[Libre-soc-dev] Simulator speed status

[Luke Kenneth Casson Leighton]

On Tue, May 18, 2021 at 4:05 PM Lauri Kasanen <cand at gmx.com> wrote:
>
> On Tue, 18 May 2021 11:23:22 +0100
> Luke Kenneth Casson Leighton <lkcl at lkcl.net> wrote:
>
> > after investigation, i had in
> > fact updated local nmigen (12th april 2021) and i've just pushed that with
> > a forced master update
>
> I'm now able to run the fp tests successfully.

hurrah.

> 53s for 10 tests / 5.3s per 1-2 instructions is very slow, I
> probably won't even try the other existing tests at this speed. And that
> was with >/dev/null to cut the massive prints out.

rright.  okay.  so this is due to loading up a huge Power ISA decoder,
which takes several seconds to initialise.

in the ISA HDL tests we converted over to a different way of doing
things, by using the *same* PowerDecoder2, and running "subTests".

i just haven't had time to convert the *simulator* unit tests over to
that same infrastructure.

> Hm.. If that's ~3.5s per instruction,

it's not: it's a huge initialisation setup per test, followed by
reaaasonably fast execution.

> I'm wary of volunteering, since the existing simulator is python which
> I don't do. Converting it to C/C++ might take me quite a long time,
> making it unviable.

we do have funding for doing power-gem5 to add SVP64 to it.
power-gem5 64-bit is not yet upstream though
https://gem5-review.googlesource.com/c/public/gem5/+/40896/5

i was wary of suggesting doing patches-on-patches.

> Getting new volunteers doesn't appear easy either.

at the listed rates, even though the NLnet donations are
tax-deductible, it's proven... difficult, yes.

l.