omp_pause_resource() warning on every SparseArray.Call() with the clang/libomp toolchain (Windows arm64)

Ouverte
#29 0 commentaires 0 réactions 0 personnes assignées Voir sur GitHub

Personne n'a encore pris cette issue.

Évaluation

Difficulté
3/5
Temps estimé
1-2 jours
Accessibilité débutants
35/100
Type d'issue
Bug
Clarté
Plutôt claire
Activité
Active
Stack technique
c, r
Domaine
backend

Piste de recherche

Start with .pause_resource() in R/thread-control.R, which warns on a non-zero return, and its C counterpart in src/thread_control.c, which calls omp_pause_resource(). Check where SparseArray.Call() invokes it in its on.exit(). Reproduce by running the NaArray-Arith-methods example, which should warn on a clang/libomp build. Done means the example and the test suite run without the warning, and the choice between the four proposed fixes is agreed on in the thread.

Rédigé par le modèle d'indexation à partir du texte de l'issue.

Description

On the Windows arm64 builder at R-Universe, SparseArray 1.13.4 emits the following warning on almost every call into its C code (see log here):

Warning in .pause_resource() :
  omp_pause_resource() failed to relinquish resources on device 0

The x86_64 Windows build of the same version is silent. The results themselves are correct and only the warnings differ. However, the warnings reach downstream packages. For example, GSVA's unit tests for version 2.7.23 failed on Windows arm64 because one test was turning such a warning into an error. I have tweaked the test in GSVA to make it pass, but other packages may run into it too.

Minimal reproducible example

I don't have a Windows machine, so this is more of a guess based on the outputs of the logs from R-Universe and suggestions by Claude Code, taking SparseArray's own NaArray-Arith-methods example, which warns in the arm64 example log:

library(SparseArray)
nam1 <- NaArray(dim=c(15, 6), type="integer")
nam1[cbind(1:15, 2)] <- 100:114
## Warning in .pause_resource() :
##   omp_pause_resource() failed to relinquish resources on device 0
nam1[cbind(1:15, 5)] <- -(114:100)
## Warning in .pause_resource() : ...
nam1
## Warning in .pause_resource() : ...

Calling the internal wrapper directly should isolate the cause:

SparseArray:::.pause_resource()  # expected to return 0
SparseArray:::.pause_resource()  # expected to warn: runtime already paused
Possible cause

SparseArray.Call() always runs .pause_resource() in its on.exit(). That calls omp_pause_resource(omp_pause_soft, omp_get_initial_device()) (src/thread_control.c) and warns whenever the return value is non-zero (R/thread-control.R).

The return value depends on the OpenMP runtime:

  • libgomp (GCC): pausing the host always succeeds and returns 0, unless it's called inside a parallel region. That's why Linux and Windows x86_64 builds are silent. macOS builds without OpenMP are unaffected.
  • LLVM libomp (clang): as far as Claude can tell from the libomp sources (__kmp_pause_resource() in kmp_runtime.cpp), a pause request returns non-zero if the runtime is already paused. A soft pause is only lifted when the next parallel region starts (__kmp_resume_if_soft_paused()).

Only a few C files contain #pragma omp parallel regions (SparseMatrix_mult.c, SparseArray_subassignment.c, SparseArray_matrixStats.c, SparseArray_subsetting.c). Most SparseArray.Call()s never open one, so after the first soft pause the runtime stays paused. Every later omp_pause_resource() call then fails as "already paused" and triggers the warning. This should affect any clang + libomp build, such as Linux with clang or macOS with OpenMP enabled, not only Windows arm64.

Possible fixes
  1. Treat the pause as best-effort: ignore the return value of omp_pause_resource(), or report it only once or when a verbose/debug option is set.
  2. Only call .pause_resource() after C routines that actually ran a parallel region, or keep a flag in C that is set when a parallel region is entered.
  3. Skip the pause when nthread <= 1, since no worker threads were created.
  4. Check the pause state before pausing again, so an already-paused runtime isn't treated as an error.
Unrelated, minor

Both Windows builders also report an install WARNING for -Wformat at OPBufTree.c:320 (%lu) and SVT_SparseArray_class.c:677-679 (%ld). Both print an R_xlen_t (long long), but long is 32-bit on Windows, so large values could be printed wrong. Casting to (long long) with %lld, or to (double) with %.0f, would fix it.

This issue has been written with the assistance of Claude Code (Anthropic).

Langage dominant
C
Étoiles
13
Forks
3
Métriques de merge des PR
Aucune PR mergée en 30 j

Préparer son environnement

Ce projet ne fournit ni conteneur de développement, ni Dockerfile, ni guide de contribution : l'installation est à votre charge. Commencez par son README, et consultez notre guide de la première contribution pour les étapes générales.

Par où commencer

  1. Lisez l'issue en entier, puis le guide de contribution du projet.
  2. Signalez en commentaire que vous la prenez — cela évite que deux personnes fassent le même travail.
  3. Forkez le dépôt et travaillez sur une branche.
  4. Ouvrez une pull request qui référence le numéro de l'issue.

Autres issues de Bioconductor/SparseArray

Toutes les issues de Bioconductor/SparseArray

Issues similaires

Plus d'issues C

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.