zzh
/
sw_perf
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

运行大的数据会内存崩溃，还没解决

This commit is contained in:
zzh 2023-08-27 16:43:20 +08:00
parent 78f791f3f2
commit f81871eff5
10 changed files with 394 additions and 419 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
Makefile
View File

@ -1,7 +1,7 @@
CC= gcc CC= gcc
#CFLAGS= -g -Wall -Wno-unused-function -mavx2 #CFLAGS= -g -Wall -Wno-unused-function -mavx2
CFLAGS= -Wall -Wno-unused-function -O2 -mavx2 CFLAGS= -Wall -Wno-unused-function -O2 -mavx2
DFLAGS= -DSHOW_PERF -DDEBUG_RETURN_VALUE DFLAGS= -DSHOW_PERF
#DFLAGS= -DSHOW_PERF -DDEBUG_OUT -DDEBUG_RETURN_VALUE #DFLAGS= -DSHOW_PERF -DDEBUG_OUT -DDEBUG_RETURN_VALUE
PROG= sw_perf PROG= sw_perf
INCLUDES= INCLUDES=

36
ksw_ext_avx2.c
View File

@ -109,13 +109,15 @@ static const uint16_t h_vec_int_mask[SIMD_WIDTH][SIMD_WIDTH] = {
fn_vec = _mm256_max_epi16(fn_vec, zero_vec); \ fn_vec = _mm256_max_epi16(fn_vec, zero_vec); \
mn_vec = _mm256_max_epi16(mn_vec, zero_vec); \ mn_vec = _mm256_max_epi16(mn_vec, zero_vec); \
hn_vec = _mm256_max_epi16(hn_vec, zero_vec); hn_vec = _mm256_max_epi16(hn_vec, zero_vec);
//int16_t *t_ptr = (int16_t *)&ts_vec; \
//fprintf(stderr, "D: %d, ibeg: %d, iend: %d, jbeg: %d, jend: %d, %d %d %d %d %d %d %d %d %d %d %d %d %d %d %d %d \n", \ /* int16_t *t_ptr = (int16_t *)&ts_vec; \
// D, ibeg, iend, beg, end, \ fprintf(stderr, "D: %d, ibeg: %d, iend: %d, jbeg: %d, jend: %d, %d %d %d %d %d %d %d %d %d %d %d %d %d %d %d %d \n", \
// t_ptr[0], t_ptr[1], t_ptr[2], t_ptr[3], \ D, ibeg, iend, beg, end, \
// t_ptr[4], t_ptr[5], t_ptr[6], t_ptr[7], \ t_ptr[0], t_ptr[1], t_ptr[2], t_ptr[3], \
// t_ptr[8], t_ptr[9], t_ptr[10], t_ptr[11], \ t_ptr[4], t_ptr[5], t_ptr[6], t_ptr[7], \
// t_ptr[12], t_ptr[13], t_ptr[14], t_ptr[15]); t_ptr[8], t_ptr[9], t_ptr[10], t_ptr[11], \
t_ptr[12], t_ptr[13], t_ptr[14], t_ptr[15]);
*/
// 存储向量化结果 // 存储向量化结果
#define SIMD_STORE \ #define SIMD_STORE \
@ -220,7 +222,7 @@ int ksw_extend_avx2(thread_mem_t *tmem,
for (i = 0; i < qlen; ++i) for (i = 0; i < qlen; ++i)
seq[i] = query[qlen - 1 - i]; seq[i] = query[qlen - 1 - i];
for (i = 0; i < tlen; ++i) for (i = 0; i < tlen; ++i)
ref[i + SIMD_WIDTH] = target[tlen - 1 - i]; ref[i + SIMD_WIDTH - 1] = target[tlen - 1 - i];
} }
else else
{ {
@ -305,16 +307,16 @@ int ksw_extend_avx2(thread_mem_t *tmem,
end1 = D; // 闭区间 end1 = D; // 闭区间
else else
end1 = qlen; end1 = qlen;
// beg1 = MAX(D - window_size, beg1); beg1 = MAX(D - window_size, beg1);
// end1 = MIN(D + window_size, end1); end1 = MIN(D + window_size, end1);
beg = MAX(beg1, beg); beg = MAX(beg1, beg);
end = MIN(end1, end); end = MIN(end1, end);
// if (beg > end) if (beg > end)
// break; break;
beg = beg1; // beg = beg1;
end = end1; // end = end1;
iend = D - beg; // ref开始计算的位置，倒序 iend = D - beg; // ref开始计算的位置，倒序
span = end - beg; span = end - beg;
@ -413,7 +415,7 @@ int ksw_extend_avx2(thread_mem_t *tmem,
} }
// 调整计算的边界 // 调整计算的边界
/*for (j = beg; LIKELY(j <= end); ++j) for (j = beg; LIKELY(j <= end); ++j)
{ {
int has_val = hA1[j - 1] | hA2[j]; int has_val = hA1[j - 1] | hA2[j];
if (has_val) if (has_val)
@ -425,13 +427,11 @@ int ksw_extend_avx2(thread_mem_t *tmem,
int has_val = hA1[j - 1] | hA2[j]; int has_val = hA1[j - 1] | hA2[j];
if (has_val) if (has_val)
break; break;
else
hA0[j - 1] = 0;
} }
end = j + 1 <= qlen ? j + 1 : qlen; end = j + 1 <= qlen ? j + 1 : qlen;
// beg = 0; // beg = 0;
// end = qlen; // uncomment this line for debugging // end = qlen; // uncomment this line for debugging
*/
// swap m, h, e, f // swap m, h, e, f
SWAP_DATA_POINTER; SWAP_DATA_POINTER;
} }

