Revert '[SWDEV-544298] AINIC implementation [#2798]' by JeniferC99 · Pull Request #3418 · ROCm/rocm-systems

JeniferC99 · 2026-02-20T19:58:02Z

This reverts commit 00e9b8d.

Motivation

ROCm/TheRock#3500
#2798 This pr contain 75 files change for amdsmi and caused failure on TheRock Linux Ci run across all gfx.

Technical Details

Reverted PR2798

Test Plan

https://github.com/ROCm/TheRock/actions/runs/22244137887

Test Result

Submission Checklist

Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

This reverts commit 00e9b8d.

ScottTodd · 2026-02-20T20:44:25Z

This has build failures on it. Was it a clean revert? If new changes landed since the original commit we may need to revert them too. See https://github.com/ROCm/TheRock/blob/main/docs/rfcs/RFC0002-MonoRepo-Gardener-Rotations.md#reverting-vs-fixing-forward

If multiple commits are implicated in broken workflows, cleanly reverting them all sequentially is less risky than reverting manually.

JeniferC99 · 2026-02-20T22:36:24Z

There's no other PR relating to the NIC other than #2798.

…not completing (#3289)" This reverts commit 3c85365.

marifamd · 2026-02-20T22:53:14Z

There's no other PR relating to the NIC other than #2798.

You missed #3289, but that was the only other one that had nic changes

dmitrii-galantsev · 2026-02-20T23:17:36Z

changed name of the pr slightly because we don't sanitize strings in "Trigger ROCm CI"

JeniferC99 · 2026-02-20T23:39:00Z

The test build went pass amdsmi. https://github.com/ROCm/TheRock/actions/runs/22244137887/job/64354403909.

@ScottTodd please advise if I should merge now or wait till the build finish. or if any other test is needed.

ScottTodd · 2026-02-20T23:44:27Z

Looks okay to merge without waiting for further workflows to complete.

* Revert "[SWDEV-544298] AINIC implementation (#2798)" This reverts commit 00e9b8d. * Revert "[ROCM-2942] Fix AMDSMI_STATUS_NO_DATA on NIC devices & tests not completing (#3289)" This reverts commit 3c85365. --------- Co-authored-by: Jenkins <jenkins-compute@amd.com> Co-authored-by: Maisam Arif <Maisam.Arif@amd.com>

This change undoes the revert: bfe5604 Revert '[SWDEV-544298] AINIC implementation [#2798]' (#3418) 1f7eb27 SWDEV-544298_AINIC_implementation - Handle case where machine has no NIC and we are doing 'amd-smi static' (#3250) 00e9b8d [SWDEV-544298] AINIC implementation (#2798) The reason why it was reverted was that it was that the external CMake project, amdsmi_unified, was doing in-source build. We cannot change that CMakeLists.txt file to make it do out of source dir builds. This changes fixes that problem by copying the amdsmi_unified project into the build dir, then building it there, ensuring that even if it builds in its source dir, it is still in the parent project's build dir.

* Revert "[SWDEV-544298] AINIC implementation (#2798)" This reverts commit 00e9b8d. * Revert "[ROCM-2942] Fix AMDSMI_STATUS_NO_DATA on NIC devices & tests not completing (#3289)" This reverts commit 3c85365. --------- Co-authored-by: Jenkins <jenkins-compute@amd.com> Co-authored-by: Maisam Arif <Maisam.Arif@amd.com>

This change undoes the revert: bfe5604 Revert '[SWDEV-544298] AINIC implementation [#2798]' (#3418) 1f7eb27 SWDEV-544298_AINIC_implementation - Handle case where machine has no NIC and we are doing 'amd-smi static' (#3250) 00e9b8d [SWDEV-544298] AINIC implementation (#2798) The reason why it was reverted was that it was that the external CMake project, amdsmi_unified, was doing in-source build. We cannot change that CMakeLists.txt file to make it do out of source dir builds. This changes fixes that problem by copying the amdsmi_unified project into the build dir, then building it there, ensuring that even if it builds in its source dir, it is still in the parent project's build dir.

## Motivation Trying to manually sanitize PR strings is painful. Use github-script to handle it all for us. This should prevent issues like #3418 (comment) where double quotes broke the script. ## Test Plan - [ ] Make a PR with double quotes in the name and see if script breaks. - [x] Run the workflow via act to test syntax Signed-off-by: Galantsev, Dmitrii <dmitrii.galantsev@amd.com>

