void bli_trmv_l_blk_var2( obj_t* alpha,
obj_t* a,
obj_t* x,
trmv_t* cntl )
{
obj_t a11, a11_pack;
obj_t a21;
obj_t x1, x1_pack;
obj_t x2;
dim_t mn;
dim_t ij;
dim_t b_alg;
// Initialize objects for packing.
bli_obj_init_pack( &a11_pack );
bli_obj_init_pack( &x1_pack );
// Query dimension.
mn = bli_obj_length( *a );
// Partition diagonally.
for ( ij = 0; ij < mn; ij += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_b( ij, mn, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A11, A21, x1, and x2.
bli_acquire_mpart_br2tl( BLIS_SUBPART11,
ij, b_alg, a, &a11 );
bli_acquire_mpart_br2tl( BLIS_SUBPART21,
ij, b_alg, a, &a21 );
bli_acquire_vpart_b2f( BLIS_SUBPART1,
ij, b_alg, x, &x1 );
bli_acquire_vpart_b2f( BLIS_SUBPART2,
ij, b_alg, x, &x2 );
// Initialize objects for packing A11 and x1 (if needed).
bli_packm_init( &a11, &a11_pack,
cntl_sub_packm_a11( cntl ) );
bli_packv_init( &x1, &x1_pack,
cntl_sub_packv_x1( cntl ) );
// Copy/pack A11, x1 (if needed).
bli_packm_int( &a11, &a11_pack,
cntl_sub_packm_a11( cntl ),
&BLIS_PACKM_SINGLE_THREADED );
bli_packv_int( &x1, &x1_pack,
cntl_sub_packv_x1( cntl ) );
// x2 = x2 + alpha * A21 * x1;
bli_gemv_int( BLIS_NO_TRANSPOSE,
BLIS_NO_CONJUGATE,
alpha,
&a21,
&x1_pack,
&BLIS_ONE,
&x2,
cntl_sub_gemv_cp( cntl ) );
// x1 = alpha * tril( A11 ) * x1;
bli_trmv_int( alpha,
&a11_pack,
&x1_pack,
cntl_sub_trmv( cntl ) );
// Copy/unpack x1 (if x1 was packed).
bli_unpackv_int( &x1_pack, &x1,
cntl_sub_unpackv_x1( cntl ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
bli_obj_release_pack( &a11_pack );
bli_obj_release_pack( &x1_pack );
}
void bli_trsv_u_blk_var2( obj_t* alpha,
obj_t* a,
obj_t* x,
cntx_t* cntx,
trsv_t* cntl )
{
obj_t a11, a11_pack;
obj_t a01;
obj_t x1, x1_pack;
obj_t x0;
dim_t mn;
dim_t ij;
dim_t b_alg;
// Initialize objects for packing.
bli_obj_init_pack( &a11_pack );
bli_obj_init_pack( &x1_pack );
// Query dimension.
mn = bli_obj_length( a );
// x = alpha * x;
bli_scalv_int( alpha,
x,
cntx, bli_cntl_sub_scalv( cntl ) );
// Partition diagonally.
for ( ij = 0; ij < mn; ij += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_b( ij, mn, a,
bli_cntl_bszid( cntl ), cntx );
// Acquire partitions for A11, A01, x1, and x0.
bli_acquire_mpart_br2tl( BLIS_SUBPART11,
ij, b_alg, a, &a11 );
bli_acquire_mpart_br2tl( BLIS_SUBPART01,
ij, b_alg, a, &a01 );
bli_acquire_vpart_b2f( BLIS_SUBPART1,
ij, b_alg, x, &x1 );
bli_acquire_vpart_b2f( BLIS_SUBPART0,
ij, b_alg, x, &x0 );
// Initialize objects for packing A11 and x1 (if needed).
bli_packm_init( &a11, &a11_pack,
cntx, bli_cntl_sub_packm_a11( cntl ) );
bli_packv_init( &x1, &x1_pack,
cntx, bli_cntl_sub_packv_x1( cntl ) );
// Copy/pack A11, x1 (if needed).
bli_packm_int( &a11, &a11_pack,
cntx, bli_cntl_sub_packm_a11( cntl ),
&BLIS_PACKM_SINGLE_THREADED );
bli_packv_int( &x1, &x1_pack,
cntx, bli_cntl_sub_packv_x1( cntl ) );
// x1 = x1 / tril( A11 );
bli_trsv_int( &BLIS_ONE,
&a11_pack,
&x1_pack,
cntx,
bli_cntl_sub_trsv( cntl ) );
// x0 = x0 - A01 * x1;
bli_gemv_int( BLIS_NO_TRANSPOSE,
BLIS_NO_CONJUGATE,
&BLIS_MINUS_ONE,
&a01,
&x1_pack,
&BLIS_ONE,
&x0,
cntx,
bli_cntl_sub_gemv_cp( cntl ) );
// Copy/unpack x1 (if x1 was packed).
bli_unpackv_int( &x1_pack, &x1,
cntx, bli_cntl_sub_unpackv_x1( cntl ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
bli_packm_release( &a11_pack, bli_cntl_sub_packm_a11( cntl ) );
bli_packv_release( &x1_pack, bli_cntl_sub_packv_x1( cntl ) );
}