* Data imputation strategy for iteration multiplexing
* Implement data imputation methodology to handle missing counter values
in case of iteration multiplexing
* Enable dispatch filtering with iteration multiplexing since we are no
longer merging dispatches
* Bugfix to prevent check for missing counter values when using csv
format when profiling with iteration multiplexing
* Move warning and info message in case of iteration multiplexing to
sanitize function which comes earlier in analyze mode
* Address review comments
* Fix typo in documentation
* Move profiling config init. after path check in sanitize()
* Graceful handling of dispatches with all counters empty within data
imputation logic
* Improve info message for iteration multiplexing based analysis
* Ensure proper error message when trying to run iteration multiplexing with attach/detach
* fix test case
Implements automatic device wake using getDRMDeviceId() DRM call when GPUs
are detected in low-power state. This ensures rocm-smi can access device
information on suspended GPUs.
Signed-off-by: Bindhiya Kanangot Balakrishnan <Bindhiya.KanangotBalakrishnan@amd.com>
Problem:
The existing SDMA engine selection logic had several issues:
1. Same VirtualGPU/stream could use different SDMA engines for consecutive
async copies since copy_engine_status may report engines as busy
2. Busy and Preferred engine check for every copy
3. No global tracking of which VirtualGPU uses which engine, leading to
suboptimal resource allocation
Solution:
Implemented a global SDMA engine allocator with per-stream affinity:
- Added Device::SdmaEngineAllocator to manage VirtualGPU → engine assignments
* Maintains global map of active assignments
* Enforces exclusivity: different streams use different engines (except
inter-GPU copies where preferred engines are prioritized for optimal
hardware paths like XGMI links)
* Thread-safe allocation/release with Monitor lock
- Modified VirtualGPU to cache assigned engine locally (assigned_sdma_engine_)
for fast lookup without map access on hot path
- Refactored rocrCopyBuffer() to:
1. Check local cached engine first → use if assigned
2. Call AllocateSdmaEngine() if not assigned → cache result
- Moved HSA API queries (memory_copy_engine_status, memory_get_preferred_copy_engine)
into AllocateEngine() for cleaner separation of concerns
- Engine release on HostQueue::finish() instead of only VirtualGPU destruction
* Improves engine utilization by releasing earlier
* Added virtual ReleaseSdmaEngines() method to device::VirtualDevice
- Added future path for simple round-robin allocation (kUseSimpleRR) for
next-gen GPUs with uniform SDMA bandwidth (disabled by default)
Cleanup:
- Removed selectSdmaEngine() helper (logic moved to allocator)
- Removed getSdmaRWMasks() (allocator accesses maxSdmaReadMask_/WriteMask_ directly)
- Removed unused sdmaEngineReadMask_/WriteMask_ member variables from DmaBlitManager
Benefits:
- Ensures consistent per-stream SDMA engine usage
- Prevents cross-stream contention and engine thrashing
- Prioritizes hardware-optimal paths for inter-GPU transfers
- Better resource utilization through earlier release
- Cleaner, more maintainable code structure
- Fixes SWDEV-559349
- Fix build failure caused by correct libunwind not being found in some environments.
- Updated the `timemory` submodule to commit `24407d37ab85c46ba6c18fba9498320f825ee4e4 `.
* Use static catch2.lib instead of catch2.dll
Using catch2.dll incraeses execution time by 12x
* handle debug option for static catch2
* SWDEV-573539 - skip atomics on windows since its taking a very long time to execute
mlsejenkins needs newer cmake but compiler breaks with newer versions
so skipping on windows can be a workaround for now
---------
Co-authored-by: Joseph Macaranas <145489236+jayhawk-commits@users.noreply.github.com>
**Thread limit configuration and enforcement: **
* Added a check in `CMakeLists.txt` to ensure `ROCPROFSYS_MAX_THREADS` is at least 128, automatically setting it to 128 with a warning if a lower value is provided.
* Replaced hardcoded thread limit (`allowed_max_threads`) in `pthread_create_gotcha.cpp` with the configurable `ROCPROFSYS_MAX_THREADS` value, ensuring all runtime checks and warnings use the actual configured limit.
**Documentation improvements: **
* Updated the development guide to explain the new thread limit behavior, including how exceeding the limit is handled gracefully, how to configure it, and the build-time validation rules.
**Test updates: **
* Modified thread limit tests to use the configurable `ROCPROFSYS_MAX_THREADS` value instead of a hardcoded limit and expanded the range of tested thread values.
* Increased test timeouts to accommodate larger thread counts and ensure reliability with higher limits.
