#pragma once
#include <stdint.h>
#include <stdlib.h>
#include <sys/time.h>

// #define SHOW_PERF

#ifdef __cplusplus
extern "C" {
#endif

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

#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_PERF
#define PROF_G_BEG(time_name) uint64_t prof_G_tmp_##time_name = realtimeMsec()
#define PROF_G_BEG_AGAIN(time_name) prof_G_tmp_##time_name = realtimeMsec()
#define PROF_G_END(time_name) gprof[GP_##time_name] += realtimeMsec() - prof_G_tmp_##time_name
#define PROF_T_BEG(time_name) uint64_t prof_T_tmp_##time_name = realtimeMsec()
#define PROF_T_BEG_AGAIN(time_name) prof_T_tmp_##time_name = realtimeMsec()
#define PROF_T_END(tid, time_name) tprof[TP_##time_name][tid] += realtimeMsec() - prof_T_tmp_##time_name
#define PROF_PRINT_BEG(tmp_time) uint64_t tmp_time = realtimeMsec()
#define PROF_PRINT_END(tmp_time)          \
    tmp_time = realtimeMsec() - tmp_time; \
    fprintf(stderr, "time %-15s:     %0.2lfs\n", #tmp_time, tmp_time * 1.0 / proc_freq)
#else
#define PROF_G_BEG(time_name)
#define PROF_G_BEG_AGAIN(time_name)
#define PROF_G_END(time_name)
#define PROF_T_BEG(time_name)
#define PROF_T_BEG_AGAIN(time_name)
#define PROF_T_END(tid, time_name)
#endif



// GLOBAL
enum { GP_0 = 0, GP_1, GP_2, GP_3, GP_4, GP_5, GP_6, GP_7, GP_8, GP_9, GP_10 };
enum {
    GP_mid_all = 11,
    GP_read,
    GP_mem_copy,
    GP_parse_block,
    GP_uncompress,
    GP_sort,
    GP_merge,
    GP_compress,
    GP_write_mid,
    GP_read_mid,
    GP_write_final
};
// THREAD
enum { TP_0 = 0, TP_1, TP_2, TP_3, TP_4, TP_5, TP_6, TP_7, TP_8, TP_9, TP_10 };
enum { TP_sort = 11, TP_mem_copy};

uint64_t realtimeMsec(void);

int displayProfiling(int);

#ifdef __cplusplus
}
#endif