• Home
  • Features
  • Pricing
  • Docs
  • Announcements
  • Sign In

IntelPython / dpnp / 44750113898
84%

Build:
DEFAULT BRANCH: master
Ran 20 Nov 2026 06:42PM UTC
Jobs 1
Files 268
Run time 1min
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst

20 Nov 2026 02:05PM UTC coverage: 84.255% (-0.01%) from 84.269%
44750113898

push

github

web-flow
Fix the OpenCL CPU compiler crash in the one-work-group radix sort (#3191)

This PR fixes a crash of the OpenCL CPU compiler in the one-work-group
radix sort, which `dpnp.sort`, `dpnp.argsort` and their `dpnp.tensor`
counterparts take for rows of 2049 to 16384 elements. Radix sort is the
default for `bool`, `int8`, `uint8`, `int16` and `uint16`, so a plain
`dpnp.sort` of such rows can segfault on a CPU device, for example
`dpt.sort(dpt.zeros((1, 4096), dtype="i1", device="opencl:cpu"))`.

The kernel puts a `group_barrier` after every unrolled counter update
when it runs on a CPU device whose kernel gets sub-groups shorter than
16 (8 on an AVX2 CPU, as checked for rows of 4096 elements). That is the
work-around for CMPLRLLVM-64592, added in IntelPython/dpctl@5125e11eb4.
The sub-group size is only known after the kernel is built, so the
barriers are compiled in on every CPU device, including AVX-512 CPUs
with sub-group 16 that never run them. With 16 or more elements per
work-item, they crash the compiler in its
`sycl-kernel-reduce-cross-barrier-values` pass (`copyAndReplaceUses`),
reported as CMPLRLLVM-79305. Every radix data type can crash, and which
ones do depends on the compiler's heap state; the same pass also hangs
at times, and master has returned a wrong sort without crashing. The
crash shows up with every OpenCL CPU runtime from 2024.18
(`intel-opencl-rt` 2025.0.4) to 2026.22 (2026.2.0).

The work-around is still needed: without it, the `int8` radix argsort is
unstable on the same runtimes. This PR replaces the barrier with a
work-group `atomic_fence`, which adds no barrier region and so no longer
crashes the compiler. In this kernel the fence keeps the `int8` argsort
stable. In the older copy of the kernel in the CMPLRLLVM-64592
reproducer it doesn't, and only the barrier does, so the fence avoids
that bug in this kernel rather than fixing it in general.

With a standalone SYCL reproducer, sort and stable argsort of `bool`,
`int8`, `uint8`, `... (continued)

1838 of 3094 branches covered (59.41%)

Branch coverage included in aggregate %.

30051 of 34754 relevant lines covered (86.47%)

40083.93 hits per line

Jobs
ID Job ID Ran Files Coverage
1 44750113898.1 20 Nov 2026 06:42PM UTC 268
84.26
GitHub Action Run
Source Files on build 44750113898
  • Tree
  • List 268
  • Changed 2
  • Source Changed 0
  • Coverage Changed 2
Coverage ∆ File Lines Relevant Covered Missed Hits/Line Branch Hits Branch Misses
  • Back to Repo
  • Github Actions Build #44750113898
  • d56955c0 on github
  • Prev Build on master (#43626763452)
  • Next Build on master (#44781361409)
STATUS · Troubleshooting · Open an Issue · Sales · Support · CAREERS · ENTERPRISE · START FREE TRIAL · SCHEDULE DEMO
ANNOUNCEMENTS · TWITTER · TOS & SLA · Supported CI Services · What's a CI service? · Automated Testing

© 2026 Coveralls, Inc