1 /* compare258_avx.c -- AVX2 version of compare258
2 * Copyright Mika T. Lindqvist <postmaster@raasu.org>
3 * For conditions of distribution and use, see copyright notice in zlib.h
4 */
5
6 #include "../../zbuild.h"
7 #include "../../zutil.h"
8
9 #include "fallback_builtins.h"
10
11 #if defined(X86_AVX2) && defined(HAVE_BUILTIN_CTZ)
12
13 #include <immintrin.h>
14 #ifdef _MSC_VER
15 # include <nmmintrin.h>
16 #endif
17
18 /* UNALIGNED_OK, AVX2 intrinsic comparison */
compare256_unaligned_avx2_static(const unsigned char * src0,const unsigned char * src1)19 static inline uint32_t compare256_unaligned_avx2_static(const unsigned char *src0, const unsigned char *src1) {
20 uint32_t len = 0;
21
22 do {
23 __m256i ymm_src0, ymm_src1, ymm_cmp;
24 ymm_src0 = _mm256_loadu_si256((__m256i*)src0);
25 ymm_src1 = _mm256_loadu_si256((__m256i*)src1);
26 ymm_cmp = _mm256_cmpeq_epi8(ymm_src0, ymm_src1); /* non-identical bytes = 00, identical bytes = FF */
27 unsigned mask = (unsigned)_mm256_movemask_epi8(ymm_cmp);
28 if (mask != 0xFFFFFFFF) {
29 uint32_t match_byte = (uint32_t)__builtin_ctz(~mask); /* Invert bits so identical = 0 */
30 return len + match_byte;
31 }
32
33 src0 += 32, src1 += 32, len += 32;
34
35 ymm_src0 = _mm256_loadu_si256((__m256i*)src0);
36 ymm_src1 = _mm256_loadu_si256((__m256i*)src1);
37 ymm_cmp = _mm256_cmpeq_epi8(ymm_src0, ymm_src1);
38 mask = (unsigned)_mm256_movemask_epi8(ymm_cmp);
39 if (mask != 0xFFFFFFFF) {
40 uint32_t match_byte = (uint32_t)__builtin_ctz(~mask);
41 return len + match_byte;
42 }
43
44 src0 += 32, src1 += 32, len += 32;
45 } while (len < 256);
46
47 return 256;
48 }
49
compare258_unaligned_avx2_static(const unsigned char * src0,const unsigned char * src1)50 static inline uint32_t compare258_unaligned_avx2_static(const unsigned char *src0, const unsigned char *src1) {
51 if (*(uint16_t *)src0 != *(uint16_t *)src1)
52 return (*src0 == *src1);
53
54 return compare256_unaligned_avx2_static(src0+2, src1+2) + 2;
55 }
56
compare258_unaligned_avx2(const unsigned char * src0,const unsigned char * src1)57 Z_INTERNAL uint32_t compare258_unaligned_avx2(const unsigned char *src0, const unsigned char *src1) {
58 return compare258_unaligned_avx2_static(src0, src1);
59 }
60
61 #define LONGEST_MATCH longest_match_unaligned_avx2
62 #define COMPARE256 compare256_unaligned_avx2_static
63 #define COMPARE258 compare258_unaligned_avx2_static
64
65 #include "match_tpl.h"
66
67 #endif
68