As the title says: I use the eclipse plugins that ship with CUDA 11. I have downloaded the latest Eclipse, but I get no syntax highlighting and no code completion. Building/compiling works, and it throws errors when I make some dummy mistakes, so the build process seems to be fine. But no code assist / code completion, and no syntax highlighting at all. I tested it by configuring all matching braces in green and bold, the editor doesn’t do that. Nothing, just the black on white text. Very frustrating.
Update: I got it to work when using a new, fresh, separate Eclipse installation – one that is “pure” and doesn’t have the Scala plugins. It seems best to have one Eclipse for Java, one for Scala, and one for C/CUDA. They can all still use the same workspace, although using separate workspaces then is the next logical step. Still don’t have yellow background for device code, but at least I have (most of) syntax highlighting and command completion / code assist back.
Are you having the indexer problem that seems to plague the C/C++ IDE? It’s to the point where I have to turn off the indexer occasionally as well as exit and start back up otherwise it just hangs…
I never had Eclipse as a C/C++ IDE, i. e. without the additional stuff from the CUDA plugin. And I don’t know what the indexer problem is. Sorry I can’t be helpful.
That’s what I’m talking about. “Eclipse with the C/CUDA stuff” is the C/C++ IDE with CUDA plugins.
a) it has a memory leak, the heap will eventually blow up and
b) the indexer eventually stops working if you have it enabled, i.e., it will attempt to index your project but sit at 0% for eternity, ultimately running multiple instances that take all of your CPU cores down.
Version 11.1 of the CUDA-Toolkit and the latest Eclipse CDT, btw.
I was on Eclipse 2020-06 when I started this thread, I’m now on 2020-09, and it’s the same problem. The interesting thing is that the kernel call line compute<<< >>>(…) is in yellow background, but the kernel code itself should also be in yellow background, which it isn’t.
And upon launch I also get a Load QML Analyzer error. The details then say “An internal error occurred during: “Load QML Analyzer”. Cannot invoke “javax.script.ScriptEngine.getContext()” because “this.engine” is null”
I’m still on CUDA 11.0, not 11.1, but I don’t think that should matter for the IDE.
The errors we’re talking about here sound so strange, and I should add that I had a lot of Eclipse errors with previous CUDA version before (I believe going back to CUDA 6 when I first started using CUDA), it’s amazing that NVidia cannot fix those things and make it reliable once and for all. And this although I heard recently in an NVidia webinar that the CEO puts special emphasis on having everything work well on all supported platforms and on new releases quickly. My guess is that he’s simply not aware of the fact that on Linux all that Eclipse stuff never works well out of the box. You always need to ask for help on this forum or on #fedora or #cuda or #nvidia on the IRC. Nothing works immediately.
How can there be indexer or memory problems in a plugin provided by such a top player company like NVidia? Sounds pretty slapdash to me. Which is why I think we need to report all those things, so NVidia starts realizing that bad press is also a negative. The company has SO MUCH money, and they can’t afford a Linux deployment specialist? It’s the o/s that runs in all the data centers, and they don’t use a Linux deployment specialist … I think just one full-time employee would be sufficient. But noooooooooooo … keep it slapdash.
I think the indexer/memory leak issue is built-in with Eclipse. I don’t know. It doesn’t do it with the older Nsight (10.1 and earlier).
Anyway, here’s my setup:
Ubuntu 20.04
Installed CUDA 10.1 toolkit update 1 (may be 2, I don’t recall)
Installed CUDA 11.1 toolkit (not an override)
Installed CUDA 11.1 Nsight EE plugins into Nsight 10.1
changed symbolic links for builders within Nsight to reference 11.1 toolchain
It works - mostly well, but not completely. I have the highlighting (no yellow) issue in mixed C++/CUDA files, but not in “pure” files (I have header files that are entirely device functions). Using C++14, though I have some of the strange “method could not be resolved” errors that I turn into infos because they’re not errors to the builder.
You might be right about reporting everything. Maybe they’ll give us back Nsight? ;)