[Milkymist-devel] Mico32 migration

[sebastien.bourdeauducq at lekernel.net]

Hi,

On the FIR filters that were causing the performance problems, I'd say
Mico32 is roughly twice as fast as AEMB. I guess that's only because Mico32
is single threaded; the multiple way caches of Mico32 do not seem to help a
lot there.

It would be interesting to do precise benchmarks, though.

Sebastien


On Fri, 3 Jul 2009 23:07:54 +0800, Shawn Tan <shawn.tan at aeste.net> wrote:
> Hi Sebastien,
> 
> Could you give me a rough idea of how much of a speed up is observed? Is
> it in 
> the order of 10%/20%/50%/70%?? It would be useful to know this. (=
> 
> Cheers.
> 
> On Friday 03 July 2009 02:45:20 am Sébastien Bourdeauducq wrote:
>> Hi,
>>
>> I just finished migrating and debugging the design with the Mico32 soft
>> processor (instead of AEMB).
>>
>> I did not do a lot of testing yet, but it seems that the software is
>> running much faster now (AEMB was only running at 50% speed because of
> the
>> threading).
>>
>> The only trick was that the processor does not work when configured for
>> using direct mapped caches while it works perfectly when switching to
> 2-way
>> set associative caches - bug in Xst or Mico32?!
>>
>> Also, synthesis with ISE 11 uncovered a problem with the DCMs that
> generate
>> the SDRAM clock - the FACTORY_JF parameter was incorrect and that could
> be
>> the source of the problems seen on some boards.
>>
>> Sébastien
>>
>> _______________________________________________
>> Devel mailing list
>> Devel at lists.milkymist.org
>> http://lists.milkymist.org/listinfo.cgi/devel-milkymist.org