321
ksw_ext_avx2_aligned.c
View File

@ -85,14 +85,14 @@ static const uint16_t h_vec_int_mask[SIMD_WIDTH][SIMD_WIDTH] = {
* h 表示最大值 * h 表示最大值
*/ */
// load向量化数据 // load向量化数据
#define SIMD_LOAD \ #define SIMD_LOAD \
__m256i m1 = _mm256_loadu_si256((__m256i *)(&cur_match_arr[j])); \ __m256i m1 = _mm256_loadu_si256((__m256i *)(&mA1[j])); \
__m256i e1 = _mm256_loadu_si256((__m256i *)(&cur_del_arr[j])); \ __m256i e1 = _mm256_loadu_si256((__m256i *)(&eA1[j])); \
__m256i m1j1 = _mm256_loadu_si256((__m256i *)(&cur_match_arr[j - 1])); \ __m256i m1j1 = _mm256_loadu_si256((__m256i *)(&mA1[j - 1])); \
__m256i f1j1 = _mm256_loadu_si256((__m256i *)(&cur_ins_arr[j - 1])); \ __m256i f1j1 = _mm256_loadu_si256((__m256i *)(&fA1[j - 1])); \
__m256i h0j1 = _mm256_loadu_si256((__m256i *)(&last_max_arr[j - 1])); \ __m256i h0j1 = _mm256_loadu_si256((__m256i *)(&hA0[j - 1])); \
__m256i qs_vec = _mm256_loadu_si256((__m256i *)(&read_seq[j])); \ __m256i qs_vec = _mm256_loadu_si256((__m256i *)(&seq[j])); \
__m256i ts_vec = _mm256_loadu_si256((__m256i *)(&ref_seq[i])); __m256i ts_vec = _mm256_loadu_si256((__m256i *)(&ref[i]));
// 比对ref和seq的序列，计算罚分 // 比对ref和seq的序列，计算罚分
#define SIMD_CMP_SEQ \ #define SIMD_CMP_SEQ \
@ -130,74 +130,74 @@ static const uint16_t h_vec_int_mask[SIMD_WIDTH][SIMD_WIDTH] = {
mn_vec = _mm256_max_epi16(mn_vec, zero_vec); \ mn_vec = _mm256_max_epi16(mn_vec, zero_vec); \
hn_vec = _mm256_max_epi16(hn_vec, zero_vec); hn_vec = _mm256_max_epi16(hn_vec, zero_vec);
#define SIMD_STORE \ #define SIMD_STORE \
max_vec = _mm256_max_epu8(max_vec, hn_vec); \ max_vec = _mm256_max_epu8(max_vec, hn_vec); \
_mm256_storeu_si256((__m256i *)&next_del_arr[j], en_vec); \ _mm256_storeu_si256((__m256i *)&eA2[j], en_vec); \
_mm256_storeu_si256((__m256i *)&next_ins_arr[j], fn_vec); \ _mm256_storeu_si256((__m256i *)&fA2[j], fn_vec); \
_mm256_storeu_si256((__m256i *)&next_match_arr[j], mn_vec); \ _mm256_storeu_si256((__m256i *)&mA2[j], mn_vec); \
_mm256_storeu_si256((__m256i *)&next_max_arr[j], hn_vec); _mm256_storeu_si256((__m256i *)&hA2[j], hn_vec);
// 去除多余的部分 // 去除多余的部分
#define SIMD_REMOVE_EXTRA \ #define SIMD_REMOVE_EXTRA \
en_vec = _mm256_and_si256(en_vec, h_vec_mask[read_end_pos - j]); \ en_vec = _mm256_and_si256(en_vec, h_vec_mask[end - j]); \
fn_vec = _mm256_and_si256(fn_vec, h_vec_mask[read_end_pos - j]); \ fn_vec = _mm256_and_si256(fn_vec, h_vec_mask[end - j]); \
mn_vec = _mm256_and_si256(mn_vec, h_vec_mask[read_end_pos - j]); \ mn_vec = _mm256_and_si256(mn_vec, h_vec_mask[end - j]); \
hn_vec = _mm256_and_si256(hn_vec, h_vec_mask[read_end_pos - j]); hn_vec = _mm256_and_si256(hn_vec, h_vec_mask[end - j]);
// 找最大值和位置 // 找最大值和位置
#define SIMD_FIND_MAX \ #define SIMD_FIND_MAX \
__m256i cmp_max = _mm256_cmpgt_epi16(max_vec, last_max_vec); \ __m256i cmp_max = _mm256_cmpgt_epi16(max_vec, last_max_vec); \
uint32_t cmp_result = _mm256_movemask_epi8(cmp_max); \ uint32_t cmp_result = _mm256_movemask_epi8(cmp_max); \
if (cmp_result > 0) \ if (cmp_result > 0) \
{ \ { \
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, 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, 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, 6)); \
max_vec = _mm256_max_epu16(max_vec, _mm256_alignr_epi8(max_vec, max_vec, 8)); \ 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)); \ max_vec = _mm256_max_epu16(max_vec, _mm256_permute2x128_si256(max_vec, max_vec, 0x01)); \
int16_t *maxVal = (int16_t *)&max_vec; \ int16_t *maxVal = (int16_t *)&max_vec; \
m = maxVal[0]; \ m = maxVal[0]; \
for (j = aligned_read_start_pos, i = aligned_ref_end_pos; j <= read_end_pos; j += SIMD_WIDTH, i -= SIMD_WIDTH) \ for (j = aligned_read_start_pos, i = aligned_ref_end_pos; j <= end; j += SIMD_WIDTH, i -= SIMD_WIDTH) \
{ \ { \
__m256i h2_vec = _mm256_loadu_si256((__m256i *)(&next_max_arr[j])); \ __m256i h2_vec = _mm256_loadu_si256((__m256i *)(&hA2[j])); \
__m256i vcmp = _mm256_cmpeq_epi16(h2_vec, max_vec); \ __m256i vcmp = _mm256_cmpeq_epi16(h2_vec, max_vec); \
uint32_t mask = _mm256_movemask_epi8(vcmp); \ uint32_t mask = _mm256_movemask_epi8(vcmp); \
if (mask > 0) \ if (mask > 0) \
{ \ { \
int pos = SIMD_WIDTH - 1 - ((__builtin_clz(mask)) >> 1); \ int pos = SIMD_WIDTH - 1 - ((__builtin_clz(mask)) >> 1); \
mj = j - 1 + pos; \ mj = j - 1 + pos; \
mi = i - 1 - pos; \ mi = i - 1 - pos; \
for (; mj + 1 < qlen && mi + 1 < tlen; mj++, mi++) \ for (; mj + 1 < qlen && mi + 1 < tlen; mj++, mi++) \
{ \ { \
if (read_seq[mj + 2] == ref_seq[mi + 1 + SIMD_WIDTH]) \ if (seq[mj + 2] == ref[mi + 1 + SIMD_WIDTH]) \
{ \ { \
m += base_match_score; \ m += base_match_score; \
} \ } \
else \ else \
{ \ { \
break; \ break; \
} \ } \
} \ } \
} \ } \
} \ } \
last_max_vec = _mm256_set1_epi16(m); \ last_max_vec = _mm256_set1_epi16(m); \
} }
// 每轮迭代后，交换数组 // 每轮迭代后，交换数组
