forked from lh3/minimap2
-
Notifications
You must be signed in to change notification settings - Fork 0
/
ksw2_ll_sse.c
160 lines (150 loc) · 4.63 KB
/
ksw2_ll_sse.c
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
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
#include <stdlib.h>
#include <stdint.h>
#include <string.h>
#include "ksw2.h"
#if defined( __aarch64__ ) && !defined(USE_SIMDE)
#include <sse2neon.h>
#define __SSE4_1__
#define __SSE2__
#else
#ifdef USE_SIMDE
#include <simde/x86/sse2.h>
#else
#include <emmintrin.h>
#endif
#endif
#ifdef __GNUC__
#define LIKELY(x) __builtin_expect((x),1)
#define UNLIKELY(x) __builtin_expect((x),0)
#else
#define LIKELY(x) (x)
#define UNLIKELY(x) (x)
#endif
typedef struct {
int qlen, slen;
uint8_t shift, mdiff, max, size;
__m128i *qp, *H0, *H1, *E, *Hmax;
} kswq_t;
/**
* Initialize the query data structure
*
* @param size Number of bytes used to store a score; valid valures are 1 or 2
* @param qlen Length of the query sequence
* @param query Query sequence
* @param m Size of the alphabet
* @param mat Scoring matrix in a one-dimension array
*
* @return Query data structure
*/
void *ksw_ll_qinit(void *km, int size, int qlen, const uint8_t *query, int m, const int8_t *mat)
{
kswq_t *q;
int slen, a, tmp, p;
size = size > 1? 2 : 1;
p = 8 * (3 - size); // # values per __m128i
slen = (qlen + p - 1) / p; // segmented length
q = (kswq_t*)kmalloc(km, sizeof(kswq_t) + 256 + 16 * slen * (m + 4)); // a single block of memory
q->qp = (__m128i*)(((size_t)q + sizeof(kswq_t) + 15) >> 4 << 4); // align memory
q->H0 = q->qp + slen * m;
q->H1 = q->H0 + slen;
q->E = q->H1 + slen;
q->Hmax = q->E + slen;
q->slen = slen; q->qlen = qlen; q->size = size;
// compute shift
tmp = m * m;
for (a = 0, q->shift = 127, q->mdiff = 0; a < tmp; ++a) { // find the minimum and maximum score
if (mat[a] < (int8_t)q->shift) q->shift = mat[a];
if (mat[a] > (int8_t)q->mdiff) q->mdiff = mat[a];
}
q->max = q->mdiff;
q->shift = 256 - q->shift; // NB: q->shift is uint8_t
q->mdiff += q->shift; // this is the difference between the min and max scores
// An example: p=8, qlen=19, slen=3 and segmentation:
// {{0,3,6,9,12,15,18,-1},{1,4,7,10,13,16,-1,-1},{2,5,8,11,14,17,-1,-1}}
if (size == 1) {
int8_t *t = (int8_t*)q->qp;
for (a = 0; a < m; ++a) {
int i, k, nlen = slen * p;
const int8_t *ma = mat + a * m;
for (i = 0; i < slen; ++i)
for (k = i; k < nlen; k += slen) // p iterations
*t++ = (k >= qlen? 0 : ma[query[k]]) + q->shift;
}
} else {
int16_t *t = (int16_t*)q->qp;
for (a = 0; a < m; ++a) {
int i, k, nlen = slen * p;
const int8_t *ma = mat + a * m;
for (i = 0; i < slen; ++i)
for (k = i; k < nlen; k += slen) // p iterations
*t++ = (k >= qlen? 0 : ma[query[k]]);
}
}
return q;
}
int ksw_ll_i16(void *q_, int tlen, const uint8_t *target, int _gapo, int _gape, int *qe, int *te)
{
kswq_t *q = (kswq_t*)q_;
int slen, i, gmax = 0, qlen8;
__m128i zero, gapoe, gape, *H0, *H1, *E, *Hmax;
uint16_t *H8;
#define __max_8(ret, xx) do { \
(xx) = _mm_max_epi16((xx), _mm_srli_si128((xx), 8)); \
(xx) = _mm_max_epi16((xx), _mm_srli_si128((xx), 4)); \
(xx) = _mm_max_epi16((xx), _mm_srli_si128((xx), 2)); \
(ret) = _mm_extract_epi16((xx), 0); \
} while (0)
// initialization
*qe = *te = -1;
zero = _mm_set1_epi32(0);
gapoe = _mm_set1_epi16(_gapo + _gape);
gape = _mm_set1_epi16(_gape);
H0 = q->H0; H1 = q->H1; E = q->E; Hmax = q->Hmax;
slen = q->slen, qlen8 = slen * 8;
memset(E, 0, slen * sizeof(__m128i));
memset(H0, 0, slen * sizeof(__m128i));
memset(Hmax, 0, slen * sizeof(__m128i));
// the core loop
for (i = 0; i < tlen; ++i) {
int j, k, imax;
__m128i e, h, f = zero, max = zero, *S = q->qp + target[i] * slen; // s is the 1st score vector
h = _mm_load_si128(H0 + slen - 1); // h={2,5,8,11,14,17,-1,-1} in the above example
h = _mm_slli_si128(h, 2);
for (j = 0; LIKELY(j < slen); ++j) {
h = _mm_adds_epi16(h, *S++);
e = _mm_load_si128(E + j);
h = _mm_max_epi16(h, e);
h = _mm_max_epi16(h, f);
max = _mm_max_epi16(max, h);
_mm_store_si128(H1 + j, h);
h = _mm_subs_epu16(h, gapoe);
e = _mm_subs_epu16(e, gape);
e = _mm_max_epi16(e, h);
_mm_store_si128(E + j, e);
f = _mm_subs_epu16(f, gape);
f = _mm_max_epi16(f, h);
h = _mm_load_si128(H0 + j);
}
for (k = 0; LIKELY(k < 8); ++k) {
f = _mm_slli_si128(f, 2);
for (j = 0; LIKELY(j < slen); ++j) {
h = _mm_load_si128(H1 + j);
h = _mm_max_epi16(h, f);
_mm_store_si128(H1 + j, h);
h = _mm_subs_epu16(h, gapoe);
f = _mm_subs_epu16(f, gape);
if(UNLIKELY(!_mm_movemask_epi8(_mm_cmpgt_epi16(f, h)))) goto end_loop_i16;
}
}
end_loop_i16:
__max_8(imax, max);
if (imax >= gmax) {
gmax = imax; *te = i;
memcpy(Hmax, H1, slen * sizeof(__m128i));
}
S = H1; H1 = H0; H0 = S;
}
for (i = 0, H8 = (uint16_t*)Hmax; i < qlen8; ++i)
if ((int)H8[i] == gmax) *qe = i / 8 + i % 8 * slen;
return gmax;
}