sw_perf/ksw_ext_avx2.c

#include <stdlib.h>
#include <stdint.h>
#include <assert.h>
#include <emmintrin.h>
#include <stdio.h>
#include <immintrin.h>
#include <emmintrin.h>
#include "thread_mem.h"
#include "common.h"

#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

#define SIMD_WIDTH 16

static const uint16_t h_vec_int_mask[SIMD_WIDTH][SIMD_WIDTH] = {
    {0xffff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
    {0xffff, 0xffff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
    {0xffff, 0xffff, 0xffff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
    {0xffff, 0xffff, 0xffff, 0xffff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
    {0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
    {0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
    {0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0, 0, 0, 0, 0, 0, 0, 0, 0},
    {0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0, 0, 0, 0, 0, 0, 0, 0},
    {0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0, 0, 0, 0, 0, 0, 0},
    {0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0, 0, 0, 0, 0, 0},
    {0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0, 0, 0, 0, 0},
    {0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0, 0, 0, 0},
    {0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0, 0, 0},
    {0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0, 0},
    {0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0},
    {0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff}};

// #define permute_mask _MM_SHUFFLE(0, 1, 2, 3)
#define permute_mask 27
// 初始化变量
#define SIMD_INIT                                               \
    int oe_del = o_del + e_del, oe_ins = o_ins + e_ins;         \
    __m256i zero_vec;                                           \
    __m256i max_vec;                                            \
    __m256i oe_del_vec;                                         \
    __m256i oe_ins_vec;                                         \
    __m256i e_del_vec;                                          \
    __m256i e_ins_vec;                                          \
    __m256i h_vec_mask[SIMD_WIDTH];                             \
    zero_vec = _mm256_setzero_si256();                          \
    oe_del_vec = _mm256_set1_epi16(-oe_del);                    \
    oe_ins_vec = _mm256_set1_epi16(-oe_ins);                    \
    e_del_vec = _mm256_set1_epi16(-e_del);                      \
    e_ins_vec = _mm256_set1_epi16(-e_ins);                      \
    __m256i match_sc_vec = _mm256_set1_epi16(base_match_score); \
    __m256i mis_sc_vec = _mm256_set1_epi16(-base_mis_score);    \
    __m256i amb_sc_vec = _mm256_set1_epi16(-1);                 \
    __m256i amb_vec = _mm256_set1_epi16(4);                     \
    for (i = 0; i < SIMD_WIDTH; ++i)                            \
        h_vec_mask[i] = _mm256_loadu_si256((__m256i *)(&h_vec_int_mask[i]));

/*
 * e 表示当前ref的碱基被删除
 * f 表示当前seq的碱基插入
 * m 表示当前碱基匹配（可以相等，也可以不想等）
 * h 表示最大值
 */
// load向量化数据
#define SIMD_LOAD                                                  \
    __m256i m1 = _mm256_loadu_si256((__m256i *)(&mA1[j]));         \
    __m256i e1 = _mm256_loadu_si256((__m256i *)(&eA1[j]));         \
    __m256i m1j1 = _mm256_loadu_si256((__m256i *)(&mA1[j - 1]));   \
    __m256i f1j1 = _mm256_loadu_si256((__m256i *)(&fA1[j - 1]));   \
    __m256i h0j1 = _mm256_loadu_si256((__m256i *)(&hA0[j - 1]));   \
    __m256i qs_vec = _mm256_loadu_si256((__m256i *)(&seq[j - 1])); \
    __m256i ts_vec = _mm256_loadu_si256((__m256i *)(&ref[i]));

// 比对ref和seq的序列，计算罚分
#define SIMD_CMP_SEQ                                                         \
    ts_vec = _mm256_permute4x64_epi64(ts_vec, permute_mask);                 \
    ts_vec = _mm256_shufflelo_epi16(ts_vec, permute_mask);                   \
    ts_vec = _mm256_shufflehi_epi16(ts_vec, permute_mask);                   \
    __m256i match_mask_vec = _mm256_cmpeq_epi16(qs_vec, ts_vec);             \
    __m256i mis_score_vec = _mm256_andnot_si256(match_mask_vec, mis_sc_vec); \
    __m256i score_vec = _mm256_and_si256(match_sc_vec, match_mask_vec);      \
    score_vec = _mm256_or_si256(score_vec, mis_score_vec);                   \
    __m256i q_amb_mask_vec = _mm256_cmpeq_epi16(qs_vec, amb_vec);            \
    __m256i t_amb_mask_vec = _mm256_cmpeq_epi16(ts_vec, amb_vec);            \
    __m256i amb_mask_vec = _mm256_or_si256(q_amb_mask_vec, t_amb_mask_vec);  \
    score_vec = _mm256_andnot_si256(amb_mask_vec, score_vec);                \
    __m256i amb_score_vec = _mm256_and_si256(amb_mask_vec, amb_sc_vec);      \
    score_vec = _mm256_or_si256(score_vec, amb_score_vec);