#define SWAP_DATA_POINTER \ #define SWAP_DATA_POINTER \
int16_t *tmp = last_max_arr; \ int16_t *tmp = hA0; \
last_max_arr = cur_max_arr; \ hA0 = hA1; \
cur_max_arr = next_max_arr; \ hA1 = hA2; \
next_max_arr = tmp; \ hA2 = tmp; \
tmp = cur_del_arr; \ tmp = eA1; \
cur_del_arr = next_del_arr; \ eA1 = eA2; \
next_del_arr = tmp; \ eA2 = tmp; \
tmp = cur_ins_arr; \ tmp = fA1; \
cur_ins_arr = next_ins_arr; \ fA1 = fA2; \
next_ins_arr = tmp; \ fA2 = tmp; \
tmp = cur_match_arr; \ tmp = mA1; \
cur_match_arr = next_match_arr; \ mA1 = mA2; \
next_match_arr = tmp; mA2 = tmp;
int ksw_extend_avx2_aligned(thread_mem_t *tmem, int ksw_extend_avx2_aligned(thread_mem_t *tmem,
int qlen, // query length 待匹配段碱基的query长度 int qlen, // query length 待匹配段碱基的query长度
@ -220,11 +220,11 @@ int ksw_extend_avx2_aligned(thread_mem_t *tmem,
int *_gscore, // query的端到端匹配得分 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 int *_max_off) // 取得最大得分时在query和reference上位置差的 最大值
{ {
int16_t *cur_match_arr, *next_match_arr, int16_t *mA1, *mA2,
*last_max_arr, *cur_max_arr, *next_max_arr, *hA0, *hA1, *hA2,
*cur_del_arr, *next_del_arr, *eA1, *eA2,
*cur_ins_arr, *next_ins_arr; // hA0保存上上个col的H，其他的保存上个H E F M *fA1, *fA2; // hA0保存上上个col的H，其他的保存上个H E F M
int16_t *read_seq, *ref_seq; int16_t *seq, *ref;
uint8_t *mem_addr; uint8_t *mem_addr;
int read_size = align_number(qlen * BASE_BYTES + MEM_ALIGN_BYTES); int read_size = align_number(qlen * BASE_BYTES + MEM_ALIGN_BYTES);
@ -235,10 +235,10 @@ int ksw_extend_avx2_aligned(thread_mem_t *tmem,
int score_mem_size = score_array_size * TMP_SCORE_ARRAY_NUM; int score_mem_size = score_array_size * TMP_SCORE_ARRAY_NUM;
int request_mem_size = read_size + ref_size + score_mem_size + MEM_ALIGN_BYTES * 3; // 左侧内存地址对齐 + 数据向左偏移一个元素 + 末尾SIMD补齐 int request_mem_size = read_size + ref_size + score_mem_size + MEM_ALIGN_BYTES * 3; // 左侧内存地址对齐 + 数据向左偏移一个元素 + 末尾SIMD补齐
int i, ref_start_pos, di, j, read_start_pos, read_end_pos, max, max_i, max_j, max_ins, max_del, max_ie, gscore, max_off; int i, ibeg, D, j, beg, end, max, max_i, max_j, max_ins, max_del, max_ie, gscore, max_off;
int span, beg1, end1; // 边界条件计算 int span, beg1, end1; // 边界条件计算
int aligned_read_start_pos, aligned_ref_end_pos; int aligned_read_start_pos, aligned_ref_end_pos;
int ref_end_pos; int iend;
SIMD_INIT; // 初始化simd用的数据 SIMD_INIT; // 初始化simd用的数据
@ -247,21 +247,21 @@ int ksw_extend_avx2_aligned(thread_mem_t *tmem,
// allocate memory // allocate memory
mem_addr = thread_mem_request(tmem, request_mem_size); mem_addr = thread_mem_request(tmem, request_mem_size);
mem_addr = (void *)align_mem((uint64_t)mem_addr); mem_addr = (void *)align_mem((uint64_t)mem_addr);
ref_seq = (int16_t *)&mem_addr[0]; ref = (int16_t *)&mem_addr[0];
read_seq = (int16_t *)(mem_addr + ref_size + SIMD_BYTES - BASE_BYTES); seq = (int16_t *)(mem_addr + ref_size + SIMD_BYTES - BASE_BYTES);
if (extend_left) if (extend_left)
{ {
for (i = 0; i < qlen; ++i) for (i = 0; i < qlen; ++i)
read_seq[i + 1] = query[qlen - 1 - i]; seq[i + 1] = query[qlen - 1 - i];
for (i = 0; i < tlen; ++i) for (i = 0; i < tlen; ++i)
ref_seq[i + SIMD_WIDTH] = target[tlen - 1 - i]; ref[i + SIMD_WIDTH - 1] = target[tlen - 1 - i];
} }
else else
{ {
for (i = 0; i < qlen; ++i) for (i = 0; i < qlen; ++i)
read_seq[i + 1] = query[i]; seq[i + 1] = query[i];
for (i = 0; i < tlen; ++i) for (i = 0; i < tlen; ++i)
ref_seq[i + SIMD_WIDTH] = target[i]; ref[i + SIMD_WIDTH - 1] = target[i];
} }
mem_addr += read_size + ref_size + (SIMD_BYTES - SCORE_BYTES); mem_addr += read_size + ref_size + (SIMD_BYTES - SCORE_BYTES);
@ -270,15 +270,15 @@ int ksw_extend_avx2_aligned(thread_mem_t *tmem,
_mm256_storeu_si256((__m256i *)&mem_addr[i], zero_vec); _mm256_storeu_si256((__m256i *)&mem_addr[i], zero_vec);
} }
last_max_arr = (int16_t *)&mem_addr[0]; hA0 = (int16_t *)&mem_addr[0];
cur_max_arr = &last_max_arr[score_element_num]; hA1 = &hA0[score_element_num];
next_max_arr = &cur_max_arr[score_element_num]; hA2 = &hA1[score_element_num];
cur_match_arr = &next_max_arr[score_element_num]; mA1 = &hA2[score_element_num];
next_match_arr = &cur_match_arr[score_element_num]; mA2 = &mA1[score_element_num];
cur_del_arr = &next_match_arr[score_element_num]; eA1 = &mA2[score_element_num];
next_del_arr = &cur_del_arr[score_element_num]; eA2 = &eA1[score_element_num];
cur_ins_arr = &next_del_arr[score_element_num]; fA1 = &eA2[score_element_num];
next_ins_arr = &cur_ins_arr[score_element_num]; fA2 = &fA1[score_element_num];
// adjust $window_size if it is too large // adjust $window_size if it is too large
// get the max score // get the max score
@ -297,10 +297,13 @@ int ksw_extend_avx2_aligned(thread_mem_t *tmem,
max_ie = -1, gscore = -1; max_ie = -1, gscore = -1;
; ;
max_off = 0; max_off = 0;
read_start_pos = 1; beg = 1;
read_end_pos = qlen; end = qlen;