commit b3aabb45064e63ec721f13a48146c0fe76d76c02 Author: alexxu-amd <159800977+alexxu-amd@users.noreply.github.com> Date: Thu Apr 23 10:23:25 2026 -0400 [rocr-debug-agent] Correct .readthedocs path for monorepo migration (#5340) ## Motivation .readthedocs.yml contains outdated path. ## Technical Details Correct the path to match monorepo setup. ## JIRA ID   ## Test Plan  ## Test Result  ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit b5cd2c0530096f57e56cf2bebdfe2b7a7d09c060 Author: jamessiddeley-amd <James.Siddeley@amd.com> Date: Thu Apr 23 10:10:59 2026 -0400 [rocprof-compute] Removed --exit-zero from ruff formatting workflow (#5308) ## Motivation The formatting workflow uses`ruff check . --fix --exit-zero`. This PR drops `--exit-zero` from the ruff CI step so the workflow actually fails on lint errors that aren't auto fixable instead of silently passing them ## Technical Details N/A ## JIRA ID N/A ## Test Plan N/A ## Test Result N/A ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: Vignesh Edithal <Vignesh.Edithal@amd.com> commit 8508fd9b5e9a42a39929eabfc5d4b5e12c90f92a Author: Ammar ELWazir <ammar.elwazir@amd.com> Date: Thu Apr 23 09:04:20 2026 -0500 feat(perfxpert): consolidate agentic redesign and validation stack (#4576) ## Summary This PR is the consolidation point for the current `experimental/python/perfxpert` stack. It is no longer just the original rename branch: it now carries the agentic redesign, provider/runtime hardening, launcher/backend integration, parity/regression coverage, and the docs/media refresh on top of the `ROCInsight` -> `PerfXpert` rename. ## Scope In scope: - `experimental/python/perfxpert/**` Explicitly out of scope for this PR: - `.github/**` - CI/workflow-only changes (handled separately) ## Design / Feature Areas ### 1. Rename + package surface - rename from `rocinsight` to `perfxpert` - updated package/import/CLI/env-var/doc branding - preserved the Python and CLI entry surfaces under the new package name ### 2. Agentic analysis stack - layered agent hierarchy with Root, Analysis, Recommendation, Correctness, and specialist agents - shared agent brain across batch `perfxpert analyze`, Python API, MCP, and interactive launcher flows - fence-sliced prompt layering retained across the hierarchy ### 3. Provider + runtime hardening - provider registry/config loading and public provider surface fixes - typed provider error taxonomy and provider routing cleanup - fallback-chain support in the runtime session layer - recursion-guard and gate-cascade fixes carried forward into the consolidated branch ### 4. Launcher + backend integration - multi-backend `perfxpert-code` surface (`opencode`, `claude`, `codex`, `gemini`) - MCP registration + staged prompt wiring for backend adapters - source-checkout preference for the locally built patched `opencode` binary before bundled/PATH fallbacks - bundled build script now tolerates Bun lockfile-format churn by retrying without `--frozen-lockfile` ### 5. CLI + user workflows - top-level `init`, `diff`, `ci`, `doctor`, `providers`, and analyze flows present on the consolidated branch - source-only, trace-only, and combined analyze paths preserved - diff/CI/reporting surfaces merged into the same branch state ### 6. Validation / safety stack - parity fixtures and parity harness added to the branch - red-team and regression-gate coverage retained - exit-dashboard and audit-gate docs/tests carried into the consolidated branch - MCP registry/tool exposure checks preserved ### 7. Docs / schemas / media - guide and architecture docs updated to reflect the live product surface - refreshed guide GIF set carried on the branch - schema/version docs updated with the additive JSON formatter evolution - stale unsupported examples removed or rewritten to match the actual CLI/runtime surface ## Current Validation On Branch Focused validation completed on the live branch state: - `214 passed` on the shared runtime / CLI / MCP / setup / docs / launcher slice - `25 passed` on the launcher + setup-build + docs-links slice - local patched `opencode` build completed successfully from the pinned submodule - source-checkout launcher resolution now prefers the locally built patched binary: - `/tmp/rocm-systems-pr4576-consolidated/experimental/python/perfxpert/opencode/packages/opencode/dist/opencode-linux-x64/bin/opencode` Representative commands run on this branch state: ```bash PYTHONPATH=experimental/python/perfxpert pytest -q \ experimental/python/perfxpert/tests/test_agents/test_runtime.py \ experimental/python/perfxpert/tests/test_agents/test_framework.py \ experimental/python/perfxpert/tests/test_cli/test_analyze_args.py \ experimental/python/perfxpert/tests/test_cli/test_provider_error_surface.py \ experimental/python/perfxpert/tests/test_cli/test_backend_dispatch.py \ experimental/python/perfxpert/tests/test_cli/test_launcher_subcommands.py \ experimental/python/perfxpert/tests/test_tools/test_agents/test_mcp_registry_has_agent_tools.py \ experimental/python/perfxpert/tests/test_integration/test_cli_dispatch.py \ experimental/python/perfxpert/tests/test_setup/test_setup_build.py \ experimental/python/perfxpert/tests/test_integration/test_docs_links.py \ experimental/python/perfxpert/tests/test_tools/test_opencode_launcher.py ``` ```bash PYTHONPATH=experimental/python/perfxpert pytest -q \ experimental/python/perfxpert/tests/test_tools/test_opencode_launcher.py \ experimental/python/perfxpert/tests/test_setup/test_setup_build.py \ experimental/python/perfxpert/tests/test_integration/test_docs_links.py ``` ## Notes - This PR description is intentionally scoped to the consolidated `perfxpert` branch state, not the original rename-only PR. - `.github` changes are intentionally excluded from this PR. - Final ROCm/manual parity validation and the remaining doc/spec closeout continue on top of this branch. commit e7d104593dabd87674c40e5fc2f6cb0d87e46d8f Author: Aurelien Bouteiller <abouteil@amd.com> Date: Thu Apr 23 09:53:32 2026 -0400 [rocshmem] rocgdb debugging scripts (#5122) ## Motivation Improve debugability of deadlocked rocshmem application by providing rocgdb scripting that analyzes live deadlocked processes and summarizes and groups wavefront backtraces and highlight wavefronts that are stuck in a rocshmem api. ## Technical Details Use python scripting to drive rocgdb, add a small script to ease attaching and summarizing for multiprocesses. ## JIRA ID AIROCSHMEM-377 ## Test Plan * [x] Ran the debug script on the pr-provided deadlock example * [x] Ran the debug script on a modified functional test that deadlocks intentionally, witnessed wg groupings * [x] Ran the debug script on modified deepep that deadlocks intentionally, witnessed wg groupings in a useful way ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Signed-off-by: Aurelien Bouteiller <abouteil@amd.com> Signed-off-by: Aurelien Bouteiller <aurelien.bouteiller@amd.com> Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> commit d192aa1e21d1f220d0c3eb06c5730a3d75f0b81f Author: marantic-amd <marantic@amd.com> Date: Thu Apr 23 15:16:09 2026 +0200 [rocprofiler-systems] Adapt CPU metrics to PMC collector approach (#4612) ## Motivation - Add CPU metrics collector with policy-based PMC architecture - Replace the legacy cpu_freq module with a new CPU collector that follows the same policy-based design as the GPU and NIC collectors. The CPU collector integrates into the unified PMC sampler via collector_slice, sampling per-CPU frequency, load, and process-level metrics. ## Technical Details - New components: - procfs driver: reads /proc/stat, /proc/cpuinfo, getrusage() with persistent FDs and zero-copy parsing for minimal overhead - procfs provider: lightweight wrapper with factory pattern for DI - CPU device: manages per-CPU state with delta-based load computation - CPU traits: adapts base::collector for single-device multi-core model - CPU cache/perfetto policies: trace cache and Perfetto output - Mock driver and 17 unit tests covering init, freq, load, process metrics, filtered CPU sets, and repeated sampling - Integration: - Added to g_collector_slices alongside GPU/NIC in pmc::sampler - New ROCPROFSYS_CATEGORY_CPU_LOAD category for Perfetto traces - Renamed cpu_freq_sample to cpu_pmc_sample in trace cache (0x0007) - Updated perfetto_processor and rocpd_processor to handle CPU load deserialization and emit per-CPU counter tracks - Pause support: cache_policy emits zero entries for all monitored CPUs when metrics are empty (pause path) - Removed: - library/cpu_freq.cpp, cpu_freq.hpp - library/components/cpu_freq.cpp, cpu_freq.hpp - cpu_freq instance block from process_sampler.cpp ## JIRA ID - AIPROFSYST-224 - AIPROFSYST-79 ## Test Plan - Added additional unit tests to the feature - Tested with existing test suite ## Test Result - All test passed ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit 34bef1e6accfef4c5a3e01542b638229a09eb7a1 Author: accauble <accauble@users.noreply.github.com> Date: Thu Apr 23 09:03:54 2026 -0400 [rocjitsu] Add pre-commit hook to check for clang-format in rocjitsu files (#5311) ## Motivation I keep forgetting to run the clang_format.sh script so I'm adding a pre-commit hook that rocjitsu developers can install to check if they need to run the script. ## Technical Details Heavily commented because I used claude and I have never written a git-hook. Adding two basic scripts: -- A script to "install" the hook in rocm-systems. Only need to do this once per clone. -- The hook: checks staged files to see if they are formatted correctly. If not, halts the commit and tells the user what went wrong. ## JIRA ID None ## Test Plan Tested: 1. Installation script works as expected 2. Checked that a poorly-formatted staged file was caught by the hook 3. Checked that a well-formatted staged file is allowed 4. Checked that a poorly-formatted cpp file outside the rocjitsu directory doesn't get caught by the hook ## Test Result They work! ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Co-authored-by: Rosa, Ada <arosa@amd.com> commit 35534db977d5e5cbdd5ec88b7bd8b586ea2077a1 Author: Milan Radosavljevic <milan.radosavljevic@amd.com> Date: Thu Apr 23 13:09:43 2026 +0200 [rocprofiler-systems] Restructure unit tests to improve test isolation (#4853) ## Motivation Unit tests hang under Thread Sanitizer (TSan) due to a deadlock during atexit static destruction. The thread_deleter calls rocprofsys_finalize_hidden() which logs via spdlog, triggering a TSan race report whose symbolizer deadlocks on its own slot lock while closing a file descriptor. This causes the sanitizer CI workflow to time out. ## Technical Details Root cause: unit test executable was including `rocprofsys-object-library` which pulled `library.cpp`. with `library.cpp` static variables were included which should trigger `ensure_finalization(true)`. `ensure_finalization(true)` was calling `thread_deleter`. Unit test should include only isolated code, where this is breaking this rule. ## JIRA ID   ## Test Plan  - Run full unit test suite: ./bin/rocprof-sys-unit-tests — all 480 tests pass - Run logger fork/stress tests under TSan — no deadlock - Verify rocprofiler-systems-object-library is not linked into the test binary- ## Test Result All unit tests pass under TSan without hanging during atexit. The test binary no longer links library.cpp's static initializers, preventing thread_data instantiation and thread_deleter finalization entirely. ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit 783707e8cb0a210e78d1723542c42b61eb95f63d Author: Jonathan L'Work <113400649+Jonathan03ant@users.noreply.github.com> Date: Thu Apr 23 06:17:30 2026 -0400 [rocprofiler-sdk]: Fix crash on iGPUs without PMC support #148 (#5294) commit f89f4135b732c7bb83b6d25363160a847762e000 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Wed Apr 22 17:22:29 2026 +0000 Bump cryptography in /projects/rocr-debug-agent/docs/sphinx Bumps [cryptography](https://github.com/pyca/cryptography) from 46.0.6 to 46.0.7. - [Changelog](https://github.com/pyca/cryptography/blob/main/CHANGELOG.rst) - [Commits](https://github.com/pyca/cryptography/compare/46.0.6...46.0.7) --- updated-dependencies: - dependency-name: cryptography dependency-version: 46.0.7 dependency-type: indirect ... Signed-off-by: dependabot[bot] <support@github.com> commit 7321d441181f08e8c53f48e7b580f8dd2434a73d Author: cfallows-amd <Carrie.Fallows@amd.com> Date: Wed Apr 22 22:12:49 2026 -0400 [rocprofiler-compute] Bugfix for memchart Matrix Ops value (#5310) ## Motivation Fix Matrix Ops value pickup in memcharts- was trying to read for MFMA but recent roof updates have converted it to general matrix operations label which forced it to N/A value. ## Technical Details Change MFMA metric title to Matrix Ops to reflect db entry. Only affects CDNA memcharts ## JIRA ID AIPROFCOMP-559 ## Test Plan Manual testing on MI350 system - profiled rocflop and occupancy samples to test for workload that captures # of MFMA ops, and workload that has 0 MFMA ops, respectively. Also regenerated hashes. ## Test Result CLI analysis for block 3 shows Matrix Ops reflecting accurate MFMA values: <img width="1485" height="653" alt="Screenshot 2026-04-22 165029" src="https://github.com/user-attachments/assets/2b3d8d59-e038-4a1a-bdac-0db98bd94da1" /> <img width="1448" height="645" alt="Screenshot 2026-04-22 165104" src="https://github.com/user-attachments/assets/1c8423e0-4af8-44c8-b9d7-3a3203f39900" /> GUI mode also reflects these changes to the value outputs, no N/A. ## Submission Checklist - [X] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com> commit 4a65e8d0efcdcd584778d8b5d86d25563891afad Author: Dmitrii <dmitrii.galantsev@amd.com> Date: Wed Apr 22 20:01:57 2026 -0500 [RDC] Fix embedded profiler by loading librdc_rocp.so with RTLD_GLOBAL (#3922) rocprofiler-sdk discovers tools by searching for rocprofiler_configure in the global symbol namespace during hsa_init(). librdc_rocp.so was loaded with RTLD_LAZY (implicit RTLD_LOCAL), hiding the symbol and silently preventing profiler counter collection in embedded mode. commit cbfa12fb0ea9a8c39b9e0f3232f3e821e2ce239f Author: marandje <marko.arandjelovic@amd.com> Date: Thu Apr 23 01:28:40 2026 +0200 Fix false queue-idle detection for first unsignaled packet (#5073) ## Motivation - `Unit_hipGetProcAddress_MemoryApisMemset` is failing because `hipStreamSynchronize` may return before `hipMemset*Async` operation has actually completed when it is the only operation enqueued on the stream. ## Technical Details - If a single command is enqueued on a stream without an attached completion signal, `IsQueueIdle()` can incorrectly report the queue as idle. In the original logic, the write index starts at `0`, so an empty queue and a queue containing exactly one command at index `0` are indistinguishable. - When `hipStreamSynchronize` is called on stream A, the finish path may call `ReleaseAllHwQueues()`, which can release an idle hardware queue belonging to stream B. If stream B has only a single enqueued command without a completion signal, that queue can be released prematurely. A later synchronization on stream B may then reacquire a different queue that is already idle, causing `hipStreamSynchronize` to return too early. - `last_write_index_` and `last_packet_with_signal_index_` still starts counting from zero but are preinitialized with an invalid queue index. ## JIRA ID ROCM-20999 ## Test Plan Run hip-tests. ## Test Result All hip-tests passed. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit 4a47fe94289240c48cd005cf0499c71221d29a0c Author: akolliasAMD <99202231+akolliasAMD@users.noreply.github.com> Date: Wed Apr 22 16:28:06 2026 -0600 removed akolliasAMD from codeowners on rocSHMEM repo (#5250) ## Motivation Removing akolliasAMD from CODEOWNERS for team changes ## Test Plan No test needed ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit c08c171dbb0d974eee4c42dc9e7709e2cb13cf56 Author: marandje <marko.arandjelovic@amd.com> Date: Thu Apr 23 00:18:39 2026 +0200 Remove the queue from a set of active queues only when last enqueued command is cleared (#5193) ## Motivation Unit_hipMultiThreadStreams1_AsyncAsync fails when GPU_ENABLE_PAL=0. In HostQueue::finish(), lastEnqueueCommand_ is preserved, but the queue is still removed from activeQueues. When the next command is enqueued, the queue is not added back to activeQueues because lastEnqueueCommand_ is still non-null, which prevents the null stream from implicitly synchronizing with that queue. ## Technical Details This change removes the queue from the set of active queues only when last enqueued command is cleared. ## JIRA ID ROCM-20999 ## Test Plan Run Unit_hipMultiThreadStreams1_AsyncAsync 100x. ## Test Result All iterations passed. