【C/C++】性能优化之打点计时（纳秒级）

目的：加时间戳打印是为了方便快速对性能耗时进行测试

比如想要测试printf这个函数调用的耗时

bind_core(4);
for(int i=0;i<MAX_SAMPLES;i++){
	timing_now(now_tsc0);
	printf("11111111");
	timing_now(now_tsc1);
	GATHERING_NS(series0,now_tsc0,now_tsc1,NULL);
}

操作步骤：

1. 隔核
2. 绑核
3. 加一个循环
4. 在printf前加开始时间戳
5. 在printf后加结束时间戳
6. 结束时间戳后加：GATHERING_NS(series0,now_tsc0,now_tsc1,NULL);

bind_core

#include <sched.h>

void bind_core(int core_i){
	cpu_set_t cpus;
	CPU_ZERO(&cpus);
	CPU_SET(core_i, &cpus);
	if(sched_setaffinity(0,sizeof(cpus),&cpus) == -1){
		ERROR("failed bind core %u",core_i);
	} else {
		INFO("success to core %u",core_i);
	}
}

MAX_SAMPLES

#define MAX_SAMPLES 512

timing_now()

#define timing_now(val) do{ \
struct timespec abc; \
clock_gettime(CLOCK_REALTIME, &abc); \
(val) = (unsigned long)abc.tv_sec * 1000000000lu + (unsigned long)abc.tv_nsec; \
} while(0)

now_tsc0、now_tsc1

typedef uint64_t timing_t;
static timing_t now_tsc0 , now_tsc1;

GATHERING_NS()

#define CYCLE_SAMPLES_MASK (MAX_SAMPLES - 1)
#define TXT(T) #T

#define GATHERING_NS(SERIES, SEQ, TSC, LEN)\
	do { \
		SERIES.samples[SERIES.count].seq = (SEQ); \
		SERIES.samples[SERIES.count].tsc = (TSC); \
		SERIES.samples[SERIES.count].len = (LEN); \
		SERIES.count++; \
		if( SERIES.count > CYCLE_SAMPLES_MASK ) { \
			perf_latency_dump_ns(TXT(SERIES), SERIES.samples, MAX_SAMPLES); \
			} \
			SERIES.count = SERIES.count & CYCLE_SAMPLES_MASK; \
		} while(0)

SERIES

typedef struct tsc_sample_struct {
	timing_t seq;
	timing_t tsc;
	int len;
} tsc_sample_struct_t;

typedef struct series_struct {
	tsc_sample_struct_t samples[MAX_SAMPLES];
	uint32_t count;
} series_struct_t;

static series_struct_t series0;

perf_latency_dump_ns()

void perf_latency_dump_ns(const char *tag,tsc_sample_struct_t *stats, int count) {
	int i;
	for (i=0; i<count; i++) {
		printf("perl:%s:%p seq=%04lu, value[%03d]=%luns, len=%d, balance=%ld\n",
		tag, stats, stats[i].seq, i, stats[i].tsc, stats[i].len, stats[i].tsc - stats[i].seq);	
	}
}