Closed rasolca closed 10 months ago
This afternoon I investigated a bit, and this is the report.
Tested https://github.com/eth-cscs/DLA-Future/commit/40e4530991500011301bb420cc9bcfb98cd2d8c9, so before the fix https://github.com/eth-cscs/DLA-Future/pull/1045
gcc | cuda | umpire | note |
---|---|---|---|
gcc@9.3.0 |
cuda@11.2.0 |
umpire@2022.03.1 |
❌ |
gcc@9.3.0 |
cuda@11.4.4 |
umpire@2022.03.1 |
✅ |
gcc@9.3.0 |
cuda@11.3.1 |
umpire@2022.03.1 |
✅ |
gcc@9.3.0 |
cuda@11.3.1 |
umpire@2022.10.0 |
❌ |
gcc | cuda | umpire | note |
---|---|---|---|
gcc@10.5.0 |
cuda@11.4.4 |
umpire@2022.10.0 |
✅ |
gcc@10.5.0 |
cuda@11.4.4 |
umpire@2022.03.1 |
✅ |
gcc | cuda | umpire | note |
---|---|---|---|
gcc@10.3.0 |
cuda@11.5.2 |
umpire@2022.03.1 |
✅ |
gcc@10.3.0 |
cuda@11.5.2 |
umpire@2022.03.0 |
✅ |
gcc@10.3.0 |
cuda@11.5.2 |
umpire@6.0.0 |
✅ |
gcc@10.3.0 |
cuda@11.5.2 |
umpire@2022.10.0 |
✅ |
In order to obtain this results I had to temporarily remove some conflicts. In particular this work triggered another investigation which led me concluding that gcc@10.3.0
conflict in umpire was not correct (for more details see https://github.com/spack/spack/pull/41254).
Overall, from this set of builds I would infer that:
umpire@2022.10.0
is not working with gcc@9
umpire@2022.10.0
works with gcc@9
starting from cuda@11.3
gcc@10
works with both umpire@2022.03
and umpire@2022.10.0
(but gcc@10 is supported from cuda@10.4.1:
)I tested some reasonable configurations, if you need any more info o you have a specific configuration you'd like to see the result for (e.g. a specific patch version of CUDA), let me know and I can test it for you quite easily now that I have the setup.
FIY @aurianer @toxa81
I think with https://github.com/spack/spack/pull/41317 this can be closed.
I found a couple of issues on daint-gpu,
1)
-pedantic-errors
gcc9 and gcc10 (tested in compiler explorer) complains about multiple ; after function declaration:2) umpire 2022.10 doesn't work with gcc9 (umpire 2022.03.1 is fine)
3) cuda 11.2 doesn't work (cuda 11.8 does)
Update:
1) #1043
2) should we add a conflict in our packages? Or should it be done in umpire? Note: umpire builds fine, but generates error when included by memory_chunk.h (https://github.com/spack/spack/pull/41317)
3) in some of the .cu units we include pika which include fmt which creates problem with nvcc 11.2. I will try to clean includes. #1045