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit c74978d8d5f3c7f8732da86659f08acf71b9427f Author: Sajina PK <sputhala@amd.com> Date: Wed Apr 22 18:14:48 2026 -0400 [rocprof-systems]: Fix makeinfo build failures on systems without texinfo (#5131) ## Motivation Bundled `binutils` bfd build can invoke `makeinfo` to refresh GNU Info manuals. On images without `Texinfo`, that fails with `makeinfo: not found` even though we only need static libraries (libbfd, libiberty, etc.). ## Technical Details - Set `MAKEINFO` to a no-op (true, via `find_program`) in the configure environment for the timemory `binutils-external` and rocprofiler-systems `rocprofiler-systems-libiberty-build` ExternalProjects. - Pass `MAKEINFO=…` on the make command line so existing build trees and `CONFIGURE_HANDLED_BY_BUILD` flows still override the value baked into generated Makefiles. - [x] Merge the timemory update: https://github.com/ROCm/timemory/pull/30 - [x] Once the timemory PR is merged update the submodule pointer to head of rocprofiler-systems branch ## JIRA ID ## Test Plan Manual test building `rocprofiler-systems` project. ## Test Result Pass ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit 4c8dc1a07aaec317ed9bf7da1f7854cac74da62a Author: spolifroni-amd <Sandra.Polifroni@amd.com> Date: Wed Apr 22 16:55:44 2026 -0400 [rocdecode, rocjpeg] added env var doc (#5247) ## Motivation rocdecode and rocjpeg both needed a list of env vars that affected their runtimes commit 08cec46e16354c0b34a404acb5c3e94d8035695d Author: Dmitrii <dmitrii.galantsev@amd.com> Date: Wed Apr 22 15:42:26 2026 -0500 [ci] Use github-script instead of curl (#3424) ## Motivation Trying to manually sanitize PR strings is painful. Use github-script to handle it all for us. This should prevent issues like https://github.com/ROCm/rocm-systems/pull/3418#issuecomment-3937597408 where double quotes broke the script. ## Test Plan - [ ] Make a PR with double quotes in the name and see if script breaks. - [x] Run the workflow via act to test syntax Signed-off-by: Galantsev, Dmitrii <dmitrii.galantsev@amd.com> commit ef7520eccdbe4658a2864ac2ff78bfdb7a1a529c Author: Scott Tsai <scottt.tw@gmail.com> Date: Thu Apr 23 04:32:28 2026 +0800 Fix rocprofiler-sdk build on gcc 15 (#934) This PR fixes building rocprofiler-sdk on gcc 15 by updating two of its external submodules: * rocprofiler-sdk/external/yaml-cpp * rocprofiler-sdk//external/elfio to include `cstdint.h` where it needs them instead of implicitly through other C++ standard headers. commit fc3df3bf950cc16c6bbddd95735e9b9ee4758835 Merge: 1c40b0c84e b5210bf473 Author: Zach Byrne <zbyrne@amd.com> Date: Wed Apr 22 14:07:44 2026 -0600 [hipfile] Add io size, bandwidth, and latency stats broken down by gpu to ais-stats (#4991) [AIHIPFILE-159] HipFile stats stores counters in a series of buckets broken up by backend, IO operation, and IO size. This lets ais-stats show histograms for different IO sizes. Also enables stats collection by default. commit 1c40b0c84e60006d154ec08710fd8e0c54fb3a09 Merge: eb5dee8755 5412a62035 Author: James Newling <james.newling@gmail.com> Date: Wed Apr 22 12:43:21 2026 -0700 [RocJitsu] Basic CI workflow (#5216) Add basic CI that runs on free github runners. Builds in ~10 mins. Tests run in ~5 seconds. See 'Checks' to see what tests ran. commit eb5dee875516bf2fd001a3781938036947fbdb66 Author: Adam Pryor <61172547+adam360x@users.noreply.github.com> Date: Wed Apr 22 13:38:00 2026 -0500 [RDC] Use longer sampling window when using SIMD rocprof metric (#5251) ## Motivation SIMD util from rocprof needs a longer sampling window to give meaningful results ## Test Verified rocprof metrics behavior correctly with longer sampling window Signed-off-by: Adam360x <Adam.pryor@amd.com> commit 8d566d526860fea2822d57e5084ffd14c65ac031 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Sat Mar 28 15:30:54 2026 +0000 Bump cryptography in /projects/rocr-debug-agent/docs/sphinx Bumps [cryptography](https://github.com/pyca/cryptography) from 43.0.0 to 46.0.6. - [Changelog](https://github.com/pyca/cryptography/blob/main/CHANGELOG.rst) - [Commits](https://github.com/pyca/cryptography/compare/43.0.0...46.0.6) --- updated-dependencies: - dependency-name: cryptography dependency-version: 46.0.6 dependency-type: indirect ... Signed-off-by: dependabot[bot] <support@github.com> commit 07a818e54c4a7f1fd24dd851c98b51730627edf2 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Fri Apr 3 03:38:20 2026 +0000 build(deps): bump pygments in /projects/rocr-debug-agent/docs/sphinx Bumps [pygments](https://github.com/pygments/pygments) from 2.18.0 to 2.20.0. - [Release notes](https://github.com/pygments/pygments/releases) - [Changelog](https://github.com/pygments/pygments/blob/master/CHANGES) - [Commits](https://github.com/pygments/pygments/compare/2.18.0...2.20.0) --- updated-dependencies: - dependency-name: pygments dependency-version: 2.20.0 dependency-type: indirect ... Signed-off-by: dependabot[bot] <support@github.com> commit fb4aa4aac81f5f0099480ad35cf8b330ac8547b1 Author: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Date: Tue Apr 7 07:38:17 2026 +0000 Bump tornado in /projects/rocr-debug-agent/docs/sphinx Bumps [tornado](https://github.com/tornadoweb/tornado) from 6.4.2 to 6.5.5. - [Changelog](https://github.com/tornadoweb/tornado/blob/master/docs/releases.rst) - [Commits](https://github.com/tornadoweb/tornado/compare/v6.4.2...v6.5.5) --- updated-dependencies: - dependency-name: tornado dependency-version: 6.5.5 dependency-type: indirect ... Signed-off-by: dependabot[bot] <support@github.com> commit f3cabf66b29ae140410bd432841d1a8833f50202 Author: Lancelot SIX <lancelot.six@amd.com> Date: Fri Apr 17 16:23:11 2026 +0100 Cleanup unused arch. commit cdbf9cb8ba938b1990afbd929170291df73b5856 Author: Lancelot SIX <lancelot.six@amd.com> Date: Fri Apr 17 16:20:15 2026 +0100 gfx12: Fix writing to pseudo_state_priv I noticed that the debugger can't write the $state_priv register, writes are eventually skipped. It turns out this is because gfx12_architecture_t::write_pseudo_register fails to memcpy the value to the register. Fix this. commit 1dd1c995a1181ecf8356d04846c35dccccbdb9d5 Author: Yiltan <ytemucin@amd.com> Date: Wed Apr 22 11:59:22 2026 -0400 [rocshmem] Remove warnings: local memory global used by non-kernel function (#4948) ## Motivation In an effort to cleanup rocSHMEM, we are fixing the warnings. The specific warning this PR fixes is ``` warning: <unknown>:0:0: in function xyz: local memory global used by non-kernel function ``` ## Technical Details `__shared__` cannot be declared with in a `__device__` function ## JIRA ID AIROCSHMEM-332 commit c6708bf70c9a47954916551644ff38992c622010 Author: Siwei He <siweihe.eng@gmail.com> Date: Wed Apr 22 11:03:34 2026 -0400 kfdtest: Use CopyDwordIsa in Ptrace test (#5233) Fix: Switch to CopyDwordIsa, which is simpler and appropriate for this test's purpose (testing ptrace access to invisible VRAM, not scratch memory). This avoids the scratch space conversion logic entirely. ## Motivation PtraceAccessInvisibleVram test failure due to change in ScratchCopyDwordIsa. This test is better using CopyDwordIsa ## Technical Details ## JIRA ID ROCM-23360 ROCM-20237 ## Test Plan Able to compile the test and pass the PtraceAccessInvisibleVram test. ## Test Result Test passed. Since this change only affects PtraceAccessInvisibleVram test. It's safe to test only this sub test case. ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit 8821e322291b57fe68aa64891ccb836a77f9da77 Author: Edgar Gabriel <edgargabriel@users.noreply.github.com> Date: Wed Apr 22 09:51:45 2026 -0500 [rocSHMEM] remove static amdsmi-loader instance (#5236) ## Motivation the constructor in a static amdsmi-loader instance tried to open the amdsmi lib and dlsym the functions used even if we do not compile with pod-detection. The static instance however not actually required. ## Test Result  ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit 0b6b84f76d0e69c311b49a25b48bff0a8c3aa7ff Author: Shadi Dashmiz <94885391+shadidashmiz@users.noreply.github.com> Date: Wed Apr 22 09:11:21 2026 -0400 ROCM-20929 - fix hiptests issues (#5159) - check for proper peer enablement - remove constexpr evalution ## Motivation 2 tests are failing due to error in the test code ## Technical Details Constexpr caused the block of code to not run so that p2p check was incomplete. Check if P2P is supported before testing the error code. If P2P isn't supported at all, skip the test. ## JIRA ID ROCM-20929 ## Test Plan Hip test in the ticket ## Test Result PSDB - passed ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. Signed-off-by: sdashmiz <shadi.dashmiz@amd.com> commit 00238f6c365df815d616fa9c6928fe5d58a48cc6 Author: Aurelien Bouteiller <abouteil@amd.com> Date: Tue Apr 21 22:21:16 2026 -0400 Workaround for the DIEExpression failure to build with Debug (#5205) RelWithDebInfo and Debug configurations fail to build with DIEExpression error on certain versions of hip/clang. THis is a workaround, a complete solution is pending on the compiler side of the house we should reconsider when this is deployed. ## Motivation Being able to debug is critical. ## Technical Details Disable expanded dwarf decorators, rely on normal dwarf. Also reenable -O0 in debug since that now also compiles. This disables some debug info but that cannot be avoided until the compiler gets updated. ## JIRA ID AIROCSHMEM-227 ## Test Plan * [x] Build with debug * [x] can rocgdb the code without crashing rocgdb * [x] some debug info avail (not all, its not possible) ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit 2ea2e50ea976340213f399bc7f6a7f6df0ff59d4 Author: jamessiddeley-amd <James.Siddeley@amd.com> Date: Tue Apr 21 21:57:42 2026 -0400 [rocprof-compute] Move intermediate csv deprecation warning to 7.13 upcoming changes (#5255) ## Motivation Move deprecation warning out of 'Unreleased' section and move to changes to ROCm 7.13 upcoming changes section instead ## Technical Details No functional changes ## JIRA ID N/A ## Test Plan N/A ## Test Result N/A ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit 1140233ffee37996820dcfaa710273fd2061a331 Author: Mark Meserve <mark.meserve@amd.com> Date: Tue Apr 21 17:48:00 2026 -0500 rocprofiler-sdk: attach: multiprocess support (#4804) - Adds the ability to attach to a process and all of its descendant processes in a single call - Adds rocattach_attach_tree and rocattach_detach_tree to rocattach - Adds --attach-children option to rocprofv3 - Defaults to True - Adds test to verify attach tool loading through a process tree - Adds test to verify tool data through rocprofv3 - Some logging and documentation cleanup commit e4029c3fed5478c48a3f5c4c36924b19a8c093cd Author: Aryan Salmanpour <aryan.salmanpour@amd.com> Date: Tue Apr 21 14:22:06 2026 -0700 [rocjpeg][samples] - Add total wall time for jpeg decoding (#5206) ## Motivation This PR adds overall wall time for the jpegDecodePerf sample. ## Technical Details  ## JIRA ID   ## Test Plan  ## Test Result  ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit bd00380ee6b9d90e17a7282b763ee7af22f18682 Author: Aryan Salmanpour <aryan.salmanpour@amd.com> Date: Tue Apr 21 14:17:51 2026 -0700 [rocdecode][ctest] - Add option to enable extended ctest tests (#5004) ## Motivation Add a CMake option `ENABLE_EXTENDED_TESTS` to enable extended ctest tests for rcodecode. ## Technical Details  ## JIRA ID   ## Test Plan  ## Test Result  ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: kiritigowda <kiritigowda@gmail.com> commit c5245d48f55fbca612be4fc6fc1c5813ae559da4 Author: Benjamin Welton <Benjamin.Welton@amd.com> Date: Tue Apr 21 13:32:08 2026 -0700 [rocprofiler-sdk] Add runtime guards to ROCP_* log macros (#5222) ## Summary - Add runtime log level checks to `ROCP_INFO`, `ROCP_WARNING`, `ROCP_ERROR` and their `_IF` variants - Prevents evaluation of stream arguments (including `fmt::format` calls) when the severity is below the configured minimum log level - Uses abseil's `LOG_IF` which internally short-circuits via a ternary pattern, ensuring zero argument evaluation overhead when logging is suppressed - `ROCP_TRACE` already guards via `VLOG_IS_ON`; `ROCP_FATAL`/`ROCP_DFATAL` are never suppressed ## Test plan - [ ] CI builds pass across all OS variants - [ ] Verify log output is unchanged at default log level (WARNING) - [ ] Verify no output from `ROCP_INFO`/`ROCP_WARNING` when `ROCP_LOG_LEVEL=error` 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-authored-by: Claude Opus 4 <noreply@anthropic.com> commit 5412a620359b09593cfd3e0b0959c7a1a443c56e Author: James Newling <james.newling@gmail.com> Date: Mon Apr 20 21:37:22 2026 +0000 add libdrm-dev dep in CI commit 8a03fbe617fa1f2441a08f772471824c2a0f9ffb Author: James Newling <james.newling@gmail.com> Date: Mon Apr 20 21:32:11 2026 +0000 add most basic of vanilla CIs commit 58d180e1f39538c7bf724cf225aca18bef944297 Author: Maisam Arif <Maisam.Arif@amd.com> Date: Tue Apr 21 14:52:48 2026 -0500 [AMD-SMI] Upgrade precommit workflow to have diff highlighting (#4534) Signed-off-by: Maisam Arif <Maisam.Arif@amd.com> commit d9b2e09246116d69de08e8108df838b8f7f67324 Author: Darren Lao <Darren.Lao@amd.com> Date: Tue Apr 21 15:52:01 2026 -0400 Fix amdsmi_pcie_info_t speed unit (#5245) Fixes https://github.com/ROCm/rocm-systems/issues/5200 This is purely a documentation change, and corrects the units for `max_pcie_speed` to be in MT/s, which I confirmed is returned from: Signed-off-by: Maisam Arif <Maisam.Arif@amd.com> commit 9fb1dff90e1d4bf73777d4eb93059a270a40a13d Author: Avinash Kethineedi <avinash.kethineedi@amd.com> Date: Tue Apr 21 13:43:45 2026 -0500 [rocSHMEM] Implement ROCSHMEM_TEAM_SHARED in all the backends (#4914) ## Motivation Implement the `ROCSHMEM_TEAM_SHARED` predefined team per the OpenSHMEM specification. TEAM_SHARED comprises all PEs sharing a common memory domain (same node), enabling node-scoped collective and point-to-point operations. ## Technical Details - Declare and define `ROCSHMEM_TEAM_SHARED` in host/device namespaces with device symbol copy support. - Add team_shared_ to TeamTracker with `NUM_RESERVED_TEAMS` so predefined teams don't consume user-creatable team slots. - Protect TEAM_SHARED from destruction in `rocshmem_team_destroy()`. - Implement `setup_team_shared()` in all three backends (IPC, RO, GDA). In RO and GDA, TEAM_SHARED is only created when IPC is enabled at compile-time and runtime; otherwise set to `ROCSHMEM_TEAM_INVALID`. - Update `team_ctx_infra_tester` to use runtime `ROCSHMEM_MAX_NUM_TEAMS` with dynamic sizing. - Add `TeamCtxSharedInfraTestType` (test 94) to validate `ROCSHMEM_TEAM_SHARED` PE identity, team size, and context creation on device. Gracefully skips when IPC is disabled. ## JIRA ID - AIROCSHMEM-355 ## Test Plan - Run teamctxsharedinfra (test 94) with 2 ranks on the same node. - Run hipmodule_init test to verify symbol copy. - Verify graceful skip when IPC is disabled. ## Test Result - [x] Tested with IPC backend - [x] Tested with RO backend (IPC enabled) - [x] Tested with RO backend (IPC disabled, verifies graceful skip) - [x] Tested with GDA backend (IPC enabled) - [x] Tested with GDA backend (IPC disabled, verifies graceful skip) commit a8eef4083e76a2b8a78221510023da85b88af144 Author: Jason Bonnell <166553723+jbonnell-amd@users.noreply.github.com> Date: Tue Apr 21 13:47:16 2026 -0400 Update rocprofiler workflows to use latest actions with node 24 (#5231) ## Motivation  Noticed that older versions of Docker actions used Node 20, and Node 20 actions are being deprecated in June. ## Technical Details  - Update `rocprofiler-*` workflow files to use the latest versions of Docker and GitHub actions (moved to Node 24) - `docker/login-action` -> v4.1.0 - `docker/metadata-action` -> v6.0.0 - `docker/build-push-action` -> v7.1.0 - `docker/setup-buildx-action` -> v4.0.0 - `actions/checkout` -> v6 - `actions/upload-artifact` -> v7 - `nick-fields/retry` -> v4 ## JIRA ID   ## Test Plan  Ensure no regression issues with action updates, workflows pass ## Test Result  TBA ## Submission Checklist - [X] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit 774ff6e7eec8fb0e79aaab408191635857c8367e Author: Victor Zhang <111778801+victzhan@users.noreply.github.com> Date: Tue Apr 21 13:06:25 2026 -0400 [AIRUNTIME-1929] HIP memory stress test fixes and improvements (#4069) ## Motivation Fix instability and build issues in HIP memory stress tests. ## Technical Details - Fix grid-stride loop in `KrnlWth2MemTypes` kernel in `hipMallocManagedStress.cc`. - Fix resource leaks (`hipFree`, `hipStreamDestroy`) in `LaunchKrnl4`. - Re-initialize managed memory between prefetch iterations in `hipMemPrftchAsyncStressTst.cc`. - Fix `hold_memory` build: correct include and add missing link library. - Simplify `stress_test` custom target. - Add `hipMallocConcurrency.cc` to the build. ## JIRA ID AIRUNTIME-1929 ## Test Plan Build and run `stress_test` target. Verify stress tests pass consistently. ## Test Result Pending ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit 9deaac2945c4ef2cb9ab787b40619f96849d41e5 Author: Victor Zhang <111778801+victzhan@users.noreply.github.com> Date: Tue Apr 21 13:05:17 2026 -0400 ROCM-2903 - Default DEBUG_HIP_MEM_POOL_VMHEAP to IS_LINUX (#5139) ## Motivation rocBLAS graph tests hang on Windows 11 after sustained stream-ordered allocation / free / trim cycles (~500-1000 cycles in). The hang is in the VM heap path: awaitCompletion on the synchronous VirtualMapCommand blocks indefinitely with the GPU event never signaling, leaving the command stuck in CL_SUBMITTED. ## Technical Details Default DEBUG_HIP_MEM_POOL_VMHEAP to IS_LINUX so that allocations on Windows are routed through SvmBuffer::malloc / free instead of VirtualMapCommand. This is a temporary workaround until the ROCR/HSA backend becomes the default on Windows, which brings direct dispatch and avoids this path entirely. ## JIRA ID ROCM-2903 ## Test Plan Run rocblas-test *graph* filter on Navi3x / Windows 11; verify the hang is resolved. ## Test Result Confirmed on W7900 (gfx1100) + Windows 11: PAL previously hung at trmv_graph_test_f32_r_U_N_N_128_2_1; with this change all 1637 graph tests pass. ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: FirstName LastName (Public) <your_public_email@example.com> commit 2ecbb87a314b60d7b992cb4f275714e8b09a6c9b Author: Kian Cossettini <Kian.Cossettini@amd.com> Date: Tue Apr 21 13:01:22 2026 -0400 [rocprofiler-systems] Install python dependencies at the Dockerfile level (#5197) ## Motivation  Part of https://github.com/ROCm/rocm-systems/pull/3962 Make the dockerfiles install the requirements.txt. Drop support for python 3.6 and 3.7 (test suite requires python 3.8 and above) ## Technical Details  Remove 3.6 and 3.7 from Dockerfiles and workflows. Install requirements.txt in Dockerfiles ## JIRA ID   ## Test Plan  ## Test Result  ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit c54a0e85fa6727ce1ffdc5ccc3695af0950365d8 Author: Dmitrii <dmitrii.galantsev@amd.com> Date: Tue Apr 21 11:54:10 2026 -0500 [RDC][AMD-SMI][ROCM-SMI] Drop POST_BUILD strip on shared libs (#5237) Defer stripping to packagers and TheRock's THEROCK_SPLIT_DEBUG_INFO. Fixes SONAME-symlink corruption under llvm-strip (https://github.com/ROCm/TheRock/issues/4640) commit e22df2da076973fb0a347c868edd2beea7f1fb7d Author: Venkateshwar Reddy Kandula <Venkateshwarreddy.Kandula@amd.com> Date: Tue Apr 21 11:17:29 2026 -0500 [rocprofiler-sdk] Add reusable composite action rocprofiler-sdk CI (#5095) ## Motivation This PR introduces a reusable local composite GitHub Action to consolidate repeated “install ROCm + build dependencies” logic across rocprofiler-sdk workflows, aiming to reduce duplication and simplify maintenance. ## Technical Details - Added a new composite action: .github/actions/rocprofiler-sdk-util. - Updated rocprofiler-sdk workflows (docs, CI, CodeQL, code coverage) to call the composite action instead of inlining dependency build steps. - Added workflow env toggles to enable/disable building specific dependencies (ROCR, HIP/CLR, rocprofiler-register, aqlprofile, rocDecode, rocJPEG) and to control nightly ROCm installation. ## JIRA ID   ## Test Plan CI checks ## Test Result Passing ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> commit 7a234bbaf6ba32af3b64aacdbe27eaaac8348823 Author: Aurelien Bouteiller <abouteil@amd.com> Date: Tue Apr 21 11:01:40 2026 -0400 [rocSHMEM] Fix a warning when building config gda_ionic (#5218) ## Motivation Fix warnings ## Technical Details missing a [[maybe-unused]] in gda_ionic configuration. ## JIRA ID AIROCSHMEM-300 commit b44f6c789daaf102492bb3572005318599abd656 Author: German Andryeyev <56892148+gandryey@users.noreply.github.com> Date: Tue Apr 21 10:59:12 2026 -0400 clr: Expose some HSA interfaces in hip dll (#4748) ## Motivation These calls are required to enable AQL profiler functionality on Windows. ## Technical Details Currently AQL profiler requires direct calls to ROCR interfaces ## JIRA ID OCM-1667 ## Test Plan Aql profiler tests ## Test Result Passed ## Submission Checklist N/A commit 1f21d5b1a5273a9237ab3594bb56ebf2e64f2fc4 Author: Jason Bonnell <166553723+jbonnell-amd@users.noreply.github.com> Date: Tue Apr 21 09:28:51 2026 -0400 [rocprofiler-compute] Update git version in Ubuntu GHCR CI images (#5208) ## Motivation  #5144 adds some new submodules, and the current CI image is failing to clone the repo with the submodules due to 22.04 using an older version of git. ## Technical Details  - Update `Dockerfile.ubuntu.ci` to include the git-core ppa repository so we can grab the latest version of git - Previous version installed was 2.34.1, which was the latest version available with default ubuntu 22.04 repos ## JIRA ID   ## Test Plan  - Ensure Dockers build properly - Once merged, verify that the updated images are able to pass the checkout step in #5144 checks ## Test Result  TBA ## Submission Checklist - [X] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit 4753424bc943669bb2a67c29783717042b457f9c Author: Alexey Sachkov <alexey.sachkov@amd.com> Date: Tue Apr 21 14:30:51 2026 +0200 [CLR][NFC] Optimize tracking of active queues (#4945) ## Motivation Performance optimization of the runtime. ## Technical Details `std::unordered_set::[insert|erase]` allocate/deallocate memory which is an unnecessary action considerng that a list of all command queues which are present throughout an application lifetime is mostly stable, i.e. the assumption is made that HIP streams are not short-lived, but instead mostly span the whole application lifetime. ## JIRA ID N/A ## Test Plan N/A, this change is non-functional ## Test Result N/A, this change is non-functional ## Submission Checklist - [X] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit 019054fdb5070b663ba741b6f3c8e715d4816a87 Author: Aleksandar Djordjevic <aleksandar.djordjevic@amd.com> Date: Tue Apr 21 14:06:28 2026 +0200 [rocprofiler-systems] filter VCN/JPEG sentinel values during metric collection (#5145) ## Motivation When only some VCN/JPEG engines are present on a GPU, the AMD SMI `vcn_activity`/`jpeg_activity` arrays contain `0xFFFF` sentinel values for absent engines. These sentinel values were being passed through to Perfetto and RocPD output as `65535`, appearing as bogus `device_vcn_activity_1`, `_2`, `_3` metrics. commit b6285c0a549a85c825d5e00fd51b3fb79f42c71d Author: Pedro Alves <pedro@palves.net> Date: Fri Apr 17 16:06:55 2026 +0100 Add missing AMD_DBGAPI_EXPORT_DECORATOR fallback The AMD_DBGAPI_EXPORT_DECORATOR definition block is missing the fallback case. Fix that. Change-Id: Idcc785b59a48b5cc17c64d4b70e534ba6e912eaf commit-id:8ab06a04 commit b1ac485c994985b67f1dd712693c81b614d0bf12 Author: Pedro Alves <pedro@palves.net> Date: Fri Apr 17 16:06:55 2026 +0100 Fix building dbgapi as a static library Since 4bfddf5494 ("dbgapi/win32: __declspec(dllexport) instead of --version-script"), building dbgapi as a static library on Windows fails with errors such as: .../rocm-dbgapi/src/initialization.cpp:41:1: error: function 'amd_dbgapi_status_t amd_dbgapi_initialize(amd_dbgapi_callbacks_s*)' definition is marked dllimport 41 | amd_dbgapi_initialize (struct amd_dbgapi_callbacks_s *callbacks) | ^~~~~~~~~~~~~~~~~~~~~ The symbols end up "marked dllimport" because with: set_target_properties(amd-dbgapi PROPERTIES ... DEFINE_SYMBOL "AMD_DBGAPI_EXPORTS" ... ... DEFINE_SYMBOL is a CMake convenience that only applies when (and while) building a shared library. When BUILD_SHARED_LIBS is OFF, DEFINE_SYMBOL does nothing. So AMD_DBGAPI_EXPORTS ends up not defined in amd-dbgapi.h, which results in AMD_DBGAPI mapping to AMD_DBGAPI_IMPORT, here, in amd-dbgapi.h.in: #define AMD_DBGAPI AMD_DBGAPI_IMPORT AMD_DBGAPI_IMPORT in turn is unconditionally defined to __declspec(dllimport) on Windows. So there are two issues: - The dbgapi build itself is treated as importing symbols (as a consumer) instead of exporting them (as a producer). - The import/export decorations are applied even when dbgapi is built as a static library, where they should not be used at all. Fix this by teaching amd-dbgapi.h about whether dbgapi was built as a static or shared library. Since the header is generated, define AMD_DBGAPI_STATIC_LIB (0 or 1) at configure time and use it to skip using import/export decorations. This also ensures that clients, including those which do not use CMake (e.g. GDB), automatically see consistent definitions without needing to know how dbgapi was built. Change-Id: I31c8c947791551cc4e20218826561c1d242f298d commit-id:12deacf2 commit e42ed4ca16c7713233161740b0a22d9c50fd2fba Author: Luis Machado <luis.machado@amd.com> Date: Mon Apr 13 09:50:10 2026 -0500 Add Python executable paths to all TheRock builds Configure shared and dist Python executables (3.10-3.14) for all build configurations to ensure consistent Python environment across CI builds. Also update the manylinux container hash to the latest. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> commit da8621629e24cabc4c92d43e2cbeb6cebc751c1c Author: vstojilj <vstojilj@amd.com> Date: Tue Apr 21 10:56:59 2026 +0200 Accept only valid device allocations for hipTexObjectCreate (#4814) ## Motivation Unit_TexObjectCreate_TypePitch2D_EdgeCases is failing on APUs when host pointer is passed to hipTexObjectCreate and it should fail with hipErrorInvalidValue. ## Technical Details Check if buffer retrieved by devptr is a valid allocation before creating tex object. ## JIRA ID Resolves AIRUNTIME-72. ## Test Plan Retest with Unit_TexObjectCreate_TypePitch2D_EdgeCases. ## Test Result test passes. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit d4570465bfb80c34171815318ec822f1dbcc07be Author: systems-assistant[bot] <221163467+systems-assistant[bot]@users.noreply.github.com> Date: Mon Apr 20 22:31:38 2026 -0700 Update TheRock reference to (702d158) (#5117) Updated TheRock ref to `702d158` due to submodule bump update - Rename ccache presets as per therock pr https://github.com/ROCm/TheRock/pull/4402 --------- Co-authored-by: therockbot <therockbot@amd.com> Co-authored-by: rahulc-gh <rahul.choudhary@amd.com> Co-authored-by: RahulC <168588218+rahulc-gh@users.noreply.github.com> commit e5a46360d85ac27fc2d028a1f094eb0c9292ec59 Author: Shweta Khatri <shweta.khatri@amd.com> Date: Thu Apr 2 14:56:05 2026 -0400 rocr: PC Sampling : Add reserved_wg field to PC sampling structs Add missing reserved_wg field to perf_sample_hosttrap_v1_t and perf_sample_snapshot_v1_t. This aligns with the profiler which expects chiplet at bits [10:8]. commit b768f127de8c219e1604615293aa68d6d2220e85 Author: Jeff Jiang <142832361+jeffqjiangNew@users.noreply.github.com> Date: Mon Apr 20 22:21:55 2026 -0400 [rocDecode] Merge debug logging into the new logging mechanism. (#5210) ## Motivation This PR merges debug logging into the new logging mechanism. ## Technical Details - rocDecode debug message was a compile option. - Now debug logging can be