/*
Description: profiling related data

Copyright : All right reserved by ICT

Author : Zhang Zhonghai
Date : 2024/04/06
*/
#ifndef PROFILING_H_
#define PROFILING_H_

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

#define USE_RDTSC 1

#define LIM_THREAD 128
#define LIM_THREAD_PROF_TYPE 128
#define LIM_GLOBAL_PROF_TYPE 128
#define LIM_THREAD_DATA_TYPE 256
#define LIM_GLOBAL_DATA_TYPE 256

#ifdef SHOW_PERF
extern uint64_t proc_freq;
extern uint64_t tprof[LIM_THREAD_PROF_TYPE][LIM_THREAD];
extern uint64_t gprof[LIM_GLOBAL_PROF_TYPE];
#endif

#ifdef SHOW_DATA_PERF
extern int64_t tdat[LIM_THREAD_DATA_TYPE][LIM_THREAD];
extern int64_t t_sd[LIM_THREAD_DATA_TYPE][LIM_THREAD];
extern int64_t gdat[LIM_GLOBAL_DATA_TYPE];
extern int64_t gd1[LIM_GLOBAL_DATA_TYPE];
#endif

#ifdef SHOW_PERF
#if USE_RDTSC
#define PROF_START(tmp_time) const uint64_t prof_tmp_##tmp_time = __rdtsc()
#define PROF_END(result, tmp_time) result += __rdtsc() - prof_tmp_##tmp_time
#else
#define PROF_START(tmp_time) uint64_t prof_tmp_##tmp_time = realtime_msec()
#define PROF_END(result, tmp_time) result += realtime_msec() - prof_tmp_##tmp_time
#endif
#else
#define PROF_START(tmp_time)
#define PROF_END(result, tmp_time)
#endif

// tdat
enum {
    TD_SEED_1_0 = 0,
    TD_SEED_1_1,
    TD_SEED_1_2,
    TD_SEED_1_3,
    TD_SEED_1_4,
    TD_SEED_1_5,
    TD_SEED_CNT,
    TD_ALIGN_1_CNT,
    TD_ALIGN_2_CNT,
    TD_MSW_CNT,
};

// gdat
enum {
    GD_READ_CNT = 0,
};

// GLOBAL
enum {
    G_ALL = 0,
    G_LOAD_IDX,
    G_PIPELINE,
    G_READ,
    G_COMPUTE,
    G_WRITE,
    G_SEED_AND_EXT,
    G_MEM_PESTAT,
    G_GEN_SAM,
    G_UNCOMPRESS,
    G_gather_matesw_task,
    G_calc_matesw,
    G_get_matesw_data,
    G_gen_sam,
    G_update_stats_cache,
};

// THREAD
enum {
    T_SEEDING = 0,
    T_GEN_SEQ,
    T_SMEM,
    T_SEED_1,
    T_SEED_2,
    T_SEED_3,
    T_CHAIN,
    T_GEN_CHAIN,
    T_SA,
    T_FLT_CHAIN,
    T_FLT_CHANNED_SEEDS,
    T_EXTENSION,
    T_MEM_CHAIN2ALN,
    T_BSW_EXT,
    T_SORT_DEDUP,
    T_GEN_SAM,
    T_MEM_REG2ALN,
    T_MSW_1,
    T_MSW_2,
    T_MSW_GET_REF,
    T_MSW_PACK_SEQ,

    T_CHAIN_ALL,
    T_ALN_ALL,
    T_INS_SIZE,
    T_SAL,
    T_READ_SA,
    T_BSW,
    T_BSW_ALL,
    T_SAM_MATESW,
    T_KSW_ALIGN2,
    T_KSW_REVERSE,
    T_SAM_REG2ALN,
    T_SAM_ALN2SAM,
    T_SAM_GEN_ALT,
    T_KSW_LOOP,
    T_SEED_LEN,
    T_SEED_1_ALL,
    T_SEED_1_0,
    T_SEED_1_1,
    T_SEED_1_2,
    T_SEED_1_3,
    T_SEED_1_3_1,
    T_SEED_1_3_2,
    T_SEED_1_3_3,
    T_SEED_1_3_4,
    T_SEED_1_3_5,
    T_SEED_1_3_6,
    T_SEED_1_3_7,
    T_SEED_2_ALL,
    T_SEED_2_0,
    T_SEED_2_1,
    T_SEED_2_2,
    T_SEED_2_2_0,
    T_SEED_2_2_1,
    T_SEED_2_2_2,
    T_SEED_2_2_3,
    T_SEED_3_ALL,
    T_SEED_3_0,
    T_SEED_3_1,
    T_SEED_3_2,
    T_SEED_3_3,
    T_SEED_3_3_0,
    T_SEED_3_3_1,
    T_SEED_3_3_2,
};

int display_stats(int);

#endif