Add ppc64le (IBM POWER) architecture support by viniciusferrao · Pull Request #2637 · openhpc/ohpc

viniciusferrao · 2026-06-29T14:29:01Z

Summary

ppc64le (IBM POWER) architecture enablement for OpenHPC 4.x / EL10. @adrianreber suggested bringing this upstream. Sponsored by VersatusHPC; validated on POWER9 hardware (AlmaLinux 10 and openEuler 24.03).

Restructured per review: one commit per spec file, single-line subjects. The gotcha ccache handling is brought in by cherry-picking the three relevant 2.x commits (git cherry-pick -x), with the ppc64le-specific gotcha change applied on top of them — not folded into a single commit.

Scope

Component specs that needed architecture handling: gotcha, gnu-compilers, boost, fftw, openblas, likwid (+ a perf_event uncore patch), pdtoolkit, numpy, mpich, and the meta-package closure. 11 files.

The GNU compiler POWER9 baseline already landed in OHPC_setup_compiler upstream, so it is no longer part of this PR.

Design

All ppc64le-specific changes are gated by %ifarch ppc64le (or %ifnarch … ppc64le in the meta-package), so the existing x86_64 and aarch64 builds are unaffected. For gotcha, rpmspec --parse differs by exactly one line between targets (the ppc64le CMAKE_INSTALL_LIBDIR=lib).
The baseline is -mcpu=power9, matching the EL9+ ppc64le distro minimum (redhat-rpm-config's %__cflags_arch_ppc64le) and forward-compatible with POWER10/POWER11.
Per-package highlights: openblas TARGET=POWER9, fftw --enable-vsx, boost architecture=power, likwid GCCPOWER + a perf_event uncore patch, pdtoolkit ibm64linux, a numpy VSX3 build fix, gotcha CMAKE_INSTALL_LIBDIR=lib (cmake otherwise defaults to lib64, which breaks the module's lib path — matching the existing scalapack idiom), and the meta-package groups scoped to the MPI families built on POWER (openmpi5).

Verification

Built every touched package from source on both x86_64 and ppc64le (clean containers; AlmaLinux 10 and openEuler 24.03). All compile, with no regression on the existing architectures — the gated hunks are inert off-POWER, confirmed by both rpmspec --parse diffs and real RPM builds. gotcha 1.0.10 (the version this PR cherry-picks from 2.x) builds clean on ppc64le with all 18 ctest cases passing.

Notes

Independent of the ucx/hwloc spec fixes at the merge level. A from-source build of the full ppc64le stack also wants the ucx Source0 fix to fetch ucx, but that's a build-time convenience, not a merge dependency.

📦 Package Count Analysis Results

Environment: UBI 10 Container
Target Branch: 4.x
Version: OpenHPC 4.x
Overall Status: ✅ Analysis completed successfully

🏭 Factory Repositories

Status: ✅ Factory repositories analysis completed successfully

Release	Base OS	aarch64	x86_64	noarch	Total
4.2	EL_10	17	29	12	58
4.2	openEuler_24.03	17	19	12	48
Total	All	34	48	24	106

Analysis performed by OpenHPC Package Count CI
Commit: 6b7fa191
Workflow: 28473398768

github-actions · 2026-06-29T14:29:47Z

📊 Build Order Analysis Results

Environment: AlmaLinux 10 Container
Total Spec Files Analyzed: 89
Status: ✅ Build order analysis completed successfully

📋 Build Order Output

📦 Build Order (84 packages)

➕ Show remaining 69 packages

Analysis performed by OpenHPC Build Order CI
Commit: 6b7fa191
Workflow: 28473398751

adrianreber · 2026-06-29T15:25:28Z

For easier re-use on our different branches it would be good to have one commit for each spec file. Please change that, so that each commit changes only one spec file. The ccache changes in gotcha.spec should also not be part of the same commit. There are three commits on the 2.x branch for gotcha which can be brought to 4.x with git cherry-pick -x. Then apply your changes on top of it.

viniciusferrao · 2026-06-29T16:14:12Z

For easier re-use on our different branches it would be good to have one commit for each spec file. Please change that, so that each commit changes only one spec file. The ccache changes in gotcha.spec should also not be part of the same commit. There are three commits on the 2.x branch for gotcha which can be brought to 4.x with git cherry-pick -x. Then apply your changes on top of it.

