#include #include #include #include #include #include #include "sys/time.h" #include "thread_mem.h" #include "ksw_ext.h" #include "utils.h" #define SW_NORMAL 0 #define SW_AVX2 1 #define SW_CUDA 2 #define SW_ALL 3 #define BLOCK_BUF_SIZE 1048576 #define READ_BUF_SIZE 2048 #define SEQ_BUF_SIZE (BLOCK_BUF_SIZE + READ_BUF_SIZE) #define INIT_ALLOC_SIZE 4096 #define DIVIDE_BY (CLOCKS_PER_SEC * 1.0) #define KERNEL_NUM 7 #ifdef SHOW_PERF // 用来调试，计算感兴趣部分的运行时间 // 获取当前毫秒数 int64_t get_mseconds() { // struct timeval tv; // gettimeofday(&tv, NULL); // return (int64_t)1000 * (tv.tv_sec + ((1e-6) * tv.tv_usec)); return clock(); } int64_t time_sw[KERNEL_NUM] = {0}; #endif #define _PERFORMANCE_TEST_NORMAL(kernel_num, func) \ cur_query_pos = 0; \ cur_target_pos = 0; \ for (i = 0; i < block_line_num; ++i) \ { \ score[kernel_num] = func( \ &tmem[kernel_num], \ info_arr[i][0], \ (uint8_t *)query_arr + cur_query_pos, \ info_arr[i][1], \ (uint8_t *)target_arr + cur_target_pos, \ 5, mat, 6, 1, 6, 1, 100, 5, 100, \ info_arr[i][2], \ &qle, &tle, >le, &gscore, &max_off[0]); \ score_total[kernel_num] += score[kernel_num]; \ cur_query_pos += info_arr[i][0]; \ cur_target_pos += info_arr[i][1]; \ } #define _PERFORMANCE_TEST_AVX2(kernel_num, func) \ cur_query_pos = 0; \ cur_target_pos = 0; \ for (i = 0; i < block_line_num; ++i) \ { \ score[kernel_num] = func( \ &tmem[kernel_num], \ info_arr[i][0], \ (uint8_t *)query_arr + cur_query_pos, \ info_arr[i][1], \ (uint8_t *)target_arr + cur_target_pos, \ 0, 6, 1, 6, 1, \ 1, 4, \ 100, 5, \ info_arr[i][2], \ &qle, &tle, >le, &gscore, &max_off[0]); \ score_total[kernel_num] += score[kernel_num]; \ cur_query_pos += info_arr[i][0]; \ cur_target_pos += info_arr[i][1]; \ } #ifdef SHOW_PERF #define PERFORMANCE_TEST_NORMAL(kernel_num, func) \ start_time = get_mseconds(); \ _PERFORMANCE_TEST_NORMAL(kernel_num, func); \ time_sw[kernel_num] += get_mseconds() - start_time #define PERFORMANCE_TEST_AVX2(kernel_num, func) \ start_time = get_mseconds(); \ _PERFORMANCE_TEST_AVX2(kernel_num, func); \ time_sw[kernel_num] += get_mseconds() - start_time #else #define PERFORMANCE_TEST_NORMAL(kernel_num, func) _PERFORMANCE_TEST_NORMAL(kernel_num, func) #define PERFORMANCE_TEST_AVX2(kernel_num, func) _PERFORMANCE_TEST_AVX2(kernel_num, func) #endif // 读取一行序列数据 int read_seq_line(char *read_buf, FILE *f_ptr, char *out_arr) { if (fgets(read_buf, READ_BUF_SIZE, f_ptr) == NULL) return 0; int line_size = strlen(read_buf); assert(line_size < READ_BUF_SIZE); if (read_buf[line_size - 1] == '\n') { read_buf[line_size - 1] = '\0'; line_size--; } convert_char_to_2bit(read_buf); strncpy(out_arr, read_buf, line_size); return line_size; } // 程序执行入口 int main(int argc, char *argv[]) { const char *qf_path = argv[1]; const char *tf_path = argv[2]; const char *if_path = argv[3]; // 初始化一些全局参数 int8_t mat[25] = {1, -4, -4, -4, -1, -4, 1, -4, -4, -1, -4, -4, 1, -4, -1, -4, -4, -4, 1, -1, -1, -1, -1, -1, -1}; int max_off[2]; int qle, tle, gtle, gscore; thread_mem_t tmem[KERNEL_NUM]; int i, j; for (i = 0; i < KERNEL_NUM; ++i) { thread_mem_init_alloc(tmem + i, INIT_ALLOC_SIZE); } // 记录计算出的分数 int score[KERNEL_NUM] = {0}; int score_total[KERNEL_NUM] = {0}; // 读取测试数据 char *query_arr = (char *)malloc(SEQ_BUF_SIZE); char *target_arr = (char *)malloc(SEQ_BUF_SIZE); int *info_buf = (int *)malloc(SEQ_BUF_SIZE * sizeof(int)); int **info_arr = (int **)malloc(SEQ_BUF_SIZE * sizeof(int *)); FILE *query_f = 0, *target_f = 0, *info_f = 0; FILE *normal_out_f = 0, *avx2_out_f = 0, *avx2_u8_out_f = 0; query_f = fopen(qf_path, "r"); target_f = fopen(tf_path, "r"); info_f = fopen(if_path, "r"); // 将每次比对的得分等信息写入文件，进行debug // normal_out_f = fopen("normal_out.txt", "w"); // avx2_out_f = fopen("avx2_out.txt", "w"); // avx2_u8_out_f = fopen("avx2_u8_out.txt", "w"); // 每次读取一定量的数据，然后执行，直到处理完所有数据 int total_line_num = 0; // 目前处理的总的数据行数 int block_line_num = 0; // 当前循环包含的数据行数 int cur_query_pos, cur_target_pos; int64_t start_time; char read_buf[READ_BUF_SIZE]; // 读文件缓存 // 初始化info_arr数组 i = 0; j = 0; while (1) { if (j > BLOCK_BUF_SIZE) break; info_arr[i] = &info_buf[j]; i += 1; j += 3; } while (!feof(target_f)) { block_line_num = 0; // 记录每次读取的行数 // target序列一般占用存储最多，先读取target，看一个buf能读多少行，query和info就按照这个行数来读 int cur_read_size = 0; while (!feof(target_f) && cur_read_size < BLOCK_BUF_SIZE) { int line_size = read_seq_line(read_buf, target_f, target_arr + cur_read_size); if (line_size == 0) break; cur_read_size += line_size; ++block_line_num; ++total_line_num; } // 读query cur_read_size = 0; for (i = 0; i < block_line_num; ++i) { int line_size = read_seq_line(read_buf, query_f, query_arr + cur_read_size); if (line_size == 0) break; cur_read_size += line_size; } // 读info cur_read_size = 0; for (i = 0; i < block_line_num; ++i) { if (fgets(read_buf, READ_BUF_SIZE, info_f) == NULL) break; const int line_size = strlen(read_buf); assert(line_size < READ_BUF_SIZE); sscanf(read_buf, "%d %d %d\n", &info_arr[i][0], &info_arr[i][1], &info_arr[i][2]); cur_read_size += line_size; // fprintf(stderr, "%-8d%-8d%-8d\n", info_arr[i][0], info_arr[i][1], info_arr[i][2]); // fprintf(stderr, "%s\n", read_buf); } // 性能测试 // normal sw PERFORMANCE_TEST_NORMAL(0, ksw_extend_normal); // avx2 PERFORMANCE_TEST_AVX2(1, ksw_extend_avx2); // avx2 u8 PERFORMANCE_TEST_AVX2(2, ksw_extend_avx2_u8); // avx2 heuristics PERFORMANCE_TEST_AVX2(3, ksw_extend_avx2_heuristics); // avx2 u8 heuristics PERFORMANCE_TEST_AVX2(4, ksw_extend_avx2_u8_heuristics); // avx2 mem aligned PERFORMANCE_TEST_AVX2(5, ksw_extend_avx2_aligned); // avx2 u8 mem aligned PERFORMANCE_TEST_AVX2(6, ksw_extend_avx2_u8_aligned); } #ifdef SHOW_PERF char *kernel_names[7] = { "normal", "avx2", "avx2_u8", "avx2_heuristics", "avx2_u8_heuristics", "avx2_aligned", "avx2_u8_aligned"}; for (i = 0; i < KERNEL_NUM; ++i) { fprintf(stderr, "[%18s] time: %9.6f s; score: %d\n", kernel_names[i], time_sw[i] / DIVIDE_BY, score_total[i]); } #endif if (query_f != 0) fclose(query_f); if (target_f != 0) fclose(target_f); if (info_f != 0) fclose(info_f); if (avx2_out_f != 0) fclose(avx2_out_f); if (avx2_u8_out_f != 0) fclose(avx2_u8_out_f); if (normal_out_f != 0) fclose(normal_out_f); }