// init init_score // init init_score
last_max_arr[0] = init_score; // 左上角 hA0[0] = init_score; // 左上角
fA1[1] = MAX(0, init_score - (o_ins + e_ins));
eA2[0] = init_score;
hA1[1] = fA1[1];
if (qlen == 0 || tlen == 0) if (qlen == 0 || tlen == 0)
back_diagnal_num = 0; // 防止意外情况 back_diagnal_num = 0; // 防止意外情况
@ -310,64 +313,51 @@ int ksw_extend_avx2_aligned(thread_mem_t *tmem,
gscore = 0; gscore = 0;
} }
for (di = 1; LIKELY(di < back_diagnal_num); ++di) for (D = 1; LIKELY(D < back_diagnal_num); ++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)
if (di > tlen) beg1 = 1;
{
span = MIN(back_diagnal_num - di, window_size); // 计算的窗口，或者说范围
beg1 = MAX(di - tlen + 1, ((di - window_size) / 2) + 1);
}
else else
{ beg1 = D - tlen + 1;
span = MIN(di - 1, window_size); if (D < qlen)
beg1 = MAX(1, ((di - window_size) / 2) + 1); end1 = D; // 闭区间
} else
end1 = MIN(qlen, beg1 + span); end1 = qlen;
beg1 = MAX(D - window_size, beg1);
end1 = MIN(D + window_size, end1);
// if (read_start_pos < beg1) beg = MAX(beg1, beg);
// read_start_pos = beg1; end = MIN(end1, end);
// if (read_end_pos > end1) if (beg > end)
// read_end_pos = end1; break;
// if (read_start_pos > read_end_pos)
// 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
read_start_pos = 1; // beg = beg1;
read_end_pos = qlen; // end = end1;
ref_end_pos = di - (read_start_pos - 1); // ref开始计算的位置，倒序
span = read_end_pos - read_start_pos; iend = D - beg; // ref开始计算的位置，倒序
ref_start_pos = ref_end_pos - span - 1; // 0开始的ref索引位置 span = end - beg;
ibeg = iend - span; // 0开始的ref索引位置
// 每一轮需要记录的数据 // 每一轮需要记录的数据
int m = 0, mj = -1, mi = -1; int m = 0, mj = -1, mi = -1;
max_vec = zero_vec; max_vec = zero_vec;
// 处理左边界
// 要处理边界 if (beg == 1)
// 左边界 处理f (insert)
if (ref_start_pos == 0)
{ {
cur_max_arr[read_end_pos] = MAX(0, init_score - (o_ins + e_ins * read_end_pos)); hA0[0] = eA2[0];
} mA1[0] = 0;
// 上边界 delete eA1[0] = MAX(0, init_score - (o_del + e_del * (iend + 1)));
if (read_start_pos == 1)
{
cur_max_arr[0] = MAX(0, init_score - (o_del + e_del * ref_end_pos));
}
else
{
cur_max_arr[read_start_pos - 1] = 0;
cur_del_arr[read_start_pos - 1] = 0;
} }
// aligned_read_start_pos = (read_start_pos >> ALIGN_SHIFT_BITS << ALIGN_SHIFT_BITS) + 1; // aligned_read_start_pos = (beg >> ALIGN_SHIFT_BITS << ALIGN_SHIFT_BITS) + 1;
// aligned_ref_end_pos = ref_end_pos + (read_start_pos - aligned_read_start_pos); // aligned_ref_end_pos = iend + (beg - aligned_read_start_pos);
aligned_read_start_pos = read_start_pos; aligned_read_start_pos = beg;
aligned_ref_end_pos = ref_end_pos; aligned_ref_end_pos = iend;
// fprintf(stderr, "%d\t%d\n", read_start_pos, aligned_read_start_pos); // fprintf(stderr, "%d\t%d\n", beg, aligned_read_start_pos);
for (j = aligned_read_start_pos, i = aligned_ref_end_pos; j <= read_end_pos + 1 - SIMD_WIDTH; j += SIMD_WIDTH, i -= SIMD_WIDTH) for (j = aligned_read_start_pos, i = aligned_ref_end_pos; j <= end + 1 - SIMD_WIDTH; j += SIMD_WIDTH, i -= SIMD_WIDTH)
{ {
// 取数据 // 取数据
SIMD_LOAD; SIMD_LOAD;
@ -379,7 +369,7 @@ int ksw_extend_avx2_aligned(thread_mem_t *tmem,
SIMD_STORE; SIMD_STORE;
} }
// 剩下的计算单元 // 剩下的计算单元
if (j <= read_end_pos) if (j <= end)
{ {
// 取数据 // 取数据
SIMD_LOAD; SIMD_LOAD;
@ -392,16 +382,22 @@ int ksw_extend_avx2_aligned(thread_mem_t *tmem,
// 存储结果 // 存储结果
SIMD_STORE; SIMD_STORE;
} }
// 处理上边界
if (ibeg == 0)
{
fA2[end + 1] = MAX(0, init_score - (o_ins + e_ins * (end + 1)));
hA2[end + 1] = fA2[end + 1];
mA2[end + 1] = 0;
}
SIMD_FIND_MAX; SIMD_FIND_MAX;
// 注意最后跳出循环j的值 // 注意最后跳出循环j的值
j = read_end_pos + 1; j = end + 1;
if (j == qlen + 1) // 遍历到了query最后一个碱基，此时next_max_arr[qlen]为全局匹配的最大分值 if (j == qlen + 1) // 遍历到了query最后一个碱基，此时next_max_arr[qlen]为全局匹配的最大分值
{ {
max_ie = gscore > next_max_arr[qlen] ? max_ie : ref_start_pos; max_ie = gscore > hA2[qlen] ? max_ie : ibeg;
gscore = gscore > next_max_arr[qlen] ? gscore : next_max_arr[qlen]; gscore = gscore > hA2[qlen] ? gscore : hA2[qlen];
} }
if (m > max) if (m > max)
{ {
@ -410,26 +406,25 @@ int ksw_extend_avx2_aligned(thread_mem_t *tmem,
} }
// 调整计算的边界 // 调整计算的边界
// for (j = read_start_pos; LIKELY(j <= read_end_pos); ++j) for (j = beg; LIKELY(j <= end); ++j)
//{ {
// int has_val = cur_max_arr[j - 1] | next_max_arr[j]; int has_val = hA1[j - 1] | hA2[j];
// if (has_val) if (has_val)
// { {
// break; break;
// } }
//} }
// read_start_pos = j; beg = j;
//
// next_max_arr[read_end_pos + 1] = 0; for (j = end + 1; LIKELY(j >= beg); --j)
// for (j = read_end_pos + 1; LIKELY(j >= read_start_pos); --j) {
//{ int has_val = hA1[j - 1] | hA2[j];
// int has_val = cur_max_arr[j - 1] | next_max_arr[j]; if (has_val)
// if (has_val) {
// { break;
// break; }
// } }
//} end = j + 1 <= qlen ? j + 1 : qlen;
// read_end_pos = j + 1 <= qlen ? j + 1 : qlen;
// swap m, h, e, f // swap m, h, e, f
SWAP_DATA_POINTER; SWAP_DATA_POINTER;
} }

