mirror of
https://github.com/intel/isa-l.git
synced 2025-10-28 11:31:51 +01:00
all: Revamp performance testing to be time based
Change-Id: I6260d28e4adc974d8db0a1c770e3eb922d87f8e4 Signed-off-by: Roy Oursler <roy.j.oursler@intel.com>
This commit is contained in:
Roy Oursler
259
include/test.h
259
include/test.h
@@ -34,70 +34,245 @@
|
||||
extern "C" {
|
||||
#endif
|
||||
|
||||
// Use sys/time.h functions for time
|
||||
#if defined (__unix__) || (__APPLE__) || (__MINGW32__)
|
||||
# include <sys/time.h>
|
||||
#endif
|
||||
|
||||
#ifdef _MSC_VER
|
||||
# define inline __inline
|
||||
# include <time.h>
|
||||
# include <Windows.h>
|
||||
#endif
|
||||
|
||||
#include <stdio.h>
|
||||
#include <stdint.h>
|
||||
|
||||
struct perf{
|
||||
struct timeval tv;
|
||||
#ifdef _MSC_VER
|
||||
# define inline __inline
|
||||
#endif
|
||||
|
||||
/* Decide wether to use benchmark time as an approximation or a minimum. Fewer
|
||||
* calls to the timer are required for the approximation case.*/
|
||||
#define BENCHMARK_MIN_TIME 0
|
||||
#define BENCHMARK_APPROX_TIME 1
|
||||
#ifndef BENCHMARK_TYPE
|
||||
#define BENCHMARK_TYPE BENCHMARK_MIN_TIME
|
||||
#endif
|
||||
|
||||
#ifdef USE_RDTSC
|
||||
/* The use of rtdsc is nuanced. On many processors it corresponds to a
|
||||
* standardized clock source. To obtain a meaningful result it may be
|
||||
* necessary to fix the CPU clock to match the rtdsc tick rate.
|
||||
*/
|
||||
# include <inttypes.h>
|
||||
# include <x86intrin.h>
|
||||
# define USE_CYCLES
|
||||
#else
|
||||
# include <time.h>
|
||||
#define USE_SECONDS
|
||||
#endif
|
||||
|
||||
#ifdef USE_RDTSC
|
||||
#ifndef BENCHMARK_TIME
|
||||
# define BENCHMARK_TIME 6
|
||||
#endif
|
||||
# define GHZ 1000000000
|
||||
# define UNIT_SCALE (GHZ)
|
||||
# define CALLIBRATE_TIME (UNIT_SCALE / 2)
|
||||
static inline long long get_time(void) {
|
||||
unsigned int dummy;
|
||||
return __rdtscp(&dummy);
|
||||
}
|
||||
|
||||
static inline long long get_res(void) {
|
||||
return 1;
|
||||
}
|
||||
#else
|
||||
#ifndef BENCHMARK_TIME
|
||||
# define BENCHMARK_TIME 3
|
||||
#endif
|
||||
#ifdef _MSC_VER
|
||||
#define UNIT_SCALE get_res()
|
||||
#define CALLIBRATE_TIME (UNIT_SCALE / 4)
|
||||
static inline long long get_time(void) {
|
||||
long long ret = 0;
|
||||
QueryPerformanceCounter(&ret);
|
||||
return ret;
|
||||
}
|
||||
|
||||
static inline long long get_res(void) {
|
||||
long long ret = 0;
|
||||
QueryPerformanceFrequency(&ret);
|
||||
return ret;
|
||||
}
|
||||
#else
|
||||
# define NANO_SCALE 1000000000
|
||||
# define UNIT_SCALE NANO_SCALE
|
||||
# define CALLIBRATE_TIME (UNIT_SCALE / 4)
|
||||
#ifdef __FreeBSD__
|
||||
# define CLOCK_ID CLOCK_MONOTONIC_PRECISE
|
||||
#else
|
||||
# define CLOCK_ID CLOCK_MONOTONIC
|
||||
#endif
|
||||
|
||||
static inline long long get_time(void) {
|
||||
struct timespec time;
|
||||
long long nano_total;
|
||||
clock_gettime(CLOCK_ID, &time);
|
||||
nano_total = time.tv_sec;
|
||||
nano_total *= NANO_SCALE;
|
||||
nano_total += time.tv_nsec;
|
||||
return nano_total;
|
||||
