opencv

1
/* compare256_neon.c - NEON version of compare256
2
 * Copyright (C) 2022 Nathan Moinvaziri
3
 * For conditions of distribution and use, see copyright notice in zlib.h
4
 */
5

6
#include "../../zbuild.h"
7

8
#include "fallback_builtins.h"
9

10
#if defined(ARM_NEON) && defined(HAVE_BUILTIN_CTZLL)
11
#include "neon_intrins.h"
12

13
static inline uint32_t compare256_neon_static(const uint8_t *src0, const uint8_t *src1) {
14
    uint32_t len = 0;
15

16
    do {
17
        uint8x16_t a, b, cmp;
18
        uint64_t lane;
19

20
        a = vld1q_u8(src0);
21
        b = vld1q_u8(src1);
22

23
        cmp = veorq_u8(a, b);
24

25
        lane = vgetq_lane_u64(vreinterpretq_u64_u8(cmp), 0);
26
        if (lane) {
27
            uint32_t match_byte = (uint32_t)__builtin_ctzll(lane) / 8;
28
            return len + match_byte;
29
        }
30
        len += 8;
31
        lane = vgetq_lane_u64(vreinterpretq_u64_u8(cmp), 1);
32
        if (lane) {
33
            uint32_t match_byte = (uint32_t)__builtin_ctzll(lane) / 8;
34
            return len + match_byte;
35
        }
36
        len += 8;
37

38
        src0 += 16, src1 += 16;
39
    } while (len < 256);
40

41
    return 256;
42
}
43

44
Z_INTERNAL uint32_t compare256_neon(const uint8_t *src0, const uint8_t *src1) {
45
    return compare256_neon_static(src0, src1);
46
}
47

48
#define LONGEST_MATCH       longest_match_neon
49
#define COMPARE256          compare256_neon_static
50

51
#include "match_tpl.h"
52

53
#define LONGEST_MATCH_SLOW
54
#define LONGEST_MATCH       longest_match_slow_neon
55
#define COMPARE256          compare256_neon_static
56

57
#include "match_tpl.h"
58

59
#endif
60