Repository navigation
refactor(rocrtst): remove hwloc dependency from memory async copy tests - #8723
Conversation
✅ All Policy Checks Passed
📖 Need help? See the Policy FAQ for details on every check and how to fix failures. |
|
🚫 Please fix the failed policies before requesting reviews. The following policy checks failed:
The |
cfreeamd
left a comment
cfreeamd
left a comment
There was a problem hiding this comment.
🤖 AI-assisted pre-review (for @cfreeamd)
Reviewed the ROCr-scoped changes (rocrtst memory_async_copy). The hwloc removal, CMake cleanup, and dropping the old hwloc hsa_shut_down()/hsa_init() teardown hack in Close() all look correct. One correctness concern on the NUMA local/remote classification left inline.
|
This pull request has been inactive for 25 days and will be marked as stale. If you would like to keep this PR open, please:
This PR will be automatically closed in 5 days if no further activity occurs. |
|
This pull request has been automatically closed due to inactivity (30 days with no updates). If you'd like to continue working on this, feel free to reopen the PR or create a new one. |
Use HSA_AGENT_INFO_NODE for NUMA topology instead of hwloc PCIe lookup, drop vendored hwloc headers/libs, and switch the optional NUMA test to libnuma. Co-authored-by: Cursor <cursoragent@cursor.com>
If no GPU shares a NUMA node with any CPU agent, assign the first CPU and GPU agents so Memory_Async_Copy can run on single-socket systems. Co-authored-by: Cursor <cursoragent@cursor.com>
1ebcfeb to
78a98bd
Compare
HSA_AGENT_INFO_NODE returns the KFD topology node id, which is unique per agent, so a CPU and a GPU never share it. The local-vs-remote comparison in Memory_Async_Copy was therefore always unequal: the "local GPU" branch was dead and classification always fell through to the iteration-order fallback, silently losing the H2D-vs-H2DRemote distinction. Read the GPU's OS NUMA node from its PCI numa_node in sysfs and compare it against the CPU agent's NUMA node, restoring the real locality test that the removed hwloc PCI-ancestor lookup provided. No hwloc dependency is reintroduced. Co-Authored-By: Claude <noreply@anthropic.com>
TheRock Submodule Bump ActivityNewest first
|
Summary
HSA_AGENT_INFO_NODEfor NUMA topology inMemory_Async_Copyinstead of hwloc PCIe lookup.hwlocCMake discovery block.ENABLE_COPY_NUMA) to libnuma APIs when enabled.JIRA ID
JIRA ID : ROCM-1282
Test plan
rocrtst64on dayatsin-test-01 (gfx1100,/opt/rocm) without hwlocrocrtst64on e3-2b (gfx950, TheRock toolchain) without hwlocrocrtstPerf.Memory_Async_CopyandrocrtstPerf.Memory_Async_Copy_On_Engineon both systems — 2/2 passed on each host