This has started failing on various developer build systems. Looking at it, it is not precisely clear how this ever worked given that nothing appears to be adding the DRM include dirs.
I'd prefer that we remove this delay loading (at least for TheRock builds where it is never needed), but in the meantime, this does fix the issue and is verified on an affected system.
Fixes https://github.com/ROCm/TheRock/issues/2744
* SWDEV-508225 - do not assert() after calling digestFatBinary() if it fails. Otherwise this causes assertions to trigger easily in systems that have an APU and a discrete GPU and the code was compiled for the discrete one
* SWDEV-508225 - fix that when using a non-existent ordinal in HIP_VISIBLE_DEVICES, getCurrentArch() would crash
fix: Add gpu_metrics 1.0 support which is still used by some hardware
Code changes related to the following:
* APIs
* Unit tests
Signed-off-by: Oliveira, Daniel <daniel.oliveira@amd.com>
* Add support for get and set APIs for CPUISOFreqPolicy and DFCState Control
- Add support for get and set APIs for CPUISOFreqPolicy and DFCState Control
in AMD SMI and also in the CLI tool
* CHANGELOG.md file updated
* SWDEV-562837: Update amdsmi-py-api.md as per the new APIs
Updated amdsmi-py-api.md as per the new APIs added.
---------
Signed-off-by: Soumya <sranjanr@amd.com>
Signed-off-by: gabrpham <Gabriel.Pham@amd.com>
Co-authored-by: Saka Sitharammurthy <SitharamMurthy.Saka@amd.com>
* SWDEV-569319 Replace ScopedAcquire with stdcpp wrappers
* Remove KernelMutex and KernelSharedMutex abstractions with std::mutex and std::shared_mutex
* Replaced unique_locks with lock_guards
* More changes
* Replace new and deletes with smart pointers
* Replaced some more with shared ptrs
* Replacements with smart pointers - pt 2
* missed change
* Remove MFMA functionality in rocflop sample since its not supported in MI50
* Add gfx arc based support for MFMA and SMFMAC in rocflop.cpp
* Add --int32 usage doc
* Address review comments
* Fix buffer tracing synchronization lock
- PR #529 (in rocprofiler-sdk-internal) introduced waiting on the syncer flag when emplacing in a buffer to prevent the overwriting buffer records currently being processed in a buffer flush callback
- The above fix introduced a block on the both buffers when a buffer flush callback was being executed instead of a block on the buffer being flushed.
* Add rocpd tests for duplicate records
* Address code review comments
* Update rocprofiler workflows to use new runner naming for mi325
* Add input options to workflow_dispatch for rocprofiler-systems CI workflow
* Update runner name on therock-ci-linux.yml as well
`tests/rocprofiler/rocprofiler.cpp` uses `std::string` without including `<string>` directly.
This works with libstdc++ due to transitive includes, but fails with libc++.
Closes#1240
* Fix merging logic for multi process
* Fix dispatch id reset logic in case of rocpd format
* Fix kernel id reset logic in case of csv format
* Revert correlation logic change in csv format
* Do inner join instead of left join
* Added tool for dumping counter and metric values
* Skip Linting
* Added support for iteration multiplexing
* Remove subparser and supress compute options
* Specify output dir
* Add kernel info
* csv name change
* Added comments
* Support dispatch id-less dataframes
* Formatting fix
* Add default for path
* Print help with no args
* Support only single workload
Fixed incorrect error code expectation in FrequenciesRead
test when calling amdsmi_get_gpu_pci_bandwidth() with nullptr
parameter.
Signed-off-by: Bindhiya Kanangot Balakrishnan <Bindhiya.KanangotBalakrishnan@amd.com>
vedithal-amd
systems-assistant[bot]
Bindhiya Kanangot Balakrishnan
koushikbillakanti-amd
Alexandra Sidorova
SaleelK
Flora Cui
David Galiffi
anujshuk-amd
Ajay GunaShekar
Alysa Liu
David Yat Sin
Aleksandar Djordjevic
Godavarthy Surya, Anusha
Stella Laurenzo
Yazen AL Musaffar
Gerardo Hernandez
Daniel Oliveira
SakaSitharammurthy
pghoshamd
Jonathan R. Madsen
habajpai-amd
Jason Bonnell
AidanBeltonS
Joseph Macaranas
Benjamin Welton
Julia Jiang
Shadi Dashmiz
jamessiddeley-amd
Swati Rawat
Sv. Lockal
Ioannis Assiouras
Joseph Narlo
Jimbo
Adam Pryor
German Andryeyev
abchoudh-amd
marantic-amd