Alright. I'll look at it.

github-actions · 2026-06-30T00:07:41Z

🚀 CCache Statistics

Configuration	🐧 RHEL	🦊 openEuler
🖥️ x86_64 / gnu15	🟡 28.94%	🟡 28.83%
🖥️ x86_64 / intel	🟢 79.13%	-
💪 aarch64 / gnu15	🟡 30.92%	🟡 38.03%

📊 Detailed Statistics

openEuler (aarch64)

Cacheable calls:    77344 / 97357 (79.44%)
  Hits:             29411 / 77344 (38.03%)
    Direct:         28147 / 29411 (95.70%)
    Preprocessed:    1264 / 29411 ( 4.30%)
  Misses:           47933 / 77344 (61.97%)
Uncacheable calls:  20013 / 97357 (20.56%)
Local storage:
  Cache size (GiB):   1.6 /   5.0 (32.23%)
  Hits:             29411 / 77344 (38.03%)
  Misses:           47933 / 77344 (61.97%)

openEuler (x86_64)

Cacheable calls:    31863 / 36310 (87.75%)
  Hits:              9187 / 31863 (28.83%)
    Direct:          8989 /  9187 (97.84%)
    Preprocessed:     198 /  9187 ( 2.16%)
  Misses:           22676 / 31863 (71.17%)
Uncacheable calls:   4447 / 36310 (12.25%)
Local storage:
  Cache size (GiB):   0.7 /   5.0 (14.59%)
  Hits:              9187 / 31863 (28.83%)
  Misses:           22676 / 31863 (71.17%)

RHEL (aarch64/gnu15)

Cacheable calls:    80520 / 101697 (79.18%)
  Hits:             24894 /  80520 (30.92%)
    Direct:         23468 /  24894 (94.27%)
    Preprocessed:    1426 /  24894 ( 5.73%)
  Misses:           55626 /  80520 (69.08%)
Uncacheable calls:  21177 / 101697 (20.82%)
Local storage:
  Cache size (GiB):   1.8 /    5.0 (35.15%)
  Hits:             24894 /  80520 (30.92%)
  Misses:           55626 /  80520 (69.08%)

RHEL (x86_64/gnu15)

Cacheable calls:    133872 / 156798 (85.38%)
  Hits:              38739 / 133872 (28.94%)
    Direct:          36974 /  38739 (95.44%)
    Preprocessed:     1765 /  38739 ( 4.56%)
  Misses:            95133 / 133872 (71.06%)
Uncacheable calls:   22926 / 156798 (14.62%)
Local storage:
  Cache size (GiB):    2.5 /    5.0 (49.60%)
  Hits:              38739 / 133872 (28.94%)
  Misses:            95133 / 133872 (71.06%)

RHEL (x86_64/intel)

Cacheable calls:    82849 / 114210 (72.54%)
  Hits:             65561 /  82849 (79.13%)
    Direct:         63647 /  65561 (97.08%)
    Preprocessed:    1914 /  65561 ( 2.92%)
  Misses:           17288 /  82849 (20.87%)
Uncacheable calls:  31361 / 114210 (27.46%)
Local storage:
  Cache size (GiB):   0.9 /    5.0 (17.17%)
  Hits:             65561 /  82849 (79.13%)
  Misses:           17288 /  82849 (20.87%)

🤖 Generated from workflow run 28473398738

viniciusferrao · 2026-06-30T02:15:19Z

Hi @adrianreber, PR was amended to have the separate commits per file.

The other issues were also addressed.

Thanks.

viniciusferrao · 2026-06-30T04:33:59Z

Those aarch64 failures were unexpected. I will look at it.

adrianreber · 2026-06-30T04:47:38Z

Those aarch64 failures were unexpected. I will look at it.

Don't worry, this is because of the meta packages for the arm HPC compiler. I know how to fix it.

github-actions · 2026-06-30T06:27:40Z

Test Results

111 files + 82 111 suites +82 5m 37s ⏱️ + 4m 47s
338 tests +279 332 ✅ +279 5 💤 - 1 1 ❌ +1
522 runs +417 516 ✅ +419 5 💤 - 3 1 ❌ +1

For more details on these failures, see this check.

Results for commit 5e4bd48. ± Comparison against base commit 680abc0.

