Author: Gerd Isenberg
Date: 12:35:34 07/06/03
Go up one level in this thread
On July 06, 2003 at 14:27:51, Dieter Buerssner wrote:
>On July 06, 2003 at 13:29:56, Gerd Isenberg wrote:
>
>>What about unrolling the loop a bit, eg. repeat the body statement 2..10 times.
>
>I unrolled by 4; I changed DECLARE_TEST_FUNCTION to
>
>#define DECLARE_TEST_FUNC(name) \
>unsigned long tfunc_##name(void) \
>{ \
> int a; \
> unsigned long n=N_ITERATIONS/4; \
> unsigned long sum=0; \
> do \
> { \
> a = RAND_VAL(); \
> sum += name(a); \
> a = RAND_VAL(); \
> sum += name(a); \
> a = RAND_VAL(); \
> sum += name(a); \
> a = RAND_VAL(); \
> sum += name(a); \
> } while (--n != 0); \
> n = N_ITERATIONS%4; \
> if (n) \
> do \
> { \
> a = RAND_VAL(); \
> sum += name(a); \
> } while (--n != 0); \
> return sum; \
>}
>
>MSVC -Ox2 -Ob2 -G6 -Gr -GF, library rand()
> nothing 3951541892 12.578
> abs 1713113360 13.699
> simple_abs 1713113360 17.395
> omid_abs 1713113360 13.900
> sbb_abs 1713113360 18.116
> cdq_abs 1713113360 20.009
> fish_abs 1713113360 19.969
> sar_abs 1713113360 18.186
> cmovl_abs 1713113360 17.165
> cmovs_abs 1713113360 18.436
>
>Omid abs is back to normal speed. Unrolling has little effect otherwise. I think
>this can be expected, when assuming that rand() is the slowest part of
>everything, and that the branches of the long loop are always get predicted
>correctly. The strangeness remains. Without unrolling I can reproduce the 7.x
>seconds always.
>
>MSVC -Ox2 -Ob2 -G6 -Gr -GF, sr32_rand:
> nothing 1305123480 5.758
> abs 2955546426 6.159
> simple_abs 2955546426 12.918
> omid_abs 2955546426 6.119
> sbb_abs 2955546426 15.713
> cdq_abs 2955546426 17.605
> fish_abs 2955546426 20.600
> sar_abs 2955546426 18.186
> cmovl_abs 2955546426 17.635
> cmovs_abs 2955546426 17.826
>
>Gcc -O3, sr32_rand:
>
> nothing 1305123480 5.714
> abs 2955546426 12.747
> simple_abs 2955546426 12.747
> omid_abs 2955546426 5.879
> sbb_abs 2955546426 6.593
> cdq_abs 2955546426 6.099
> fish_abs 2955546426 12.747
> sar_abs 2955546426 6.154
> cmovl_abs 2955546426 5.769
> cmovs_abs 2955546426 5.714
>
>Gcc -O3 -march=pentium4, sr32_rand:
>
> nothing 1305123480 5.714
> abs 2955546426 5.714
> simple_abs 2955546426 5.714
> omid_abs 2955546426 6.044
> sbb_abs 2955546426 6.538
> cdq_abs 2955546426 6.044
> fish_abs 2955546426 12.802
> sar_abs 2955546426 6.154
> cmovl_abs 2955546426 5.769
> cmovs_abs 2955546426 5.769
>
>Library abs() and simple_abs() are fast now, because conditional movs are used:
>
>In ATT-Synthax:
> movl %eax, %ecx
> negl %ecx
> cmpl $-1, %eax
> cmovle %ecx, %eax
>
>>Doubling the speed of a function by adding additional abs code - not bad ;-)
>
>Indeed. Or is it the other way around. For some strange reason, rand() does not
>run at normal speed here? I cannot get a loop with only rand() (no sum, etc.) as
>fast as together with omid_abs.
He, He, may you should try to speedup Yace a bit...
But i guess for some strange reasons this speedup only occurs in this loop.
May be due to some pipelength or microcode alignment reason some internal
hyperthreading like unrolling occurs. Using all pipes perfectly, processing two
loop bodies simultaniously with "different" register sets?
Regards,
Gerd
>
>Regards,
>Dieter
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.