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_gemm_packb
(
obj_t* a,
obj_t* b,
obj_t* c,
cntx_t* cntx,
cntl_t* cntl,
thrinfo_t* thread
)
{
obj_t b_pack;
// Pack matrix B according to the control tree node.
bli_l3_packm
(
b,
&b_pack,
cntx,
cntl,
thread
);
// Proceed with execution using packed matrix B.
bli_gemm_int
(
&BLIS_ONE,
a,
&b_pack,
&BLIS_ONE,
c,
cntx,
bli_cntl_sub_node( cntl ),
bli_thrinfo_sub_node( thread )
);
}
void bli_l3_thrinfo_free
(
thrinfo_t* thread
)
{
if ( thread == NULL ||
thread == &BLIS_PACKM_SINGLE_THREADED ||
thread == &BLIS_GEMM_SINGLE_THREADED
) return;
thrinfo_t* thrinfo_sub_node = bli_thrinfo_sub_node( thread );
// Free the communicators, but only if the current thrinfo_t struct
// is marked as needing them to be freed. The most common example of
// thrinfo_t nodes NOT marked as needing their comms freed are those
// associated with packm thrinfo_t nodes.
if ( bli_thrinfo_needs_free_comm( thread ) )
{
// The ochief always frees his communicator, and the ichief free its
// communicator if we are at the leaf node.
if ( bli_thread_am_ochief( thread ) )
bli_thrcomm_free( bli_thrinfo_ocomm( thread ) );
}
// Free all children of the current thrinfo_t.
bli_l3_thrinfo_free( thrinfo_sub_node );
// Free the thrinfo_t struct.
bli_free_intl( thread );
}
void bli_gemm_blk_var2
(
obj_t* a,
obj_t* b,
obj_t* c,
cntx_t* cntx,
rntm_t* rntm,
cntl_t* cntl,
thrinfo_t* thread
)
{
obj_t b1, 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_n( a, b, c, cntl );
// Determine the current thread's subpartition range.
bli_thread_range_ndim
(
direct, thread, a, b, c, cntl, cntx,
&my_start, &my_end
);
// Partition along the n 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, b,
bli_cntl_bszid( cntl ), cntx );
// Acquire partitions for B1 and C1.
bli_acquire_mpart_ndim( direct, BLIS_SUBPART1,
i, b_alg, b, &b1 );
bli_acquire_mpart_ndim( direct, BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Perform gemm subproblem.
bli_gemm_int
(
&BLIS_ONE,
a,
&b1,
&BLIS_ONE,
&c1,
cntx,
rntm,
bli_cntl_sub_node( cntl ),
bli_thrinfo_sub_node( thread )
);
}
}
void bli_l3_thrinfo_print_paths
(
thrinfo_t** threads
)
{
dim_t n_threads = bli_thread_num_threads( threads[0] );
dim_t gl_comm_id;
thrinfo_t* jc_info = threads[0];
thrinfo_t* pc_info = bli_thrinfo_sub_node( jc_info );
thrinfo_t* pb_info = bli_thrinfo_sub_node( pc_info );
thrinfo_t* ic_info = bli_thrinfo_sub_node( pb_info );
thrinfo_t* pa_info = bli_thrinfo_sub_node( ic_info );
thrinfo_t* jr_info = bli_thrinfo_sub_node( pa_info );
thrinfo_t* ir_info = bli_thrinfo_sub_node( jr_info );
dim_t jc_way = bli_thread_n_way( jc_info );
dim_t pc_way = bli_thread_n_way( pc_info );
dim_t pb_way = bli_thread_n_way( pb_info );
dim_t ic_way = bli_thread_n_way( ic_info );
dim_t pa_way = bli_thread_n_way( pa_info );
dim_t jr_way = bli_thread_n_way( jr_info );
dim_t ir_way = bli_thread_n_way( ir_info );
dim_t gl_nt = bli_thread_num_threads( jc_info );
dim_t jc_nt = bli_thread_num_threads( pc_info );
dim_t pc_nt = bli_thread_num_threads( pb_info );
dim_t pb_nt = bli_thread_num_threads( ic_info );
dim_t ic_nt = bli_thread_num_threads( pa_info );
dim_t pa_nt = bli_thread_num_threads( jr_info );
dim_t jr_nt = bli_thread_num_threads( ir_info );
printf( " gl jc kc pb ic pa jr ir\n" );
printf( "xx_nt: %4lu %4lu %4lu %4lu %4lu %4lu %4lu %4lu\n",
( unsigned long )gl_nt,
( unsigned long )jc_nt,
( unsigned long )pc_nt,
( unsigned long )pb_nt,
( unsigned long )ic_nt,
( unsigned long )pa_nt,
( unsigned long )jr_nt,
( unsigned long )1 );
printf( "\n" );
printf( " jc kc pb ic pa jr ir\n" );
printf( "xx_way: %4lu %4lu %4lu %4lu %4lu %4lu %4lu\n",
( unsigned long )jc_way,
( unsigned long )pc_way,
( unsigned long )pb_way,
( unsigned long )ic_way,
( unsigned long )pa_way,
( unsigned long )jr_way,
( unsigned long )ir_way );
printf( "=================================================\n" );
for ( gl_comm_id = 0; gl_comm_id < n_threads; ++gl_comm_id )
{
jc_info = threads[gl_comm_id];
pc_info = bli_thrinfo_sub_node( jc_info );
pb_info = bli_thrinfo_sub_node( pc_info );
ic_info = bli_thrinfo_sub_node( pb_info );
pa_info = bli_thrinfo_sub_node( ic_info );
jr_info = bli_thrinfo_sub_node( pa_info );
ir_info = bli_thrinfo_sub_node( jr_info );
dim_t gl_comm_id = bli_thread_ocomm_id( jc_info );
dim_t jc_comm_id = bli_thread_ocomm_id( pc_info );
dim_t pc_comm_id = bli_thread_ocomm_id( pb_info );
dim_t pb_comm_id = bli_thread_ocomm_id( ic_info );
dim_t ic_comm_id = bli_thread_ocomm_id( pa_info );
dim_t pa_comm_id = bli_thread_ocomm_id( jr_info );
dim_t jr_comm_id = bli_thread_ocomm_id( ir_info );
dim_t jc_work_id = bli_thread_work_id( jc_info );
dim_t pc_work_id = bli_thread_work_id( pc_info );
dim_t pb_work_id = bli_thread_work_id( pb_info );
dim_t ic_work_id = bli_thread_work_id( ic_info );
dim_t pa_work_id = bli_thread_work_id( pa_info );
dim_t jr_work_id = bli_thread_work_id( jr_info );
dim_t ir_work_id = bli_thread_work_id( ir_info );
printf( " gl jc pb kc pa ic jr \n" );
printf( "comm ids: %4lu %4lu %4lu %4lu %4lu %4lu %4lu\n",
( unsigned long )gl_comm_id,
( unsigned long )jc_comm_id,
( unsigned long )pc_comm_id,
( unsigned long )pb_comm_id,
( unsigned long )ic_comm_id,
( unsigned long )pa_comm_id,
( unsigned long )jr_comm_id );
printf( "work ids: %4ld %4ld %4lu %4lu %4ld %4ld %4ld\n",
( unsigned long )jc_work_id,
( unsigned long )pc_work_id,
( unsigned long )pb_work_id,
( unsigned long )ic_work_id,
( unsigned long )pa_work_id,
( unsigned long )jr_work_id,
( unsigned long )ir_work_id );
printf( "---------------------------------------\n" );
}
}
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 );
}
}
}