opencv

1
/* compare256_avx2.c -- AVX2 version of compare256
2
 * Copyright Mika T. Lindqvist  <postmaster@raasu.org>
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(X86_AVX2) && defined(HAVE_BUILTIN_CTZ)
11

12
#include <immintrin.h>
13
#ifdef _MSC_VER
14
#  include <nmmintrin.h>
15
#endif
16

17
static inline uint32_t compare256_avx2_static(const uint8_t *src0, const uint8_t *src1) {
18
    uint32_t len = 0;
19

20
    do {
21
        __m256i ymm_src0, ymm_src1, ymm_cmp;
22
        ymm_src0 = _mm256_loadu_si256((__m256i*)src0);
23
        ymm_src1 = _mm256_loadu_si256((__m256i*)src1);
24
        ymm_cmp = _mm256_cmpeq_epi8(ymm_src0, ymm_src1); /* non-identical bytes = 00, identical bytes = FF */
25
        unsigned mask = (unsigned)_mm256_movemask_epi8(ymm_cmp);
26
        if (mask != 0xFFFFFFFF) {
27
            uint32_t match_byte = (uint32_t)__builtin_ctz(~mask); /* Invert bits so identical = 0 */
28
            return len + match_byte;
29
        }
30

31
        src0 += 32, src1 += 32, len += 32;
32

33
        ymm_src0 = _mm256_loadu_si256((__m256i*)src0);
34
        ymm_src1 = _mm256_loadu_si256((__m256i*)src1);
35
        ymm_cmp = _mm256_cmpeq_epi8(ymm_src0, ymm_src1);
36
        mask = (unsigned)_mm256_movemask_epi8(ymm_cmp);
37
        if (mask != 0xFFFFFFFF) {
38
            uint32_t match_byte = (uint32_t)__builtin_ctz(~mask);
39
            return len + match_byte;
40
        }
41

42
        src0 += 32, src1 += 32, len += 32;
43
    } while (len < 256);
44

45
    return 256;
46
}
47

48
Z_INTERNAL uint32_t compare256_avx2(const uint8_t *src0, const uint8_t *src1) {
49
    return compare256_avx2_static(src0, src1);
50
}
51

52
#define LONGEST_MATCH       longest_match_avx2
53
#define COMPARE256          compare256_avx2_static
54

55
#include "match_tpl.h"
56

57
#define LONGEST_MATCH_SLOW
58
#define LONGEST_MATCH       longest_match_slow_avx2
59
#define COMPARE256          compare256_avx2_static
60

61
#include "match_tpl.h"
62

63
#endif
64