}
|
||||
|
||||
static inline long long get_res(void) {
|
||||
struct timespec time;
|
||||
long long nano_total;
|
||||
clock_getres(CLOCK_ID, &time);
|
||||
nano_total = time.tv_sec;
|
||||
nano_total *= NANO_SCALE;
|
||||
nano_total += time.tv_nsec;
|
||||
return nano_total;
|
||||
}
|
||||
#endif
|
||||
#endif
|
||||
struct perf {
|
||||
long long start;
|
||||
long long stop;
|
||||
long long run_total;
|
||||
long long iterations;
|
||||
};
|
||||
|
||||
|
||||
#if defined (__unix__) || (__APPLE__) || (__MINGW32__)
|
||||
static inline int perf_start(struct perf *p)
|
||||
{
|
||||
return gettimeofday(&(p->tv), 0);
|
||||
}
|
||||
static inline int perf_stop(struct perf *p)
|
||||
{
|
||||
return gettimeofday(&(p->tv), 0);
|
||||
static inline void perf_init(struct perf *p) {
|
||||
p->start = 0;
|
||||
p->stop = 0;
|
||||
p->run_total = 0;
|
||||
}
|
||||
|
||||
static inline void perf_print(struct perf stop, struct perf start, long long dsize)
|
||||
{
|
||||
long long secs = stop.tv.tv_sec - start.tv.tv_sec;
|
||||
long long usecs = secs * 1000000 + stop.tv.tv_usec - start.tv.tv_usec;
|
||||
static inline void perf_continue(struct perf *p) {
|
||||
p->start = get_time();
|
||||
}
|
||||
|
||||
static inline void perf_pause(struct perf *p) {
|
||||
p->stop = get_time();
|
||||
p->run_total = p->run_total + p->stop - p->start;
|
||||
p->start = p->stop;
|
||||
}
|
||||
|
||||
static inline void perf_start(struct perf *p) {
|
||||
perf_init(p);
|
||||
perf_continue(p);
|
||||
}
|
||||
|
||||
static inline void perf_stop(struct perf *p) {
|
||||
perf_pause(p);
|
||||
}
|
||||
|
||||
static inline double get_time_elapsed(struct perf *p) {
|
||||
return 1.0 * p->run_total / UNIT_SCALE;
|
||||
}
|
||||
|
||||
static inline long long get_base_elapsed(struct perf *p) {
|
||||
return p->run_total;
|
||||
}
|
||||
|
||||
static inline int estimate_perf_iterations(struct perf *p, unsigned long long runs,
|
||||
unsigned long long total) {
|
||||
total = total * runs;
|
||||
if (get_base_elapsed(p) > 0)
|
||||
return (total + get_base_elapsed(p) - 1) / get_base_elapsed(p);
|
||||
else
|
||||
return (total + get_res() - 1) / get_res();
|
||||
}
|
||||
|
||||
#define CALLIBRATE(PERF, FUNC_CALL) { \
|
||||
unsigned long long _iter = 1; \
|
||||
perf_start(PERF); \
|
||||
FUNC_CALL; \
|
||||
perf_pause(PERF); \
|
||||
\
|
||||
while (get_base_elapsed(PERF) < CALLIBRATE_TIME) { \
|
||||
_iter = estimate_perf_iterations(PERF, _iter, \
|
||||
2 * CALLIBRATE_TIME); \
|
||||
perf_start(PERF); \
|
||||
for (int _i = 0; _i < _iter; _i++) { \
|
||||
FUNC_CALL; \
|
||||
} \
|
||||
perf_stop(PERF); \
|
||||
} \
|
||||
(PERF)->iterations=_iter; \
|
||||
}
|
||||
|
||||
#define PERFORMANCE_TEST(PERF, RUN_TIME, FUNC_CALL) { \
|
||||
unsigned long long _iter = (PERF)->iterations; \
|
||||
unsigned long long _run_total = RUN_TIME; \
|
||||
_run_total *= UNIT_SCALE; \
|
||||
_iter = estimate_perf_iterations(PERF, _iter, _run_total);\
|
||||
(PERF)->iterations = 0; \
|
||||
perf_start(PERF); \
|
||||
for (int _i = 0; _i < _iter; _i++) { \
