void bli_trmv_u_blk_var1( obj_t* alpha,
obj_t* a,
obj_t* x,
trmv_t* cntl )
{
obj_t a11, a11_pack;
obj_t a12;
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_f( ij, mn, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A11, A12, x1, and x2.
bli_acquire_mpart_tl2br( BLIS_SUBPART11,
ij, b_alg, a, &a11 );
bli_acquire_mpart_tl2br( BLIS_SUBPART12,
ij, b_alg, a, &a12 );
bli_acquire_vpart_f2b( BLIS_SUBPART1,
ij, b_alg, x, &x1 );
bli_acquire_vpart_f2b( 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 ) );
// x1 = alpha * triu( A11 ) * x1;
bli_trmv_int( alpha,
&a11_pack,
&x1_pack,
cntl_sub_trmv( cntl ) );
// x1 = x1 + alpha * A12 * x2;
bli_gemv_int( BLIS_NO_TRANSPOSE,
BLIS_NO_CONJUGATE,
alpha,
&a12,
&x2,
&BLIS_ONE,
&x1_pack,
cntl_sub_gemv_rp( 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_packm_release( &a11_pack, cntl_sub_packm_a11( cntl ) );
bli_packv_release( &x1_pack, cntl_sub_packv_x1( cntl ) );
}
void bli_hemv_blk_var1( conj_t conjh,
obj_t* alpha,
obj_t* a,
obj_t* x,
obj_t* beta,
obj_t* y,
cntx_t* cntx,
hemv_t* cntl )
{
obj_t a11, a11_pack;
obj_t a10;
obj_t x1, x1_pack;
obj_t x0;
obj_t y1, y1_pack;
obj_t y0;
dim_t mn;
dim_t ij;
dim_t b_alg;
// Even though this blocked algorithm is expressed only in terms of the
// lower triangular case, the upper triangular case is still supported:
// when bli_acquire_mpart_tl2br() is passed a matrix that is stored in
// in the upper triangle, and the requested subpartition resides in the
// lower triangle (as is the case for this algorithm), the routine fills
// the request as if the caller had actually requested the corresponding
// "mirror" subpartition in the upper triangle, except that it marks the
// subpartition for transposition (and conjugation).
// Initialize objects for packing.
bli_obj_init_pack( &a11_pack );
bli_obj_init_pack( &x1_pack );
bli_obj_init_pack( &y1_pack );
// Query dimension.
mn = bli_obj_length( a );
// y = beta * y;
bli_scalv_int( beta,
y,
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_f( ij, mn, a,
bli_cntl_bszid( cntl ), cntx );
// Acquire partitions for A11, A10, x1, x0, y1, and y0.
bli_acquire_mpart_tl2br( BLIS_SUBPART11,
ij, b_alg, a, &a11 );
bli_acquire_mpart_tl2br( BLIS_SUBPART10,
ij, b_alg, a, &a10 );
bli_acquire_vpart_f2b( BLIS_SUBPART1,
ij, b_alg, x, &x1 );
bli_acquire_vpart_f2b( BLIS_SUBPART0,
ij, b_alg, x, &x0 );
bli_acquire_vpart_f2b( BLIS_SUBPART1,
ij, b_alg, y, &y1 );
bli_acquire_vpart_f2b( BLIS_SUBPART0,
ij, b_alg, y, &y0 );
// Initialize objects for packing A11, x1, and y1 (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 ) );
bli_packv_init( &y1, &y1_pack,
cntx, bli_cntl_sub_packv_y1( cntl ) );
// Copy/pack A11, x1, y1 (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 ) );
bli_packv_int( &y1, &y1_pack,
cntx, bli_cntl_sub_packv_y1( cntl ) );
// y0 = y0 + alpha * A10' * x1;
bli_gemv_int( bli_apply_conj( conjh, BLIS_TRANSPOSE ),
BLIS_NO_CONJUGATE,
alpha,
&a10,
&x1_pack,
&BLIS_ONE,
&y0,
cntx,
bli_cntl_sub_gemv_t_rp( cntl ) );
// y1 = y1 + alpha * A11 * x1;
bli_hemv_int( conjh,
alpha,
&a11_pack,
&x1_pack,
&BLIS_ONE,
&y1_pack,
cntx,
bli_cntl_sub_hemv( cntl ) );
// y1 = y1 + alpha * A10 * x0;
bli_gemv_int( BLIS_NO_TRANSPOSE,
BLIS_NO_CONJUGATE,
alpha,
&a10,
&x0,
&BLIS_ONE,
&y1_pack,
cntx,
bli_cntl_sub_gemv_n_rp( cntl ) );
// Copy/unpack y1 (if y1 was packed).