114
ksw_ext_avx2_heuristics.c
View File

@ -222,7 +222,7 @@ int ksw_extend_avx2_heuristics(thread_mem_t *tmem,
uint8_t *mem; uint8_t *mem;
int16_t *qtmem, *vmem; int16_t *qtmem, *vmem;
int seq_size = qlen + SIMD_WIDTH, ref_size = tlen + SIMD_WIDTH; 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 i, ibeg, iend, 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 D从1开始遍历 int Dloop = tlen + qlen; // 循环跳出条件 D从1开始遍历
int span, beg1, end1; // 边界条件计算 int span, beg1, end1; // 边界条件计算
int col_size = qlen + 2 + SIMD_WIDTH; int col_size = qlen + 2 + SIMD_WIDTH;
@ -243,14 +243,14 @@ int ksw_extend_avx2_heuristics(thread_mem_t *tmem,
for (i = 0; i < qlen; ++i) for (i = 0; i < qlen; ++i)
seq[i] = query[qlen - 1 - i]; seq[i] = query[qlen - 1 - i];
for (i = 0; i < tlen; ++i) for (i = 0; i < tlen; ++i)
ref[i + SIMD_WIDTH] = target[tlen - 1 - i]; ref[i + SIMD_WIDTH - 1] = target[tlen - 1 - i];
} }
else else
{ {
for (i = 0; i < qlen; ++i) for (i = 0; i < qlen; ++i)
seq[i] = query[i]; seq[i] = query[i];
for (i = 0; i < tlen; ++i) for (i = 0; i < tlen; ++i)
ref[i + SIMD_WIDTH] = target[i]; ref[i + SIMD_WIDTH - 1] = target[i];
} }
vmem = &ref[ref_size]; vmem = &ref[ref_size];
@ -288,12 +288,15 @@ int ksw_extend_avx2_heuristics(thread_mem_t *tmem,
// DP loop // DP loop
max = init_score, max_i = max_j = -1; max = init_score, max_i = max_j = -1;
max_ie = -1, gscore = -1; max_ie = -1, gscore = -1;
;
max_off = 0; max_off = 0;
beg = 1; beg = 1;
end = qlen; end = qlen;
// init init_score // init init_score
last_max_arr[0] = init_score; // 左上角 last_max_arr[0] = init_score; // 左上角
cur_ins_arr[1] = MAX(0, init_score - (o_ins + e_ins));
next_del_arr[0] = init_score;
cur_max_arr[1] = cur_ins_arr[1];
if (qlen == 0 || tlen == 0) if (qlen == 0 || tlen == 0)
Dloop = 0; // 防止意外情况 Dloop = 0; // 防止意外情况
@ -303,56 +306,41 @@ int ksw_extend_avx2_heuristics(thread_mem_t *tmem,
gscore = 0; gscore = 0;
} }
int iend;
for (D = 1; LIKELY(D < Dloop); ++D) 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)
if (D > tlen) beg1 = 1;
{
span = MIN(Dloop - D, window_size); // 计算的窗口，或者说范围
beg1 = MAX(D - tlen + 1, ((D - window_size) / 2) + 1);
}
else else
{ beg1 = D - tlen + 1;
span = MIN(D - 1, window_size); if (D < qlen)
beg1 = MAX(1, ((D - window_size) / 2) + 1); end1 = D; // 闭区间
} else
end1 = MIN(qlen, beg1 + span); end1 = qlen;
beg1 = MAX(D - window_size, beg1);
end1 = MIN(D + window_size, end1);
beg = 1; beg = MAX(beg1, beg);
end = qlen; end = MIN(end1, end);
if (beg > end)
break;
// if (beg < beg1) // beg = beg1;
// beg = beg1; // end = end1;
// 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 iend = D - beg; // ref开始计算的位置，倒序
span = end - beg; span = end - beg;
iStart = iend - span - 1; // 0开始的ref索引位置 ibeg = iend - span; // 0开始的ref索引位置
// 每一轮需要记录的数据 // 每一轮需要记录的数据
int m = 0, mj = -1, mi = -1; int m = 0, mj = -1, mi = -1;
max_vec = zero_vec; max_vec = zero_vec;
// 要处理边界 // 处理左边界
// 左边界 处理f (insert)
if (iStart == 0)
{
cur_max_arr[end] = MAX(0, init_score - (o_ins + e_ins * end));
}
// 上边界
if (beg == 1) if (beg == 1)
{ {
cur_max_arr[0] = MAX(0, init_score - (o_del + e_del * iend)); last_max_arr[0] = next_del_arr[0];
} cur_match_arr[0] = 0;
else cur_del_arr[0] = MAX(0, init_score - (o_del + e_del * (iend + 1)));
{
cur_max_arr[beg - 1] = 0;
cur_del_arr[beg - 1] = 0;
} }
for (j = beg, i = iend; j <= end + 1 - SIMD_WIDTH; j += SIMD_WIDTH, i -= SIMD_WIDTH) for (j = beg, i = iend; j <= end + 1 - SIMD_WIDTH; j += SIMD_WIDTH, i -= SIMD_WIDTH)
@ -380,6 +368,13 @@ int ksw_extend_avx2_heuristics(thread_mem_t *tmem,
// 存储结果 // 存储结果
SIMD_STORE; SIMD_STORE;
} }
// 处理上边界
if (ibeg == 0)
{
next_ins_arr[end + 1] = MAX(0, init_score - (o_ins + e_ins * (end + 1)));
next_max_arr[end + 1] = next_ins_arr[end + 1];
next_match_arr[end + 1] = 0;
}
SIMD_FIND_MAX; SIMD_FIND_MAX;
@ -388,7 +383,7 @@ int ksw_extend_avx2_heuristics(thread_mem_t *tmem,
if (j == qlen + 1) if (j == qlen + 1)
{ {
max_ie = gscore > next_max_arr[qlen] ? max_ie : iStart; max_ie = gscore > next_max_arr[qlen] ? max_ie : ibeg;
gscore = gscore > next_max_arr[qlen] ? gscore : next_max_arr[qlen]; gscore = gscore > next_max_arr[qlen] ? gscore : next_max_arr[qlen];
} }
if (m > max) if (m > max)
@ -397,25 +392,24 @@ int ksw_extend_avx2_heuristics(thread_mem_t *tmem,
max_off = max_off > abs(mj - mi) ? max_off : abs(mj - mi); max_off = max_off > abs(mj - mi) ? max_off : abs(mj - mi);
} }
// for (j = beg; LIKELY(j <= end); ++j) for (j = beg; LIKELY(j <= end); ++j)
//{ {
// int has_val = cur_max_arr[j - 1] | next_max_arr[j]; int has_val = cur_max_arr[j - 1] | next_max_arr[j];
// if (has_val) if (has_val)
// { {
// break; break;
// } }
// } }
// beg = j; beg = j;
// next_max_arr[end + 1] = 0; for (j = end + 1; LIKELY(j >= beg); --j)
// for (j = end + 1; LIKELY(j >= beg); --j) {
//{ int has_val = cur_max_arr[j - 1] | next_max_arr[j];
// int has_val = cur_max_arr[j - 1] | next_max_arr[j]; if (has_val)
// if (has_val) {
// { break;
// break; }
// } }
// } end = j + 1 <= qlen ? j + 1 : qlen;
// end = j + 1 <= qlen ? j + 1 : qlen;
// swap m, h, e, f // swap m, h, e, f
SWAP_DATA_POINTER; SWAP_DATA_POINTER;

46
ksw_ext_avx2_u8.c
View File

@ -243,7 +243,7 @@ int ksw_extend_avx2_u8(thread_mem_t *tmem,
for (i = 0; i < qlen; ++i) for (i = 0; i < qlen; ++i)
seq[i] = query[qlen - 1 - i]; seq[i] = query[qlen - 1 - i];
for (i = 0; i < tlen; ++i) for (i = 0; i < tlen; ++i)
ref[i + SIMD_WIDTH] = target[tlen - 1 - i]; ref[i + SIMD_WIDTH - 1] = target[tlen - 1 - i];
} }
else else
{ {
@ -317,16 +317,16 @@ int ksw_extend_avx2_u8(thread_mem_t *tmem,
end1 = D; // 闭区间 end1 = D; // 闭区间
else else
end1 = qlen; end1 = qlen;
// beg1 = MAX(D - window_size, beg1); beg1 = MAX(D - window_size, beg1);
// end1 = MIN(D + window_size, end1); end1 = MIN(D + window_size, end1);
beg = MAX(beg1, beg); beg = MAX(beg1, beg);
end = MIN(end1, end); end = MIN(end1, end);
// if (beg > end) if (beg > end)
// break; break;
beg = beg1; // beg = beg1;
end = end1; // end = end1;
iend = D - beg; // ref开始计算的位置，倒序 iend = D - beg; // ref开始计算的位置，倒序
span = end - beg; span = end - beg;
@ -394,22 +394,22 @@ int ksw_extend_avx2_u8(thread_mem_t *tmem,
} }
// 调整计算的边界 // 调整计算的边界
// for (j = beg; LIKELY(j <= end); ++j) for (j = beg; LIKELY(j <= end); ++j)
//{ {
// int has_val = hA1[j - 1] | hA2[j]; int has_val = hA1[j - 1] | hA2[j];
// if (has_val) if (has_val)
// break; break;
//} }
// beg = j; beg = j;
// for (j = end + 1; LIKELY(j >= beg); --j) for (j = end + 1; LIKELY(j >= beg); --j)
//{ {
// int has_val = hA1[j - 1] | hA2[j]; int has_val = hA1[j - 1] | hA2[j];
// if (has_val) if (has_val)
// break; break;
// else else
// hA0[j - 1] = 0; hA0[j - 1] = 0;
//} }
// end = j + 1 <= qlen ? j + 1 : qlen; end = j + 1 <= qlen ? j + 1 : qlen;
// swap m, h, e, f // swap m, h, e, f
SWAP_DATA_POINTER; SWAP_DATA_POINTER;

