Public bug reported:
[ Impact ]
* rocFFT 7.2.4 (library version 1.0.36) fixes three correctness issues present
in 7.1.1:
- A potential division-by-zero crash when constructing FFT plans using
dimensions of length 1. Affected callers would see a hard crash or
undefined behaviour at plan creation time rather than a clean error.
- Incorrect result scaling on multi-device transforms. Applications using
rocFFT across multiple GPUs could silently produce wrong numerical
results without any error being raised.
- Broken callbacks on multi-device transforms. Custom load/store callbacks
registered by the caller were not being invoked correctly when the
transform spanned multiple devices, leading to silent data corruption
for any application relying on that feature.
* Additionally, two performance improvements are included: removal of an
unnecessary global transpose from MPI 3D multi-GPU pencil decompositions,
and enablement of the same optimisation for single-process multi-GPU
transforms. These reduce latency and memory traffic for HPC workloads
using distributed FFTs.
* Without this update, users on resolute running multi-GPU or length-1
FFT workloads may experience crashes, silent wrong results, or missed
callbacks. The fixes are present only in the 7.2.x upstream branch and
cannot be cherry-picked without pulling in the full release.
* Reverse dependencies: librocfft-dev (headers only); no other Ubuntu
archive package links against librocfft0 at this time.
[ Test Plan ]
1. Build:
- dpkg-buildpackage -S succeeds (source package builds cleanly).
- dpkg --compare-versions 7.2.4-0ubuntu1 gt 7.1.1-0ubuntu1 returns true.
- No symbols file is shipped (library uses a version script to restrict
exported symbols to the rocfft_ prefix); ABI was verified with abidiff:
abidiff librocfft.so.0.1 (7.1.1) vs librocfft.so.0.1 (7.2.4)
→ exit code 0, no output. SOVERSION unchanged at 0.
2. Installability:
- apt install librocfft0 librocfft-dev succeeds.
- No reverse dependencies require a rebuild.
3. Autopkgtest (test: librocfft0-tests):
Full run against ppa:igorluppi/rocfft-7.2.3, LXD rocm-gpu profile,
Ubuntu resolute, amd64 (2026-06-01):
Number of successful tests: 95436
Number of skipped tests: 120356
Number of runtime issues: 0
2 FAILED TESTS
Test suite took 3 hours 18 minutes
half precision max l-inf epsilon: 0.000746231
single precision max l-inf epsilon: 0.0391119
double precision max l-inf epsilon: 4.58786e-16
autopkgtest summary:
librocfft0-tests FAIL non-zero exit status 1
The 2 failures are GPU out-of-memory conditions on extreme test sizes
(buffer requirements of 1.5–2.1 GB), not correctness failures:
- pow2_1D complex forward len 268435456, single, batch 1 → 2.1 GB
- pow5_1D complex forward len 48828125, single, batch 4 → 1.56 GB
These failures are non-deterministic (different tests fail per run due
to random seed and GPU memory state) and predate this version bump:
journalctl on the test machine shows rocfft-test OOM-killed by the
kernel on May 19 and May 26 (anon-rss ~22 GB) during 7.1.x testing.
A targeted re-run of the previously failing tests with --gtest_filter
passed cleanly: 20/20, 0 failures (run: 2026-05-29, 8m 54s, PASS).
[ Where problems could occur ]
* FFT correctness regression: if the upstream fixes for multi-device
scaling or callbacks introduced a new bug, applications running
multi-GPU transforms could produce wrong numerical results. Symptom:
output values differ from single-GPU reference or from FFTW, detectable
by the accuracy test suite.
* Plan creation regression: if the division-by-zero fix interacted badly
with other plan construction paths, rocfft_plan_create could return an
unexpected error status or crash for previously working problem sizes.
Symptom: applications fail at plan creation with a non-zero return code.
* Callback regression: if the multi-device callback fix broke the
single-device callback path, applications using load/store callbacks
would see their callback never invoked or invoked with wrong arguments.
Symptom: silent wrong output in callback-dependent code paths.
* ABI break: although abidiff reports no changes, a missed symbol removal
would manifest as an immediate SIGILL or undefined-symbol error when
loading the library. SOVERSION is unchanged so the package manager will
not flag this; it would only appear at runtime.
[ Other Info ]
* ABI: abidiff 7.1.1 → 7.2.4 reports no changes (exit 0, no output).
SOVERSION unchanged at 0 (librocfft.so.0). No debian/*.symbols file;
symbol visibility is controlled by a version script (patch
Add-version-script-to-control-exposed-symbols.patch).
* This update is part of the coordinated ROCm 7.2.x stack release for
Ubuntu stonking (26.10) and resolute (26.04 LTS).
* PPA: https://launchpad.net/~igorluppi/+archive/ubuntu/rocfft-7.2.3
* Upstream version comparison:
https://github.com/ROCm/rocm-libraries/compare/rocm-7.1.1...rocm-7.2.4
** Affects: rocfft (Ubuntu)
Importance: Undecided
Status: New
--
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2155392
Title:
New upstream version 7.2.4
To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/rocfft/+bug/2155392/+subscriptions
--
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
