1 | /* slide_neon.c -- Optimized hash table shifting for ARM with support for NEON instructions |
2 | * Copyright (C) 2017-2020 Mika T. Lindqvist |
3 | * |
4 | * Authors: |
5 | * Mika T. Lindqvist <postmaster@raasu.org> |
6 | * Jun He <jun.he@arm.com> |
7 | * |
8 | * For conditions of distribution and use, see copyright notice in zlib.h |
9 | */ |
10 | |
11 | #if defined(ARM_NEON_SLIDEHASH) |
12 | #ifdef _M_ARM64 |
13 | # include <arm64_neon.h> |
14 | #else |
15 | # include <arm_neon.h> |
16 | #endif |
17 | #include "../../zbuild.h" |
18 | #include "../../deflate.h" |
19 | |
20 | /* SIMD version of hash_chain rebase */ |
21 | static inline void slide_hash_chain(Pos *table, unsigned int entries, uint16_t window_size) { |
22 | Z_REGISTER uint16x8_t v, *p; |
23 | Z_REGISTER size_t n; |
24 | |
25 | size_t size = entries*sizeof(table[0]); |
26 | Assert((size % sizeof(uint16x8_t) * 8 == 0), "hash table size err" ); |
27 | |
28 | Assert(sizeof(Pos) == 2, "Wrong Pos size" ); |
29 | v = vdupq_n_u16(p0: window_size); |
30 | |
31 | p = (uint16x8_t *)table; |
32 | n = size / (sizeof(uint16x8_t) * 8); |
33 | do { |
34 | p[0] = vqsubq_u16(p0: p[0], p1: v); |
35 | p[1] = vqsubq_u16(p0: p[1], p1: v); |
36 | p[2] = vqsubq_u16(p0: p[2], p1: v); |
37 | p[3] = vqsubq_u16(p0: p[3], p1: v); |
38 | p[4] = vqsubq_u16(p0: p[4], p1: v); |
39 | p[5] = vqsubq_u16(p0: p[5], p1: v); |
40 | p[6] = vqsubq_u16(p0: p[6], p1: v); |
41 | p[7] = vqsubq_u16(p0: p[7], p1: v); |
42 | p += 8; |
43 | } while (--n); |
44 | } |
45 | |
46 | Z_INTERNAL void slide_hash_neon(deflate_state *s) { |
47 | unsigned int wsize = s->w_size; |
48 | |
49 | slide_hash_chain(table: s->head, HASH_SIZE, window_size: wsize); |
50 | slide_hash_chain(table: s->prev, entries: wsize, window_size: wsize); |
51 | } |
52 | #endif |
53 | |