opencv

1
/* chunkset_neon.c -- NEON inline functions to copy small data chunks.
2
 * For conditions of distribution and use, see copyright notice in zlib.h
3
 */
4

5
#ifdef ARM_NEON
6
#include "neon_intrins.h"
7
#include "../../zbuild.h"
8
#include "../generic/chunk_permute_table.h"
9

10
typedef uint8x16_t chunk_t;
11

12
#define CHUNK_SIZE 16
13

14
#define HAVE_CHUNKMEMSET_2
15
#define HAVE_CHUNKMEMSET_4
16
#define HAVE_CHUNKMEMSET_8
17
#define HAVE_CHUNK_MAG
18

19
static const lut_rem_pair perm_idx_lut[13] = {
20
    {0, 1},      /* 3 */
21
    {0, 0},      /* don't care */
22
    {1 * 32, 1}, /* 5 */
23
    {2 * 32, 4}, /* 6 */
24
    {3 * 32, 2}, /* 7 */
25
    {0 * 32, 0}, /* don't care */
26
    {4 * 32, 7}, /* 9 */
27
    {5 * 32, 6}, /* 10 */
28
    {6 * 32, 5}, /* 11 */
29
    {7 * 32, 4}, /* 12 */
30
    {8 * 32, 3}, /* 13 */
31
    {9 * 32, 2}, /* 14 */
32
    {10 * 32, 1},/* 15 */
33
};
34

35
static inline void chunkmemset_2(uint8_t *from, chunk_t *chunk) {
36
    uint16_t tmp;
37
    memcpy(&tmp, from, sizeof(tmp));
38
    *chunk = vreinterpretq_u8_u16(vdupq_n_u16(tmp));
39
}
40

41
static inline void chunkmemset_4(uint8_t *from, chunk_t *chunk) {
42
    uint32_t tmp;
43
    memcpy(&tmp, from, sizeof(tmp));
44
    *chunk = vreinterpretq_u8_u32(vdupq_n_u32(tmp));
45
}
46

47
static inline void chunkmemset_8(uint8_t *from, chunk_t *chunk) {
48
    uint64_t tmp;
49
    memcpy(&tmp, from, sizeof(tmp));
50
    *chunk = vreinterpretq_u8_u64(vdupq_n_u64(tmp));
51
}
52

53
#define CHUNKSIZE        chunksize_neon
54
#define CHUNKCOPY        chunkcopy_neon
55
#define CHUNKUNROLL      chunkunroll_neon
56
#define CHUNKMEMSET      chunkmemset_neon
57
#define CHUNKMEMSET_SAFE chunkmemset_safe_neon
58

59
static inline void loadchunk(uint8_t const *s, chunk_t *chunk) {
60
    *chunk = vld1q_u8(s);
61
}
62

63
static inline void storechunk(uint8_t *out, chunk_t *chunk) {
64
    vst1q_u8(out, *chunk);
65
}
66

67
static inline chunk_t GET_CHUNK_MAG(uint8_t *buf, uint32_t *chunk_rem, uint32_t dist) {
68
    lut_rem_pair lut_rem = perm_idx_lut[dist - 3];
69
    *chunk_rem = lut_rem.remval;
70

71
    /* See note in chunkset_ssse3.c for why this is ok */
72
    __msan_unpoison(buf + dist, 16 - dist);
73

74
    /* This version of table is only available on aarch64 */
75
#if defined(_M_ARM64) || defined(_M_ARM64EC) || defined(__aarch64__)
76
    uint8x16_t ret_vec = vld1q_u8(buf);
77

78
    uint8x16_t perm_vec = vld1q_u8(permute_table + lut_rem.idx);
79
    return vqtbl1q_u8(ret_vec, perm_vec);
80
#else
81
    uint8x8_t ret0, ret1, a, b, perm_vec0, perm_vec1;
82
    perm_vec0 = vld1_u8(permute_table + lut_rem.idx);
83
    perm_vec1 = vld1_u8(permute_table + lut_rem.idx + 8);
84
    a = vld1_u8(buf);
85
    b = vld1_u8(buf + 8);
86
    ret0 = vtbl1_u8(a, perm_vec0);
87
    uint8x8x2_t ab = {{a, b}};
88
    ret1 = vtbl2_u8(ab, perm_vec1);
89
    return vcombine_u8(ret0, ret1);
90
#endif
91
}
92

93
#include "chunkset_tpl.h"
94

95
#define INFLATE_FAST     inflate_fast_neon
96

97
#include "inffast_tpl.h"
98

99
#endif
100