| 26 Jun 2026 |
 emily | that's presumably going to need to be fixed | 15:38:02 |
 hexa | removed | 15:39:51 |
| hexa | because what we have in staging as a default looks fine | 15:40:03 |
 Grimmauld (any/all) | Okay, i have a hook for OMPNUMTHREADS almost ready | 16:33:55 |
| hexa | tested? | 16:34:22 |
| Grimmauld (any/all) | Okay, i have a hook for OMP_NUM_THREADS almost ready | 16:35:27 |
| Grimmauld (any/all) | It sets the env var, it doesn't override the env var if it is already set, and numpy/scipy builds | 16:35:47 |
| Grimmauld (any/all) | doing a bit more testing rn | 16:36:10 |
| hexa | you should check, e.g. with --option cores 2 and perf top that the host isn't dominated by omp threads/gomp calls | 16:37:48 |
| emily | ok, had to push a fix but everything https://github.com/NixOS/nixpkgs/pull/535508 touches builds | 16:38:35 |
| emily | and perhaps someone might like to hit the button on https://github.com/NixOS/nixpkgs/pull/535513 and https://github.com/NixOS/nixpkgs/pull/535511 as well (though those can land any time) | 16:39:16 |
| emily | should be good to start this cycle with that merged from my PoV | 16:39:27 |
| Grimmauld (any/all) | I mean, not much more we can do than setting OMP_NUM_THREADS, right? If the env var is there, it should be fine. And confirming the env var is there in the check processes is trivial and that i did do | 16:39:34 |
| hexa | true | 16:39:55 |
| emily | there's some logic somewhere that caps it to 64 | 16:40:05 |
| emily | not sure why | 16:40:07 |
| emily | I also saw one package dividing it (maybe because parallel processes using OpenMP or something?) | 16:40:20 |
| hexa | lmao | 16:40:29 |
| hexa | some set it to $NIX_BUILD_CORES | 16:40:48 |
| emily | export OMP_NUM_THREADS=$(( $NIX_BUILD_CORES / 4 ))
if [ $OMP_NUM_THREADS -eq 0 ]; then
export OMP_NUM_THREADS=1
fi
| 16:40:52 |
| hexa | so they have to check if that is 0 | 16:40:55 |
| hexa | I have a part in that | 16:41:02 |
| hexa | feel free to blame | 16:41:05 |
| emily | and NumPy and SciPy do export OMP_NUM_THREADS=$((NIX_BUILD_CORES > 64 ? 64 : NIX_BUILD_CORES)) | 16:41:11 |
| emily | which, how does that handle the 0 case actually | 16:41:24 |
| hexa | not | 16:41:31 |
| emily | oh SciPy does the division+check, NumPy doesn't | 16:41:37 |
| Grimmauld (any/all) | numpy tests quite quickly, even with OMPNUMTHREADS=1 | 16:41:40 |
| hexa | https://github.com/OpenMathLib/OpenBLAS/issues/2993 | 16:41:41 |
| hexa | https://gitlab.freedesktop.org/pixman/pixman/-/work_items/62 | 16:41:51 |
