On Fri, Oct 22, 2010 at 11:22 AM, Glenn Ko <[email protected]> wrote:
>
>>Good question... I don't know the answer off the top of my head, but
>>you should be able to do it without adding more .cc and .hh files, but
>>just by adding stages to the time buffer so that it takes the buffered
>>values additional cycles to get from one stage to the next.
>>
>>At least that's the theory of how it's supposed to work...
>>
>>Steve
>
>
> What do you mean by adding stages to the time buffer?
> Do you mean increasing the delay value for the time buffer between two
> stages?
Ummm... OK :-). I really don't have any firsthand experience with the
O3 code, I just was there when it was architected. So as I said, I
know the theory about how it's supposed to work, but not much more.
Maybe Nate or Kevin can speak up on this. Also if you haven't looked
at the ASPLOS tutorial slides or our MICRO06 paper I think there is
some discussion in both of those places.
> I have several questions regarding these values in O3CPU.py.
> 1. What is the difference between {stage}To{nextstage}Delay (e.g.
> fetchToDecodeDelay) variables and {back/forward}ComSize variables?
> The default value for the fetchToDecodeDelay values are 1.
> I thought it is to implement 1 cycle delay between the stages.
> In that case, what does the default value 5 of the back/forwardComSize
> variables mean?
> I see that timeBuffer and *Queues are instantiated with
> back/forwardComSize = 5, but not sure what that means.
Me neither...
> 2. I'm also not entirely sure how to distinguish between the skidbuffer and
> time buffers.
Skid buffers are microarchitectural structures that store instructions
that get partway through the pipeline but then have to stall. It's
particularly useful for SMT since it lets instructions from
non-stalled threads pass those from stalled threads.
Time buffers are purely simulator artifacts to model signals passing
between pipestages that are potentially multiple cycles apart.
> 3. Also, why is the rename stage's skidbuffer size different from one for the
> decode stage?
>
> decode_impl.hh: 58
> skidBufferMax = (fetchToDecodeDelay * params->fetchWidth) +
> decodeWidth;
> rename_imple.hh: 76
> skidBufferMax = (2 * (iewToRenameDelay * params->decodeWidth)) +
> renameWidth;
Can't say off the top of my head...
Steve
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
