1 /* chunkset_sse2.c -- SSE2 inline functions to copy small data chunks.
2 * For conditions of distribution and use, see copyright notice in zlib.h
3 */
4
5 #include "zbuild.h"
6
7 #ifdef X86_SSE2
8 #include <immintrin.h>
9
10 typedef __m128i 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
chunkmemset_2(uint8_t * from,chunk_t * chunk)18 static inline void chunkmemset_2(uint8_t *from, chunk_t *chunk) {
19 int16_t tmp;
20 zmemcpy_2(&tmp, from);
21 *chunk = _mm_set1_epi16(tmp);
22 }
23
chunkmemset_4(uint8_t * from,chunk_t * chunk)24 static inline void chunkmemset_4(uint8_t *from, chunk_t *chunk) {
25 int32_t tmp;
26 zmemcpy_4(&tmp, from);
27 *chunk = _mm_set1_epi32(tmp);
28 }
29
chunkmemset_8(uint8_t * from,chunk_t * chunk)30 static inline void chunkmemset_8(uint8_t *from, chunk_t *chunk) {
31 int64_t tmp;
32 zmemcpy_8(&tmp, from);
33 *chunk = _mm_set1_epi64x(tmp);
34 }
35
loadchunk(uint8_t const * s,chunk_t * chunk)36 static inline void loadchunk(uint8_t const *s, chunk_t *chunk) {
37 *chunk = _mm_loadu_si128((__m128i *)s);
38 }
39
storechunk(uint8_t * out,chunk_t * chunk)40 static inline void storechunk(uint8_t *out, chunk_t *chunk) {
41 _mm_storeu_si128((__m128i *)out, *chunk);
42 }
43
44 #define CHUNKSIZE chunksize_sse2
45 #define CHUNKCOPY chunkcopy_sse2
46 #define CHUNKCOPY_SAFE chunkcopy_safe_sse2
47 #define CHUNKUNROLL chunkunroll_sse2
48 #define CHUNKMEMSET chunkmemset_sse2
49 #define CHUNKMEMSET_SAFE chunkmemset_safe_sse2
50
51 #include "chunkset_tpl.h"
52
53 #endif
54