1 /* chunkset_sse.c -- SSE 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 #include "zutil.h"
7
8 #ifdef X86_SSE2
9 #include <immintrin.h>
10
11 typedef __m128i chunk_t;
12
13 #define HAVE_CHUNKMEMSET_1
14 #define HAVE_CHUNKMEMSET_2
15 #define HAVE_CHUNKMEMSET_4
16 #define HAVE_CHUNKMEMSET_8
17
chunkmemset_1(uint8_t * from,chunk_t * chunk)18 static inline void chunkmemset_1(uint8_t *from, chunk_t *chunk) {
19 *chunk = _mm_set1_epi8(*(int8_t *)from);
20 }
21
chunkmemset_2(uint8_t * from,chunk_t * chunk)22 static inline void chunkmemset_2(uint8_t *from, chunk_t *chunk) {
23 *chunk = _mm_set1_epi16(*(int16_t *)from);
24 }
25
chunkmemset_4(uint8_t * from,chunk_t * chunk)26 static inline void chunkmemset_4(uint8_t *from, chunk_t *chunk) {
27 *chunk = _mm_set1_epi32(*(int32_t *)from);
28 }
29
chunkmemset_8(uint8_t * from,chunk_t * chunk)30 static inline void chunkmemset_8(uint8_t *from, chunk_t *chunk) {
31 *chunk = _mm_set1_epi64x(*(int64_t *)from);
32 }
33
loadchunk(uint8_t const * s,chunk_t * chunk)34 static inline void loadchunk(uint8_t const *s, chunk_t *chunk) {
35 *chunk = _mm_loadu_si128((__m128i *)s);
36 }
37
storechunk(uint8_t * out,chunk_t * chunk)38 static inline void storechunk(uint8_t *out, chunk_t *chunk) {
39 _mm_storeu_si128((__m128i *)out, *chunk);
40 }
41
42 #define CHUNKSIZE chunksize_sse2
43 #define CHUNKCOPY chunkcopy_sse2
44 #define CHUNKCOPY_SAFE chunkcopy_safe_sse2
45 #define CHUNKUNROLL chunkunroll_sse2
46 #define CHUNKMEMSET chunkmemset_sse2
47 #define CHUNKMEMSET_SAFE chunkmemset_safe_sse2
48
49 #include "chunkset_tpl.h"
50
51 #endif
52