[Libre-soc-dev] ls180 update

Luke Kenneth Casson Leighton lkcl at lkcl.net
Mon Apr 19 01:24:37 BST 2021


On Monday, April 19, 2021, Jean-Philippe Turcotte <jp2k5 at hotmail.com> wrote:

> > yes - the specs for the first phase (yosys turns HDL into gate-level
> > BLIF, coriolis2 turns that back into VHDL, yosys turns that into
> > Verilog, verilator compiles that to c++) [...]
>
> First of all, thank you for your detailed answer. I know you are quite
> busy right now.
>
> It is not clear on the HDL workflow page, but to get up and running I
> would need to run the dev-env-setup script,


>
that gets you the nmigen stuff yes


>

 and then install yosys and coriolis2, correct?


 using coriolis2_chroot script yes


>  Can I use Debian 10 for this?


yes.

btw i can just simply send you the VST (VHDL) files.

oh, you need verilator v4.106 not later than that


>
> > [...] would be that each machine
> > would require at least 128 GB of RAM, and to complete in a reasonable
> > time-frame at least 100 machines would be required.
> >
> > for the second phase (coriolis2 completes place-and-route, coriolis2
> > extracts P&R as VHDL, yosys turns that into Verilog, verilog compiles
> > that to c++) would require machines with 5x those resources and
> > require 10x those numbers of machines.
> >
> > that's 1,000 machines, with 512 GB of RAM in each machine (1 TB of RAM
> > may be needed).
>
> Okay, here is my idea:
>
> First, I need to start with a smaller example to reduce the compilation
> time,


experiments10_verilog, soclayout repository, designed for exactly that.



> and check that the method works. I can use a high-memory cloud VM (I do
> not know what is available to trial users, but 512 GB seems easily
> accessible, potentially up to 5 TB) to generate the C++.


i have reached out to a test company through nlnet, you should not have to
pay.


> Second, it seems I could use gg (https://github.com/stanfordsnr/gg) to
> compile the C++ on tens of thousands of transient VMs to get the executable
> in a human timeframe.


what does it provide that distcc does not?



>
> Third, I suppose we can somehow test that the simulation produces accurate
> results.


indeed.  the executable once compiled needs to be run locally through
cocotb.

thus distcc realistically needs to be used in order to get the object files
back.



>
>
> > i will ask NLnet if they have anything.
> >
> > l.
>
> That would be great!
>
> By the way, sorry for the weird list reply. I was in digest mode, so I did
> not have an email to which I could reply.


doh :)


>
>

-- 
---
crowd-funded eco-conscious hardware: https://www.crowdsupply.com/eoma68


More information about the Libre-soc-dev mailing list