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_CHUNKSET
6 #ifdef _M_ARM64
7 # include <arm64_neon.h>
8 #else
9 # include <arm_neon.h>
10 #endif
11 #include "../../zbuild.h"
12
13 typedef uint8x16_t chunk_t;
14
15 #define CHUNK_SIZE 16
16
17 #define HAVE_CHUNKMEMSET_2
18 #define HAVE_CHUNKMEMSET_4
19 #define HAVE_CHUNKMEMSET_8
20
chunkmemset_2(uint8_t * from,chunk_t * chunk)21 static inline void chunkmemset_2(uint8_t *from, chunk_t *chunk) {
22 uint16_t tmp;
23 zmemcpy_2(&tmp, from);
24 *chunk = vreinterpretq_u8_u16(vdupq_n_u16(tmp));
25 }
26
chunkmemset_4(uint8_t * from,chunk_t * chunk)27 static inline void chunkmemset_4(uint8_t *from, chunk_t *chunk) {
28 uint32_t tmp;
29 zmemcpy_4(&tmp, from);
30 *chunk = vreinterpretq_u8_u32(vdupq_n_u32(tmp));
31 }
32
chunkmemset_8(uint8_t * from,chunk_t * chunk)33 static inline void chunkmemset_8(uint8_t *from, chunk_t *chunk) {
34 uint64_t tmp;
35 zmemcpy_8(&tmp, from);
36 *chunk = vreinterpretq_u8_u64(vdupq_n_u64(tmp));
37 }
38
39 #define CHUNKSIZE chunksize_neon
40 #define CHUNKCOPY chunkcopy_neon
41 #define CHUNKCOPY_SAFE chunkcopy_safe_neon
42 #define CHUNKUNROLL chunkunroll_neon
43 #define CHUNKMEMSET chunkmemset_neon
44 #define CHUNKMEMSET_SAFE chunkmemset_safe_neon
45
loadchunk(uint8_t const * s,chunk_t * chunk)46 static inline void loadchunk(uint8_t const *s, chunk_t *chunk) {
47 *chunk = vld1q_u8(s);
48 }
49
storechunk(uint8_t * out,chunk_t * chunk)50 static inline void storechunk(uint8_t *out, chunk_t *chunk) {
51 vst1q_u8(out, *chunk);
52 }
53
54 #include "chunkset_tpl.h"
55
56 #endif
57