|
||||
FUNC_CALL; \
|
||||
} \
|
||||
perf_pause(PERF); \
|
||||
(PERF)->iterations += _iter; \
|
||||
\
|
||||
if(get_base_elapsed(PERF) < _run_total && \
|
||||
BENCHMARK_TYPE == BENCHMARK_MIN_TIME) { \
|
||||
_iter = estimate_perf_iterations(PERF, _iter, \
|
||||
_run_total - get_base_elapsed(PERF) + \
|
||||
(UNIT_SCALE / 16)); \
|
||||
perf_continue(PERF); \
|
||||
for (int _i = 0; _i < _iter; _i++) { \
|
||||
FUNC_CALL; \
|
||||
} \
|
||||
perf_pause(PERF); \
|
||||
(PERF)->iterations += _iter; \
|
||||
} \
|
||||
}
|
||||
|
||||
#define BENCHMARK(PERF, RUN_TIME, FUNC_CALL) { \
|
||||
if((RUN_TIME) > 0) { \
|
||||
CALLIBRATE(PERF, FUNC_CALL); \
|
||||
PERFORMANCE_TEST(PERF, RUN_TIME, FUNC_CALL); \
|
||||
\
|
||||
} else { \
|
||||
(PERF)->iterations = 1; \
|
||||
perf_start(PERF); \
|
||||
FUNC_CALL; \
|
||||
perf_stop(PERF); \
|
||||
} \
|
||||
}
|
||||
|
||||
#ifdef USE_CYCLES
|
||||
static inline void perf_print(struct perf p, long long unit_count) {
|
||||
long long total_units = p.iterations * unit_count;
|
||||
|
||||
printf("runtime = %10lld ticks", get_base_elapsed(&p));
|
||||
if (total_units != 0) {
|
||||
printf(", bandwidth %lld MB in %.4f GC = %.2f ticks/byte",
|
||||
total_units / (1000000), get_time_elapsed(&p),
|
||||
get_base_elapsed(&p) / (double)total_units);
|
||||
}
|
||||
printf("\n");
|
||||
}
|
||||
#else
|
||||
static inline void perf_print(struct perf p, double unit_count) {
|
||||
long long total_units = p.iterations * unit_count;
|
||||
long long usecs = (long long)(get_time_elapsed(&p) * 1000000);
|
||||
|
||||
printf("runtime = %10lld usecs", usecs);
|
||||
if (dsize != 0) {
|
||||
#if 1 // not bug in printf for 32-bit
|
||||
printf(", bandwidth %lld MB in %.4f sec = %.2f MB/s\n", dsize/(1024*1024),
|
||||
((double) usecs)/1000000, ((double) dsize) / (double)usecs);
|
||||
#else
|
||||
printf(", bandwidth %lld MB ", dsize/(1024*1024));
|
||||
printf("in %.4f sec ",(double)usecs/1000000);
|
||||
printf("= %.2f MB/s\n", (double)dsize/usecs);
|
||||
#endif
|
||||
if (total_units != 0) {
|
||||
printf(", bandwidth %lld MB in %.4f sec = %.2f MB/s",
|
||||
total_units / (1000000), get_time_elapsed(&p),
|
||||
((double)total_units) / (1000000 * get_time_elapsed(&p)));
|
||||
}
|
||||
else
|
||||
printf("\n");
|
||||
printf("\n");
|
||||
}
|
||||
#endif
|
||||
|
||||
static inline uint64_t get_filesize(FILE *fp)
|
||||
{
|
||||
static inline uint64_t get_filesize(FILE * fp) {
|
||||
uint64_t file_size;
|
||||
fpos_t pos, pos_curr;
|
||||
|
||||
fgetpos(fp, &pos_curr); /* Save current position */
|
||||
fgetpos(fp, &pos_curr); /* Save current position */
|
||||
#if defined(_WIN32) || defined(_WIN64)
|
||||
_fseeki64(fp, 0, SEEK_END);
|
||||
#else
|
||||
fseeko(fp, 0, SEEK_END);
|
||||
#endif
|
||||
fgetpos(fp, &pos);
|
||||
file_size = *(uint64_t *)&pos;
|
||||
fsetpos(fp, &pos_curr); /* Restore position */
|
||||
file_size = *(uint64_t *) & pos;
|
||||
fsetpos(fp, &pos_curr); /* Restore position */
|
||||
|
||||
return file_size;
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user