This pull request removes 3 and adds 282 tests. Note that renamed tests count towards both.

test_harness ‑ [rms-harness] Verify long-running MPI job terminates with timeout parameter (slurm/gnu15/mvapich2)
test_harness ‑ [rms-harness] Verify non-zero exit code from MPI job detected as failure (slurm/gnu15/mvapich2)
test_harness ‑ [rms-harness] Verify zero exit code from MPI job runs OK (slurm/gnu15/mvapich2)

rm_execution ‑ [Boost/Accumulators] max-test under resource manager (slurm/gnu15/mpich)
rm_execution ‑ [Boost/Accumulators] max-test under resource manager (slurm/gnu15/mvapich2)
rm_execution ‑ [Boost/Accumulators] max-test under resource manager (slurm/gnu15/openmpi5)
rm_execution ‑ [Boost/Accumulators] max-test under resource manager (slurm/intel/mpich)
rm_execution ‑ [Boost/Accumulators] max-test under resource manager (slurm/intel/mvapich2)
rm_execution ‑ [Boost/Accumulators] max-test under resource manager (slurm/intel/openmpi5)
rm_execution ‑ [Boost/Accumulators] min-test under resource manager (slurm/gnu15/mpich)
rm_execution ‑ [Boost/Accumulators] min-test under resource manager (slurm/gnu15/mvapich2)
rm_execution ‑ [Boost/Accumulators] min-test under resource manager (slurm/gnu15/openmpi5)
rm_execution ‑ [Boost/Accumulators] min-test under resource manager (slurm/intel/mpich)
…

♻️ This comment has been updated with latest results.

adrianreber · 2026-06-30T08:12:31Z

Please rebase. That should solve the problems.

adrianreber · 2026-06-30T13:24:56Z

I see you make mpich on ppc64le depend on hwloc from OpenHPC. Any reason for that?

viniciusferrao · 2026-06-30T13:58:58Z

I see you make mpich on ppc64le depend on hwloc from OpenHPC. Any reason for that?

I tough that was the correct thing to do since OpenHPC ships hwloc. Should it depends on the distro hwloc?

adrianreber · 2026-06-30T14:01:40Z

I see you make mpich on ppc64le depend on hwloc from OpenHPC. Any reason for that?

I tough that was the correct thing to do since OpenHPC ships hwloc. Should it depends on the distro hwloc?

Probably correct. But then we should do it for all architectures. Let's leave this PR for now. It is already large enough. But maybe something for a future change.

viniciusferrao · 2026-06-30T14:10:24Z

I see you make mpich on ppc64le depend on hwloc from OpenHPC. Any reason for that?

I tough that was the correct thing to do since OpenHPC ships hwloc. Should it depends on the distro hwloc?

Probably correct. But then we should do it for all architectures. Let's leave this PR for now. It is already large enough. But maybe something for a future change.

Ok. I can remove that and we do afterwards.

viniciusferrao · 2026-06-30T15:39:26Z

Done @adrianreber

I'm still running tests here. Recompiling everything takes a while.

adrianreber · 2026-06-30T20:18:41Z

Sorry, but this needs another rebase. There were still things missing from other branches.

viniciusferrao · 2026-06-30T20:20:49Z

Sorry, but this needs another rebase. There were still things missing from other branches.

Yeah, I was looking at it. There was permission errors in the CI. Let me fix.

Signed-off-by: Adrian Reber <areber@redhat.com> (cherry picked from commit b4591a8)

- gotcha: 1.0.8 -> 1.0.10 upstream: llnl/gotcha Command: `misc/check_for_package_updates.py -v -o markdown gotcha --update` Signed-off-by: Adrian Reber <areber@redhat.com> (cherry picked from commit e0a7b0b)

Signed-off-by: Adrian Reber <areber@redhat.com> (cherry picked from commit abff307)

Sponsored-by: VersatusHPC Signed-off-by: Vinícius Ferrão <ferrao@versatushpc.com.br>

adrianreber · 2026-07-01T06:22:22Z

I merged this for now. For further changes please try to do smaller PRs for easier testing of the changes.

viniciusferrao force-pushed the pr/ppc64le-support branch from bc90e57 to e5aea32 Compare June 29, 2026 21:43

viniciusferrao force-pushed the pr/ppc64le-support branch from e5aea32 to 28f9407 Compare June 30, 2026 11:43

viniciusferrao force-pushed the pr/ppc64le-support branch from 28f9407 to 7926887 Compare June 30, 2026 14:46

viniciusferrao force-pushed the pr/ppc64le-support branch from 7926887 to efa3b26 Compare June 30, 2026 19:37

adrianreber and others added 9 commits June 30, 2026 17:22

gotcha: enable ccache if the build system supports it

4d9646b

Signed-off-by: Adrian Reber <areber@redhat.com> (cherry picked from commit b4591a8)

