c( cntl ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
bli_obj_release_pack( &a1_pack );
bli_obj_release_pack( &b_pack );
bli_obj_release_pack( &c1_pack );
}