🔒 Locked

CPUID timing

Started by drslush Jul 21, 2003 at 10:58 AM 6 replies 1.4k views

Original Post

Jul 21, 2003 at 10:58 AM

I was just timing CPUID (because i need a serializing instruction for a timing class) and it came out to 548 cycles! Can this possibly be right? The following was my code to time CPUID:

__asm {
	CPUID 
	CPUID
	CPUID  // Call CPUID three times, since the first couple of times can take longer

	RDTSC
	PUSH EAX
	CPUID
	RDTSC
	POP EBX
	SUB EAX, EBX
}

The number of cycles should be in EAX now.

drslush

Jul 21, 2003 at 11:01 AM

Correction, I think its 460 cycles...

darookie

Jul 21, 2003 at 11:04 AM

No matter if it really takes so many cycles or not -
it simply doesn't matter as you won't call CPUID
inside your main loop.

[edited by - darookie on July 21, 2003 12:04:48 PM]

drslush

Jul 21, 2003 at 1:47 PM

I was just wondering if it was accurate. And yes, you might want to call CPUID inside your main loop, if you need a serializing instruction. For timing routines, this is quite important.

Sneftel

Jul 21, 2003 at 1:58 PM

quote:
Original post by drslush
I was just wondering if it was accurate. And yes, you might want to call CPUID inside your main loop, if you need a serializing instruction. For timing routines, this is quite important.

No, you''d never do this (if you were designing it well). You''d call CPUID once, near the beginning of your program, and store whatever information you needed from it in your own variables. Then, functions which depend on CPU type would use those variables. (You could also use runtime code modification... but that''s a bit more tricky.)

How appropriate. You fight like a cow.

drslush

Jul 21, 2003 at 2:07 PM

Sneftl, you missed my point. I''m using it as a serializing function. Modern Pentium archetecture can execute multiple instructions at the same time, as long as they arent working with the same data. So in other words if you had an fdiv in between two timing calls, you would have to use a serializing instruction (like CPUID) in order for there to be a garuntee that the instructions would be executed in order. Otherwise you cannot accurately time certain operations.

Jan Wassenberg

Jul 21, 2003 at 7:10 PM

Yeah, that''s possible - I believe it is that slow on Intel CPUs. My Athlon dox call it 42 clocks.
That''s probably the best serializing instruction, too - forget iret and rsm; the rest are priviledged.
Do you really need to serialize the counter? I suppose that''d only be necessary when timing long, single instructions (e.g. div), not for a general time reference.

drslush

Jul 21, 2003 at 9:36 PM

I need it to be extremely accurate, but you are right that it probably isn''t really necessary. But I don''t see any reason why not to.

Topic Locked

This topic has been locked by a moderator. New replies are not allowed.