// 向量化计算h, e, f, m
#define SIMD_COMPUTE                                      \
    __m256i en_vec0 = _mm256_add_epi16(m1, oe_del_vec);   \
    __m256i en_vec1 = _mm256_add_epi16(e1, e_del_vec);    \
    __m256i en_vec = _mm256_max_epi16(en_vec0, en_vec1);  \
    __m256i fn_vec0 = _mm256_add_epi16(m1j1, oe_ins_vec); \
    __m256i fn_vec1 = _mm256_add_epi16(f1j1, e_ins_vec);  \
    __m256i fn_vec = _mm256_max_epi16(fn_vec0, fn_vec1);  \
    __m256i mn_vec0 = _mm256_add_epi16(h0j1, score_vec);  \
    __m256i mn_mask = _mm256_cmpgt_epi16(h0j1, zero_vec); \
    __m256i mn_vec = _mm256_and_si256(mn_vec0, mn_mask);  \
    __m256i hn_vec0 = _mm256_max_epi16(en_vec, fn_vec);   \
    __m256i hn_vec = _mm256_max_epi16(hn_vec0, mn_vec);   \
    en_vec = _mm256_max_epi16(en_vec, zero_vec);          \
    fn_vec = _mm256_max_epi16(fn_vec, zero_vec);          \
    mn_vec = _mm256_max_epi16(mn_vec, zero_vec);          \
    hn_vec = _mm256_max_epi16(hn_vec, zero_vec);

// 存储向量化结果
#define SIMD_STORE                                   \
    max_vec = _mm256_max_epi16(max_vec, hn_vec);     \
    _mm256_storeu_si256((__m256i *)&eA2[j], en_vec); \
    _mm256_storeu_si256((__m256i *)&fA2[j], fn_vec); \
    _mm256_storeu_si256((__m256i *)&mA2[j], mn_vec); \
    _mm256_storeu_si256((__m256i *)&hA2[j], hn_vec);

// 去除多余的部分
#define SIMD_REMOVE_EXTRA                                   \
    en_vec = _mm256_and_si256(en_vec, h_vec_mask[end - j]); \
    fn_vec = _mm256_and_si256(fn_vec, h_vec_mask[end - j]); \
    mn_vec = _mm256_and_si256(mn_vec, h_vec_mask[end - j]); \
    hn_vec = _mm256_and_si256(hn_vec, h_vec_mask[end - j]);

// 找最大值和位置
#define SIMD_FIND_MAX                                                                       \
    max_vec = _mm256_max_epu16(max_vec, _mm256_alignr_epi8(max_vec, max_vec, 2));           \
    max_vec = _mm256_max_epu16(max_vec, _mm256_alignr_epi8(max_vec, max_vec, 4));           \
    max_vec = _mm256_max_epu16(max_vec, _mm256_alignr_epi8(max_vec, max_vec, 6));           \
    max_vec = _mm256_max_epu16(max_vec, _mm256_alignr_epi8(max_vec, max_vec, 8));           \
    max_vec = _mm256_max_epu16(max_vec, _mm256_permute2x128_si256(max_vec, max_vec, 0x01)); \
    int16_t *maxVal = (int16_t *)&max_vec;                                                  \
    m = maxVal[0];                                                                          \
    if (m > 0)                                                                              \
    {                                                                                       \
        for (j = beg, i = iend; j <= end; j += SIMD_WIDTH, i -= SIMD_WIDTH)                 \
        {                                                                                   \
            __m256i h2_vec = _mm256_loadu_si256((__m256i *)(&hA2[j]));                      \
            __m256i vcmp = _mm256_cmpeq_epi16(h2_vec, max_vec);                             \
            uint32_t mask = _mm256_movemask_epi8(vcmp);                                     \
            if (mask > 0)                                                                   \
            {                                                                               \
                int pos = SIMD_WIDTH - 1 - ((__builtin_clz(mask)) >> 1);                    \
                mj = j - 1 + pos;                                                           \
                mi = i - 1 - pos;                                                           \
            }                                                                               \
        }                                                                                   \
    }

