Hacker News Reader

Comment by IshKebab | Hacker News Reader

The problems with all not-SV HDLs are:

1. None of the commercial tools support them. All other HDLs compile to SV (or plain Verilog) and then you're wasting hours and hours debugging generated code. Not fun. Ask me how I know...

2. SV has an absolute mountain of features and other HDLs rarely come close. Especially when it comes to multi-clock designs (which are annoying and awkward but very common), and especially verification.

The only glimpse of hope I see on the horizon is Veryl, which hews close enough to SV that interop is going to be easy and the generated code is going to be very readable. Plus it's made by very experienced people. It's kind of the Typescript of SystemVerilog.

View on HN

What are the benefits of SV for multi-clock design? I found migen (and amaranth) to be much nicer for multi-clock designs, providing a stdlib for CDCs and async FIFOs and keeping track of clock domains seperately from normal signals.

My issue with systemverilog is the multitude of implementation with widely varying degrees of support and little open source. Xsim poorly supports more advanced constructs and crashes with them, leaving you to figure out which part causes issues. Vivado only supports a subset. Toolchains for smaller FPGAs (lattice, chinese, ...) are much worse. The older Modelsim versions I used were also not great. You really have to figure out the basic common subset of all the tools and for synthesis, that basically leaves interfaces and logic . Interfaces are better than verilog, but much worse than equivalents in these neo-HDLs(?).

While tracing back compiled verilog is annoying, you are also only using one implementation of the HDL, without needing to battle multiple buggy, poorly documented implementation. There is only one, usually less buggy, poorly documented implementation.

Looking forward, it seems possible for Amaranth to be a full fledged language unto itself without needing python. One could maybe use python as an embedded macro language still -- which could be very powerful.

One of the reasons amaranth (and other neo-HDLs) is so great is the full-fleged seamless integration with the host language. Generating DSP filters using the numpy for all parameters, creating CRC structures, diffent logic for different word widths, ... .

This is all feasible with SV or an embedded Macro language as well, but you'll either have to live with a poorly documented meta language (as not a whole lot of people are using it) or heavy missmatches between the meta language and the "real" language. Cocotb very much suffers from this for simulation usage.

And, tbh, if it can be nicely implemented in the host language (which IMHO is the case with amaranth, less so with migen), I don't think there are many benefits by being standalone.

I've always respected PonyORM because of their level of hacking of the python language.

    select(p for p in Product if p.name.startswith('A') and p.cost <= 1000)

PonyORM inspects the AST of the generator, and then translates that to SQL.

I feel like maybe something similar could happen to make the syntax more native for Amaranth.

SpinalHDL's multiple clock domain support via lexical scoping is excellent.

Save for things like SV interfaces (which are equivalently implemented in a far better way using Scala's type system), SpinalHDL can emit pretty much any Verilog you can imagine.

Reply on news.ycombinator.com