void bli_ger_blk_var1( obj_t* alpha,
obj_t* x,
obj_t* y,
obj_t* a,
ger_t* cntl )
{
obj_t a1, a1_pack;
obj_t x1, x1_pack;
dim_t i;
dim_t b_alg;
dim_t m_trans;
// Initialize objects for packing.
bli_obj_init_pack( &a1_pack );
bli_obj_init_pack( &x1_pack );
// Query dimension in partitioning direction.
m_trans = bli_obj_length_after_trans( *a );
// Partition along the m dimension.
for ( i = 0; i < m_trans; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_f( i, m_trans, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and x1.
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_vpart_f2b( BLIS_SUBPART1,
i, b_alg, x, &x1 );
// Initialize objects for packing A1 and x1 (if needed).
bli_packm_init( &a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
bli_packv_init( &x1, &x1_pack,
cntl_sub_packv_x( cntl ) );
// Copy/pack A1, x1 (if needed).
bli_packm_int( &a1, &a1_pack,
cntl_sub_packm_a( cntl ),
&BLIS_PACKM_SINGLE_THREADED );
bli_packv_int( &x1, &x1_pack,
cntl_sub_packv_x( cntl ) );
// A1 = A1 + alpha * x1 * y;
bli_ger_int( BLIS_NO_CONJUGATE,
BLIS_NO_CONJUGATE,
alpha,
&x1_pack,
y,
&a1_pack,
cntl_sub_ger( cntl ) );
// Copy/unpack A1 (if A1 was packed).
bli_unpackm_int( &a1_pack, &a1,
cntl_sub_unpackm_a( cntl ),
&BLIS_PACKM_SINGLE_THREADED );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
bli_packm_release( &a1_pack, cntl_sub_packm_a( cntl ) );
bli_packv_release( &x1_pack, cntl_sub_packv_x( cntl ) );
}
void bli_gemv_blk_var2( obj_t* alpha,
obj_t* a,
obj_t* x,
obj_t* beta,
obj_t* y,
gemv_t* cntl )
{
obj_t a1, a1_pack;
obj_t x1, x1_pack;
dim_t n_trans;
dim_t i;
dim_t b_alg;
// Initialize objects for packing.
bli_obj_init_pack( &a1_pack );
bli_obj_init_pack( &x1_pack );
// Query dimension in partitioning direction.
n_trans = bli_obj_width_after_trans( *a );
// y = beta * y;
bli_scalv_int( beta,
y,
cntl_sub_scalv( cntl ) );
// Partition along the "k" dimension (n dimension of A).
for ( i = 0; i < n_trans; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_f( i, n_trans,
a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and x1.
bli_acquire_mpart_l2r( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_vpart_f2b( BLIS_SUBPART1,
i, b_alg, x, &x1 );
// Initialize objects for packing A1 and x1 (if needed).
bli_packm_init( &a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
bli_packv_init( &x1, &x1_pack,
cntl_sub_packv_x( cntl ) );
// Copy/pack A1, x1 (if needed).
bli_packm_int( alpha,
&a1,
&a1_pack,
cntl_sub_packm_a( cntl ) );
bli_packv_int( &x1,
&x1_pack,
cntl_sub_packv_x( cntl ) );
// y = y + alpha * A1 * x1;
bli_gemv_int( BLIS_NO_TRANSPOSE,
BLIS_NO_CONJUGATE,
alpha,
&a1_pack,
&x1_pack,
&BLIS_ONE,
y,
cntl_sub_gemv( 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( &x1_pack );
}