// 每轮迭代后，交换数组
#define SWAP_DATA_POINTER \
    int16_t *tmp = hA0;   \
    hA0 = hA1;            \
    hA1 = hA2;            \
    hA2 = tmp;            \
    tmp = eA1;            \
    eA1 = eA2;            \
    eA2 = tmp;            \
    tmp = fA1;            \
    fA1 = fA2;            \
    fA2 = tmp;            \
    tmp = mA1;            \
    mA1 = mA2;            \
    mA2 = tmp;

int ksw_extend_avx2(thread_mem_t *tmem,
                    int qlen,              // query length  待匹配段碱基的query长度
                    const uint8_t *query,  // read碱基序列
                    int tlen,              // target length reference的长度
                    const uint8_t *target, // reference序列
                    int extend_left,       // 是不是向左扩展
                    int o_del,             // deletion 错配开始的惩罚系数
                    int e_del,             // deletion extension的惩罚系数
                    int o_ins,             // insertion 错配开始的惩罚系数
                    int e_ins,             // insertion extension的惩罚系数SIMD_BTYES
                    int base_match_score,  // 碱基match时的分数
                    int base_mis_score,    // 碱基mismatch时的惩罚分数（正数）
                    int window_size,       // 提前剪枝系数，w =100   匹配位置和beg的最大距离
                    int end_bonus,         // 如果query比对到了最后一个字符，额外奖励分值
                    int init_score,        // 该seed的初始得分（完全匹配query的碱基数）
                    int *_qle,             // 匹配得到全局最大得分的碱基在query的位置
                    int *_tle,             // 匹配得到全局最大得分的碱基在reference的位置
                    int *_gtle,            // query全部匹配上的target的长度
                    int *_gscore,          // query的端到端匹配得分
                    int *_max_off)         // 取得最大得分时在query和reference上位置差的 最大值
{
    int16_t *mA, *hA, *eA, *fA, *mA1, *mA2, *hA0, *hA1, *eA1, *fA1, *hA2, *eA2, *fA2; // hA0保存上上个col的H，其他的保存上个H E F M
    int16_t *seq, *ref;
    uint8_t *mem;
    int16_t *qtmem, *vmem;
    int seq_size = qlen + SIMD_WIDTH, ref_size = tlen + SIMD_WIDTH;
    int i, iStart, D, j, beg, end, max, max_i, max_j, max_ins, max_del, max_ie, gscore, max_off;
    int Dloop = tlen + qlen; // 循环跳出条件
    int span, beg1, end1;    // 边界条件计算
    int col_size = qlen + 2 + SIMD_WIDTH;
    int val_mem_size = (col_size * 9 * 2 + 31) >> 5 << 5; // 32字节的整数倍
    int mem_size = (seq_size + ref_size) * 2 + val_mem_size;

    SIMD_INIT; // 初始化simd用的数据

    assert(init_score > 0);

    // allocate memory
    mem = thread_mem_request(tmem, mem_size);
    qtmem = (int16_t *)&mem[0];
    seq = &qtmem[0];
    ref = &qtmem[seq_size];
    if (extend_left)
    {
        for (i = 0; i < qlen; ++i)
            seq[i] = query[qlen - 1 - i];
        for (i = 0; i < tlen; ++i)
            ref[i + SIMD_WIDTH] = target[tlen - 1 - i];
    }
    else
    {
        for (i = 0; i < qlen; ++i)
            seq[i] = query[i];
        for (i = 0; i < tlen; ++i)
            ref[i + SIMD_WIDTH] = target[i];
    }

    vmem = &ref[ref_size];
    for (i = 0; i < (val_mem_size >> 1); i += SIMD_WIDTH)
    {
        _mm256_storeu_si256((__m256i *)&vmem[i], zero_vec);
    }
    hA = &vmem[0];
    mA = &vmem[col_size * 3];
    eA = &vmem[col_size * 5];
    fA = &vmem[col_size * 7];

    hA0 = &hA[0];
    hA1 = &hA[col_size];
    hA2 = &hA1[col_size];
    mA1 = &mA[0];
    mA2 = &mA[col_size];
    eA1 = &eA[0];
    eA2 = &eA[col_size];
    fA1 = &fA[0];
    fA2 = &fA[col_size];

    // adjust $window_size if it is too large
    // get the max score
    max = base_match_score;
    max_ins = (int)((double)(qlen * max + end_bonus - o_ins) / e_ins + 1.);
    max_ins = max_ins > 1 ? max_ins : 1;
    window_size = window_size < max_ins ? window_size : max_ins;
    max_del = (int)((double)(qlen * max + end_bonus - o_del) / e_del + 1.);
    max_del = max_del > 1 ? max_del : 1;
    window_size = window_size < max_del ? window_size : max_del; // TODO: is this necessary?
    if (tlen < qlen)
        window_size = MIN(tlen - 1, window_size);