bli_unpackv_int( &y1_pack, &y1,
cntx, bli_cntl_sub_unpackv_y1( 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 ) );
bli_packv_release( &y1_pack, bli_cntl_sub_packv_y1( cntl ) );
}
void bli_her2_blk_var3( conj_t conjh,
obj_t* alpha,
obj_t* alpha_conj,
obj_t* x,
obj_t* y,
obj_t* c,
her2_t* cntl )
{
obj_t c11, c11_pack;
obj_t c10;
obj_t c21;
obj_t x1, x1_pack;
obj_t y1, y1_pack;
obj_t y0;
obj_t y2;
dim_t mn;
dim_t ij;
dim_t b_alg;
// Even though this blocked algorithm is expressed only in terms of the
// lower triangular case, the upper triangular case is still supported:
// when bli_acquire_mpart_tl2br() is passed a matrix that is stored in
// in the upper triangle, and the requested subpartition resides in the
// lower triangle (as is the case for this algorithm), the routine fills
// the request as if the caller had actually requested the corresponding
// "mirror" subpartition in the upper triangle, except that it marks the
// subpartition for transposition (and conjugation).
// Initialize objects for packing.
bli_obj_init_pack( &c11_pack );
bli_obj_init_pack( &x1_pack );
bli_obj_init_pack( &y1_pack );
// Query dimension.
mn = bli_obj_length( *c );
// Partition diagonally.
for ( ij = 0; ij < mn; ij += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_f( ij, mn, c,
cntl_blocksize( cntl ) );
// Acquire partitions for C11, C10, C21, x1, y1, y0, and y2.
bli_acquire_mpart_tl2br( BLIS_SUBPART11,
ij, b_alg, c, &c11 );
bli_acquire_mpart_tl2br( BLIS_SUBPART10,
ij, b_alg, c, &c10 );
bli_acquire_mpart_tl2br( BLIS_SUBPART21,
ij, b_alg, c, &c21 );
bli_acquire_vpart_f2b( BLIS_SUBPART1,
ij, b_alg, x, &x1 );
bli_acquire_vpart_f2b( BLIS_SUBPART1,
ij, b_alg, y, &y1 );
bli_acquire_vpart_f2b( BLIS_SUBPART0,
ij, b_alg, y, &y0 );
bli_acquire_vpart_f2b( BLIS_SUBPART2,
ij, b_alg, y, &y2 );
// Initialize objects for packing C11, x1, and y1 (if needed).
bli_packm_init( &c11, &c11_pack,
cntl_sub_packm_c11( cntl ) );
bli_packv_init( &x1, &x1_pack,
cntl_sub_packv_x1( cntl ) );
bli_packv_init( &y1, &y1_pack,
cntl_sub_packv_y1( cntl ) );
// Copy/pack C11, x1, y1 (if needed).
bli_packm_int( &c11, &c11_pack,
cntl_sub_packm_c11( cntl ),
&BLIS_PACKM_SINGLE_THREADED );
bli_packv_int( &x1, &x1_pack,
cntl_sub_packv_x1( cntl ) );
bli_packv_int( &y1, &y1_pack,
cntl_sub_packv_y1( cntl ) );
// C10 = C10 + alpha * x1 * y0';
bli_ger_int( BLIS_NO_CONJUGATE,
conjh,
alpha,
&x1_pack,
&y0,
&c10,
cntl_sub_ger_rp( cntl ) );
// C21 = C21 + conj(alpha) * y2 * x1';
bli_ger_int( BLIS_NO_CONJUGATE,
conjh,
alpha_conj,
&y2,
&x1_pack,
&c21,
cntl_sub_ger_cp( cntl ) );
// C11 = C11 + alpha * x1 * y1' + conj(alpha) * y1 * x1';
bli_her2_int( conjh,
alpha,
alpha_conj,
&x1_pack,
&y1_pack,
&c11_pack,
cntl_sub_her2( cntl ) );
// Copy/unpack C11 (if C11 was packed).
bli_unpackm_int( &c11_pack, &c11,
cntl_sub_unpackm_c11( cntl ),
&BLIS_PACKM_SINGLE_THREADED );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
bli_obj_release_pack( &c11_pack );
bli_obj_release_pack( &x1_pack );
bli_obj_release_pack( &y1_pack );
}
