void bli_syr2_front
(
obj_t* alpha,
obj_t* x,
obj_t* y,
obj_t* c,
cntx_t* cntx
)
{
her2_t* her2_cntl;
num_t dt_targ_x;
num_t dt_targ_y;
//num_t dt_targ_c;
bool_t x_has_unit_inc;
bool_t y_has_unit_inc;
bool_t c_has_unit_inc;
obj_t alpha_local;
num_t dt_alpha;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_syr2_check( alpha, x, y, c );
// Query the target datatypes of each object.
dt_targ_x = bli_obj_target_dt( x );
dt_targ_y = bli_obj_target_dt( y );
//dt_targ_c = bli_obj_target_dt( c );
// Determine whether each operand with unit stride.
x_has_unit_inc = ( bli_obj_vector_inc( x ) == 1 );
y_has_unit_inc = ( bli_obj_vector_inc( y ) == 1 );
c_has_unit_inc = ( bli_obj_is_row_stored( c ) ||
bli_obj_is_col_stored( c ) );
// Create an object to hold a copy-cast of alpha. Notice that we use
// the type union of the datatypes of x and y.
dt_alpha = bli_dt_union( dt_targ_x, dt_targ_y );
bli_obj_scalar_init_detached_copy_of( dt_alpha,
BLIS_NO_CONJUGATE,
alpha,
&alpha_local );
// If all operands have unit stride, we choose a control tree for calling
// the unblocked implementation directly without any blocking.
if ( x_has_unit_inc &&
y_has_unit_inc &&
c_has_unit_inc )
{
// We use two control trees to handle the four cases corresponding to
// combinations of upper/lower triangular storage and row/column-storage.
// The row-stored lower triangular and column-stored upper triangular
// trees are identical. Same for the remaining two trees.
if ( bli_obj_is_lower( c ) )
{
if ( bli_obj_is_row_stored( c ) ) her2_cntl = her2_cntl_bs_ke_lrow_ucol;
else her2_cntl = her2_cntl_bs_ke_lcol_urow;
}
else // if ( bli_obj_is_upper( c ) )
{
if ( bli_obj_is_row_stored( c ) ) her2_cntl = her2_cntl_bs_ke_lcol_urow;
else her2_cntl = her2_cntl_bs_ke_lrow_ucol;
}
}
else
{
// Mark objects with unit stride as already being packed. This prevents
// unnecessary packing from happening within the blocked algorithm.
if ( x_has_unit_inc ) bli_obj_set_pack_schema( BLIS_PACKED_VECTOR, x );
if ( y_has_unit_inc ) bli_obj_set_pack_schema( BLIS_PACKED_VECTOR, y );
if ( c_has_unit_inc ) bli_obj_set_pack_schema( BLIS_PACKED_UNSPEC, c );
// Here, we make a similar choice as above, except that (1) we look
// at storage tilt, and (2) we choose a tree that performs blocking.
if ( bli_obj_is_lower( c ) )
{
if ( bli_obj_is_row_stored( c ) ) her2_cntl = her2_cntl_ge_lrow_ucol;
else her2_cntl = her2_cntl_ge_lcol_urow;
}
else // if ( bli_obj_is_upper( c ) )
{
if ( bli_obj_is_row_stored( c ) ) her2_cntl = her2_cntl_ge_lcol_urow;
else her2_cntl = her2_cntl_ge_lrow_ucol;
}
}
// Invoke the internal back-end with the copy-cast scalar and the
// chosen control tree. Set conjh to BLIS_NO_CONJUGATE to invoke the
// symmetric (and not Hermitian) algorithms.
bli_her2_int( BLIS_NO_CONJUGATE,
&alpha_local,
&alpha_local,
x,
y,
c,
cntx,
her2_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 );
}
