opencv

1
/* slide_hash_armv6.c -- Optimized hash table shifting for ARMv6 with support for SIMD instructions
2
 * Copyright (C) 2023 Cameron Cawley
3
 * For conditions of distribution and use, see copyright notice in zlib.h
4
 */
5

6
#if defined(ARM_SIMD)
7
#include "acle_intrins.h"
8
#include "../../zbuild.h"
9
#include "../../deflate.h"
10

11
/* SIMD version of hash_chain rebase */
12
static inline void slide_hash_chain(Pos *table, uint32_t entries, uint16_t wsize) {
13
    Z_REGISTER uint16x2_t v;
14
    uint16x2_t p0, p1, p2, p3;
15
    Z_REGISTER size_t n;
16

17
    size_t size = entries*sizeof(table[0]);
18
    Assert((size % (sizeof(uint16x2_t) * 4) == 0), "hash table size err");
19

20
    Assert(sizeof(Pos) == 2, "Wrong Pos size");
21
    v = wsize | (wsize << 16);
22

23
    n = size / (sizeof(uint16x2_t) * 4);
24
    do {
25
        p0 = *((const uint16x2_t *)(table));
26
        p1 = *((const uint16x2_t *)(table+2));
27
        p2 = *((const uint16x2_t *)(table+4));
28
        p3 = *((const uint16x2_t *)(table+6));
29
        p0 = __uqsub16(p0, v);
30
        p1 = __uqsub16(p1, v);
31
        p2 = __uqsub16(p2, v);
32
        p3 = __uqsub16(p3, v);
33
        *((uint16x2_t *)(table)) = p0;
34
        *((uint16x2_t *)(table+2)) = p1;
35
        *((uint16x2_t *)(table+4)) = p2;
36
        *((uint16x2_t *)(table+6)) = p3;
37
        table += 8;
38
    } while (--n);
39
}
40

41
Z_INTERNAL void slide_hash_armv6(deflate_state *s) {
42
    unsigned int wsize = s->w_size;
43

44
    slide_hash_chain(s->head, HASH_SIZE, wsize);
45
    slide_hash_chain(s->prev, wsize, wsize);
46
}
47
#endif
48