#pragma once

#include <emmintrin.h>
#include <immintrin.h>
#include <stdint.h>

#include "ksw.h"


#define AMBIG_ 4  // ambiguous base
// for 16 bit
#define DUMMY1_ 4
#define DUMMY2_ 5
#define DUMMY3 26
#define AMBR16 15
#define AMBQ16 16
// for 8-bit
#define DUMMY8 8
#define DUMMY5 5
#define AMBRQ 0xFF
#define AMBR 4
#define AMBQ 8

#define min_(x, y) ((x) > (y) ? (y) : (x))
#define max_(x, y) ((x) > (y) ? (x) : (y))

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_avx_t;

typedef struct {
    int seq_len;
    int ref_len;
    uint8_t* H0;
    uint8_t* H1;
    uint8_t* Hmax;
    uint8_t* F;
    uint8_t* rowMax;
    uint8_t* seqArr;
    uint8_t* refArr;
} msw_buf_t; // inter-query算法的mate sw计算过程需要用到的缓存空间

#ifdef __cplusplus
extern "C" {
#endif


void ksw_align_avx2_u8(int qlen, uint8_t* query, int tlen, uint8_t* target, int m, const int8_t* mat, int gapo, int gape, int xtra, kswq_t** qry);
void ksw_align_avx2_i16(int qlen, uint8_t* query, int tlen, uint8_t* target, int m, const int8_t* mat, int gapo, int gape, int xtra, kswq_t** qry);

void ksw_align_avx512_u8(int8_t w_match,       // match分数，正数
                         int8_t w_mismatch,    // 错配罚分，负数
                         int8_t o_ins,         // 开始一个insert罚分，正数
                         int8_t e_ins,         // 延续一个insert罚分，正数
                         int8_t o_del,         // 开始一个delete罚分，正数
                         int8_t e_del,         // 延续一个delete罚分，正数
                         msw_buf_t* cache,  // 计算用到的一些数据
                         uint8_t* seq1SoA,     // ref序列，已经pack好了
                         uint8_t* seq2SoA,     // seq序列
                         int16_t nrow,         // 最长的行数，对应ref长度
                         int16_t ncol,         // 最长的列数，对应seq长度
                         int* xtras,           // 每个seq对应一个xtra
                         int* rlenA,           // ref真实长度
                         kswr_avx_t* alns,     // 存放结果
                         int phase);           // 正向阶段0，反向阶段1

void ksw_align_avx512_i16(int qlen, uint8_t* query, int tlen, uint8_t* target, int m, const int8_t* mat, int gapo, int gape, int xtra, kswq_t** qry);

#ifdef __cplusplus
}
#endif