* [Draft]: Add support for RCCL tracing
Address comments
* [Draft]: Add support for RCCL tracing
Address PR comments, changes from RCCL upstream
* Add RCCL library table registration
Working on adding support to rocprofiler-register
* Support compilation w/o <rccl/amd_detail/api_trace.h>
- dummy api_trace.h header
- return ROCPROFILER_STATUS_ERROR_NOT_IMPLEMENTED when RCCL does not have api_trace.h header
* RCCL API tracing tool support
- add to rocprofv3
- add to json-tool
---------
Co-authored-by: Jonathan R. Madsen <jonathanrmadsen@gmail.com>
* LD_PRELOAD librocprofiler-sdk-roctx.so when marker-trace enabled
- this enables apps to link against old ROCTx (libroctx64.so) but get marker tracing in rocprofv3
* Update CHANGELOG
* Validation test for app linked to old (roctracer) ROCTx library
* Tweak scope of tool_counter_info
- causing "signal-unsafe call inside of a signal" error for ThreadSanitizer on mi200
* Fix handling of missing transpose-roctracer-roctx
* Disable rocprofv3 aborted-app test (ThreadSanitizer)
- ThreadSanitizer + mi200/mi300 + aborted-app results in a signal-unsafe call inside a signal that cannot be specifically suppressed as usual via rocprofv3_error_signal_handler for some unknown reason
* Add UndefinedBehaviorSanitizer job
* Fix -d option broken by hostname
* Fix rocprofv3 output filename containing directory
* Fix TID handling in Perfetto and OTF2 output
* Revert changes which removed hostname
* Revise tests/rocprofv3/tracing output filenames
- specify an output filename for tests which include a subdirectory
---------
Co-authored-by: Jonathan R. Madsen <jonathanrmadsen@gmail.com>
* Move include/rocprofiler-sdk/cxx/details/delimit.hpp to tokenize.hpp
* Update docs/how-to/using-rocprofv3.rst
- fix code block indents
- reorder rocprofv3 options, limit them to important options
- add docs for `--runtime-trace`
* Update rocprofv3.py
- parser argument groups
- new `--runtime-trace` option
- new `--summary` option
- new `--summary-per-domain` option
- new `--summary-groups` option
- new `--summary-output-file` option
- new `--summary-units` option
* Update lib/rocprofiler-sdk/hsa/async_copy.cpp
- fix async copy operation names: add "MEMORY_COPY_" prefix
* lib/rocprofiler-sdk-tool: update statistics.{hpp,cpp}
- statistics<>::get_percent function
- stats_entry_t struct
- stats_formatter struct
- percentage struct
- std::to_string(::rocprofiler::tool::percentage)
* lib/rocprofiler-sdk-tool: update domain_type.{hpp,cpp}
- reorder domain_type enum values
* lib/rocprofiler-sdk-tool: update generateCSV.{hpp,cpp}
- separate writing CSV from accumulating statistics
- a lot of functionality was moved to statistics.{hpp,cpp}
* lib/rocprofiler-sdk-tool: update output_file.{hpp,cpp}
- output_stream_t struct
- get_output_stream(...) returns output_stream_t instance
* lib/rocprofiler-sdk-tool: update generateJSON.cpp
- update get_output_stream usage to output_stream_t
* lib/rocprofiler-sdk-tool: update generateOTF2.cpp
- header include order tweak
* lib/rocprofiler-sdk-tool: update buffered_output.hpp
- stats_data_t was renamed to stats_entry_t
* lib/rocprofiler-sdk-tool: update generatePerfetto.cpp
- header include tweak
* lib/rocprofiler-sdk-tool: update tmp_file_buffer.hpp
- emit warning message if write_ring_buffer fails after offloading instead of aborting
- prefer placement new instead of assignment in write_ring_buffer
* lib/rocprofiler-sdk-tool: add generateStats.{hpp,cpp}
- functions for accumulating statistics
* Update tests/rocprofv3/tracing-hip-in-libraries/CMakeLists.txt
- accommodate tweak to CSV output file name for HIP and HSA traces
* lib/rocprofiler-sdk-tool: update config.{hpp,cpp}
- new config variables
- stats_summary
- stats_summary_per_domain
- summary_output
- stats_summary_unit_value