enabled by setting env var ROCDEC_LOG_LEVEL to 4 . This helps issue triage on the release builds. ## JIRA ID   ## Test Plan Run CTests and decode conformance tests to catch any regressions. Run rocdecdecode samples to verify user set output. ## Test Result All CTests and conformance tests pass. Example summary of the test results: CTests: Start 1: video_decodeRaw-HEVC 1/6 Test #1: video_decodeRaw-HEVC ............. Passed 2.10 sec Start 2: video_decodeRaw-AVC 2/6 Test #2: video_decodeRaw-AVC .............. Passed 1.96 sec Start 3: video_decodeRaw-AV1 3/6 Test #3: video_decodeRaw-AV1 .............. Passed 2.03 sec Start 4: video_decodeRaw-VP9 4/6 Test #4: video_decodeRaw-VP9 .............. Passed 2.05 sec Start 5: rocdec_Decode-HEVC 5/6 Test #5: rocdec_Decode-HEVC ............... Passed 0.91 sec Start 6: rocDecode_Negative_API_Tests 6/6 Test #6: rocDecode_Negative_API_Tests ..... Passed 0.73 sec Conformance tests: AVC conformance test: Conformance test completed on the 127 streams: The number of passing streams is 127 The number of failing streams is 0 The number of streams that did not finish decoding is 0 HEVC conformance test: Conformance test completed on the 135 streams: The number of passing streams is 135 The number of failing streams is 0 The number of streams that did not finish decoding is 0 AV1 conformance test: Conformance test completed on the 162 streams: The number of passing streams is 162 The number of failing streams is 0 The number of streams that did not finish decoding is 0 VP9 conformance test: Conformance test completed on the 108 streams: The number of passing streams is 108 The number of failing streams is 0 The number of streams that did not finish decoding is 0  ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit ff1db549b2d5061ff0026ee9933c0e83b5533fef Author: koushikbillakanti-amd <koushik.billakanti@amd.com> Date: Mon Apr 20 21:06:48 2026 -0500 Add NPU device type support (#4952) ## Motivation Add NPU device support to CUID so RyzenAI/amdxdna devices are discoverable and assigned unique identifiers. ## Technical Details Introduce CuidNpu class with discovery via accel, primary/derived CUID generation, PCI config space fingerprinting, and wire NPU through the daemon, CLI, device manager, and file serialization paths. Co-authored-by: Koushik Billakanti <kbillaka@amd.com> commit 4e965940280c5b567d4ad36dc41418f6b2da8593 Author: Kian Cossettini <Kian.Cossettini@amd.com> Date: Mon Apr 20 18:12:14 2026 -0400 [rocprofiler-systems] Improve troubleshooting messages for Dyninst errors (#4947) ## Motivation  During instrumentation, it is possible for Dyninst to fail to process certain sequences of instructions in a function. When this occurs, a vague error message would be printed. This PR adds more details to the error message and adds information to `instrumenting-rewriting-binary-applications.rst` to inform the user to exclude the problematic function/module/object in subsequent runs (which may prevent errors). With this PR: ``` # Before Failed to transform trace 9105 # After Failed to transform trace 9105 in function 'llvm::msgpack::Reader::read' in module '/opt/rocm-7.1.1/lib/llvm/lib/libomptarget.so.20.0git' in object 'libomptarget.so.20.0git' at address 0x7f4c3f179120 ``` ## Technical Details  See the original issue: https://github.com/dyninst/dyninst/issues/2203 Merge https://github.com/ROCm/dyninst/pull/13. Update submodule pointer to https://github.com/ROCm/dyninst/commit/77203e0b4f917aec534ac2717643c17838fded7e. Added `Failed to transform trace error` section to `projects/rocprofiler-systems/docs/how-to/instrumenting-rewriting-binary-application.rst`. ## JIRA ID   ROCM-21219 ## Test Plan  N/A ## Test Result  N/A ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. commit e0fa30b1465526e3255df950df5edc8dfd83b631 Author: Bindhiya Kanangot Balakrishnan <Bindhiya.KanangotBalakrishnan@amd.com> Date: Mon Apr 20 16:56:59 2026 -0500 [ROCM-1553] GPU metrics v1.9 support with new attributes (#5125) [ROCM-1553] GPU metrics v1.9 support with new attributes Added support for updated dynamic GPU metrics version 1.9 by introducing new metric attributes based on latest amdgpu header. Updated python interface and CLI outputs for new gpu_metric fields. Unified CLI output for amdsmi metric. Updated Changelog. Signed-off-by: Bindhiya Kanangot Balakrishnan <Bindhiya.KanangotBalakrishnan@amd.com> commit 6c571bb686712f9366fa8ee8646e48fdf25222c5 Author: Jeff Jiang <142832361+jeffqjiangNew@users.noreply.github.com> Date: Mon Apr 20 14:43:02 2026 -0400 [rocDecode] Logging format in utility classes now matches RocDecLogger format. (#4997) ## Motivation Unify logging format in utility classes with core library logging format. ## Technical Details This PR unifies logging format in utility classes with core library logging format. - Logs from utility classes are now more informative. ## JIRA ID   ## Test Plan Run CTests and decode conformance tests to catch any regressions. Run rocdecdecode samples to verify user set output. ## Test Result All CTests and conformance tests pass. Example summary of the test results: CTests: Start 1: video_decodeRaw-HEVC 1/6 Test https://github.com/ROCm/rocm-systems/pull/1: video_decodeRaw-HEVC ............. Passed 1.96 sec Start 2: video_decodeRaw-AVC 2/6 Test https://github.com/ROCm/rocm-systems/pull/2: video_decodeRaw-AVC .............. Passed 1.84 sec Start 3: video_decodeRaw-AV1 3/6 Test https://github.com/ROCm/rocm-systems/pull/3: video_decodeRaw-AV1 .............. Passed 1.88 sec Start 4: video_decodeRaw-VP9 4/6 Test https://github.com/ROCm/rocm-systems/pull/4: video_decodeRaw-VP9 .............. Passed 1.93 sec Start 5: rocdec_Decode-HEVC 5/6 Test https://github.com/ROCm/rocm-systems/pull/5: rocdec_Decode-HEVC ............... Passed 0.85 sec Start 6: rocDecode_Negative_API_Tests 6/6 Test https://github.com/ROCm/rocm-systems/pull/6: rocDecode_Negative_API_Tests ..... Passed 0.72 sec 100% tests passed, 0 tests failed out of 6 AVC conformance test: Conformance test completed on the 127 streams: The number of passing streams is 127 The number of failing streams is 0 The number of streams that did not finish decoding is 0 HEVC conformance test: Conformance test completed on the 135 streams: The number of passing streams is 135 The number of failing streams is 0 The number of streams that did not finish decoding is 0 AV1 conformance test: Conformance test completed on the 162 streams: The number of passing streams is 162 The number of failing streams is 0 The number of streams that did not finish decoding is 0 VP9 conformance test: Conformance test completed on the 108 streams: The number of passing streams is 108 The number of failing streams is 0 The number of streams that did not finish decoding is 0 ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com> commit 2e1eb97edb6d9b0545c2ce3f5e900131540da0c7 Author: Avinash Kethineedi <avinash.kethineedi@amd.com> Date: Mon Apr 20 13:27:03 2026 -0500 [rocSHMEM] Add library metadata constants, version/name query APIs, and documentation (#5151) ## Motivation Add library metadata constants and version/name query APIs to rocSHMEM, aligning with the OpenSHMEM specification. ## Technical Details - Replace version parsing from the C++ header with a direct CMake variable and add backend/vendor string construction to CMakeLists.txt - Add vendor version, vendor string, and backend metadata to rocshmem_config.h and surface them in rocshmem_info output - Add ROCSHMEM_MAJOR_VERSION, ROCSHMEM_MINOR_VERSION, ROCSHMEM_MAX_NAME_LEN to rocshmem.hpp - Add host APIs: rocshmem_info_get_version, rocshmem_info_get_name, rocshmem_vendor_get_version_info - Update rocshmem_init_attr_test example to use the new query APIs - Add Sphinx documentation for library constants and query APIs - Templatize print_entry in build_info.cpp to support int arguments - Update docs/conf.py to read version from CMakeLists.txt - Add Linux build instructions and custom output directory usage to docs README ## JIRA ID - AIROCSHMEM-359 ## Test Plan - Build with all backend configurations (GDA with IONIC/BNXT/MLX5, RO, IPC, and combinations) and verify rocshmem_info output shows the correct vendor string, version, and backend for each configuration - Run rocshmem_init_attr_test and verify version/name output on rank 0 - Build the Sphinx documentation and verify the new library constants and query API pages render correctly ## Test Result - All backend configurations produce the expected vendor strings (e.g. `rocSHMEM_v3.3.0_GDA(IONIC/BNXT/MLX5)+RO+IPC` when compiled with all backends and NICs enabled) - rocshmem_info output displays correct version, vendor string, and backend metadata for all tested configurations - Sphinx documentation builds without warnings in the newly added sections; library constants and query API pages render correctly commit b87cea434d77e56499835c54c01db8a5fa51f13b Author: Sumanth Gavini <sumanth.gavini@amd.com> Date:…

