1 /* chunkset_avx.c -- AVX inline functions to copy small data chunks.
2 * For conditions of distribution and use, see copyright notice in zlib.h
3 */
4 #include "zbuild.h"
5 #include "zutil.h"
6
7 #ifdef X86_AVX_CHUNKSET
8 #include <immintrin.h>
9
10 typedef __m256i chunk_t;
11
12 #define HAVE_CHUNKMEMSET_1
13 #define HAVE_CHUNKMEMSET_2
14 #define HAVE_CHUNKMEMSET_4
15 #define HAVE_CHUNKMEMSET_8
16
chunkmemset_1(uint8_t * from,chunk_t * chunk)17 static inline void chunkmemset_1(uint8_t *from, chunk_t *chunk) {
18 *chunk = _mm256_set1_epi8(*(int8_t *)from);
19 }
20
chunkmemset_2(uint8_t * from,chunk_t * chunk)21 static inline void chunkmemset_2(uint8_t *from, chunk_t *chunk) {
22 *chunk = _mm256_set1_epi16(*(int16_t *)from);
23 }
24
chunkmemset_4(uint8_t * from,chunk_t * chunk)25 static inline void chunkmemset_4(uint8_t *from, chunk_t *chunk) {
26 *chunk = _mm256_set1_epi32(*(int32_t *)from);
27 }
28
chunkmemset_8(uint8_t * from,chunk_t * chunk)29 static inline void chunkmemset_8(uint8_t *from, chunk_t *chunk) {
30 *chunk = _mm256_set1_epi64x(*(int64_t *)from);
31 }
32
loadchunk(uint8_t const * s,chunk_t * chunk)33 static inline void loadchunk(uint8_t const *s, chunk_t *chunk) {
34 *chunk = _mm256_loadu_si256((__m256i *)s);
35 }
36
storechunk(uint8_t * out,chunk_t * chunk)37 static inline void storechunk(uint8_t *out, chunk_t *chunk) {
38 _mm256_storeu_si256((__m256i *)out, *chunk);
39 }
40
41 #define CHUNKSIZE chunksize_avx
42 #define CHUNKCOPY chunkcopy_avx
43 #define CHUNKCOPY_SAFE chunkcopy_safe_avx
44 #define CHUNKUNROLL chunkunroll_avx
45 #define CHUNKMEMSET chunkmemset_avx
46 #define CHUNKMEMSET_SAFE chunkmemset_safe_avx
47
48 #include "chunkset_tpl.h"
49
50 #endif
51