- stats_summary_unit
- stats_summary_file
- stats_summary_groups
- support output keys for hostname: %hostname% / %h
* lib/rocprofiler-sdk-tool: update tool.cpp
- support summary output
* Documentation fixes
* Test for summary output
* Update tests/bin/transpose to use more ROCTx
- also support building with the roctracer ROCTx
* Remove roctxMark from OTF2 + fix kernel-rename tests
- following more ROCTx calls in transpose, kernel-rename validation had to be updated
* JSON metadata + JSON summary
- add serialization support for config
- add serialization support for statistics
- additions to json spec
- rocprofiler-sdk-tool/metadata/config
- rocprofiler-sdk-tool/metadata/command
- rocprofiler-sdk-tool/summary
- config output_keys support for NVIDIA %q{<ENV-VAR>} syntax
- config output_keys support keys within keys
* rocprofv3 --summary-groups warning if no domain matches
- emit warning if a regex in for summary groups did not match any domain names
* Compile fix for lib/rocprofiler-sdk-tool/tool.cpp
- get_config().scratch_memory_trace
- pass contributions to write_json
* Update rocprofv3.py to preload rocprofiler-sdk-roctx
- appended to LD_PRELOAD when args.marker_trace is enabled
* Fix ReST link errors about subtitle underline being too short
* Patch tokenization of config::stats_summary_groups
- guard against array values of empty strings
* Tweak rocprofv3 summary test
- input-summary.yaml (used by rocprofv3-test-summary-inp-yaml-execute) only provides one summary group regex
* Disable LD_PRELOAD of librocprofiler-sdk-roctx.so
- this causes problems in the sanitizers, will be addressed in another PR
* PC sampling: online partial PC sampling decoding
PC sampling service decodes a PC sample partially
by replacing the PC with an id of the loaded code object instance
containing PC and the offset of the PC within that code object instance.
* PC sampling: marker records removed
* PC sampling parser: minor doc update in mock
* PC sampling: introducing rocprofiler_pc_t
* NULL value of the code object id introduced.
* Clarifying documenation related to PC offset.
* PC offset documentation improvement
* PC sampling parser benchmark: Reducing the number of samples to recreate half of performance.
* Update HSA ABI checks for tracing
* Update lib/common/abi.hpp
- perform ABI versioning checks even when `ROCPROFILER_CI` is not defined (or ROCPROFILER_CI=0)
* Enforce versioning size for various HSA AmdExt step versions + hsa_amd_enable_logging support
* Minor HIP abi.cpp updates
- ROCPROFILER_API after function
- use rocprofiler_tracing_operation_t in lieu of uint32_t where appropriate
- rocprofiler_tracing_operation_t is not int32_t typedef (formerly uint32_t)
- use const T* instead of T* where appropriate
* Add kernel profiling time info to counter collection records
- lib/rocprofiler-sdk/kernel_dispatch
- added profiling_time.{hpp,cpp}
- restructured tracing.cpp
- updated queue.cpp AsyncSignalHandler
- gets kernel dispatch profiling time and passes to dispatch_complete and signal callbacks
- structured some header includes to reduce cyclic include probability
- originally, including kernel_dispatch/tracing.hpp in hsa/queue.hpp created a lot of cyclic includes
* Fix kernel_dispatch.cpp includes
* Fix kernel_dispatch.cpp
- include <cstring>
- replace use of ROCPROFILER_HSA_AMD_EXT_API_ID_NONE with ROCPROFILER_KERNEL_DISPATCH_LAST
* Tidying ATT dispatch API. ATT Agent to be initialized with rest of profiler. Removing read_index-based wait.
* Formatting
* Adding some input validation
* Add perf test for agent
* Removing async
* SQ Counter Documentation
Improve documentation of SQ counters. Attempts to
make what the counters are outputting (and where
applicable what the counter means in terms o
performance) more clear.
