void bli_gemm_blk_var1
(
obj_t* a,
obj_t* b,
obj_t* c,
cntx_t* cntx,
cntl_t* cntl,
thrinfo_t* thread
)
{
obj_t a1, c1;
dir_t direct;
dim_t i;
dim_t b_alg;
dim_t my_start, my_end;
// Determine the direction in which to partition (forwards or backwards).
direct = bli_l3_direct( a, b, c, cntl );
// Prune any zero region that exists along the partitioning dimension.
bli_l3_prune_unref_mparts_m( a, b, c, cntl );
// Determine the current thread's subpartition range.
bli_thread_get_range_mdim
(
direct, thread, a, b, c, cntl, cntx,
&my_start, &my_end
);
// Partition along the m dimension.
for ( i = my_start; i < my_end; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize( direct, i, my_end, a,
bli_cntl_bszid( cntl ), cntx );
// Acquire partitions for A1 and C1.
bli_acquire_mpart_mdim( direct, BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_mdim( direct, BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Perform gemm subproblem.
bli_gemm_int
(
&BLIS_ONE,
&a1,
b,
&BLIS_ONE,
&c1,
cntx,
bli_cntl_sub_node( cntl ),
bli_thrinfo_sub_node( thread )
);
}
}
void blx_gemm_blk_var3
(
obj_t* a,
obj_t* b,
obj_t* c,
cntx_t* cntx,
rntm_t* rntm,
cntl_t* cntl,
thrinfo_t* thread
)
{
obj_t a1, b1;
dim_t i;
dim_t b_alg;
dim_t k_trans;
// Query dimension in partitioning direction.
k_trans = bli_obj_width_after_trans( a );
// Partition along the k dimension.
for ( i = 0; i < k_trans; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = blx_determine_blocksize_f( i, k_trans, c,
bli_cntl_bszid( cntl ), cntx );
// Acquire partitions for A1 and B1.
bli_acquire_mpart_ndim( BLIS_FWD, BLIS_SUBPART1, i, b_alg, a, &a1 );
bli_acquire_mpart_mdim( BLIS_FWD, BLIS_SUBPART1, i, b_alg, b, &b1 );
// Perform gemm subproblem.
blx_gemm_int
(
&a1, &b1, c, cntx, rntm,
bli_cntl_sub_node( cntl ),
bli_thrinfo_sub_node( thread )
);
bli_thread_obarrier( bli_thrinfo_sub_node( thread ) );
// This variant executes multiple rank-k updates. Therefore, if the
// internal beta scalar on matrix C is non-zero, we must use it
// only for the first iteration (and then BLIS_ONE for all others).
// And since c is a locally aliased obj_t, we can simply overwrite
// the internal beta scalar with BLIS_ONE once it has been used in
// the first iteration.
if ( i == 0 ) bli_obj_scalar_reset( c );
}
}
void bli_trsm_blk_var3
(
obj_t* a,
obj_t* b,
obj_t* c,
cntx_t* cntx,
cntl_t* cntl,
thrinfo_t* thread
)
{
obj_t a1, b1;
dir_t direct;
dim_t i;
dim_t b_alg;
dim_t k_trans;
// Determine the direction in which to partition (forwards or backwards).
direct = bli_l3_direct( a, b, c, cntl );
// Prune any zero region that exists along the partitioning dimension.
bli_l3_prune_unref_mparts_k( a, b, c, cntl );
// Query dimension in partitioning direction.
k_trans = bli_obj_width_after_trans( *a );
// Partition along the k dimension.
for ( i = 0; i < k_trans; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_trsm_determine_kc( direct, i, k_trans, a, b,
bli_cntl_bszid( cntl ), cntx );
// Acquire partitions for A1 and B1.
bli_acquire_mpart_ndim( direct, BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_mdim( direct, BLIS_SUBPART1,
i, b_alg, b, &b1 );
// Perform trsm subproblem.
bli_trsm_int
(
&BLIS_ONE,
&a1,
&b1,
&BLIS_ONE,
c,
cntx,
bli_cntl_sub_node( cntl ),
bli_thrinfo_sub_node( thread )
);
//bli_thread_ibarrier( thread );
bli_thread_obarrier( bli_thrinfo_sub_node( thread ) );
// This variant executes multiple rank-k updates. Therefore, if the
// internal alpha scalars on A/B and C are non-zero, we must ensure
// that they are only used in the first iteration.
if ( i == 0 )
{
bli_obj_scalar_reset( a ); bli_obj_scalar_reset( b );
bli_obj_scalar_reset( c );
}
}
}