by

Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Repo info
Activity
Valentin Haenel
@esc
Meeting notes from yesterday, some interesting points were made about the numba "extensions" package: https://github.com/numba/numba/wiki/Minutes_2020_09_15
Graham Markall
@gmarkall
Morning
stuartarchibald
@stuartarchibald
thanks for uploading @esc
Graham Markall
@gmarkall
I'm just trying to triage numba/llvmlite#631 but don't know much/anything about the readthedocs setup - anyone have any pointers as to how I might reproduce a readthedocs build?
stuartarchibald
@stuartarchibald
that's a bit weird
can't think of anything that'd be the immediate cause of the problem
0.31 was last py27 perhaps ?
Graham Markall
@gmarkall
Ah, thanks for the link... I missed that, was looking elsewhere in their docs about the build process
looks a bit heavy just to debug a build :-/
stuartarchibald
@stuartarchibald
yeah, I couldn't see anywhere which describes how to just replicate the environment without essentially building out hosting, perhaps open an issue on their tracker/ask how to debug it there?
Valentin Haenel
@esc
@gmarkall @stuartarchibald I saw that issue too and suspected it might be case of "your pip is too old"
stuartarchibald
@stuartarchibald
Am wondering if we could put a thing into numba to stop such pip problems
Valentin Haenel
@esc
yeah, I think we were having this conversation
we've had...
an initial thought was to improve the error message
but in this case we don't even see an error message
and the sys.exit mentioned here does seem suspicious: readthedocs/readthedocs.org#7466
stuartarchibald
@stuartarchibald
that sys.exit has been there for years
also, llvmlites own RTD seems ok
Graham Markall
@gmarkall
yeah... althought IIRC the llvmlite build on readthedocs was broken so it got a bit old - maybe around 0.31? I think I mentioned this to @stuartarchibald a couple of months ago
unfortunately my memory is a bit vague / fuzzy
Graham Markall
@gmarkall
eagerly compiling a @cuda.jit decorated function at module scope within the tests (e.g in numba/cuda/tests/cudapy/test_cooperative_groups.py) seems to result in something (possibly the loaded module) getting freed (and therefore damaged) by the time any test using it actually runs - as a workaround I have to make sure I eagerly compile only inside test functions. Is this surprising, or something I should have expected?
stuartarchibald
@stuartarchibald
I'm locally surprised as everything does this, but globally not surprised given Numba has weird problems.
Presume this is PR numba/numba#6245 ?
How does it reproduce? Just run those tests?
Graham Markall
@gmarkall
yeah, but with the exception of the last commit, where I pushed the eager compilation into the test
what do you mean "everything does this"? there is no cuda test with eager compilation at global scope
stuartarchibald
@stuartarchibald
"everything" = "not CUDA target"
Graham Markall
@gmarkall
in numba/tests, grep "^@jit" * -R (and for njit) reveals very few cases where it is done
I feel like it's some oddity in the test harness, because if there were problems with eager compilation at global scope in general in cuda, it would have been noticed by now
stuartarchibald
@stuartarchibald
guess the question is... is it working for other targets by accident or is there something special about the CUDA target?
Graham Markall
@gmarkall
can you show me an example of a test for the CPU target that does this and calls the function? (I'm struggling to find one)
for me, the question is still "is it working for other targets?"
stuartarchibald
@stuartarchibald
there's @vectorize's in ufunc tests
(also, those tests need a CC guard?)
Graham Markall
@gmarkall
ok, but I think @vectorize is sufficiently different that there's not much conclusion that can be drawn from that
stuartarchibald
@stuartarchibald
hmmm, I guess I'm recalling wrongly! seems like there's far less top level forward definition than I thought
egrep -r "^[@]?[n]?jit" numba/tests
Graham Markall
@gmarkall
ah yes, I do need to add a CC guard
stuartarchibald
@stuartarchibald
I can't seem to reproduce on cards sub-Volta (and commenting the CG code)
Graham Markall
@gmarkall
what sub-volta cards are you using?
stuartarchibald
@stuartarchibald
GTX750Ti and a GTX970
Think that's CC 5 and 5.2.
Graham Markall
@gmarkall
Thanks for looking at this so far - as I'm on PTO shortly I will have to save this for another day, so I've just added numba/numba#6255 to track it for now
stuartarchibald
@stuartarchibald
:thumbsup: will try and sort out some volta hardware
GeForce 16 series is lowest consumer grade with it?
Graham Markall
@gmarkall
the 10 series is the lowest consumer grade with CC 6 - the 16 series is actually turing (cc 7.5)
stuartarchibald
@stuartarchibald
oh is it just 6 needed?
Graham Markall
@gmarkall
yeah