[ydl-gen] Programming for speed on the PS3

[Warren Nagourney]

I was very surprised that the spu was so much faster  than the ppu  
since the spu has all of the added code needed to move the data to  
and from the preferred slot.

There is some double precision in my program, but it is mostly  
single. I did this to facilitate conversion to altivec which doesn't  
support doubles (this was 3 years ago). One hopes that the next  
generation of the cell will support doubles better (I assume that the  
shrink to 65 nm will not be different in this respect).

-wn

On Mar 8, 2007, at 9:35 AM, Jonathan Bartlett wrote:

> Warren Nagourney wrote:
>> I found that the xlc compiler produces code that runs about 15%-20%
>> faster than gcc. This is on a program with a moderate amount of
>> single precision arithmetic. The interesting thing is that the same
>> (scalar) code runs about 30% faster on the spu as it does on the ppu
>> (with the same 15% advantage to xlc). Provided the code and data fit
>> in the local store of the spu, it seems better to use the spu  for
>> scalar code! Compared to a G4, the ppu is more or less equivalent to
>> a 1.2 GHz processor and the spu is equivalent to a 1.7-2 GHz G4.
>>
>>
> I'm not surprised that it is this fast with single-precision  
> arithmetic, even as scalar code.  Do you have any data for double- 
> precision?
>
> For single-precision arithmetic, the SPU can process a new value  
> each clock tick (it's a 6-cycle calculation, but fully-pipeline- 
> able).  For double-precision, not only does it take 13 cycles for  
> the calculation, but it completely stalls the processor for the  
> first 6 cycles!  It's almost as bad as a branch!
>
> Jon
> -- 
> <jon_logo.jpg>
> <jon_logo.jpg>
> _______________________________________________
> yellowdog-general mailing list
> yellowdog-general at lists.terrasoftsolutions.com
> http://lists.terrasoftsolutions.com/mailman/listinfo/yellowdog-general
> HINT: to Google archives, try  '<keywords>  
> site:terrasoftsolutions.com'