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 #include "../../zutil.h"
13
14 typedef uint8x16_t chunk_t;
15
16 #define HAVE_CHUNKMEMSET_1
17 #define HAVE_CHUNKMEMSET_2
18 #define HAVE_CHUNKMEMSET_4
19 #define HAVE_CHUNKMEMSET_8
20
chunkmemset_1(uint8_t * from,chunk_t * chunk)21 static inline void chunkmemset_1(uint8_t *from, chunk_t *chunk) {
22 *chunk = vld1q_dup_u8(from);
23 }
24
chunkmemset_2(uint8_t * from,chunk_t * chunk)25 static inline void chunkmemset_2(uint8_t *from, chunk_t *chunk) {
26 int16_t tmp;
27 memcpy(&tmp, from, 2);
28 *chunk = vreinterpretq_u8_s16(vdupq_n_s16(tmp));
29 }
30
chunkmemset_4(uint8_t * from,chunk_t * chunk)31 static inline void chunkmemset_4(uint8_t *from, chunk_t *chunk) {
32 int32_t tmp;
33 memcpy(&tmp, from, 4);
34 *chunk = vreinterpretq_u8_s32(vdupq_n_s32(tmp));
35 }
36
chunkmemset_8(uint8_t * from,chunk_t * chunk)37 static inline void chunkmemset_8(uint8_t *from, chunk_t *chunk) {
38 *chunk = vcombine_u8(vld1_u8(from), vld1_u8(from));
39 }
40
41 #define CHUNKSIZE chunksize_neon
42 #define CHUNKCOPY chunkcopy_neon
43 #define CHUNKCOPY_SAFE chunkcopy_safe_neon
44 #define CHUNKUNROLL chunkunroll_neon
45 #define CHUNKMEMSET chunkmemset_neon
46 #define CHUNKMEMSET_SAFE chunkmemset_safe_neon
47
loadchunk(uint8_t const * s,chunk_t * chunk)48 static inline void loadchunk(uint8_t const *s, chunk_t *chunk) {
49 *chunk = vld1q_u8(s);
50 }
51
storechunk(uint8_t * out,chunk_t * chunk)52 static inline void storechunk(uint8_t *out, chunk_t *chunk) {
53 vst1q_u8(out, *chunk);
54 }
55
56 #include "chunkset_tpl.h"
57
58 #endif
59