1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2017 Intel Corporation
3 */
4
5 #include <rte_mbuf.h>
6 #include "distributor_private.h"
7 #include "smmintrin.h"
8
9
10 void
find_match_vec(struct rte_distributor * d,uint16_t * data_ptr,uint16_t * output_ptr)11 find_match_vec(struct rte_distributor *d,
12 uint16_t *data_ptr,
13 uint16_t *output_ptr)
14 {
15 /* Setup */
16 __m128i incoming_fids;
17 __m128i inflight_fids;
18 __m128i preflight_fids;
19 __m128i wkr;
20 __m128i mask1;
21 __m128i mask2;
22 __m128i output;
23 struct rte_distributor_backlog *bl;
24 uint16_t i;
25
26 /*
27 * Function overview:
28 * 2. Loop through all worker ID's
29 * 2a. Load the current inflights for that worker into an xmm reg
30 * 2b. Load the current backlog for that worker into an xmm reg
31 * 2c. use cmpestrm to intersect flow_ids with backlog and inflights
32 * 2d. Add any matches to the output
33 * 3. Write the output xmm (matching worker ids).
34 */
35
36
37 output = _mm_set1_epi16(0);
38 incoming_fids = _mm_load_si128((__m128i *)data_ptr);
39
40 for (i = 0; i < d->num_workers; i++) {
41 bl = &d->backlog[i];
42
43 inflight_fids =
44 _mm_load_si128((__m128i *)&(d->in_flight_tags[i]));
45 preflight_fids =
46 _mm_load_si128((__m128i *)(bl->tags));
47
48 /*
49 * Any incoming_fid that exists anywhere in inflight_fids will
50 * have 0xffff in same position of the mask as the incoming fid
51 * Example (shortened to bytes for brevity):
52 * incoming_fids 0x01 0x02 0x03 0x04 0x05 0x06 0x07 0x08
53 * inflight_fids 0x03 0x05 0x07 0x00 0x00 0x00 0x00 0x00
54 * mask 0x00 0x00 0xff 0x00 0xff 0x00 0xff 0x00
55 */
56
57 mask1 = _mm_cmpestrm(inflight_fids, 8, incoming_fids, 8,
58 _SIDD_UWORD_OPS |
59 _SIDD_CMP_EQUAL_ANY |
60 _SIDD_UNIT_MASK);
61 mask2 = _mm_cmpestrm(preflight_fids, 8, incoming_fids, 8,
62 _SIDD_UWORD_OPS |
63 _SIDD_CMP_EQUAL_ANY |
64 _SIDD_UNIT_MASK);
65
66 mask1 = _mm_or_si128(mask1, mask2);
67 /*
68 * Now mask contains 0xffff where there's a match.
69 * Next we need to store the worker_id in the relevant position
70 * in the output.
71 */
72
73 wkr = _mm_set1_epi16(i+1);
74 mask1 = _mm_and_si128(mask1, wkr);
75 output = _mm_or_si128(mask1, output);
76 }
77
78 /*
79 * At this stage, the output 128-bit contains 8 16-bit values, with
80 * each non-zero value containing the worker ID on which the
81 * corresponding flow is pinned to.
82 */
83 _mm_store_si128((__m128i *)output_ptr, output);
84 }
85