1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
|
/* chunkset_lasx.c -- LASX inline functions to copy small data chunks, based on Intel AVX2 implementation
* Copyright (C) 2025 Vladislav Shchapov <vladislav@shchapov.ru>
* For conditions of distribution and use, see copyright notice in zlib.h
*/
#ifdef LOONGARCH_LASX
#include "zbuild.h"
#include "zsanitizer.h"
#include "zmemory.h"
#include <lasxintrin.h>
#include "lasxintrin_ext.h"
#include "lsxintrin_ext.h"
#include "arch/generic/chunk_256bit_perm_idx_lut.h"
typedef __m256i chunk_t;
typedef __m128i halfchunk_t;
#define HAVE_CHUNKMEMSET_2
#define HAVE_CHUNKMEMSET_4
#define HAVE_CHUNKMEMSET_8
#define HAVE_CHUNKMEMSET_16
#define HAVE_CHUNK_MAG
#define HAVE_HALF_CHUNK
static inline void chunkmemset_2(uint8_t *from, chunk_t *chunk) {
*chunk = __lasx_xvreplgr2vr_h(zng_memread_2(from));
}
static inline void chunkmemset_4(uint8_t *from, chunk_t *chunk) {
*chunk = __lasx_xvreplgr2vr_w(zng_memread_4(from));
}
static inline void chunkmemset_8(uint8_t *from, chunk_t *chunk) {
*chunk = __lasx_xvreplgr2vr_d(zng_memread_8(from));
}
static inline void chunkmemset_16(uint8_t *from, chunk_t *chunk) {
*chunk = lasx_broadcast_128(__lsx_vld(from, 0));
}
static inline void loadchunk(uint8_t const *s, chunk_t *chunk) {
*chunk = __lasx_xvld(s, 0);
}
static inline void storechunk(uint8_t *out, chunk_t *chunk) {
__lasx_xvst(*chunk, out, 0);
}
static inline chunk_t GET_CHUNK_MAG(uint8_t *buf, size_t *chunk_rem, size_t dist) {
lut_rem_pair lut_rem = perm_idx_lut[dist - 3];
__m256i ret_vec;
/* While technically we only need to read 4 or 8 bytes into this vector register for a lot of cases, GCC is
* compiling this to a shared load for all branches, preferring the simpler code. Given that the buf value isn't in
* GPRs to begin with the 256 bit load is _probably_ just as inexpensive */
*chunk_rem = lut_rem.remval;
/* See note in chunkset_ssse3.c for why this is ok */
__msan_unpoison(buf + dist, 32 - dist);
if (dist < 16) {
/* This simpler case still requires us to shuffle in 128 bit lanes, so we must apply a static offset after
* broadcasting the first vector register to both halves. This is _marginally_ faster than doing two separate
* shuffles and combining the halves later */
__m256i perm_vec = __lasx_xvld(permute_table+lut_rem.idx, 0);
__m128i ret_vec0 = __lsx_vld(buf, 0);
ret_vec = __lasx_concat_128(ret_vec0, ret_vec0);
ret_vec = lasx_shuffle_b(ret_vec, perm_vec);
} else {
__m128i ret_vec0 = __lsx_vld(buf, 0);
__m128i ret_vec1 = __lsx_vld(buf, 16);
/* Take advantage of the fact that only the latter half of the 256 bit vector will actually differ */
__m128i perm_vec1 = __lsx_vld(permute_table + lut_rem.idx, 0);
__m128i xlane_permutes = __lsx_vslt_b(perm_vec1, __lsx_vreplgr2vr_b(16));
__m128i xlane_res = lsx_shuffle_b(ret_vec0, perm_vec1);
/* Since we can't wrap twice, we can simply keep the later half exactly how it is instead of having to _also_
* shuffle those values */
__m128i latter_half = __lsx_vbitsel_v(ret_vec1, xlane_res, xlane_permutes);
ret_vec = __lasx_concat_128(ret_vec0, latter_half);
}
return ret_vec;
}
static inline void loadhalfchunk(uint8_t const *s, halfchunk_t *chunk) {
*chunk = __lsx_vld(s, 0);
}
static inline void storehalfchunk(uint8_t *out, halfchunk_t *chunk) {
__lsx_vst(*chunk, out, 0);
}
static inline chunk_t halfchunk2whole(halfchunk_t *chunk) {
/* We zero extend mostly to appease some memory sanitizers. These bytes are ultimately
* unlikely to be actually written or read from */
return lasx_zext_128(*chunk);
}
static inline halfchunk_t GET_HALFCHUNK_MAG(uint8_t *buf, size_t *chunk_rem, size_t dist) {
lut_rem_pair lut_rem = perm_idx_lut[dist - 3];
__m128i perm_vec, ret_vec;
__msan_unpoison(buf + dist, 16 - dist);
ret_vec = __lsx_vld(buf, 0);
*chunk_rem = half_rem_vals[dist - 3];
perm_vec = __lsx_vld(permute_table + lut_rem.idx, 0);
ret_vec = lsx_shuffle_b(ret_vec, perm_vec);
return ret_vec;
}
#define CHUNKSIZE chunksize_lasx
#define CHUNKCOPY chunkcopy_lasx
#define CHUNKUNROLL chunkunroll_lasx
#define CHUNKMEMSET chunkmemset_lasx
#define CHUNKMEMSET_SAFE chunkmemset_safe_lasx
#include "chunkset_tpl.h"
#define INFLATE_FAST inflate_fast_lasx
#include "inffast_tpl.h"
#endif
|