106
ksw_ext_avx2_u8_aligned.c
View File

@ -250,7 +250,7 @@ int ksw_extend_avx2_u8_aligned(thread_mem_t *tmem,
{ {
uint8_t *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 uint8_t *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
uint8_t *read_seq, *ref_seq; uint8_t *read_seq, *ref_seq;
int i, iStart, D, j, beg, end, max, max_i, max_j, max_ins, max_del, max_ie, gscore, max_off; int i, ibeg, iend, D, j, beg, end, max, max_i, max_j, max_ins, max_del, max_ie, gscore, max_off;
int span, beg1, end1; // 边界条件计算 int span, beg1, end1; // 边界条件计算
uint8_t *mem_addr; uint8_t *mem_addr;
@ -276,14 +276,14 @@ int ksw_extend_avx2_u8_aligned(thread_mem_t *tmem,
for (i = 0; i < qlen; ++i) for (i = 0; i < qlen; ++i)
read_seq[i + 1] = query[qlen - 1 - i]; read_seq[i + 1] = query[qlen - 1 - i];
for (i = 0; i < tlen; ++i) for (i = 0; i < tlen; ++i)
ref_seq[i + SIMD_WIDTH] = target[tlen - 1 - i]; ref_seq[i + SIMD_WIDTH - 1] = target[tlen - 1 - i];
} }
else else
{ {
for (i = 0; i < qlen; ++i) for (i = 0; i < qlen; ++i)
read_seq[i + 1] = query[i]; read_seq[i + 1] = query[i];
for (i = 0; i < tlen; ++i) for (i = 0; i < tlen; ++i)
ref_seq[i + SIMD_WIDTH] = target[i]; ref_seq[i + SIMD_WIDTH - 1] = target[i];
} }
mem_addr += read_size + ref_size; mem_addr += read_size + ref_size;
@ -318,12 +318,15 @@ int ksw_extend_avx2_u8_aligned(thread_mem_t *tmem,
// DP loop // DP loop
max = init_score, max_i = max_j = -1; max = init_score, max_i = max_j = -1;
max_ie = -1, gscore = -1; max_ie = -1, gscore = -1;
;
max_off = 0; max_off = 0;
beg = 1; beg = 1;
end = qlen; end = qlen;
// init init_score // init init_score
hA0[0] = init_score; // 左上角 hA0[0] = init_score; // 左上角
fA1[1] = MAX(0, init_score - (o_ins + e_ins));
eA2[0] = init_score;
hA1[1] = fA1[1];
if (qlen == 0 || tlen == 0) if (qlen == 0 || tlen == 0)
back_diagnal_num = 0; // 防止意外情况 back_diagnal_num = 0; // 防止意外情况
@ -333,55 +336,42 @@ int ksw_extend_avx2_u8_aligned(thread_mem_t *tmem,
gscore = 0; gscore = 0;
} }
int iend;
for (D = 1; LIKELY(D < back_diagnal_num); ++D) for (D = 1; LIKELY(D < back_diagnal_num); ++D)
{ {
// 边界条件一定要注意！ tlen 大于，等于，小于 qlen时的情况 // 边界条件一定要注意！ tlen 大于，等于，小于 qlen时的情况
if (D > tlen) if (D < tlen)
{ beg1 = 1;
span = MIN(back_diagnal_num - D, window_size);
beg1 = MAX(D - tlen + 1, ((D - window_size) / 2) + 1);
}
else else
{ beg1 = D - tlen + 1;
span = MIN(D - 1, window_size); if (D < qlen)
beg1 = MAX(1, ((D - window_size) / 2) + 1); end1 = D; // 闭区间
} else
end1 = MIN(qlen, beg1 + span); end1 = qlen;
beg1 = MAX(D - window_size, beg1);
end1 = MIN(D + window_size, end1);
// if (beg < beg1) beg = MAX(beg1, beg);
// beg = beg1; end = MIN(end1, end);
// if (end > end1) if (beg > end)
// end = end1; break;
// 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
beg = 1; // beg = beg1;
end = qlen; // end = end1;
iend = D - (beg - 1); // ref开始计算的位置，倒序
iend = D - beg; // ref开始计算的位置，倒序
span = end - beg; span = end - beg;
iStart = iend - span - 1; // 0开始的ref索引位置 ibeg = iend - span; // 0开始的ref索引位置
// 每一轮需要记录的数据 // 每一轮需要记录的数据
int m = 0, mj = -1, mi = -1; int m = 0, mj = -1, mi = -1;
max_vec = zero_vec; max_vec = zero_vec;
// 要处理边界 // 处理左边界
// 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) if (beg == 1)
{ {
hA1[0] = MAX(0, init_score - (o_del + e_del * iend)); hA0[0] = eA2[0];
} mA1[0] = 0;
else eA1[0] = MAX(0, init_score - (o_del + e_del * (iend + 1)));
{
hA1[beg - 1] = 0;
eA1[beg - 1] = 0;
} }
for (j = beg, i = iend; j <= end + 1 - SIMD_WIDTH; j += SIMD_WIDTH, i -= SIMD_WIDTH) for (j = beg, i = iend; j <= end + 1 - SIMD_WIDTH; j += SIMD_WIDTH, i -= SIMD_WIDTH)
@ -410,6 +400,13 @@ int ksw_extend_avx2_u8_aligned(thread_mem_t *tmem,
SIMD_STORE; SIMD_STORE;
} }
// 处理上边界
if (ibeg == 0)
{
fA2[end + 1] = MAX(0, init_score - (o_ins + e_ins * (end + 1)));
hA2[end + 1] = fA2[end + 1];
mA2[end + 1] = 0;
}
SIMD_FIND_MAX; SIMD_FIND_MAX;
// 注意最后跳出循环j的值 // 注意最后跳出循环j的值
@ -417,7 +414,7 @@ int ksw_extend_avx2_u8_aligned(thread_mem_t *tmem,
if (j == qlen + 1) if (j == qlen + 1)
{ {
max_ie = gscore > hA2[qlen] ? max_ie : iStart; max_ie = gscore > hA2[qlen] ? max_ie : ibeg;
gscore = gscore > hA2[qlen] ? gscore : hA2[qlen]; gscore = gscore > hA2[qlen] ? gscore : hA2[qlen];
} }
@ -428,21 +425,20 @@ int ksw_extend_avx2_u8_aligned(thread_mem_t *tmem,
} }
// 调整计算的边界 // 调整计算的边界
// for (j = beg; LIKELY(j <= end); ++j) for (j = beg; LIKELY(j <= end); ++j)
//{ {
// int has_val = hA1[j - 1] | hA2[j]; int has_val = hA1[j - 1] | hA2[j];
// if (has_val) if (has_val)
// break; break;
//} }
// beg = j; beg = j;
// hA2[end + 1] = 0; for (j = end + 1; LIKELY(j >= beg); --j)
// for (j = end + 1; LIKELY(j >= beg); --j) {
//{ int has_val = hA1[j - 1] | hA2[j];
// int has_val = hA1[j - 1] | hA2[j]; if (has_val)
// if (has_val) break;
// break; }
//} end = j + 1 <= qlen ? j + 1 : qlen;
// end = j + 1 <= qlen ? j + 1 : qlen;
// swap m, h, e, f // swap m, h, e, f
SWAP_DATA_POINTER; SWAP_DATA_POINTER;