Revert "[SWDEV-544298] AINIC implementation (#2798)"

a852dd1

This reverts commit 00e9b8d.

github-actions Bot added the project: amdsmi label Feb 20, 2026

JeniferC99 mentioned this pull request Feb 20, 2026

[SWDEV-544298] AINIC implementation #2798

Merged

1 task

systems-assistant Bot added the organization: ROCm label Feb 20, 2026

ScottTodd mentioned this pull request Feb 20, 2026

[Issue] Linux Packages nightly failed to build amdsmi with error libamdsminic.a": Invalid cross-device link ROCm/TheRock#3500

Closed

Revert "[ROCM-2942] Fix AMDSMI_STATUS_NO_DATA on NIC devices & tests …

f5e61f0

…not completing (#3289)" This reverts commit 3c85365.

marifamd marked this pull request as ready for review February 20, 2026 22:38

marifamd requested a review from a team as a code owner February 20, 2026 22:38

dmitrii-galantsev changed the title ~~Revert "[SWDEV-544298] AINIC implementation (#2798)"~~ Revert '[SWDEV-544298] AINIC implementation [#2798]' Feb 20, 2026

dmitrii-galantsev approved these changes Feb 20, 2026

View reviewed changes

JeniferC99 merged commit bfe5604 into develop Feb 20, 2026
30 of 33 checks passed

JeniferC99 deleted the jechrist/revertPR2798 branch February 20, 2026 23:59

dmitrii-galantsev mentioned this pull request Feb 21, 2026

[ci] Use github-script instead of curl #3424

Merged

2 tasks

JeniferC99 mentioned this pull request Feb 21, 2026

Bump rocm-systems from a21f844 to f0d733e ROCm/TheRock#3534

Merged

dmitrii-galantsev mentioned this pull request Feb 21, 2026

[AMD-SMI] Allow amdsmi init to succeed when no NIC hardware is present #3403

Merged

1 task

dmitrii-galantsev mentioned this pull request Mar 2, 2026

[AMD-SMI] Clean up SMI #3655

Merged

Navigation Menu

Uh oh!

Revert '[SWDEV-544298] AINIC implementation [#2798]' - #3418

Revert '[SWDEV-544298] AINIC implementation [#2798]'#3418
JeniferC99 merged 2 commits into
developfrom
jechrist/revertPR2798

JeniferC99 commented Feb 20, 2026 •

edited

Loading

Uh oh!

ScottTodd commented Feb 20, 2026

Uh oh!

JeniferC99 commented Feb 20, 2026

Uh oh!

marifamd commented Feb 20, 2026 •

edited

Loading

Uh oh!

dmitrii-galantsev commented Feb 20, 2026

Uh oh!

JeniferC99 commented Feb 20, 2026

Uh oh!

ScottTodd commented Feb 20, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

JeniferC99 commented Feb 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

Uh oh!

ScottTodd commented Feb 20, 2026

Uh oh!

JeniferC99 commented Feb 20, 2026

Uh oh!

marifamd commented Feb 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

dmitrii-galantsev commented Feb 20, 2026

Uh oh!

JeniferC99 commented Feb 20, 2026

Uh oh!

ScottTodd commented Feb 20, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

JeniferC99 commented Feb 20, 2026 •

edited

Loading

marifamd commented Feb 20, 2026 •

edited

Loading