diff --git a/bwt.cpp b/bwt.cpp index b2d2438..bd9f45e 100644 --- a/bwt.cpp +++ b/bwt.cpp @@ -155,7 +155,7 @@ bwtint_t bwt_occ(const bwt_t *bwt, bwtint_t k, uint8_t c) // 统计k行（bwt mtx行）之前4种碱基累积数量，这里的k是bwt矩阵里的行，比bwt字符串多1 void bwt_occ4(const bwt_t *bwt, bwtint_t k, bwtint_t cnt[4]) { - bwtint_t x; + bwtint_t x = 0; uint32_t *p, tmp, *end; // bwtint_t bwt_k_base_line = k >> OCC_INTV_SHIFT << OCC_INTV_SHIFT; if (k == (bwtint_t)(-1)) @@ -165,7 +165,7 @@ void bwt_occ4(const bwt_t *bwt, bwtint_t k, bwtint_t cnt[4]) } k -= (k >= bwt->primary); // because $ is not in bwt p = bwt_occ_intv(bwt, k); - //cout << "k: " << k << "; occ: " << p[0] << ' ' << p[1] << ' ' << p[2] << ' ' << p[3] << endl; + // cout << "k: " << k << "; occ: " << p[0] << ' ' << p[1] << ' ' << p[2] << ' ' << p[3] << endl; cnt[0] = p[0]; cnt[1] = p[1]; cnt[2] = p[2]; @@ -173,14 +173,14 @@ void bwt_occ4(const bwt_t *bwt, bwtint_t k, bwtint_t cnt[4]) p += 4; // sizeof(bwtint_t) = 4*(sizeof(bwtint_t)/sizeof(uint32_t)) end = p + ((k >> 4) - ((k & ~OCC_INTV_MASK) >> 4)); // this is the end point of the following loop //end = p + 1; - //cout << "k - kbase: " << k - bwt_k_base_line << endl; - for (x = 0; p < end; ++p) + // cout << "k - kbase: " << k - bwt_k_base_line << endl; + for (; p < end; ++p) { x += __occ_aux4(bwt, *p); //print_base_uint32(*p); } tmp = *p & ~((1U << ((~k & 15) << 1)) - 1); - //print_base_uint32(tmp); + // print_base_uint32(tmp); x += __occ_aux4(bwt, tmp) - (~k & 15); // 这里多算了A，要减去 cnt[0] += x & 0xff; cnt[1] += x >> 8 & 0xff; diff --git a/bwt.h b/bwt.h index e6c555a..faae948 100644 --- a/bwt.h +++ b/bwt.h @@ -7,7 +7,7 @@ using std::string; // occ间隔对应的右移位数，5表示间隔32行保存一次 -#define OCC_INTV_SHIFT 6 +#define OCC_INTV_SHIFT 7 #define OCC_INTERVAL (1LL << OCC_INTV_SHIFT) #define OCC_INTV_MASK (OCC_INTERVAL - 1) diff --git a/fmt_index.cpp b/fmt_index.cpp index 4be4fe7..e994ef8 100644 --- a/fmt_index.cpp +++ b/fmt_index.cpp @@ -120,6 +120,23 @@ void fmt_gen_cnt_occ(FMTIndex *fmt) } } +// fmt-index的count table，4对应着bwt碱基的累积量，0,1,2,3分别对应着bwt是A,C,G,T，pre-bwt的累积量 +void fmt_gen_cnt_table(uint32_t cnt_table[4][256]) +{ + int i, j, k; + uint32_t x = 0; + for (i = 0; i != 256; ++i) // 遍历单个字节的各种情况 + { + for (k = 0; k < 4; ++k) // bwt碱基 + { + x = 0; // for [A,C,G,T][A,C,G,T] + for (j = 0; j != 4; ++j) // pre-bwt碱基 + x |= (((i >> 6 & 3) == j && (i >> 4 & 3) == k) + ((i >> 2 & 3) == j && (i & 3) == k)) << (j << 3); + cnt_table[k][i] = x; + } + } +} + void dump_fmt(const char *fn, const FMTIndex *fmt) { FILE *fp; @@ -169,6 +186,13 @@ FMTIndex *create_fmt_from_bwt(bwt_t *bwt) uint32_t