- two copies of HSA table: internal and tracing
- internal is used to invoke HSA function without any possibility of triggering tracing, etc.
[ROCm/rocprofiler-sdk commit: bc9f86ec62]
The function Init() called by one of the constructors of lazy_ptr is undefined.
Replacing with reset method sets the object to an uninitialized state and assigns a new constructor function
Fix submitted on github by zhoumin2 - https://github.com/ROCm/ROCR-Runtime/pull/184
Change-Id: I7d906d526ce7fe7e2548b01810e6395b13497bf3
Add pc sampling support.
Change-Id: I08199024ba5a8eb2845c048d499fc8fcd260d2e8
Signed-off-by: David Yat Sin <David.YatSin@amd.com>
Signed-off-by: James Zhu <James.Zhu@amd.com>
Add pc sampling support
Change-Id: I2c472ce00ff8648904cf7e585687e81d3f493049
Signed-off-by: David Yat Sin <David.YatSin@amd.com>
Signed-off-by: James Zhu <James.Zhu@amd.com>
* Fix agent node id + randomize offset id
- fixes the node_id value
- randomizes a constant offset for the id.handle values
- switch to using node ids in rocprofiler-sdk-tool library
- update tests related to agents
* Logical node id
- sequential node id values from 0 to (N-1) where N is the number of agents
[ROCm/rocprofiler-sdk commit: 1addfed9f6]
* Fix agent node id + randomize offset id
- fixes the node_id value
- randomizes a constant offset for the id.handle values
- switch to using node ids in rocprofiler-sdk-tool library
- update tests related to agents
* Logical node id
- sequential node id values from 0 to (N-1) where N is the number of agents
* Handle hsa_queue_destroy after finalization
- fixes issue where hsa_queue_destroy(...) is invoked after rocprofiler-sdk has finalized
- hsa::get_queue_controller() returns pointer
- if queue controller is a null pointer, skip invoking QueueController::destroy_queue
* Update HIP/HSA/marker update_table logging
* Update rocprofv3 tests
- remove HSA_TOOLS_LIB env variable
- remove setting ROCPROFILER_LOG_LEVEL env variable
- add timeouts to tests which are missing them
* Disable thread sanitizer deadlock detection
* Update CI workflow
- rename vega20-ubuntu job to core-ci
- enable navi32 in core-ci and sanitizers
* Update run-ci.py
- set gcovr html medium and high threshold
* Update lib/rocprofiler-sdk/hsa/queue_controller.cpp
- remove this capture from enable/disable serialization
* Update lib/rocprofiler-sdk/hsa/{hsa_barrier,profile_serializer}.*
- hsa_barrier::set_barrier accepts const-ref to queue map
- profile_serializer::enable and profile_serializer::disable accept const-ref to queue map
* Logging for HIP/HSA/marker/profile_serializer
* Logging for HIP/HSA/marker/queue_controller
* Improve test_retired_correlation_ids asserts
* Fix tests/counter-collection/validate.py
- scale expected SQ_WAVES counter value based on warp size of GPU
* Tweak github comment for code coverage
* Remove gcovr html high/medium threshold args
* Fix tests/counter-collection/validate.py
- round before casting to int in test_counter_values
* operator bool for profile_serializer
- only wait on CV if profile_serializer is used
* Logging updates (profile_serializer + code_object)
* Update counter-collection validate.py
* QueueController does not wait on CV if finalizing/finalized
* Update CI workflow
- remove navi32 from core job
* Improve HIP/HSA/marker tracing get_functor/functor
- remove lambda wrapper around functor
* Update lib/rocprofiler-sdk/hsa/queue_controller.cpp
- do not acquire cvmutex lock during finalization
* Update lib/rocprofiler-sdk/hsa/hsa_barrier.*
- move ctor and dtor to implementation
- skip signal store screlease and destroy if already finalized
* Update CI workflow
- remove navi32 runners
* bwelton fixes for hangs
* CMake improvements + simplified demangle
- remove amd-comgr from common target (and thus removed from roctx DT_NEEDED)
---------
Co-authored-by: Benjamin Welton <bewelton@amd.com>
[ROCm/rocprofiler-sdk commit: 2f9b1767e9]
* Handle hsa_queue_destroy after finalization
- fixes issue where hsa_queue_destroy(...) is invoked after rocprofiler-sdk has finalized
- hsa::get_queue_controller() returns pointer
- if queue controller is a null pointer, skip invoking QueueController::destroy_queue
* Update HIP/HSA/marker update_table logging
* Update rocprofv3 tests
- remove HSA_TOOLS_LIB env variable
- remove setting ROCPROFILER_LOG_LEVEL env variable
- add timeouts to tests which are missing them
* Disable thread sanitizer deadlock detection
* Update CI workflow
- rename vega20-ubuntu job to core-ci
- enable navi32 in core-ci and sanitizers
* Update run-ci.py
- set gcovr html medium and high threshold
* Update lib/rocprofiler-sdk/hsa/queue_controller.cpp
- remove this capture from enable/disable serialization
* Update lib/rocprofiler-sdk/hsa/{hsa_barrier,profile_serializer}.*
- hsa_barrier::set_barrier accepts const-ref to queue map
- profile_serializer::enable and profile_serializer::disable accept const-ref to queue map
* Logging for HIP/HSA/marker/profile_serializer
* Logging for HIP/HSA/marker/queue_controller
* Improve test_retired_correlation_ids asserts
* Fix tests/counter-collection/validate.py
- scale expected SQ_WAVES counter value based on warp size of GPU
* Tweak github comment for code coverage
* Remove gcovr html high/medium threshold args
* Fix tests/counter-collection/validate.py
- round before casting to int in test_counter_values
* operator bool for profile_serializer
- only wait on CV if profile_serializer is used
* Logging updates (profile_serializer + code_object)
* Update counter-collection validate.py
* QueueController does not wait on CV if finalizing/finalized
* Update CI workflow
- remove navi32 from core job
* Improve HIP/HSA/marker tracing get_functor/functor
- remove lambda wrapper around functor
* Update lib/rocprofiler-sdk/hsa/queue_controller.cpp
- do not acquire cvmutex lock during finalization
* Update lib/rocprofiler-sdk/hsa/hsa_barrier.*
- move ctor and dtor to implementation
- skip signal store screlease and destroy if already finalized
* Update CI workflow
- remove navi32 runners
* bwelton fixes for hangs
* CMake improvements + simplified demangle
- remove amd-comgr from common target (and thus removed from roctx DT_NEEDED)
---------
Co-authored-by: Benjamin Welton <bewelton@amd.com>
After the dead process is detected, pthread_mutex_consistent() will
be called. After that, the pthread_mutex_unlock() should also be
called to unlock it: "It is the responsibility of the application to
recover the state so it can be reused."
Change-Id: I45d3e2e68c3b06779f3acb1e908dbec0c6a39297
Giovanni LB
gobhardw
Konstantin Zhuravlyov
Benjamin Welton
Jonathan R. Madsen
Shweta.Khatri
Mihai Preda
James Zhu
Ammar ELWazir
David Yat Sin
Bill(Shuzhou) Liu
Karl W. Schulz