Re: [f-cpu] vhdl2c 0.1

[Yann Guidon <whygee@f-cpu.org>]

hi !

Michael Riepe wrote:
> 
> On Sun, Feb 10, 2002 at 03:16:57AM +0100, Yann Guidon wrote:
> 
> > a few days ago, Cedric spoke about converting VHDL to C for
> > simulating the F-CPU faster than with a normal simulator,
> > though i do not agree with this strategy.
> 
> Some simulators do that internally. Others use a virtual machine (like
> VHDL Simili). I can't tell which one is faster; that depends on the
> implementation. If the VM is well-designed, it may actually be faster
> than pure C. And it will probably need less memory, which is the more
> important factor in our case.

Add to that : "real" VHDL software come with "optimised" libraries !
it means that there are hand-optimised versions of std_ulogic_vector
functions, which even a compiler can't compete with. Simili does that,
which alone justifies the 30% performance difference with Vanilla
on running your adder test suite :-)

ncsim goes much further in the optimisation frenzy, according to the
docs it compiles down to strips of machine code. However the CPU's instruction
cache might well appreciate less...


Back to the original reason of my post : Cedric claims that he can do
full-system simulations in his school, given the hundreds of 1GHz Athlon on
a 100mbps network. I however claim that
 - the network is the bottleneck
 - VHDL simulations are not easily parallelisable.
A globaly adressable, distributed memory multiprocessor might give better
results thanks to faster synchronisation, better internal bandwidth and
smaller OS overhead.

But franckly, if we do our job correctly, we wouln't even need that ...
a good management of the events can half the simulator's activity
(and even the power consumption due to the decrease of activity).
A good example is the multiplier : the input data latch should memorise
operands only when a multiply operation is wanted. This prevents the
whole multiply pipeline to oscillate with the Xbar's value. This thus
avoids the simulator's event queue to fill uselessly.

And the day we will need real speed, we'll ask Meta Systems or Quickturn
to use one of their machines...

>  Michael "Tired" Riepe <Michael.Riepe@stud.uni-hannover.de>
WHYGEE
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
*************************************************************
To unsubscribe, send an e-mail to majordomo@seul.org with
unsubscribe f-cpu       in the body. http://f-cpu.seul.org/