rdtsc 指令测试 cpu 频率

#ifndef __TIMING__INCLUDED__IOSA_4__
#define __TIMING__INCLUDED__IOSA_4__

#include <crtdefs.h>  
#include <stdio.h>  
#include <windows.h>  
#include <mmsystem.h>  
  
#ifdef  __cplusplus  
extern "C" {  
#endif  
    _CRTIMP void __cdecl _wassert(_In_z_ const wchar_t* _Message, _In_z_ const wchar_t* _File, _In_ unsigned _Line);  
#ifdef  __cplusplus  
}  
#endif  
  
#define ENFORCE(_Expression) (void)((!!(_Expression)) || (_wassert(_CRT_WIDE(#_Expression),_CRT_WIDE(__FILE__),__LINE__),0))

static unsigned __int64 UisHnack, CpuFreq, MilliCpuFreq;

__declspec(naked) static unsigned __int64 LoadTSC ( void ) {
	__asm rdtsc
	__asm ret
}

__declspec(naked) static void TSleep ( unsigned int dwMilliseconds ) {
	
	static unsigned __int64 tStart;
	
	__asm {
		rdtsc 
		mov dword ptr [tStart], eax
		mov dword ptr [tStart+4], edx
		mov eax, 4[esp]
		mov dword ptr [UisHnack+4], 0
		mov dword ptr [UisHnack], eax	
	} 
	
	UisHnack *= MilliCpuFreq;
		
	while ( LoadTSC () - tStart <= UisHnack );

	__asm ret
}

static unsigned __int64 LoadSpecifyCpuFreq ( unsigned int CpuCoreIndex ) {

	unsigned __int64 timeStart64, cpu_fm64, timeToRDTSC, timeToQPF, Q_Error, QsecTCnt, pfq, lpQFCount;  
    unsigned __int32 OldPriClassFlags, OldPriThreadFlags;  
  
	HANDLE CurProcessHandle = GetCurrentProcess();  
    HANDLE CurThreadHandle  = GetCurrentThread ();
	
	OldPriClassFlags  = GetPriorityClass ( CurProcessHandle );  
    OldPriThreadFlags = GetThreadPriority( CurThreadHandle ); 
	
	ENFORCE ( CpuCoreIndex < 4 ); 
	ENFORCE ( SetThreadAffinityMask  ( CurThreadHandle,  ( (unsigned int) 1 << CpuCoreIndex ) ) );
	ENFORCE ( SetProcessAffinityMask ( CurProcessHandle, ( (unsigned int) 1 << CpuCoreIndex ) ) ); 
    ENFORCE ( QueryPerformanceFrequency ( ( (LARGE_INTEGER*) &pfq) ) );  
	ENFORCE ( SetPriorityClass ( CurProcessHandle, REALTIME_PRIORITY_CLASS ) );   
    ENFORCE ( SetThreadPriority( CurThreadHandle,  THREAD_PRIORITY_HIGHEST ) ); 
     
    __asm  { 
        rdtsc  	
        mov ebx, eax /* call rdtsc * 1 */ 
        mov ecx, edx 
		
        rdtsc  
		sbb edx, ecx
        sub eax, ebx 	
        mov dword ptr [timeToRDTSC], eax  
		mov dword ptr [timeToRDTSC+4], edx 
		
		rdtsc 
		mov dword ptr [timeToQPF], eax  
		mov dword ptr [timeToQPF+4], edx 
		
		push offset UisHnack
		call dword ptr [QueryPerformanceCounter]
		
		rdtsc
		sub eax, dword ptr [timeToQPF] 
		sbb edx, dword ptr [timeToQPF+4]  
		mov dword ptr [timeToQPF], eax
		mov dword ptr [timeToQPF+4], edx
		
		push offset UisHnack
		call dword ptr [QueryPerformanceCounter]
		
		lea  eax, QsecTCnt
		push eax
		call dword ptr [QueryPerformanceCounter]
		
		mov eax, dword ptr [UisHnack]   ; low 
		mov ebx, dword ptr [UisHnack+4] ; high	
		sub dword ptr [QsecTCnt], eax
		sbb dword ptr [QsecTCnt+4], ebx	
    }  
	
    QueryPerformanceCounter((LARGE_INTEGER*)&lpQFCount);  
  
    timeStart64 = lpQFCount;  
  
    __asm {
		rdtsc  
		mov dword ptr [cpu_fm64], eax  
		mov dword ptr [cpu_fm64+4], edx  
	}
  
    while ((Q_Error = (lpQFCount - timeStart64)) <= pfq )  
    {  
        QueryPerformanceCounter((LARGE_INTEGER*)&lpQFCount);  
    }   
	__asm {
		rdtsc    
		sub eax, dword ptr [cpu_fm64] 
		sbb edx, dword ptr [cpu_fm64+4]  
		mov dword ptr [cpu_fm64], eax
		mov dword ptr [cpu_fm64+4], edx
	}   
	MilliCpuFreq = ( CpuFreq = cpu_fm64 = cpu_fm64 - timeToQPF - timeToRDTSC - (unsigned __int64)( (double) ( Q_Error - pfq ) / (double) QsecTCnt * ( double ) timeToQPF ) ) / 1000;
	
