This is important work. And the author is not wrong about the sad state of hardware description through verilog. One of the reasons I chose to use VHDL over verilog was that it was somewhat more robust in terms of complete implementations you could get (especially cheaply)
I find it interesting that ARM is doing this now versus all along. Was V8 the point where their licensees revolted and said "Hey, your spec is full of bugs, we're not going to pay a license fee for this crap." ?
Clash (Haskell) and Chisel (Scala) are vastly superior HDLs. Clash is actually a compiler plugin that compiles a large subset of Haskell to hardware, while chisel is a DSL.
Thanks for the pointers to other HDLs. I was aware of Chisel but had not seen Clash before.
Another Haskell Hardware DSL is Lava: http://ku-fpg.github.io/software/kansas-lava/
Chisel is very similar to Lava in many ways although I think Chisel has been used for larger designs.
Reduceron is written in York Lava and is not a tiny project. Lava predates Chisel, Clash, and MyHDL and in my personal opinion is superiour. There are room for more improvement though and Real Soon Now I'll start work on that.
As you can see, it's just plain old Haskell. Clash is really just a few GHC extensions and plugins. Spits out Verilog, SystemVerilog, or VHDL. You can also compile it to a regular executable, which means you don't have to use shitty HDL debugging software; you just debug the executable like normal (including error messages and stuff).
This is really cool! I have been thinking about the possibility of this for a long time.
How much hardware experience do you have? I'm curious to know how much effort there would be in porting a Haskell library consisting of pure functions, like the Elliptic Curve Cryptography library [1]. ECC is fairly CPU intensive to verify (the best hand-written assembly implementation takes 0.5 ms per signature). This makes it a bottleneck for some operations, like verifying the Bitcoin blockchain, which contains millions of them.
I can't wait for high level libraries to start forming on top of the Clash library. The only thing left is more competition in the FPGA market, I guess, but perhaps this won't happen until non-hardware folks can actually program an FPGA?
Being a pure function isn't sufficient; it also can't use general recursion or dynamically sized data structures. You could certainly write an ECC implementation using Clash, but you might have to do a bit of porting work.
> One of the reasons I chose to use VHDL over verilog was that it was somewhat more robust in terms of complete implementations you could get (especially cheaply)
A) Icarus Verilog is open source and free, and the author works with a couple of people on the Verilog committee last I checked ...
B) What VHDL implementations are free and open source?
I found the free VHDL tools from Altera and Xilinx to be quite good from a spec completeness point of view. At least for their tools what you "paid" for was a better synthesis back end. I built a 7 segment HEX decoder which should have synthesized to 2 LUTs on a Xilinx chip but it kept synthesizing to 14. I finally spec'd it in VHDL as a series of gates (practically RTL) and that worked. The paid version got it into 2 LUTs pretty much no matter how I specified it.
Altera Quartus has supported SV (SystemVerilog) quite well for some time now. I tend to work in 9.1sp2 as that was the last version of Quartus with integrated simulator (waaa!).
Xilinx requires you to use their newer tooling in order to write in SV. As these tools tend to be incredibly bloated, this is a disadvantage IMO (you want to use the earliest and therefore least bloaty tool that does the job).
I started out in VHDL and was forced by the other person on the "team" to learn verilog for a project. As with ANY language, you end up trying a bunch of stuff out to see how things are implemented, what breaks, etc. in order to get your footing. Over time I grew to appreciate the C like nature of verilog - its reduced verboseness over VHDL makes it an easier language to get actual things done in. And it makes switching between C and verilog pretty natural.
I resisted looking into SystemVerilog for forever as I was under the impression that it was more of a systems verification thing, but it is actually just plain old verilog with some incredibly useful extensions for synthesis. The package system is really, really nice. Look online for papers by Stuart Sutherland.
No language is perfect, and there are things I would change in SV (support for negative bit indices, less awkward sub vector select, etc.) but SV comes closest to the ideal HDL IMO. HW design via HDL is fascinating, and a strangely small space for the times we live in.
Using high level languages to do clock-based concurrent stuff is IMO insane as it just adds to the chain breakage, and few will be trained and able to easily use the source. Would you listen to anyone who proposed a VHDL to Haskell converter for writing everyday conventional software? HDLs are close to the metal for a reason.