* pre-format
* Address comments + YAML formatting
* More definition fixes
---------
Co-authored-by: Benjamin Welton <ben@amd.com>
Co-authored-by: Gopesh Bhardwaj <gopesh.bhardwaj@amd.com>
* look for symbols in dynsym table
* checking both symtab and dynsym
* Avoid symbol duplication in non stripped binaries
* clang-format
* Minor elf_utils.cpp updates
- use 'else if' instead of 'if'
- logging tweaks
* Update registration
- tweak logging
* Update testing
- strip the rocprofiler-sdk-c-tool library
- add test-c-tool-rocp-tool-lib-execute test which does NOT LD_PRELOAD the library (uses only ROCP_TOOL_LIBRARIES instead)
---------
Co-authored-by: Jonathan R. Madsen <jonathanrmadsen@gmail.com>
* Convert counter def format to YAML
Converts counter definition format to YAML with the
following structure:
```yaml
COUNTER_NAME:
architectures:
gfxXX: // Can be more than one, / deliminated if they share idential data
block: <Optional>
event: <Optional>
expression: <optional>
description: <Optional> // In case per arch notes are needed
gfxYY:
...
description: General counter desctiption
```
All counters (derived and hardware) are now defined
in the same file for ease of future additions/subtractions.
Removes existing XML parser. Keeps the existing XML
definitions for now (since other tools still rely on
its presence).
* Update lib/rocprofiler-sdk/agent.cpp
- use static_object wrapper for vector of agent_pair (rocp agent <-> hsa agent)
* Fix get_aql_handles() shutdown error
- use `static_object` wrapper for vector of `aqlprofile_agent_handle_t`
- missing-new-line CI job: ensures all source files end with new line
- logging updates
- add new line to the end of many files
- fix header include ordering is misc places
- transition to use hsa::get_core_table() and hsa::get_amd_ext_table() in various places instead of making copies
* General fixes to ATT, packets and event ID retrieval
* Update source/lib/rocprofiler-sdk/hsa/aql_packet.hpp
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
- source/lib/rocprofiler-sdk/hsa/queue.cpp
- Optimize WriteInterceptor to eliminate extra barrier packets causing gaps between kernels in kernel tracing
- increase timeout_hint in hsa_signal_wait in set_profiler_active_on_queue
- misc logging improvements
- source/lib/rocprofiler-sdk/counters/agent_profiling.cpp
- increase timeout_hint in hsa_signal_wait in set_profiler_active_on_queue
- tests/rocprofv3/hsa-queue-dependency/CMakeLists.txt
- add TIMEOUT for rocprofv3-test-hsa-multiqueue-execute
The following changes are introduced:
- Use functions instead of macros.
- Verify the error code when querying KFD IOCTL version.
- Skip tests and samples if KFD IOCTL < 1.16 or PC Sampling IOCTL < 0.1.
* Incremental Counter Profile Creation
Adds support for incremental counter creation. How this functions is the
behavior of rocprofiler_create_profile_config has been changed.
rocprofiler_create_profile_config(rocprofiler_agent_id_t agent_id,
rocprofiler_counter_id_t* counters_list,
size_t counters_count,
rocprofiler_profile_config_id_t* config_id)
The behavior of this function now allows an existing config_id to be
supplied via config_id. The counters contained in this config will be
copied over and used as a base for a new config along with any counters
supplied in counters_list. The new config id is returned via config_id
and can be used in future dispatch/agent counting sessions.
A new config is created over modifying an existing config since there
is no gaurentee that the existing config isn't already in use. While we
could add locks (or other mutual exclusion properties) to check if its
in use and reject an update, the benefit from doing so is minor in
comparison to just creating a new config. This also side steps a common
pattern a tool may use to add additional counters at some point later on
during execution. Now they can do that without destroying the existing
config.
---------
Co-authored-by: Benjamin Welton <ben@amd.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Jonathan R. Madsen
Mythreya
Giovanni Lenzi Baraldi
Gopesh Bhardwaj
itrowbri
Vladimir Indic
SrirakshaNag
Benjamin Welton
Ammar ELWazir