    // DP loop
    max = init_score, max_i = max_j = -1;
    max_ie = -1, gscore = -1;
    ;
    max_off = 0;
    beg = 1;
    end = qlen;
    // init init_score
    hA0[0] = init_score; // 左上角

    if (qlen == 0 || tlen == 0)
        Dloop = 0; // 防止意外情况
    if (window_size >= qlen)
    {
        max_ie = 0;
        gscore = 0;
    }

    int iend;

#ifdef DEBUG_OUT
    int16_t ins[tlen + 1][qlen + 1];
    int16_t del[tlen + 1][qlen + 1];
    int16_t score[tlen + 1][qlen + 1];
    ins[0][0] = del[0][0] = score[0][0] = init_score;
#endif

    for (D = 1; LIKELY(D < Dloop); ++D)
    {
        // 边界条件一定要注意！ tlen 大于，等于，小于 qlen时的情况
        if (D > tlen)
        {
            span = MIN(Dloop - D, window_size);
            beg1 = MAX(D - tlen + 1, ((D - window_size) / 2) + 1);
        }
        else
        {
            span = MIN(D - 1, window_size);
            beg1 = MAX(1, ((D - window_size) / 2) + 1);
        }
        end1 = MIN(qlen, beg1 + span);

        beg = 1;
        end = qlen;

        // if (beg < beg1)
        //     beg = beg1;
        // if (end > end1)
        //     end = end1;
        // if (beg > end)
        //     break; // 不用计算了，直接跳出，否则hA2没有被赋值，里边是上一轮hA0的值，会出bug

        iend = D - (beg - 1); // ref开始计算的位置，倒序
        span = end - beg;
        iStart = iend - span - 1; // 0开始的ref索引位置

        // 每一轮需要记录的数据
        int m = 0, mj = -1, mi = -1;
        max_vec = zero_vec;

        // 要处理边界
        // 左边界 处理f (insert)
        if (iStart == 0)
        {
            hA1[end] = MAX(0, init_score - (o_ins + e_ins * end));
        }
        // 上边界
        if (beg == 1)
        {
            hA1[0] = MAX(0, init_score - (o_del + e_del * iend));
        }
        else
        {
            hA1[beg - 1] = 0;
            eA1[beg - 1] = 0;
        }

        for (j = beg, i = iend; j <= end + 1 - SIMD_WIDTH; j += SIMD_WIDTH, i -= SIMD_WIDTH)
        {
            // 取数据
            SIMD_LOAD;
            // 比对seq，计算罚分
            SIMD_CMP_SEQ;
            // 计算
            SIMD_COMPUTE;
            // 存储结果
            SIMD_STORE;
        }
        // 剩下的计算单元
        if (j <= end)
        {
            // 取数据
            SIMD_LOAD;
            // 比对seq，计算罚分
            SIMD_CMP_SEQ;
            // 计算
            SIMD_COMPUTE;
            // 去除多余计算的部分
            SIMD_REMOVE_EXTRA;
            // 存储结果
            SIMD_STORE;
        }

        SIMD_FIND_MAX;

        // 注意最后跳出循环j的值
        j = end + 1;

        if (j == qlen + 1)
        {
            max_ie = gscore > hA2[qlen] ? max_ie : iStart;
            gscore = gscore > hA2[qlen] ? gscore : hA2[qlen];
        }

        if (m > max)
        {
            max = m, max_i = mi, max_j = mj;
            max_off = max_off > abs(mj - mi) ? max_off : abs(mj - mi);
        }

        // 调整计算的边界
        /*for (j = beg; LIKELY(j <= end); ++j)
        {
            int has_val = hA1[j - 1] | hA2[j];
            if (has_val)
                break;
        }
        beg = j;
        for (j = end + 1; LIKELY(j >= beg); --j)
        {
            int has_val = hA1[j - 1] | hA2[j];
            if (has_val)
                break;
            else
                hA0[j - 1] = 0;
        }
        end = j + 1 <= qlen ? j + 1 : qlen;
        // beg = 0;
        // end = qlen; // uncomment this line for debugging
        */
        // swap m, h, e, f
        SWAP_DATA_POINTER;
    }

    // free(mem);
    thread_mem_release(tmem, mem_size);
    if (_qle)
        *_qle = max_j + 1;
    if (_tle)
        *_tle = max_i + 1;
    if (_gtle)
        *_gtle = max_ie + 1;
    if (_gscore)
        *_gscore = gscore;
    if (_max_off)
        *_max_off = max_off;
    return max;
}