Comment by metaphor | Hacker News Reader

Comment by metaphor | Hacker News Reader

Parent Full thread

> There is no way FPGAs will become popular among software developers unless the tooling improves dramatically.

Care to expound in what way? I don't disagree with your opinion completely, but I suspect our reasons differ.

IMHO, the barrier isn't so much the tools, but a clashing of paradigms. All too often, traditional software developers without a solid foundational background in digital hardware architectures think they can easily pick up a HDL as if it were yet another programming language. The HD part is quickly forgotten as behavioral processes are hacked away using all too familiar constructs like if-then-else, for/while loops, etc. Inference warnings are ignored, timing constraints are shrugged off, metastability and synchronization aren't even a thing, let alone driver/receiver selection, pinmap planning, and signal integrity considerations at the PCB level...set to default and things should just work, right?

Anecdotally speaking, the only thing in common that HDLs have with traditional programming languages is a superficial charset. I think every software developer who has ever used tools like VS, Eclipse, Emacs, Vim, Make, Doxygen, Git, etc. would agree that each has its quirks and it takes a bit of use before a comfortable flow is discovered. FPGA vendor tools are no different, except--testbed simulations aside--there's that immutably distinct custom hardware integration end-game that most traditional software has the pleasure of conveniently abstracting away.

Any kernel/driver developers with HDL proficiency care to chime in?

P.S. ground-up PCIe/Ethernet FSM+datapath architectures after a few months without prior HDL experience is really impressive.

I went to CMU for a BS in EE, I used a lot of verilog, and also a bit of Cadence for layout in one class. But I've worked as a software engineer since then.

I totally agree that HDL tools are astonishingly crap. The market is just so much more specialized, expensive, locked down and closed.

Imagine that you had to use the Intel compiler to get source code compiled for intel processors. You had to use their special headers, you practically had to use their huge 8GB Intel IDE with a million lists and buttons, you had to choose what features were on your intel processor (VT-X? SSE4?). Imagine AMD is the same, but separate. This is all closed source stuff of course. And you have to use it to run on any smartphone/laptop/desktop/server class processor.

Imagine that there's no GCC, no CLANG, no GDB, no Jetbrains, no Eclipse. All the related tools we use wouldn't have much of a reason to exist because just about everyone had to use the huge IDEs for each vendor anyway.

The vendor stuff is huge and inefficient and crap, because features sell, and crap doesn't not sell because there's no meaningful competition.

To be more specific, "crap" means high bloat and low reliability. Crashes, inexplicable errors and failures. Huge amounts of bloat. But you can make a tweak and try again move on, so any sane hardware person just gets used to it, it's not like they can do anything about it.

Apologies but I really can't tell if your point is against the proprietary nature of reconfigurable devices, or vendor tool bloat.

> Imagine that there's no GCC, no CLANG, no GDB, no Jetbrains, no Eclipse.

This is quite easy to imagine when you remove the key element which allows these L7 abstractions to be meaningful: an underlying kernel with a well defined interface. What equivalence does the reconfigurable world have when every target device requires its own unique "kernel", if you will? I can't think of any...which would explain why 3rd-party tools are constrained to synthesis, while place-and-route is an explicit function of the vendor tool. Perhaps we too easily conflate the size of the tool with the size of the target?

On the flip side, supposing there were some open reconfigurable interface standard, I don't think this would fly in the current market given the high-performance nature of these devices. Top two FPGA vendors apparently change slice/CLB structure every generation, let alone agree on an open fabric interface standard.

> Crashes, inexplicable errors and failures. Huge amounts of bloat.

Putting the whole kit and kaboodle aside and focusing on just synthesis, isn't it strange that even the big, specialized 3rd party vendor tools (e.g. Cadenace, MG, Synopsis) suffer just as much? I think it's a genuinely difficult problem given the multi-disciplinary nature of the things EDA engineers have to deal with. As much as I dislike dealing with flake tools, I'm nevertheless humbled by their challenges.

You've pretty much hit the nail on the head. What people outside of the EDA field fail to understand is they are coding at a much higher level. RTL does not model all the aspects of the logic. You have different signal delays due to lots of different reasons: fanout, wire lengths etc. As you go through different stages of synthesis via these tools, a lot of decisions are made governed by the constraints files. Each decision affects the performance of the design. It's slow because it's doing a lot of work to optimize speed or area. Some designs won't fit the constraints. Then it reports errors.

The tools are trying to help. The end product is the physical device. The various models are all just abstractions of the physical device. The tools are reporting the problems on the abstractions to assist you to improve the physical device. If you can understand the reports, you can improve things, either altering the RTL or adding more constraints.

The point being that the only time RTL is actually run like a software program is during simulation. This simulation is only an approximation of how the actual thing will work. It is not like SW. The tools do a lot of other things with that RTL. Maybe if people don't throw garbage in, it wont crap itself trying to figure it all out.

> Imagine that there's no GCC, no CLANG, no GDB, no Jetbrains, no Eclipse. All the related tools we use wouldn't have much of a reason to exist because just about everyone had to use the huge IDEs for each vendor anyway.

