Re: Effectiveness of SSE on Athlon XP with gcc 3.2

["Andres Chavarria" <el_andrecillo at hotmail dot com>]

Well after some thinking and analizing the assembler code I got the answer 
to my question. The bottleneck is given by the throughout of memory reading. 
If I use a smaller array the data is read a lot faster and I got at least a 
effectiveness of 3. Pentium 4 has probably a bigger cache so it could handle 
the big amouunt better...

Andres






> From: "Andres Chavarria" <el_andrecillo@hotmail.com>
> To: gcc-help@gcc.gnu.org
> Subject: Effectiveness of SSE on Athlon XP with gcc 3.2
> Date: Thu, 28 Nov 2002 15:45:49 +0000
>
> Hi,
>
> I was wondering how effective the parallelization using SSE is, and I wrote 
> a test program that sums the values of a vector up. The SSE free code for 
> looks like this:
>
> // non SSE, numbers is an array of size cNumber
>  float sum1,sum2,sum3,sum4;
>  float val;
>
>  sum1=0.0f; sum2=0.0f;sum3=0.0f; sum4=0.0f;
>  for(i=0;i<cNumber;i+=4){
>    sum1+=*(numbers++);
>    sum2+=*(numbers++);
>    sum3+=*(numbers++);
>    sum4+=*(numbers++);
>  }
>  val=(sum1+sum2)+(sum1+sum2);
>
>
> And the one using SSE looks like this:
>
> // with SSE, numbers is an array of size cNumber used 
> __attrib__((aligned(16))) to aligne it
>  register v4sf val1,val2;
>  float __attrib__((aligned(16))) momVal[4];
>  float val;
>
>  val1=__builtin_ia32_loadups(numbers);
>
>  for(i=4;i<cNumber;i+=4){
>    val2=__builtin_ia32_loadaps(numbers+i);
>    val1=__builtin_ia32_addps(val2,val1);
>  }
>
>  __builtin_ia32_storeups(momVal,val1);
>
>  val=(momVal[0]+momVal[1])+(momVal[2]+momVal[3]);
>
> I used gcc v. 3.2 on a linux machine with kernel 2.4.19 (distribution SusE 
> 8.1) an compiled the following way:
>
> gcc -msse -march=athlon-xp -o ssetest main.cpp -lstdc++
>
>
> If I compare the calculation time of the first code with the second one I 
> get on an Athlon XP 2200+ (actually 1800 Mhz) only a speedup of about 
> (array size 307200):
>
> calc. time without SSE: 2.72 sec
> calc. time with SSE:	2.05 sec
> speedup factor 2.72/2.05 = 1.33
>
> Theoretically the speedup factor should be somewhat smaller than 4. So I'm 
> quiet dissapointed... The first time I compiled and ran a similar program I 
> did it on a Pentium 4 mobility (1800 Mhz) processor (options -msse 
> -march=pentium4). I got about a speedup factor of 3.9, which was superb.
>
> I don't think it is because of the processor (Athlon XP wouldn't have a 
> chance against Pentium 4, and a report of such a difference should be 
> present somewhere in the net...) I suspect the SSE code I wrote is far from 
> good or that gcc does something wrong for the Athlon XP.
>
> I don't have access to the Pentium 4 machine so I cant compare the above 
> code directly anymore. It would be great I someone could test it on a 
> Pentium 4 or if someone has a hint what is going on.
>
> Thanx in advance,
>
> Andres
>
>
>
> _________________________________________________________________
> MSN Fotos: la forma más fácil de compartir e imprimir fotos. 
> http://photos.msn.es/support/worldwide.aspx
_________________________________________________________________
MSN. Más Útil Cada Día http://www.msn.es/intmap/