/* 適当にmain.cppにコピペしてコンパイル 処理は下の方 */ /************************************* 簡易ベンチマーク (BMT_/bmt_) (途中) (.c/.cpp/Windows/Linux/x86/AMD64) *************************************/ #define BMT_STD 0 #define BMT_RTC 1 #define BMT_TSC 2 /* BMT_TIMER_MODE 0 or BMT_STD = 標準クロック 1 or BMT_RTC = リアルタイムクロック(-lrt) 2 or BMT_TSC = タイムスタンプカウンタ */ #define BMT_TIMER_MODE 2 /* 計測回数(BMT_MAXTRY)と処理回数(BMT_MAXLOOP) < 擬似コード > for(BMT_MAXTRY) { 計測開始 for(BMT_MAXLOOP) { 処理 } 計測終了＆ベストタイム更新 } 結果が安定するまで適当に変更する */ #if BMT_TIMER_MODE == BMT_STD #define BMT_MAXTRY 1 #define BMT_MAXLOOP 1 #elif BMT_TIMER_MODE == BMT_RTC #define BMT_MAXTRY 1 #define BMT_MAXLOOP 1 #elif BMT_TIMER_MODE == BMT_TSC #define BMT_MAXTRY 100 #define BMT_MAXLOOP 1024 #endif /* BMT_USE_THREAD 0 = スレッドを使わない 1 = スレッドを使う(-lpthread) */ #define BMT_USE_THREAD 0 /*==================================*/ #include #include #include #include #include #include #include #include #if defined(_WIN32) #if !defined(NOMINMAX) #define NOMINMAX 1 #endif #include #include #if defined(__GNUC__) #define BMT_INLINE static inline #define BMT_NOINLINE __attribute__((noinline)) #define BMT_ALIGNED(x) __attribute__((aligned(x))) #else #define BMT_INLINE __inline #define BMT_NOINLINE __declspec(noinline) #define BMT_ALIGNED(x) __declspec(align(x)) #endif #else /* Linux etc.. */ #define _tmain main #define _T(x) x #define _TCHAR char #define BMT_INLINE static inline #define BMT_NOINLINE __attribute__((noinline)) #define BMT_ALIGNED(x) __attribute__((aligned(x))) #endif /* printf */ void bmt_printf(_TCHAR const *fmt, ...) { va_list ap; va_start(ap,fmt); #if defined(_WIN32) && (defined(UNICODE)||defined(_UNICODE)) vwprintf(fmt,ap); #else vprintf(fmt,ap); #endif #if defined(__MINGW32__) || defined(__MINGW64__) fflush(stdout); #endif } /* timer */ #if BMT_TIMER_MODE == BMT_STD typedef unsigned long bmt_nt; /* normalized type */ typedef clock_t bmt_rt; /* raw type 　　　*/ #elif BMT_TIMER_MODE == BMT_RTC typedef unsigned long long bmt_nt; #if defined(_WIN32) typedef LARGE_INTEGER bmt_rt; #else typedef struct timespec bmt_rt; #endif #elif BMT_TIMER_MODE == BMT_TSC typedef unsigned long long bmt_nt; typedef unsigned long long bmt_rt; #endif typedef struct { struct BMT_TIMER_INNER { double freq; bmt_rt st,et; bmt_nt elp,best; int lc,tc; } i; char pad[64-sizeof(struct BMT_TIMER_INNER)]; } bmt_timer_t; #if BMT_TIMER_MODE == BMT_STD BMT_INLINE bmt_nt bmt_norm(bmt_rt *t) { return (bmt_nt)(*t); } BMT_INLINE void bmt_setfreq(bmt_timer_t *t) { t->i.freq=(double)CLOCKS_PER_SEC; } BMT_INLINE void bmt_setmax(bmt_timer_t *t) { t->i.best = ULONG_MAX; } BMT_INLINE void bmt_now(bmt_rt *t) { *t = clock(); } BMT_INLINE void bmt_result(_TCHAR const *title, bmt_timer_t *t) { bmt_printf(_T("%16s -> %10.6f(s) %.0f(t)\n"),title ,((double)t->i.best) / t->i.freq,(double)t->i.best); } #elif BMT_TIMER_MODE == BMT_RTC #if defined(_WIN32) BMT_INLINE bmt_nt bmt_norm(bmt_rt *t) { return (bmt_nt)(t->QuadPart); } BMT_INLINE int bmt_setfreq(bmt_timer_t *t) { int r; bmt_rt rt; r = (int)QueryPerformanceFrequency(&rt); return r?