113
ksw_ext_avx2_u8_heuristics.c
View File

@ -243,7 +243,7 @@ int ksw_extend_avx2_u8_heuristics(thread_mem_t *tmem,
uint8_t *seq, *ref; uint8_t *seq, *ref;
uint8_t *mem, *qtmem, *vmem; uint8_t *mem, *qtmem, *vmem;
int seq_size = qlen + SIMD_WIDTH, ref_size = tlen + SIMD_WIDTH; 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 i, ibeg, iend, 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 Dloop = tlen + qlen; // 循环跳出条件
int span, beg1, end1; // 边界条件计算 int span, beg1, end1; // 边界条件计算
int col_size = qlen + 2 + SIMD_WIDTH; int col_size = qlen + 2 + SIMD_WIDTH;
@ -265,14 +265,14 @@ int ksw_extend_avx2_u8_heuristics(thread_mem_t *tmem,
for (i = 0; i < qlen; ++i) for (i = 0; i < qlen; ++i)
seq[i] = query[qlen - 1 - i]; seq[i] = query[qlen - 1 - i];
for (i = 0; i < tlen; ++i) for (i = 0; i < tlen; ++i)
ref[i + SIMD_WIDTH] = target[tlen - 1 - i]; ref[i + SIMD_WIDTH - 1] = target[tlen - 1 - i];
} }
else else
{ {
for (i = 0; i < qlen; ++i) for (i = 0; i < qlen; ++i)
seq[i] = query[i]; seq[i] = query[i];
for (i = 0; i < tlen; ++i) for (i = 0; i < tlen; ++i)
ref[i + SIMD_WIDTH] = target[i]; ref[i + SIMD_WIDTH - 1] = target[i];
} }
vmem = &ref[ref_size]; vmem = &ref[ref_size];
@ -311,12 +311,15 @@ int ksw_extend_avx2_u8_heuristics(thread_mem_t *tmem,
// DP loop // DP loop
max = init_score, max_i = max_j = -1; max = init_score, max_i = max_j = -1;
max_ie = -1, gscore = -1; max_ie = -1, gscore = -1;
;
max_off = 0; max_off = 0;
beg = 1; beg = 1;
end = qlen; end = qlen;
// init init_score // init init_score
hA0[0] = init_score; // 左上角 hA0[0] = init_score; // 左上角
fA1[1] = MAX(0, init_score - (o_ins + e_ins));
eA2[0] = init_score;
hA1[1] = fA1[1];
if (qlen == 0 || tlen == 0) if (qlen == 0 || tlen == 0)
Dloop = 0; // 防止意外情况 Dloop = 0; // 防止意外情况
@ -325,62 +328,42 @@ int ksw_extend_avx2_u8_heuristics(thread_mem_t *tmem,
max_ie = 0; max_ie = 0;
gscore = 0; gscore = 0;
} }
int iend;
#ifdef SHOW_PERF
// time_bsw_init += get_mseconds() - start_time;
#endif
for (D = 1; LIKELY(D < Dloop); ++D) for (D = 1; LIKELY(D < Dloop); ++D)
{ {
#ifdef SHOW_PERF
// start_time = get_mseconds();
#endif
// 边界条件一定要注意！ tlen 大于，等于，小于 qlen时的情况 // 边界条件一定要注意！ tlen 大于，等于，小于 qlen时的情况
if (D > tlen) if (D < tlen)
{ beg1 = 1;
span = MIN(Dloop - D, window_size);
beg1 = MAX(D - tlen + 1, ((D - window_size) / 2) + 1);
}
else else
{ beg1 = D - tlen + 1;
span = MIN(D - 1, window_size); if (D < qlen)
beg1 = MAX(1, ((D - window_size) / 2) + 1); end1 = D; // 闭区间
} else
end1 = MIN(qlen, beg1 + span); end1 = qlen;
beg1 = MAX(D - window_size, beg1);
end1 = MIN(D + window_size, end1);
// if (beg < beg1) beg = MAX(beg1, beg);
// beg = beg1; end = MIN(end1, end);
// if (end > end1) if (beg > end)
// end = end1; break;
// 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
beg = 1; // beg = beg1;
end = qlen; // end = end1;
iend = D - (beg - 1); // ref开始计算的位置，倒序
iend = D - beg; // ref开始计算的位置，倒序
span = end - beg; span = end - beg;
iStart = iend - span - 1; // 0开始的ref索引位置 ibeg = iend - span; // 0开始的ref索引位置
// 每一轮需要记录的数据 // 每一轮需要记录的数据
int m = 0, mj = -1, mi = -1; int m = 0, mj = -1, mi = -1;
max_vec = zero_vec; max_vec = zero_vec;
// 要处理边界 // 处理左边界
// 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) if (beg == 1)
{ {
hA1[0] = MAX(0, init_score - (o_del + e_del * iend)); hA0[0] = eA2[0];
} mA1[0] = 0;
else eA1[0] = MAX(0, init_score - (o_del + e_del * (iend + 1)));
{
hA1[beg - 1] = 0;
eA1[beg - 1] = 0;
} }
for (j = beg, i = iend; j <= end + 1 - SIMD_WIDTH; j += SIMD_WIDTH, i -= SIMD_WIDTH) for (j = beg, i = iend; j <= end + 1 - SIMD_WIDTH; j += SIMD_WIDTH, i -= SIMD_WIDTH)
@ -408,6 +391,13 @@ int ksw_extend_avx2_u8_heuristics(thread_mem_t *tmem,
// 存储结果 // 存储结果
SIMD_STORE; SIMD_STORE;
} }
// 处理上边界
if (ibeg == 0)
{
fA2[end + 1] = MAX(0, init_score - (o_ins + e_ins * (end + 1)));
hA2[end + 1] = fA2[end + 1];
mA2[end + 1] = 0;
}
SIMD_FIND_MAX; SIMD_FIND_MAX;
@ -416,7 +406,7 @@ int ksw_extend_avx2_u8_heuristics(thread_mem_t *tmem,
if (j == qlen + 1) if (j == qlen + 1)
{ {
max_ie = gscore > hA2[qlen] ? max_ie : iStart; max_ie = gscore > hA2[qlen] ? max_ie : ibeg;
gscore = gscore > hA2[qlen] ? gscore : hA2[qlen]; gscore = gscore > hA2[qlen] ? gscore : hA2[qlen];
} }
@ -427,21 +417,20 @@ int ksw_extend_avx2_u8_heuristics(thread_mem_t *tmem,
} }
// 调整计算的边界 // 调整计算的边界
// for (j = beg; LIKELY(j <= end); ++j) for (j = beg; LIKELY(j <= end); ++j)
//{ {
// int has_val = hA1[j - 1] | hA2[j]; int has_val = hA1[j - 1] | hA2[j];
// if (has_val) if (has_val)
// break; break;
//} }
// beg = j; beg = j;
// hA2[end + 1] = 0; for (j = end + 1; LIKELY(j >= beg); --j)
// for (j = end + 1; LIKELY(j >= beg); --j) {
//{ int has_val = hA1[j - 1] | hA2[j];
// int has_val = hA1[j - 1] | hA2[j]; if (has_val)
// if (has_val) break;
// break; }
//} end = j + 1 <= qlen ? j + 1 : qlen;
// end = j + 1 <= qlen ? j + 1 : qlen;
// swap m, h, e, f // swap m, h, e, f
SWAP_DATA_POINTER; SWAP_DATA_POINTER;

