rep; Nop / -asm pause

[mark-whitener]

On Xeon or Hyper-Threading tech, what isthe cycle time for rep; nop or _asm pause? The documentation hints that it can be anywhere from a nop to some definite value. So what is it?

Thanks in advance

[Intel_Software_Netw1]

We are forwarding your question to our engineering contacts and will let you know how they respond.

Regards,

Lexi S.

IntelSoftware NetworkSupport

http://www.intel.com/software

Contact us

Message Edited by intel.software.network.support on 12-02-2005 01:14 PM

[Intel_Software_Netw1]

Here is the response we received from our Application Engineers:

NOP instruction can be between 0.4-0.5 clocks and PAUSE instruction can consume 38-40 clocks. Please refer to the whitepaper on how to measure the latency and throughput of various instructions. The REPE instruction comes in various flavors and the latency/throughput of each of them varies. Please also see below for the sample code to measure the average clocks.

#include

#define ReadTSC( x ) __asm cpuid

__asm rdtsc

__asm mov dword ptr x,eax

__asm mov dword ptr x+4,edx

#define LOOP_COUNT 160000.

#define REPEAT_25( x ) x x x x x x x x x x x x x x x x x x x x x x x x x

#define REPEAT_100(x) REPEAT_25(x) REPEAT_25(x) REPEAT_25(x) REPEAT_25(x)

#define REPEAT_1000(x) REPEAT_100(x) REPEAT_100(x) REPEAT_100(x) REPEAT_100(x) REPEAT_100(x) REPEAT_100(x) REPEAT_100(x) REPEAT_100(x) REPEAT_100(x) REPEAT_100(x)

#define FACTOR ((double)LOOP_COUNT*1000.0)

#define CLOCKSPERINSTRUCTION(start,end) ((double)end-(double)start)/(FACTOR)

void main(int argc,char **argv)

{

__int64 start, end,total;

total = 0;;

ReadTSC(start);

for (int i=0; i

{

REPEAT_1000(__asm { nop};)

}

ReadTSC(end);

total = end-start;

printf("nop: clocks per instruction %4.2f ",(double)total/(double)FACTOR);

ReadTSC(start);

total = 0;;

for (int i=0; i

{

REPEAT_1000(__asm { pause};)

}

ReadTSC(end);

total = end-start;

printf("pause: clocks per instruction %4.2f ",(double)total/(double)FACTOR);

}

==

Regards,

Lexi S.

IntelSoftware NetworkSupport

http://www.intel.com/software

Contact us

Message Edited by intel.software.network.support on 12-02-2005 01:14 PM

[mark-whitener]

Thank you. The documentation states the results vary as does the answer. I'll try this on our various 8x, 16x and32x machines and see what the variance is.

Thanks again.