(t->i.freq=(double)bmt_norm(&rt),1):0; } BMT_INLINE void bmt_now(bmt_rt *t) { QueryPerformanceCounter(t); } #else #define BMT_NS (1000000000ull) BMT_INLINE bmt_nt bmt_norm(bmt_rt *t) { return (bmt_nt)(t->tv_nsec)+((bmt_nt)(t->tv_sec)*BMT_NS); } BMT_INLINE int bmt_setfreq(bmt_timer_t *t) { int r; bmt_rt rt; r = clock_getres(CLOCK_REALTIME,&rt); return r?0:(t->i.freq=(double)BMT_NS/(double)bmt_norm(&rt),1); } BMT_INLINE void bmt_now(bmt_rt *t) { clock_gettime(CLOCK_REALTIME,t); } #endif BMT_INLINE void bmt_setmax(bmt_timer_t *t) { t->i.best = 0xffffffffffffffffull /*ULLONG_MAX*/; } BMT_INLINE void bmt_result(_TCHAR const *title, bmt_timer_t *t) { bmt_printf(_T("%16s -> %10.6f(s) %.0f(t)\n"),title ,((double)t->i.best) / t->i.freq,(double)t->i.best); } #elif BMT_TIMER_MODE == BMT_TSC #define bmt_norm(x) (*(x)) BMT_INLINE void bmt_setmax(bmt_timer_t *t) { t->i.best = 0xffffffffffffffffull /*ULLONG_MAX*/; } #if defined(_MSC_VER) && defined(_WIN32) BMT_NOINLINE void bmt_now(bmt_rt *t) { *t = (bmt_rt)__rdtsc(); } #else BMT_NOINLINE void bmt_now(bmt_rt *t) { unsigned int low, high; __asm__ volatile("rdtsc" : "=a"(low), "=d"(high)); *t = (bmt_rt)(high)<<32|low; } #endif BMT_INLINE void bmt_result(_TCHAR const *title, bmt_timer_t *t) { bmt_printf(_T("%16s -> %.0f(t)\n"),title,(double)t->i.best); } #endif BMT_INLINE void bmt_end(bmt_timer_t *t) { bmt_now(&t->i.et); t->i.elp = bmt_norm(&t->i.et)-bmt_norm(&t->i.st); if(t->i.elpi.best) t->i.best = t->i.elp; } #define BMT_TIMER_BEGIN {\ bmt_setmax(&g_bmt_tdat);\ for(g_bmt_tdat.i.tc=BMT_MAXTRY;g_bmt_tdat.i.tc>0;--g_bmt_tdat.i.tc) {\ int bmt_lc = BMT_MAXLOOP;\ for(bmt_now(&g_bmt_tdat.i.st);bmt_lc>0;--bmt_lc) { #define BMT_TIMER_END(title)\ } bmt_end(&g_bmt_tdat);\ } bmt_result(title, &g_bmt_tdat);\ } static BMT_ALIGNED(64) bmt_timer_t g_bmt_tdat; /* aligned malloc/calloc/free */ void *bmt_aligned_malloc(size_t const size,size_t const align/* 2**n */) { size_t ap = (size_t)NULL; size_t np = (size_t)malloc(size+sizeof(size_t)+align-1); if(np) { ap = ((np+sizeof(size_t)+align-1)&~(align-1)); *(size_t*)(ap-sizeof(size_t)) = np; } return (void*)ap; } void *bmt_aligned_calloc(size_t size,size_t const align/* 2**n */) { size_t ap = (size_t)bmt_aligned_malloc(size,align); if(ap) { char volatile *p = (char volatile*)ap; while(size) { *p++ = 0; size--; } } return (void*)ap; } void bmt_aligned_free(void *p) { if(p) free((void*)*(size_t*)((size_t)p-sizeof(size_t))); } #if BMT_USE_THREAD #if defined(_WIN32) #include #include typedef HANDLE bmt_thread_t; #else #include #include typedef pthread_t bmt_thread_t; #endif /* atomic */ #if defined(__GNUC__) BMT_INLINE int bmt_exchg(int volatile *pval,int newval) { return (int)__sync_lock_test_and_set(pval,newval); } BMT_INLINE int bmt_exchg_if(int volatile *pval,int newval,int oldval) { return (int)__sync_val_compare_and_swap(pval,oldval,newval); } #elif defined(_WIN32) BMT_INLINE int bmt_exchg(int volatile *pval,int newval) { return (int)_InterlockedExchange((long*)pval,newval); } BMT_INLINE int bmt_exchg_if(int volatile *pval,int newval,int oldval) { return (int)_InterlockedCompareExchange((long*)pval,newval,oldval); } #endif /* spin lock */ typedef enum { BMT_UNLOCKED=0, BMT_LOCKED=1 } bmt_stat_t; BMT_INLINE void bmt_spin_init(int volatile *p,bmt_stat_t s) { bmt_exchg(p,s); } BMT_INLINE void bmt_spin_wait(int volatile *p) { do { _mm_pause(); } while(*p); } BMT_INLINE void bmt_spin_lock(int volatile *p) { if(bmt_exchg(p,BMT_LOCKED)) do { bmt_spin_wait(p); } while(bmt_exchg(p,BMT_LOCKED)); } BMT_INLINE void bmt_spin_unlock(int volatile *p) { bmt_exchg(p,BMT_UNLOCKED); } /* thread */ typedef struct { void (*func)(void*); void *arg; int *lock; } bmt_thread_data; #if defined(_WIN32) unsigned __stdcall bmt_first(void *arg) { bmt_thread_data td = *(bmt_thread_data*)arg; bmt_spin_unlock(td.lock); (td.func)(td.arg); return 0; } int bmt_create(bmt_thread_t *th,void (*func)(void*),void *arg) { int lock; bmt_thread_data td; td.func = func; td.arg = arg; td.lock = &lock; bmt_spin_init(td.lock,BMT_LOCKED); *th = (bmt_thread_t)_beginthreadex(NULL,0,bmt_first,&td,0,NULL); return *th?(bmt_spin_wait(td.lock),1):0; } void bmt_join(bmt_thread_t th) { WaitForSingleObject(th,INFINITE); CloseHandle(th); } #else void *bmt_first(void *arg) { bmt_thread_data td = *(bmt_thread_data*)arg; bmt_spin_unlock(td.lock); (td.func)(td.arg); return 0; } int bmt_create(bmt_thread_t *th,void (*func)(void*),void *arg) { int lock; bmt_thread_data td; td.func = func; td.arg = arg; td.lock = &lock; bmt_spin_init(td.lock,BMT_LOCKED); return pthread_create(th,NULL,bmt_first,&td)?0:(bmt_spin_wait(td.lock),1); } void bmt_join(BMT_THT th) { pthread_join(th,NULL); } #endif #endif /* BMT_USE_THREAD */ /* main */ void BMT_NOINLINE run_benchmark(); int _tmain() { #if BMT_TIMER_MODE == BMT_STD bmt_setfreq(&g_bmt_tdat); bmt_printf(_T("<< start >> clock , %.0fticks/sec\n"),g_bmt_tdat.i.freq); run_benchmark(); bmt_printf(_T("<< end >>\n")); #elif BMT_TIMER_MODE == BMT_RTC if(!bmt_setfreq(&g_bmt_tdat)) { bmt_printf(_T("<< start >> rtc is not supported.\n")); } else { bmt_printf(_T("<< start >> rtc , %.0fticks/sec\n"),g_bmt_tdat.i.freq); run_benchmark(); } bmt_printf(_T("<< end >>\n")); #elif BMT_TIMER_MODE == BMT_TSC bmt_printf(_T("<< start >> tsc\n")); run_benchmark(); bmt_printf(_T("<< end >>\n")); #endif #if defined(_MSC_VER) bmt_printf(_T("Press Enter to quit.")); getchar(); #endif return 0; } /************************************* ここから *************************************/ #include struct BUF { int volatile one; int dmy1[(128/sizeof(int))-1]; int volatile zero; int dmy2[(128/sizeof(int))-1]; }; #define N (1000) #define DO_ONE p->one = 1 #define DO_ZERO p->zero = 0 /*------------------------------------ 851 ------------------------------------*/ void BMT_NOINLINE func_851(struct BUF *p) { uint_fast16_t bits = 0x49; BMT_TIMER_BEGIN // 全体で 13bit * N回 の処理 int i; for(i = 0; i<(13*N); ++i) { // 13bit * N回 if (bits & 1) { DO_ONE; // ビットが 1 の時の処理 bits |= 1 << 13; } else { DO_ZERO; // ビットが 0 の時の処理 } bits >>= 1; } BMT_TIMER_END(_T("851")) } /*------------------------------------ 864-1 ------------------------------------*/ void BMT_NOINLINE func_864_1(struct BUF *p) { uint_fast16_t const bits = 0x49; // ビット列(下位13ビットが有効) BMT_TIMER_BEGIN // 全体で 13bit * N回 の処理 int i; for(i = 0; i>=1) { // 13bit分ループ if(bits & mask) { DO_ONE; // ビットが 1 の時の処理 } else { DO_ZERO; // ビットが 0 の時の処理 } } } BMT_TIMER_END(_T("864-2")) } /************************************* run_benchmark *************************************/ void BMT_NOINLINE run_benchmark() { struct BUF *p = (struct BUF*)bmt_aligned_malloc(sizeof(struct BUF), 4096); if(p) { func_864_2(p); func_851(p); func_864_1(p); func_864_2(p); func_851(p); func_864_1(p); bmt_aligned_free(p); } }