54
ksw_ext_normal.c
View File

@ -25,11 +25,11 @@ int ksw_extend_normal(thread_mem_t *tmem, int qlen, const uint8_t *query, int tl
int8_t *qp; // query profile int8_t *qp; // query profile
int i, j, k, oe_del = o_del + e_del, oe_ins = o_ins + e_ins, beg, end, max, max_i, max_j, max_ins, max_del, max_ie, gscore, max_off; int i, j, k, oe_del = o_del + e_del, oe_ins = o_ins + e_ins, beg, end, max, max_i, max_j, max_ins, max_del, max_ie, gscore, max_off;
assert(h0 > 0); assert(h0 > 0);
// qp = malloc(qlen * m); qp = malloc(qlen * m);
// eh = calloc(qlen + 1, 8); eh = calloc(qlen + 1, 8);
qp = thread_mem_request(tmem, qlen * m); // qp = thread_mem_request(tmem, qlen * m);
eh = thread_mem_request_and_clean(tmem, (qlen + 1) * 8); // eh = thread_mem_request_and_clean(tmem, (qlen + 1) * 9);
// generate the query profile // generate the query profile
for (k = i = 0; k < m; ++k) for (k = i = 0; k < m; ++k)
{ {
const int8_t *p = &mat[k * m]; const int8_t *p = &mat[k * m];
@ -77,14 +77,14 @@ int ksw_extend_normal(thread_mem_t *tmem, int qlen, const uint8_t *query, int tl
// 对于target第i个字符，query中每个字符的分值，只有匹配和不匹配 // 对于target第i个字符，query中每个字符的分值，只有匹配和不匹配
int8_t *q = &qp[target[i] * qlen]; int8_t *q = &qp[target[i] * qlen];
// apply the band and the constraint (if provided) // apply the band and the constraint (if provided)
// if (beg < i - w) // 检查开始点是否可以缩小一些 if (beg < i - w) // 检查开始点是否可以缩小一些
// beg = i - w; beg = i - w;
// if (end > i + w + 1) // 检查终点是否可以缩小，使得整体的遍历范围缩小 if (end > i + w + 1) // 检查终点是否可以缩小，使得整体的遍历范围缩小
// end = i + w + 1; end = i + w + 1;
// if (end > qlen) // 终点不超过query长度 if (end > qlen) // 终点不超过query长度
// end = qlen; end = qlen;
beg = 0; // beg = 0;
end = qlen; // end = qlen;
// compute the first column // compute the first column
if (beg == 0) if (beg == 0)
{ {
@ -142,8 +142,8 @@ int ksw_extend_normal(thread_mem_t *tmem, int qlen, const uint8_t *query, int tl
max_ie = gscore > h1 ? max_ie : i; // max_ie表示取得全局最大分值时，target字符串的位置 max_ie = gscore > h1 ? max_ie : i; // max_ie表示取得全局最大分值时，target字符串的位置
gscore = gscore > h1 ? gscore : h1; gscore = gscore > h1 ? gscore : h1;
} }
// if (m == 0) // 遍历完query之后，当前轮次的最大分值为0，则跳出循环 if (m == 0) // 遍历完query之后，当前轮次的最大分值为0，则跳出循环
// break; break;
if (m > max) // 当前轮最大分值大于之前的最大分值 if (m > max) // 当前轮最大分值大于之前的最大分值
{ {
max = m, max_i = i, max_j = mj; // 更新取得最大值的target和query的位置 max = m, max_i = i, max_j = mj; // 更新取得最大值的target和query的位置
@ -163,19 +163,19 @@ int ksw_extend_normal(thread_mem_t *tmem, int qlen, const uint8_t *query, int tl
} }
} }
// update beg and end for the next round // update beg and end for the next round
// for (j = beg; LIKELY(j < end) && eh[j].h == 0 && eh[j].e == 0; ++j) for (j = beg; LIKELY(j < end) && eh[j].h == 0 && eh[j].e == 0; ++j)
// ; ;
// beg = j; beg = j;
// for (j = end; LIKELY(j >= beg) && eh[j].h == 0 && eh[j].e == 0; --j) for (j = end; LIKELY(j >= beg) && eh[j].h == 0 && eh[j].e == 0; --j)
// ; ;
// end = j + 2 < qlen ? j + 2 : qlen; // 剪枝没考虑f，即insert end = j + 2 < qlen ? j + 2 : qlen; // 剪枝没考虑f，即insert
// beg = 0, end = qlen; // uncomment this line for debugging // beg = 0, end = qlen; // uncomment this line for debugging
// fprintf(stderr, "\n"); // fprintf(stderr, "\n");
// fprintf(stderr, "%d\n", end); // fprintf(stderr, "%d\n", end);
} }
// free(eh); free(eh);
// free(qp); free(qp);
thread_mem_release(tmem, qlen * m + (qlen + 1) * 8); // thread_mem_release(tmem, qlen * m + (qlen + 1) * 9);
if (_qle) if (_qle)
*_qle = max_j + 1; *_qle = max_j + 1;
if (_tle) if (_tle)

20
main.c
View File

@ -242,17 +242,17 @@ int main(int argc, char *argv[])
PERFORMANCE_TEST_NORMAL(0, ksw_extend_normal); PERFORMANCE_TEST_NORMAL(0, ksw_extend_normal);
// avx2 // avx2
PERFORMANCE_TEST_AVX2(1, ksw_extend_avx2); // PERFORMANCE_TEST_AVX2(1, ksw_extend_avx2);
// avx2 heuristics // avx2 heuristics
// PERFORMANCE_TEST_AVX2(2, ksw_extend_avx2_heuristics); // PERFORMANCE_TEST_AVX2(2, ksw_extend_avx2_heuristics);
// // avx2 mem aligned // avx2 mem aligned
// PERFORMANCE_TEST_AVX2(3, ksw_extend_avx2_aligned); // PERFORMANCE_TEST_AVX2(3, ksw_extend_avx2_aligned);
//
// // avx2 u8 // avx2 u8
PERFORMANCE_TEST_AVX2(4, ksw_extend_avx2_u8); // PERFORMANCE_TEST_AVX2(4, ksw_extend_avx2_u8);
// // avx2 u8 heuristics // avx2 u8 heuristics
// PERFORMANCE_TEST_AVX2(5, ksw_extend_avx2_u8_heuristics); // PERFORMANCE_TEST_AVX2(5, ksw_extend_avx2_u8_heuristics);
// // avx2 u8 mem aligned // avx2 u8 mem aligned
// PERFORMANCE_TEST_AVX2(6, ksw_extend_avx2_u8_aligned); // PERFORMANCE_TEST_AVX2(6, ksw_extend_avx2_u8_aligned);
} }

1
thread_mem.c
View File

@ -38,6 +38,7 @@ void thread_mem_init_alloc(thread_mem_t *tmem, size_t byte_cnt)
// 请求内存 // 请求内存
void *thread_mem_request(thread_mem_t *tmem, size_t byte_cnt) void *thread_mem_request(thread_mem_t *tmem, size_t byte_cnt)
{ {
// fprintf(stderr, "capacity:%ld, occupied: %ld, byte_cnt: %ld\n", tmem->capacity, tmem->occupied, byte_cnt);
void *ret_mem = 0; void *ret_mem = 0;
if (tmem == 0) if (tmem == 0)
{ {