c[4], c2[16]; /*c用来保存原来的bwt碱基串的累积值，c2用来保存pre-bwt和bwt碱基对的累计值，如AA..TT*/ uint32_t *buf; /* 计算之后变成fmt结构中bwt部分 */ +#ifdef FMT_MID_INTERVAL + // 加入中间的check point + uint32_t mc[4] = {0}; + uint32_t cnt_table[4][256]; // 4对应原来的cnt_table，0,1,2,3,分别对应该碱基的扩展 + fmt_gen_cnt_table(cnt_table); +#endif + fmt->seq_len = bwt->seq_len; // bwt碱基序列的长度，不包含$字符，也就是该长度比bwt matrix长度少1 for (i = 0; i < 5; ++i) fmt->L2[i] = bwt->L2[i]; // 每个碱基的总累积值 @@ -177,6 +201,14 @@ FMTIndex *create_fmt_from_bwt(bwt_t *bwt) n_occ = (bwt->seq_len + FMT_OCC_INTERVAL - 1) / FMT_OCC_INTERVAL + 1; // check point 个数 fmt->bwt_size = (fmt->seq_len * 2 + 15) >> 4; // 要保存最后两列碱基 fmt->bwt_size += n_occ * 20; // A,C,G,T和AA,AC.....TG,TT共20个 + +#ifdef FMT_MID_INTERVAL + bwtint_t mn_occ = (bwt->seq_len >> FMT_OCC_INTV_SHIFT) * (FMT_OCC_INTERVAL / FMT_MID_INTERVAL - 1); + bwtint_t last_seq_len = bwt->seq_len % FMT_OCC_INTERVAL; + mn_occ += (last_seq_len + FMT_MID_INTERVAL - 1) / FMT_MID_INTERVAL - 1; + fmt->bwt_size += mn_occ * 4; +#endif + buf = (uint32_t *)calloc(fmt->bwt_size, 4); // 开辟计算fmt用到的缓存 c[0] = c[1] = c[2] = c[3] = 0; @@ -198,6 +230,7 @@ FMTIndex *create_fmt_from_bwt(bwt_t *bwt) k += 4; memcpy(buf + k, c2, sizeof(uint32_t) * 16); // pre-bwt:bwt碱基对的occ k += 16; + mc[0] = mc[1] = mc[2] = mc[3] = 0; } // 每个32位整数保存8个倒数第二列碱基（pre-bwt）和8个倒数第一列(bwt)碱基 if (i % 16 == 0) // 每个32位整数可以包含16个碱基，每次需要处理16个碱基，也就是间隔最小可以设置为16 @@ -257,6 +290,7 @@ FMTIndex *create_fmt_from_bwt(bwt_t *bwt) } // 保存bwt和pre_bwt uint32_t pre_and_bwt_seq = 0; + uint32_t pre_and_bwt_seq_2 = 0; for (m = 0; m < 8; ++m) { int lshift_bit = 30 - 2 * m; @@ -265,14 +299,28 @@ FMTIndex *create_fmt_from_bwt(bwt_t *bwt) buf[k++] = pre_and_bwt_seq; if (j > 8) { - pre_and_bwt_seq = 0; for (m = 8; m > 0; --m) { int lshift_bit = 2 * m - 2; - pre_and_bwt_seq |= (((pre_bwt_16_seq & (3 << lshift_bit)) << (m * 2)) | ((bwt_16_seq & (3 << lshift_bit)) << (m * 2 - 2))); + pre_and_bwt_seq_2 |= (((pre_bwt_16_seq & (3 << lshift_bit)) << (m * 2)) | ((bwt_16_seq & (3 << lshift_bit)) << (m * 2 - 2))); } - buf[k++] = pre_and_bwt_seq; + buf[k++] = pre_and_bwt_seq_2; } +#ifdef FMT_MID_INTERVAL + // 加入中间的check point + for (m = 0; m < 4; ++m) + { + uint32_t s1 = pre_and_bwt_seq; + mc[m] += cnt_table[m][s1 & 0xff] + cnt_table[m][s1 >> 8 & 0xff] + cnt_table[m][s1 >> 16 & 0xff] + cnt_table[m][s1 >> 24]; + s1 = pre_and_bwt_seq_2; + mc[m] += cnt_table[m][s1 & 0xff] + cnt_table[m][s1 >> 8 & 0xff] + cnt_table[m][s1 >> 16 & 0xff] + cnt_table[m][s1 >> 24]; + } + if ((i + 16) % FMT_OCC_INTERVAL != 0 && j > 8) // 序列最后的部分不用保存 + { + for (m = 0; m < 