-
Notifications
You must be signed in to change notification settings - Fork 217
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[WIP] Alpaka develop (pre-0.4.0) #2807
[WIP] Alpaka develop (pre-0.4.0) #2807
Conversation
git-subtree-dir: thirdParty/alpaka git-subtree-split: b8306fbaecb4ccddd62a8abd93c33f5817012cae
@sbastrakov uh dang, I just saw some updates need to go also into https://github.com/ComputationalRadiationPhysics/cupla |
It's actually almost exclusively cupla changes, with only a minor change to pmacc. Sorry, should have pointed that out myself. So what is the course of action then, after I try other examples and figure out the failing tests? |
Just push all here first until everything works. Then take the cupla part out and do a cupla PR against cupla dev.
When that is merged, we also update cupla to dev.
|
I guess I'm blind, but from the failing test log I actually don't see what exactly goes wrong: for all the failing examples after 45% there are some warnings and then the log stops. What am I missing here? On hypnos all standard examples compile fine. |
Probably the error message is cropped.
Can you please run:
pic-compile -q -j 16 -l $PIC_EXAMPLES
On an interactive K80 job on hypnos and run it manually?
|
Now the situation looks weird to me. I've tried your command (with the destination directory added to the end) on hypnos and everything but both configurations of FoilLCT builds there (while on the testing system all examples seem to fail). However, the error message looks cropped (strangely reminiscent of the testing system) and so not really informative to both me and @psychocoderHPC . He suggested the error might be somehow related to overly long names. He also suggested I might try to remove some particle species from this example and see if it starts compiling. The same example however builds fine on Hemera. |
Offline discussed: @sbastrakov will also try to compile it on Hemera with CUDA 9.2 |
@psychocoderHPC I already did and it compiles fine. Now trying your another idea to remove some of existing particle species. |
Just as a note, these are the software dependencies loaded on the compile suite:
(Intentionally using the oldest supported versions of things.) |
@ax3l thanks for info. Not related to the bug in question, but for future use: is there any easy way to see that versions from the logs alone? If not, maybe worth adding. |
Currently documented here and here but usually visible in the CMake output, which I crop intentionally away. |
Since I'm again stuck here is a summary of the current status with changes of this PR:
The problem in FoilLCT on hypnos k80 is cased by Thomas-Fermi ionization, build log. Modifying the example to not use this ionizer makes it build. Removing some particle species while keeping Thomas-Fermi does not help. @psychocoderHPC suggested it might be some kind of cuda 8 bug (offline discussion). However, to me it does not really explain the difference between the testing system and hypnos k80. |
A little follow-up: with hypnos laser profile FoilLCT also builds fine. |
So many new warnings in Boost/Alpaka... I increased the limit of reported lines now. (I crop them in the compile suite reports since the data base storing them gets huge over time.) Just push here again to trigger a new build and let's hope another 90 lines of warnings are enough to see the first error. |
@ax3l I've pushed yesterday evening and the testing is still pending. Wondering if everything is alright with the testing system. |
Maybe of the involved hooks failed or one of the participating services was briefly offline. Pls push again. |
The error seems to be too large to be reported to the proxy. Here it is, carved out manually:
Did you try building with CUDA 8 on Hypnos? |
cc @BenjaminW3 just in case you haven't seen: we are struggling a little as it looks with CUDA/NVCC 8. @sbastrakov is currently on it. |
@sbastrakov please leave this CUDA 8.0 issue aside so far, @psychocoderHPC wants to take care of this later on and it's not super urgent. We might switch away for CUDA/NVCC 9.0+ (C++14) beginning next year. Please open your PR with the new boundary conditions, that is of high priority. |
#2958 is the updated version of this PR. |
As discussed with @ax3l , replaced alpaka 0.3.4 with the current alpaka develop branch and made the necessary changes in the picongpu code.
Alpaka changed naming of streams to queues. I've changed variable and type names accordingly, but for now kept names in cupla/types.hpp like
AccHostStream
, as they seemed more global to me.So far only tested that the standard LWFA compiles on CPU and GPU and runs on GPU, so WIP.
Alpaka Pre-0.4.0 Version
Update Alpaka from
0.3.4
)to
Used Update Command