Re: [vhdl-200x-perf] Performance proposals


Subject: Re: [vhdl-200x-perf] Performance proposals
From: Kevin Cameron x3251 (Kevin.Cameron@nsc.com)
Date: Fri May 09 2003 - 16:48:18 PDT


> From: "john ries" <johnr@model.com>
>
> Hi All,
>
> I have posted an initial list of language changes to help performance.
> They are all just rough entries trying to capture the ideas that
> had been presented at the initial kick-off and working group
> meetings in February.
>
> What I need now is a discussion about the proposal and help
> to generate analysis of this proposal.
>
> Also any addition proposals would be warmly welcomed.
>
> Regards,
> John
> --
> -- mailto: johnr@model.com phone: (503)685-0864
> -- http://www.model.com fax: (503)685-0921

Some feedback:

PERF-01:

   Is that a request for Verilog's blocking assign?

PERF-02:

   I think you are just moving the evaluation code around
   in your example it's hard to tell where you would get
   a speed up - an optimizing compiler should give much
   the same result either way. Having said that it looks
   like a reasonable extension since it reduces complexity
   from a user's perspective.

PERF-04:

   Post elaboration optimization should handle that anyway,
   do you really need a new class of signal?

PERF-06:

   While it's good to remove unnecessary stuff, I'm not sure
   how that improves performance much.

PERF-07:

   Is there an actual proposal?

   Personnally I think waiting for a design to be stable is
   a bad idea - all you really know in a synchronous system
   is that it should be stable by the next clock edge.

   If you do PERF-01 then you can subdivide a delta by adding
   implicit signals which indicate that a signal is about to
   be updated or that it has been updated and all dependent
   processes have executed e.g.:

      clock'changing - sample data/assert last cycle done
      clock - change clock value
      clock'changed - assign sampled data etc.
   
   I.e. all the processes sensitive to a 'changing signal are
   processed before those sensitive to a regular signal, and
   those sensitive to 'changed are done last.
   

PERF-08:

   How does that enhance performance?

Regards,
Kev.



This archive was generated by hypermail 2b28 : Fri May 09 2003 - 16:49:11 PDT