/*********************************************************************************************
    Description: stripped sw align functions in bwa-mem

    Copyright : All right reserved by NCIC.ICT

    Author : Zhang Zhonghai
    Date : 2024/04/11
***********************************************************************************************/
#ifndef __ALIGN_H
#define __ALIGN_H
#include <stdio.h>
#include <emmintrin.h>
#include <immintrin.h>
#include "byte_alloc.h"

#define KSW_XBYTE 0x10000
#define KSW_XSTOP 0x20000
#define KSW_XSUBO 0x40000
#define KSW_XSTART 0x80000

typedef struct
{
    int score;       // best score
    int te, qe;      // target end and query end
    int score2, te2; // second best score and ending position on the target
    int tb, qb;      // target start and query start
} kswr_t;

extern const kswr_t g_defr;

typedef struct _kswq_t
{
    int qlen, slen;
    uint8_t shift, mdiff, max, size;
    __m128i *qp, *H0, *H1, *E, *Hmax;
} kswq_sse_t;

typedef struct _kswq_avx2_t {
    int qlen, slen;
    uint8_t shift, mdiff, max, size;
    __m256i *qp, *H0, *H1, *E, *Hmax;
} kswq_avx2_t;

#define ALIGN_FUNC_NUM 4

kswq_sse_t *aln_sse_qinit(byte_mem_t *bmem, int size, int qlen, const uint8_t *query, int m, const int8_t *mat);
kswr_t align_sse_i16(byte_mem_t *bmem, kswq_sse_t *q, int tlen, const uint8_t *target, int _o_del, int _e_del, int _o_ins, int _e_ins, int xtra); // the first gap costs -(_o+_e)
kswr_t align_sse_u8(byte_mem_t *bmem, kswq_sse_t *q, int tlen, const uint8_t *target, int _o_del, int _e_del, int _o_ins, int _e_ins, int xtra);
kswq_avx2_t *aln_avx2_qinit(byte_mem_t *bmem, int size, int qlen, const uint8_t *query, int m, const int8_t *mat);
kswr_t align_avx2_u8(byte_mem_t *bmem, kswq_avx2_t *q, int tlen, const uint8_t *target, int _o_del, int _e_del, int _o_ins,
                     int _e_ins, int xtra);
kswr_t align_avx2_i16(byte_mem_t *bmem, kswq_avx2_t *q, int tlen, const uint8_t *target, int _o_del, int _e_del, int _o_ins,
                      int _e_ins, int xtra);
int main_align(int argc, char *argv[]);
#endif