The motivation is generally that Electronic Design Automation software sucks .. for black hole levels of suck.
EDA software exists in a weird reality of a small number of people who use it and an even smaller number of people who understand, really, how to program it while being excessively valuable to a very captive market base.
A small number of users combined with an even smaller programmer base results in a large number of features per user. Consequently, if you aren't using the absolute, bog-standard central features, you are going to hit bugs. And, if your feature is even slightly esoteric, you may be 1 of 2 users of that feature worldwide.
Add in to this mix that the programmers of EDA software often don't actually know how to use that use that software, and you have an absolute nightmare of misfeatures combined with bad usability with a whole bunch of bugginess on top.
The only way around this is open-source. Sure, your open source version is worse. However, a sufficiently motivated user can pull the code and fix a bug. A sufficiently motivated user can pull the code and add a missing feature. Open source can accumulate the effort of the very small number of individuals who can both code software and design hardware over a very large block of time without interference from market pressures.
Finally, EDA tools have software lifetimes of decades. We are still using SPICE decks that trace to late 70s FORTRAN.
An example of all of this coming together is software I wrote probably 15 years ago. I once wrote an HSPICE reader in Python. No big deal you say? Well, it could read 15 years of HSPICE files--some of which had been transferred from strange things like fixed-record file machines. It was WAY better than anything any HSPICE vendor ever coughed up. I released it open source.
And about 10 years ago, somebody found it, and extended it to even MORE HSPICE formats. It is now even better than any commercial HSPICE reader will likely ever be. However, had I made the code closed source, the person who found it would have created his own code that only dealt with a couple of formats that he was interested in.
So, the fact that I made that code open source means that if I ever need to come back to it, it is now fundamentally better than anything I could have written and is available to me regardless of what company I'm working for.
So, not only does my itch get scratched, it gets scratched better than I could possibly have ever done myself.
I don't have too much knowledge of ARM, but from what I hear the errata are surprisingly few in number (compared to say, an x86 cpu).
Also, that is not to say companies like ARM are only recently verifying their chips. This is only in regards to formal validation, which others have explained the circumstances of pretty well. There are still entire teams dedicated to code and functional validation.
I'm seriously thinking of looking into MyHDL (myhdl.org). It's a way of specifying spec in python syntax and I think it automatically compiles down to VHDL (and probably more). It's especially good for programmers like me who would also be interested in building layers of abstraction to offload some of the cognitive burden.
However, full programming languages really aren't very good for specifying hardware, though. There are simply far too many constructs that don't actually map to to an implementation.
Sure, if you just try to write software and expect it to work as hardware you're going to have a bad time, but that's not really the point of these tools.
The reason to use a programming language for hardware design is that modern programming languages are monumentally better at abstraction than pretty much any available HDL or hardware design tools.
However those abstractions rarely map to hardware well.
For example, I use hash maps and vectors for almost all of the adhoc data structures in my programs. Neither of these have a useful abstraction in hardware.
At that point, why am I using a modern programming language?
Again, those are software concepts. That's not what I'm talking about.
The point is, there are many hardware level constructions (FIFOs, BRAMs, CAMs, etc.) which can be very usefully abstracted.
And you use a software language to do it, simply because the hardware languages are so bad at it. There's no reason a HDL couldn't be good at expressing these abstractions (indeed Bluespec exists and does a fairly good job), but since they don't, using a programming language simply as a metalanguage is still better than actual HDLs.
And the designs that are feasibly synthesizable usually have to be tweaked until the timing works out, which gets harder the faster your clock speed and the more different clock/power domains you have.
There's a significant risk that any high level abstraction would be leaky enough that you might have to rip up large parts of the design and start from scratch with Verilog/HDL anyway.
> It's interesting (and no coincidence?) that the one caveat with the OK Labs (now General Dynamics) verified micro-kernel was "assuming the hardware satisfies its specification".
I find it interesting that ARM is doing this now versus all along. Was V8 the point where their licensees revolted and said "Hey, your spec is full of bugs, we're not going to pay a license fee for this crap." ?