I don't need to imagine, none of them existed when I started programming.

> The vendor stuff is huge and inefficient and crap

Thankfully, we now have a pretty decent open source alternative:

http://www.clifford.at/icestorm

and you are welcome to contribute!

Interesting project, but it hardly qualifies as "a pretty decent open source alternative" considering it targets a single family of FPGAs from a vendor with comparatively insignificant market share.

P.S. The only dabblings I've had with Lattice devices was designing a custom tool to configure legacy (pre-JTAG) ispLSI family of CPLDs using our tech, not theirs. Goal was to eliminate their obsolete tool chain (ispLever Classic and piece of shit USB dongle) from the configuration management loop for existing stable applications in longterm lifecycle sustain mode.

> Crashes, inexplicable errors and failures.

I bought a xilinx spartan dev board 10 years ago and it was just like this. ISE was 4gb of java bloatware, 2 gb (!) more of "updates", crashed frequently, barely ran, imploded under its own weight. I managed to make some lights go blinky and said "eff this".

It's sad to hear that 10 years later nothing has changed.

When I started the learning process I did expect to have a hard time. I expected VHDL to be very different from software. I expected all kinds of problems related to the "clashing of paradigms". I have encountered those problems. But at the same time I've been fighting with complex config files, unhelpful warnings, horribly long debugging sessions... I feel like the tools could be better. They could have helped me instead of being in my way. I now know how to do all of it, but it took me a LOT of time to learn.

And yes, maybe my goals were too ambitious. Ethernet and PCIe have proven to be hard to master.

I always emphasize that it is hardware design, not hardware programming.

You describe a physical and electrical reality at register transfer level when writing implementation code in Verilog and VHDL to be used in FPGAs and ASICs.

I've seen quite a few implementations (esp in VHDL) by people not understanding that what they write will end up as hardware.

As an example, ff your design contains numerous wide variables written to and read by several instances, that means there will be buses going back and forth between your logic gates. There are only a few levels of wires crossing eachother (that can either be built in an ASIC or are available in a FPGA device). When that limit has been reached, you will have to route around.

Think HW (at fairly ideal RTL) and write code to efficiently describe that in a way that _all_ tools in your chain can understand and parse correctly.

I have only briefly touched FPGAs in college, but to me writing Verilog/VHDL is akin to writing C for baremetal MCU. In both cases it is not too difficult to write correct code that simply does not work, or more appropriately makes no sense, on given hardware.

As for the tooling, to me it looked too low level with too many knobs for a beginner without significant domain knowledge to grasp. Maybe that or [Footnote] was what intimidated GP. I understand the reasons why high level building blocks in high level programming are easy to use self contained modules while high level modules in HDL have that many twists and knobs to twiddle with.

I guess software developers are so much used to taking prebuilt modules and something-something-fudge-fudge-rinse-repeat-until-works'ing them together (not meant as an insult in any way; I am myself guilty of that), that we sometimes forget that `x if y else z` actually moves the magic smoke and pixies in ICs. So I think I can sum up your comment on traditional developer sloppiness and GP's comment on atrociousness of tooling in one sentence: too often we forget that HDL programs must be formally correct.

[Footnote]: I remember some fragments of my fights with FPGA when things worked intermittently because my signals were not stable (race conditions, dear past me) or whatever have I written did not even make sense on given hardware (probably no hardware would support such weird trigger conditions, but there I was staring at the screen with blank eyes and only beginning to seriously consider actually thinking), yet the tooling happily synthesised my code.

> IMHO, the barrier isn't so much the tools, but a clashing of paradigms.

I guess another issue is that most EE guys are happily using GUI tools and don't have any ideological problem with commercial tooling, as what matters is the final result.

Agreed on the closing note, but I'm not certain that's the case. I mean a good tool is a good tool, regardless of foundational discipline.

It comes down to choice...or rather lack thereof. EDA tools are very expensive and tend to have a steep learning curve. Professional divisions tend to standardize to keep costs in check. Here's an example of one that's "affordable"[1].

[1] http://www.digikey.com/products/en/development-boards-kits-p...

I think you can treat 60 - 70% of the FPGA design flow as open source. For example, I am developing a system using PCIe, 10G Base T and some logic to send and receive network packets and to design the HDL and test it, I am using two open source tools predominately (icarus verilog and cocotb). I just use the FPGA P&R tools for building the design once I am satisfied it works. You can also run these tools on the command line quite easily and automate most of the process (They all use tcl for scripting up the flow). Sure theres a few FPGa specific interfaces you have to deal with (transceivers, DDR4, pcie hard ip) but you can pretty much traet these as black boxes and write your tests to target the interfaces in and out of the logic. Also, for things like transceivers, the interface is really not that different between Xilinx and Altera (I treat them as a black box that generates 32-bits every 322MHz cycle for 10G-Base T). The flow to my mind is not that disimilar to a traditional software development flow. I have simulation tests and test cases, I use continous integration to run tests everytime something is commited, everytime I build the FPGA with the P&R tools, I kick off hardware tests automatically, etc

Reply on news.ycombinator.com