    ENFORCE ( timeToRDTSC < 800 );  
    ENFORCE ( SetPriorityClass ( CurProcessHandle, OldPriClassFlags ) );   
    ENFORCE ( SetThreadPriority( CurThreadHandle,  OldPriThreadFlags ) );  

    return cpu_fm64;  
}

static unsigned __int64 SetFrame ( unsigned int FrameTicks, unsigned int CpuCoreIndex ) {
	return ( LoadSpecifyCpuFreq ( CpuCoreIndex ) / ( unsigned __int64 ) FrameTicks );
}
#endif

		.686                      ; create 32 bit code
		.mmx
		.xmm                     
		.model flat, stdcall      ; 32 bit memory model
		option casemap :none      ; case sensitive

; marco 		
	byt equ byte ptr
	wot equ word ptr
	dwot equ dword ptr
	
; extrn Windows API ...
	extrn exit:proc ; no WindowsAPI 
	extrn MessageBoxA@16:proc
	extrn GetPriorityClass@4:proc 
	extrn SetPriorityClass@8:proc 	
	extrn GetThreadPriority@4:proc 
	extrn SetThreadPriority@8:proc 
	extrn SetThreadAffinityMask@8:proc 
	extrn SetProcessAffinityMask@8:proc 
	extrn QueryPerformanceCounter@4:proc
	extrn QueryPerformanceFrequency@4:proc 

.data?
		__rdtsc_t dd 2 dup(?)
		__per_freq dd 2 dup(?) 	
		__time_temp1 dd 2 dup(?) 
		__time_temp2 dd 2 dup(?) 
		__time_temp3 dd 2 dup(?) 
		__time_temp4 dd 2 dup(?)
		__rdtsc_clock dd 2 dup(?)
		__per_counter_clock dd 2 dup(?)
		__per_counter_clock_plck dd 2 dup(?)
.code

__load_main_cpu_ticks proc C
					  option prologue:none, epilogue:none
		push 1
		push -2
		call SetThreadAffinityMask@8 
		cmp	eax, 0
		je	__exit
		
		push 1
		push -1
		call SetProcessAffinityMask@8 
		cmp	eax, 0
		je	__exit
		
		push 256
		push -1
		call SetPriorityClass@8 
		cmp	eax, 0
		je	__exit
		
		push 2
		push -2
		call SetThreadPriority@8
		cmp eax, -1
		je __exit
		
		push offset __per_freq
		call QueryPerformanceFrequency@4
	; main 
		mov	eax, 35600
		wait_some:		
			dec eax
			jne wait_some
		
			rdtsc 
			mov ecx, eax ; low 8 bit 
			rdtsc 
			sub eax, ecx ; low 8 bit sub ... 	
			mov dwot[__rdtsc_clock], eax 
			
			rdtsc 
			mov dwot[__rdtsc_t], eax 
			push offset __time_temp1 
			call QueryPerformanceCounter@4 
			
			rdtsc 
			sub eax, dwot[__rdtsc_t]
			sub eax, dwot[__rdtsc_clock]
			mov dwot[__per_counter_clock], eax 
			
			push offset __time_temp1
			call QueryPerformanceCounter@4 
			push offset __time_temp2
			call QueryPerformanceCounter@4 
			
			mov eax, dwot[__time_temp2]
			sub eax, dwot[__time_temp1]
			mov dwot[__per_counter_clock_plck], eax 
			
			rdtsc 
			mov dwot[__time_temp3], eax 
			mov dwot[__time_temp3+4], edx 
			push offset __time_temp1
			call QueryPerformanceCounter@4 
			mov eax, dwot[__time_temp1]
			mov edx, dwot[__time_temp1+4]	
			mov dwot[__time_temp2], eax 
			mov dwot[__time_temp2+4], edx 
			
	align 16
		main_loop:
			mov eax, dwot[__time_temp2]
			mov edx, dwot[__time_temp2+4]
			sub eax, dwot[__time_temp1] 
			sbb edx, dwot[__time_temp1+4] 
			cmp edx, dwot[__per_freq+4]
			ja __out_step 
			jb __in_step 
			sub eax, dwot[__per_freq]
			jae __out_step
			__in_step:
				push offset __time_temp2
				call QueryPerformanceCounter@4
				jmp main_loop			
	align 16 
		__out_step:	
			cvtsi2sd xmm0, eax 
			rdtsc 
			sub eax, dwot[__time_temp3]
			sbb edx, dwot[__time_temp3+4]
			sub eax, dwot[__rdtsc_clock]
			sbb edx, dwot[__rdtsc_clock+4]
			mov ecx, dwot[__per_counter_clock]
			shl ecx, 1 
			sub eax, ecx 
			sbb edx, 0		
			cvtsi2sd xmm2, dwot[__per_counter_clock]
			cvtsi2sd xmm1, dwot[__per_counter_clock_plck]
			divsd xmm2, xmm1 
			mulsd xmm2, xmm0 
			cvtsd2si ecx, xmm2
			sub eax, ecx 
			sbb edx, 0
			ret
			
		__exit:
			push 0
			push 0
			push 0 
			push 0
			call MessageBoxA@16
			push -1
			call exit 
			add esp, 4
			ret 		  

__load_main_cpu_ticks endp
	end