4; ++m) + buf[k++] = mc[m]; + } +#endif } } // the last element @@ -284,15 +332,15 @@ FMTIndex *create_fmt_from_bwt(bwt_t *bwt) // cout << "size: " << k << '\t' << fmt->bwt_size << endl; xassert(k == fmt->bwt_size, "inconsistent bwt_size"); // update fmt - fmt->bwt = buf; + fmt->bwt = buf; return fmt; } // 扩展两个个碱基，计算bwt base为b的pre-bwt str中各个碱基的occ void fmt_e2_occ(const FMTIndex *fmt, bwtint_t k, int b1, int b2, bwtint_t cnt[4]) { - uint32_t x; - uint32_t *p, *q, tmp, *end; + uint32_t x = 0; + uint32_t *p, *q, tmp; bwtint_t bwt_k_line = k, bwt_k_base_line = k >> FMT_OCC_INTV_SHIFT << FMT_OCC_INTV_SHIFT; int i, ti; cnt[0] = 0; @@ -301,25 +349,80 @@ void fmt_e2_occ(const FMTIndex *fmt, bwtint_t k, int b1, int b2, bwtint_t cnt[4] if (k == (bwtint_t)(-1)) { p = fmt->bwt + 4 + b1 * 4; - for (i = 3; i > b2; --i) cnt[2] += p[i]; + for (i = b2 + 1; i < 4; ++i) cnt[2] += p[i]; cnt[3] = p[b2]; return; } ti = b1 << 2 | b2; - // _mm_prefetch((const char *)(&fmt->cnt_occ[ti]), _MM_HINT_T0); k -= (k >= fmt->primary); // k由bwt矩阵对应的行转换成bwt字符串对应的行（去掉了$，所以大于$的行，都减掉1） p = fmt_occ_intv(fmt, k); // cout << "k-base: " << k << "; occ: " << p[0] << ' ' << p[1] << ' ' << p[2] << ' ' << p[3] << endl; - for (i = 3; i > b1; --i) cnt[0] += p[i]; - cnt[1] = p[b1]; + for (i = b1 + 1; i < 4; ++i) cnt[0] += p[i]; // 大于b1的碱基的occ之和 + cnt[1] = p[b1]; // b1的occ q = p + 4 + b1 * 4; - for (i = 3; i > b2; --i) cnt[2] += q[i]; - cnt[3] = q[b2]; - p += 20; // 该地址是bwt和pre_bwt字符串数据的首地址 - end = p + ((k >> 3) - ((k & ~FMT_OCC_INTV_MASK) >> 3)); // this is the end point of the following loop - //p = end - (end - p) / 4; + for (i = b2 + 1; i < 4 ; ++i) cnt[2] += q[i]; // 大于b2的occ之和 + cnt[3] = q[b2]; // b2的occ + p += 20; +#ifdef FMT_MID_INTERVAL + // 使用mid interval信息 + int mk = k % FMT_OCC_INTERVAL; + int n_mintv = mk >> FMT_MID_INTV_SHIFT; + //if (k == 3137454504) + //{ + // for (i = 0; i < n_mintv; ++i) + // { + // q = p + i * 6; + // print_base_uint32(*q); + // print_base_uint32(*(q + 1)); + // x = *(q + 2); + // cout << ((x) >> 24 & 0xff) << ' ' << ((x) >> 16 & 0xff) << ' ' << ((x) >> 8 & 0xff) << ' ' << ((x) & 0xff) << ' ' << __fmt_mid_sum(x) << endl; + // x = *(q + 3); + // cout << ((x) >> 24 & 0xff) << ' ' << ((x) >> 16 & 0xff) << ' ' << ((x) >> 8 & 0xff) << ' ' << ((x) & 0xff) << ' ' << __fmt_mid_sum(x) << endl; + // x = *(q + 4); + // cout << ((x) >> 24 & 0xff) << ' ' << ((x) >> 16 & 0xff) << ' ' << ((x) >> 8 & 0xff) << ' ' << ((x) & 0xff) << ' ' << __fmt_mid_sum(x) << endl; + // x = *(q + 5); + // cout << ((x) >> 24 & 0xff) << ' ' << ((x) >> 16 & 0xff) << ' ' << ((x) >> 8 & 