>Replace the per-object rte_node_enqueue_x1() calls with batched >rte_node_enqueue() calls. The function now tracks runs of consecutive >objects going to the same edge and flushes them in bulk. > >When all objects go to the same edge and come from the node's own >buffer (objs == node->objs), use rte_node_next_stream_move() which >swaps pointers instead of copying. > >Signed-off-by: Robin Jarry <[email protected]> >--- > lib/graph/rte_graph_worker_common.h | 15 +++++++++++++-- > 1 file changed, 13 insertions(+), 2 deletions(-) > >diff --git a/lib/graph/rte_graph_worker_common.h >b/lib/graph/rte_graph_worker_common.h >index 4ab53a533e4c..7fda67c07169 100644 >--- a/lib/graph/rte_graph_worker_common.h >+++ b/lib/graph/rte_graph_worker_common.h >@@ -432,10 +432,21 @@ static inline void > rte_node_enqueue_next(struct rte_graph *graph, struct rte_node *node, > rte_edge_t *nexts, void **objs, uint16_t nb_objs) > { >+ rte_edge_t last = nexts[0]; >+ uint16_t run_start = 0; > uint16_t i; > >- for (i = 0; i < nb_objs; i++) >- rte_node_enqueue_x1(graph, node, nexts[i], objs[i]); >+ for (i = 1; i < nb_objs; i++) { >+ if (nexts[i] != last) {
We can probably use SIMD here for comparision, 128b would process 8 nexts at a time. But might not be worth it if nexts are too unique. >+ rte_node_enqueue(graph, node, last, &objs[run_start], >i - run_start); >+ run_start = i; >+ last = nexts[i]; >+ } >+ } >+ if (run_start == 0 && objs == node->objs) >+ rte_node_next_stream_move(graph, node, last); >+ else >+ rte_node_enqueue(graph, node, last, &objs[run_start], nb_objs >- run_start); > } > > /** -- 2.52.0