serial-libs: update gotcha to 1.0.10

0737634

- gotcha: 1.0.8 -> 1.0.10 upstream: llnl/gotcha Command: `misc/check_for_package_updates.py -v -o markdown gotcha --update` Signed-off-by: Adrian Reber <areber@redhat.com> (cherry picked from commit e0a7b0b)

gotcha: skip tests with the intel compiler

ca2b471

Signed-off-by: Adrian Reber <areber@redhat.com> (cherry picked from commit abff307)

gnu-compilers: make module whatis architecture-neutral

4d2d762

Sponsored-by: VersatusHPC Signed-off-by: Vinícius Ferrão <ferrao@versatushpc.com.br>

boost: add ppc64le (power) architecture support

fff5a90

Sponsored-by: VersatusHPC Signed-off-by: Vinícius Ferrão <ferrao@versatushpc.com.br>

fftw: enable VSX SIMD on ppc64le

b4d5280

Sponsored-by: VersatusHPC Signed-off-by: Vinícius Ferrão <ferrao@versatushpc.com.br>

openblas: set POWER9 target on ppc64le

2218ed4

Sponsored-by: VersatusHPC Signed-off-by: Vinícius Ferrão <ferrao@versatushpc.com.br>

likwid: add ppc64le support with perf_event uncore patch

6b02e89

Sponsored-by: VersatusHPC Signed-off-by: Vinícius Ferrão <ferrao@versatushpc.com.br>

pdtoolkit: add ppc64le (ibm64linux) support

6111a65

Sponsored-by: VersatusHPC Signed-off-by: Vinícius Ferrão <ferrao@versatushpc.com.br>

viniciusferrao added 3 commits June 30, 2026 17:22

numpy: add ppc64le VSX3 build fix

27a8e36

Sponsored-by: VersatusHPC Signed-off-by: Vinícius Ferrão <ferrao@versatushpc.com.br>

meta-packages: scope MPI library groups for ppc64le

4ff5ead

Sponsored-by: VersatusHPC Signed-off-by: Vinícius Ferrão <ferrao@versatushpc.com.br>

gotcha: add ppc64le (power) architecture support

5e4bd48

Sponsored-by: VersatusHPC Signed-off-by: Vinícius Ferrão <ferrao@versatushpc.com.br>

viniciusferrao force-pushed the pr/ppc64le-support branch from efa3b26 to 5e4bd48 Compare June 30, 2026 20:22

adrianreber merged commit ba5eef8 into openhpc:4.x Jul 1, 2026
24 of 27 checks passed

Uh oh!

Conversation

viniciusferrao commented Jun 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Scope

Design

Verification

Notes

Uh oh!

github-actions Bot commented Jun 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

📦 Package Count Analysis Results

🏭 Factory Repositories

Uh oh!

github-actions Bot commented Jun 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

📊 Build Order Analysis Results

📋 Build Order Output

📦 Build Order (84 packages)

Uh oh!

adrianreber commented Jun 29, 2026

Uh oh!

viniciusferrao commented Jun 29, 2026

Uh oh!

github-actions Bot commented Jun 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🚀 CCache Statistics

openEuler (aarch64)

openEuler (x86_64)

RHEL (aarch64/gnu15)

RHEL (x86_64/gnu15)

RHEL (x86_64/intel)

Uh oh!

viniciusferrao commented Jun 30, 2026

Uh oh!

viniciusferrao commented Jun 30, 2026

Uh oh!

adrianreber commented Jun 30, 2026

Uh oh!

github-actions Bot commented Jun 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Test Results

Uh oh!

adrianreber commented Jun 30, 2026

Uh oh!

adrianreber commented Jun 30, 2026

Uh oh!

viniciusferrao commented Jun 30, 2026

Uh oh!

adrianreber commented Jun 30, 2026

Uh oh!

viniciusferrao commented Jun 30, 2026

Uh oh!

viniciusferrao commented Jun 30, 2026

Uh oh!

adrianreber commented Jun 30, 2026

Uh oh!

viniciusferrao commented Jun 30, 2026

Uh oh!

Uh oh!

adrianreber commented Jul 1, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

viniciusferrao commented Jun 29, 2026 •

edited

Loading

github-actions Bot commented Jun 29, 2026 •

edited

Loading

github-actions Bot commented Jun 29, 2026 •

edited

Loading

github-actions Bot commented Jun 30, 2026 •

edited

Loading

github-actions Bot commented Jun 30, 2026 •

edited

Loading