0xff) << ' ' << ((x) & 0xff) << ' ' << __fmt_mid_sum(x) << endl; + // } + // x = 0; + //} + if (n_mintv > 0) // 至少超过了第一个mid interval + { + p += n_mintv * 6 - 4; // 对应的mid interval check point的首地址，即A C G T的局部累积量 + q = p + b1; + for (i = b1 + 1; i < 4; ++i) + x += p[i]; // 大于b1的碱基的occ之和 + cnt[0] += __fmt_mid_sum(x); + x = *q; + // if (k == 3137454504) + // { + // cout << ((x) >> 24 & 0xff) << '\t' << ((x) >> 16 & 0xff) + // << '\t' << ((x) >> 8 & 0xff) << '\t' << ((x) & 0xff) << endl; + // cout << __fmt_mid_sum(x) << endl; + // } + cnt[1] += __fmt_mid_sum(x); // b1的occ + for (i = 3; i > b2; --i) + cnt[2] += x >> (i << 3) & 0xff; // 大于b2的occ之和 + cnt[3] += x >> (b2 << 3) & 0xff; // b2的occ + x = 0; + p += 4; + } + + // cout << "k: " << k << ' ' << cnt[1] << endl; + + if ((mk & FMT_MID_INTV_MASK) >> 3) + { + x += __fmt_occ_e2_aux2(fmt, ti, *p); + // print_base_uint32(*p); + ++p; + } + tmp = *p & ~((1U << ((~k & 7) << 2)) - 1); + // print_base_uint32(tmp); + x += __fmt_occ_e2_aux2(fmt, ti, tmp); +#else // 该地址是bwt和pre_bwt字符串数据的首地址 + uint32_t *end = p + ((k >> 3) - ((k & ~FMT_OCC_INTV_MASK) >> 3)); // this is the end point of the following loop + // p = end - (end - p) / 8; // cout << "k - kbase: " << k - bwt_k_base_line << endl; - for (x = 0; p < end; ++p) + for (; p < end; ++p) { x += __fmt_occ_e2_aux2(fmt, ti, *p); // print_base_uint32(*p); @@ -327,14 +430,7 @@ void fmt_e2_occ(const FMTIndex *fmt, bwtint_t k, int b1, int b2, bwtint_t cnt[4] tmp = *p & ~((1U << ((~k & 7) << 2)) - 1); // print_base_uint32(tmp); x += __fmt_occ_e2_aux2(fmt, ti, tmp); - // end = p + (end - p) / 4; - // end = p + 2; - // p = end - (end - p) / 32; - // if (k % 2 == 0) - // x += __fmt_occ_e2_aux2(fmt, ti, *p); - // p += 1; - // tmp = *p & ~((1U << ((~k & 7) << 2)) - 1); - // x += __fmt_occ_e2_aux2(fmt, ti, tmp); +#endif if (b1 == 0) { @@ -345,10 +441,10 @@ void fmt_e2_occ(const FMTIndex *fmt, bwtint_t k, int b1, int b2, bwtint_t cnt[4] // 如果跨过了second_primary,那么可能需要减掉一次累积值 if (b1 == fmt->first_base && bwt_k_base_line < fmt->sec_primary && bwt_k_line >= fmt->sec_primary) { - if (b2 == fmt->last_base) - x -= 1 << 24; - else if (b2 < fmt->last_base) + if (b2 < fmt->last_base) x -= 1 << 16; + else if (b2 == fmt->last_base) + x -= 1 << 24; } cnt[0] += x & 0xff; cnt[1] += x >> 8 & 0xff; @@ -405,13 +501,12 @@ bwtintv_t fmt_search(FMTIndex *fmt, const string &q) fmt_set_intv(fmt, bval(q[x]), ik); ik.info = x + 1; - // cout << "fmt : " << ik.x[0] << '\t' << ik.x[1] << '\t' << ik.x[2] << endl; + // cout << "fmt : " << ik.x[0] << '\t' << ik.x[1] << '\t' << ik.x[2] << endl; for (i = x + 1; i + 1 < qlen; i += 2) { if (bval(q[i]) < 4 && bval(q[i + 1]) < 4) { - c1 = cbval(q[i]); c2 = cbval(q[i + 1]); //double tm_t = realtime(); @@ -440,6 +535,14 @@ bwtintv_t fmt_search(FMTIndex *fmt, const string &q) return ik; } +uint32_t str2bit(string &str) +{ + uint32_t pac = 0; + for (int i = 0; i < 16; ++i) + pac = (pac << 2) | bval(str[i]); + return pac; +} + int main_fmtidx(int argc, char **argv) { // string seq("ACCCTAACCCTAACCCTAACCCTAACCCTAACCCTAACCCTAACCCTAACCCTAACCCTA"); @@ -449,9 +552,31 @@ int main_fmtidx(int argc, char **argv) //create_bwt_mtx(seq); //cout << seq << endl; - //string bwt_str = string(argv[1]) + ".bwt.str"; - // string bwt_idx = string(argv[1]) + ".128.bwt"; - string bwt_idx = string(argv[1]) + ".64.bwt"; + // uint32_t cnt_table[4][256]; + // fmt_gen_cnt_table(cnt_table); + // string ss1("ATATATCAATTCTCTT"); + // string ss2("ATTGCCTCTGAATTAC"); + // uint32_t bs1 = str2bit(ss1), bs2 = str2bit(ss2); + // uint32_t s1 = bs1; + // uint32_t x = 0; + // for (int m = 0; m < 4; ++m) + // { + // x = cnt_table[m][s1 & 0xff] + cnt_table[m][s1 >> 8 & 0xff] + cnt_table[m][s1 >> 16 & 0xff] + cnt_table[m][s1 >> 24]; + // cout << ((x) >> 24 & 0xff) << ' ' << ((x) >> 16 & 0xff) << ' ' << ((x) >> 8 & 0xff) << ' ' << ((x) & 0xff) << ' ' << __fmt_mid_sum(x) << endl; + // } + // x = 0; + // s1 = bs2; + // for (int m = 0; m < 4; ++m) + // { + // x = cnt_table[m][s1 & 0xff] + cnt_table[m][s1 >> 8 & 0xff] + cnt_table[m][s1 >> 16 & 0xff] + cnt_table[m][s1 >> 24]; + // cout << ((x) >> 24 & 0xff) << ' ' << ((x) >> 16 & 0xff) << ' ' << ((x) >> 8 & 0xff) << ' ' << ((x) & 0xff) << ' ' << __fmt_mid_sum(x) << endl; + // } + + // return 0; + + // string bwt_str = string(argv[1]) + ".bwt.str"; + string bwt_idx = string(argv[1]) + ".128.bwt"; + //string bwt_idx = string(argv[1]) + ".64.bwt"; // string bwt_idx = string(argv[1]) + ".16.bwt"; // string bwt_idx = string(argv[1]) + ".bwt"; //bwt_t *bwt = restore_bwt(bwt_str.c_str()); @@ -459,17 +584,21 @@ int main_fmtidx(int argc, char **argv) //dump_bwt(bwt_idx.c_str(), bwt); //string fmt_idx = string(argv[1]) + ".fmt"; - // string fmt_idx = string(argv[1]) + ".256.fmt"; + string fmt_idx = string(argv[1]) + ".256.fmt"; // string fmt_idx = string(argv[1]) + ".128.fmt"; - string fmt_idx = string(argv[1]) + ".64.fmt"; + // string fmt_idx = string(argv[1]) + ".64.fmt"; // string fmt_idx = string(argv[1]) + ".32.fmt"; - - + t11 = realtime(); bwt_t *bwt = restore_bwt(bwt_idx.c_str()); + t11 = realtime() - t11; + cout << "[time read bwt:] " << t11 << "s" << endl; + t10 = realtime(); FMTIndex *fmt = restore_fmt(fmt_idx.c_str()); // FMTIndex *fmt = create_fmt_from_bwt(bwt); // dump_fmt(fmt_idx.c_str(), fmt); + t10 = realtime() - t10; + cout << "[time gen fmt:] " << t10 << "s" << endl; vector seed_arr(10000000); seed_arr[0] = "GCGATACTAAGA"; @@ -492,6 +621,8 @@ int main_fmtidx(int argc, char **argv) t3 = realtime() - t3; cout << "[time fmt search:] " << t3 << "s" << endl; + return 0; + t4 = realtime(); for (int i = 0; i < (int)seed_arr.size(); ++i) { diff --git a/fmt_index.h b/fmt_index.h index 490557f..e825505 100644 --- a/fmt_index.h +++ b/fmt_index.h @@ -1,16 +1,23 @@ #ifndef FMT_INDEX_H_ #define FMT_INDEX_H_ -#define FMT_OCC_INTV_SHIFT 6 +#define FMT_OCC_INTV_SHIFT 8 #define FMT_OCC_INTERVAL (1LL << FMT_OCC_INTV_SHIFT) #define FMT_OCC_INTV_MASK (FMT_OCC_INTERVAL - 1) -#define FMT_MID_INTV 16 +#define FMT_MID_INTV_SHIFT 4 +#define FMT_MID_INTERVAL (1LL << FMT_MID_INTV_SHIFT) +#define FMT_MID_INTV_MASK (FMT_MID_INTERVAL - 1) // 获取碱基c（待查找序列的首个碱基）和对应的互补碱基对应的行，以及间隔 #define fmt_set_intv(fmt, c, ik) ((ik).x[0] = (fmt)->L2[(int)(c)] + 1, (ik).x[2] = (fmt)->L2[(int)(c) + 1] - (fmt)->L2[(int)(c)], (ik).x[1] = (fmt)->L2[3 - (c)] + 1, (ik).info = 0) // k行（bwt str行（不包含$））对应的check point occ数据起始地址（小于k且是OCC_INTERVAL的整数倍） +#ifdef FMT_MID_INTERVAL +// 包含mid interval +#define fmt_occ_intv(b, k) ((b)->bwt + (k) / FMT_OCC_INTERVAL * (FMT_OCC_INTERVAL / 8 + 80)) +#else #define fmt_occ_intv(b, k) ((b)->bwt + (k) / FMT_OCC_INTERVAL * (FMT_OCC_INTERVAL / 8 + 20)) +#endif // 字节val中包含bwt base为b的pre-bwt中T G C A（按顺序保存在32位整数里（每个占8bit））的数量 #define __fmt_occ_e2_aux4(fmt, b, val) \ ((fmt)->cnt_table[(b)][(val) & 0xff] + (fmt)->cnt_table[b][(val) >> 8 & 0xff] + (fmt)->cnt_table[b][(val) >> 16 & 0xff] + (fmt)->cnt_table[b][(val) >> 24]) @@ -18,6 +25,9 @@ #define __fmt_occ_e2_aux2(fmt, b, val) \ ((fmt)->cnt_occ[(b)][(val) & 0xff] + (fmt)->cnt_occ[b][(val) >> 8 & 0xff] + (fmt)->cnt_occ[b][(val) >> 16 & 0xff] + (fmt)->cnt_occ[b][(val) >> 24]) +#define __fmt_mid_sum(x) \ + ((x) >> 24 & 0xff) + ((x) >> 16 & 0xff) + ((x) >> 8 & 0xff) + ((x) & 0xff) + // fm-index, extend twice in one search step (one memory access) struct FMTIndex { @@ -28,7 +38,6 @@ struct FMTIndex bwtint_t bwt_size; // size of bwt, about seq_len/4 uint32_t *bwt; // BWT // occurance array, separated to two parts - uint32_t cnt_table[5][256]; // 4对应原来的cnt_table，0,1,2,3,分别对应该碱基的扩展 uint32_t cnt_occ[16][256]; // 前16-24位表示b（碱基）的occ，8-16位表示大于b的occ，0-8表示大于a的occ，ba格式 uint8_t sec_bcp; // base couple for sec primary line, AA=>0, AC=>1 ... TT=>15 uint8_t first_base; // 序列的第一个碱基2bit的int类型，0,1,2,3