[dpdk-dev] [PATCH v4 08/10] event/dsw: let DSW event device sort events on dequeue

["Mattias Rönnblom" <mattias.ronnblom@ericsson.com>]

With this patch, the DSW event device will (optionally) sort the event
burst before giving it to the application. The sorting will primarily
be on queue id, and secondary on flow id.

The sorting is an attempt to optimize data and instruction cache usage
for the application, at the cost of additional event device overhead.

Signed-off-by: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
---
 drivers/event/dsw/dsw_evdev.h | 11 ++++++++
 drivers/event/dsw/dsw_event.c | 23 +++++++++++++++++
 drivers/event/dsw/dsw_sort.h  | 48 +++++++++++++++++++++++++++++++++++
 3 files changed, 82 insertions(+)
 create mode 100644 drivers/event/dsw/dsw_sort.h

diff --git a/drivers/event/dsw/dsw_evdev.h b/drivers/event/dsw/dsw_evdev.h
index 783c418bf..f6f8f0454 100644
--- a/drivers/event/dsw/dsw_evdev.h
+++ b/drivers/event/dsw/dsw_evdev.h
@@ -93,6 +93,17 @@
  */
 #define DSW_CTL_IN_RING_SIZE ((DSW_MAX_PORTS-1)*4)
 
+/* With DSW_SORT_DEQUEUED enabled, the scheduler will, at the point of
+ * dequeue(), arrange events so that events with the same flow id on
+ * the same queue forms a back-to-back "burst", and also so that such
+ * bursts of different flow ids, but on the same queue, also come
+ * consecutively. All this in an attempt to improve data and
+ * instruction cache usage for the application, at the cost of a
+ * scheduler overhead increase.
+ */
+
+/* #define DSW_SORT_DEQUEUED */
+
 struct dsw_queue_flow {
 	uint8_t queue_id;
 	uint16_t flow_hash;
diff --git a/drivers/event/dsw/dsw_event.c b/drivers/event/dsw/dsw_event.c
index f0347592d..a84b19c33 100644
--- a/drivers/event/dsw/dsw_event.c
+++ b/drivers/event/dsw/dsw_event.c
@@ -4,6 +4,10 @@
 
 #include "dsw_evdev.h"
 
+#ifdef DSW_SORT_DEQUEUED
+#include "dsw_sort.h"
+#endif
+
 #include <stdbool.h>
 #include <string.h>
 
@@ -1121,6 +1125,21 @@ dsw_port_record_seen_events(struct dsw_port *port, struct rte_event *events,
 				DSW_MAX_EVENTS_RECORDED);
 }
 
+#ifdef DSW_SORT_DEQUEUED
+
+#define DSW_EVENT_TO_INT(_event)				\
+	((int)((((_event)->queue_id)<<16)|((_event)->flow_id)))
+
+static inline int
+dsw_cmp_event(const void *v_event_a, const void *v_event_b)
+{
+	const struct rte_event *event_a = v_event_a;
+	const struct rte_event *event_b = v_event_b;
+
+	return DSW_EVENT_TO_INT(event_a) - DSW_EVENT_TO_INT(event_b);
+}
+#endif
+
 static uint16_t
 dsw_port_dequeue_burst(struct dsw_port *port, struct rte_event *events,
 		       uint16_t num)
@@ -1191,5 +1210,9 @@ dsw_event_dequeue_burst(void *port, struct rte_event *events, uint16_t num,
 	 *	0.
 	 */
 
+#ifdef DSW_SORT_DEQUEUED
+	dsw_stable_sort(events, dequeued, sizeof(events[0]), dsw_cmp_event);
+#endif
+
 	return dequeued;
 }
diff --git a/drivers/event/dsw/dsw_sort.h b/drivers/event/dsw/dsw_sort.h
new file mode 100644
index 000000000..609767fdf
--- /dev/null
+++ b/drivers/event/dsw/dsw_sort.h
@@ -0,0 +1,48 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2018 Ericsson AB
+ */
+
+#ifndef _DSW_SORT_
+#define _DSW_SORT_
+
+#include <string.h>
+
+#include <rte_common.h>
+
+#define DSW_ARY_ELEM_PTR(_ary, _idx, _elem_size)	\
+	RTE_PTR_ADD(_ary, (_idx) * (_elem_size))
+
+#define DSW_ARY_ELEM_SWAP(_ary, _a_idx, _b_idx, _elem_size)		\
+	do {								\
+		char tmp[_elem_size];					\
+		void *_a_ptr = DSW_ARY_ELEM_PTR(_ary, _a_idx, _elem_size); \
+		void *_b_ptr = DSW_ARY_ELEM_PTR(_ary, _b_idx, _elem_size); \
+		memcpy(tmp, _a_ptr, _elem_size);			\
+		memcpy(_a_ptr, _b_ptr, _elem_size);			\
+		memcpy(_b_ptr, tmp, _elem_size);			\
+	} while (0)
+
+static inline void
+dsw_insertion_sort(void *ary, uint16_t len, uint16_t elem_size,
+		   int (*cmp_fn)(const void *, const void *))
+{
+	uint16_t i;
+
+	for (i = 1; i < len; i++) {
+		uint16_t j;
+		for (j = i; j > 0 &&
+			     cmp_fn(DSW_ARY_ELEM_PTR(ary, j-1, elem_size),
+				    DSW_ARY_ELEM_PTR(ary, j, elem_size)) > 0;
+		     j--)
+			DSW_ARY_ELEM_SWAP(ary, j, j-1, elem_size);
+	}
+}
+
+static inline void
+dsw_stable_sort(void *ary, uint16_t len, uint16_t elem_size,
+		int (*cmp_fn)(const void *, const void *))
+{
+	dsw_insertion_sort(ary, len, elem_size, cmp_fn);
+}
+
+#endif
-- 
2.17.1