loop speed question

[a_h]

Hi!

Which loop implementation do you consider to be faster on the P4 (have no time to try it myself)

Code Select Expand


.xxx
...
sub eax, 1
jnz .xxx


or (I would prefer that version, counter in ecx instead of eax)

Code Select Expand


repeat
...
.untilcxz


Thanks for your input!

Have a nice evening, Hannes

[roticv]

Does .untilcxz uses jecxz? If it does, better to use the former.

[MichaelW]

.UNTILCXZ uses the LOOP instruction. On a P3 SUB/JNZ is almost 3x faster than .REPEAT/UNTILCXZ.

[Snouphruh]

I recommend you to use DEC and JNZ instructions.
example:

Code Select Expand


                    mov ecx, <nCount>
myLoop:
                    ...
                    dec ecx
                    jnz myLoop


try to use no this:

Code Select Expand


                    mov ecx, <nCount>
myLoop:
                    ...
                    loop myLoop

'cause LOOP instruction takes more CPU time than DEC + JNZ. DEC + JNZ are paired, e. g. these 2 instructions take 1 CPU clock (LOOP takes 3 ones)

[a_h]

Thanks for your replies! :U

Actually I thought the masm macro repeat/unitlcxz would expand to dec/jnz, but I'm wrong!

@Snouphruh: according to Intel's manuals and Agner Fogs'pdf inc/dec is not optimal for the P4, that's the reason I substitute it with sub/add (as suggested by these docs). When I have the time, I will test wether this replacement is really any good for performance!

Cheers, Hannes

[Snouphruh]

ok.
when you got the results write back to me, please. ok?

ps: I have AMD Athlon XP 2500+ @ 2244 MHz.

[roticv]

loop is a complex opcode. Intel optimisation manual for P4 recommend people not to use it.

[Eóin]

And slightly off the topic, but don't forget the neat little trick for moving up through an arry from offset 0 to the end eg;

Code Select Expand

mov ecx,-256*4
lp: mov eax,[Array+256*4+ecx] ; or something

add ecx,4
js lp

[a_h]

Ok here the results:

Using a small loop (stdcall) without memory accesses like

Code Select Expand


looptest0 proc
mov ecx, 10000
L1:
mov ebx, 1
add ebx, ecx
sub ecx,1
jnz L1
  ret
looptest0 endp


yields following results on a P4:

using jnz/sub (code above): ~30 000 cycles
using loop: 34 400 cycles
using jnz/dec: 42 250 cycles
using repeat: 43 800 cycles

all numbers are averaged from 4 runs (done 3times to check).

Thanks for reading, Hannes