* Adding --torch-operator option in rocprof-compute. Creates csv file for
each operator that has gpu activity, showing operator to counter values
mapping.
* --torch-operators flag added to rocprofiler-sdk
* Adding ctest for --torch-operators.
* Adding pytest markers.
* Corrections in ctest and message logging.
* Apply suggestion from @Copilot
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Apply suggestion from @Copilot
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Apply suggestion from @Copilot
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Apply suggestion from @Copilot
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Apply suggestion from @Copilot
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Apply suggestion from @Copilot
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Adding a check for pytorch installation only when --torch-operators is passed.
* moving inject_roctx.py into src/utils.
* rebase
* Updating docs and changelog.
* Update projects/rocprofiler-compute/src/argparser.py
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Update projects/rocprofiler-compute/src/utils/inject_roctx.py
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Removing special characters.
* Minor corrections.
* Setting default value for torch_operators_enabled.
* Updating the number of files according to the number of passes.
* Adding rocpd support.
* Adding a warning message to be shown when profiling a non-python workload.
* copilot suggestions, rocpd+native tool fix
* Fixed the incorrect usage of dispatch_id as event_id in the function update_rocpd_pmc_events()
* ruff format fix
* ruff formating
* Deleting torch_trace.csvs after consolidating the operator data.
* Removing checks since *torch_trace.csv files are deleted.
* Fixing file deletion.
* Update projects/rocprofiler-compute/src/utils/inject_roctx.py
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Update projects/rocprofiler-compute/src/rocprof_compute_profile/profiler_base.py
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Update projects/rocprofiler-compute/src/utils/utils.py
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Update projects/rocprofiler-compute/tests/test_profile_general.py
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Using default options in the testcase.
* Adding test for overhead measurement.
* Corrections in docs.
* doc updates.
* Update projects/rocprofiler-compute/src/utils/inject_roctx.py
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Handling potential empty frames.
* Corrected the test cases.
* Changing the flag to --torch-trace
* Fixed helper_app path issues
* Path issues
* process_torch_trace_output() now takes csv file paths as input + allows default usage.
* Replaced pandas with sqlite3
* Adding marker_trace extraction to rocpd_data.py
* Allowing all workloads to use --torch-trace option. Assuming the workload is user verified.
* Modified help section for the flag.
* Added difference in runtimes for longest running kernels in each profiling runs to overhead measurements.
* Update projects/rocprofiler-compute/src/rocprof_compute_profile/profiler_base.py
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Update projects/rocprofiler-compute/src/rocprof_compute_profile/profiler_base.py
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Removed the accesses to the tables.
* Ruff fixes.
* ruff
* Ruff Fixes
* Adding getattr for args.torch_trace to handle mock args.
* Fix for 'Missing guid in counter collection data - in csv mode'
* Sending output_format to process_torch_trace_output
* Warning for self contained binaries.
* Ruff
* Ruff
* Measuring longest_running_kernel_baseline instead of worst_kernel_increase, very small kernel runtimes are blowing up the worst_kernel_increase metric.
* Minor fixes in input arguments
* Ruff
* Loging PyTorch version
* Fix ruff formatting for PyTorch version logging
---------
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
*Added iteration_multiplex_impute_counters on pmc data- GUI dataframe did not implement this in the build_layout method previously
*Created a Workload() in profile mode post-processing for roofline html standalone plot to be generated- this will be removed once roofline plot is moved to analyze phase in future release
*Added iteration_multiplexing run parameter to roofline object init so that we can accurately parse dataframe if the option was used during profiling- this helps us to avoid reading nan values in certain dispatches that did not get imputed in calc_ai_profile
*Cleanup for unused legacy code, adjusted method parameters to assist in moving roofline plotting to analyze mode in future release
*Update iteration multiplexing data imputation algorithm to impute counters for ungrouped dispatches at the end based on the previous group. This however won't work if there are no dispatches that can be grouped (i.e. number of dispatches < number of counter buckets)
---------
Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>
Co-authored-by: Vignesh Edithal <Vignesh.Edithal@amd.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Removed attach tool library path
* Support new attach/detach API
* New attach/detach API was introduced in
https://github.com/ROCm/rocm-systems/pull/1653
* Provide backward compatibility with old api
* Stabilize attach/detach tests by adding sleep to help workload get
ready for attachment
* Fix typo in test name
---------
Co-authored-by: Vignesh Edithal <Vignesh.Edithal@amd.com>
Co-authored-by: Fei Zheng <44449748+feizheng10@users.noreply.github.com>
* Analysis database v1.2.0
* `pc_sampling` and `roofline_data` tables should relate to `kernel` table instead of `workload` table
* Remove `kernel_name` fields in `pc_sampling` and `roofline_data` table
* Add kernel existence check for roofline data to prevent KeyError (#2536)
* Initial plan
* Add kernel existence check for roofline data to prevent KeyError
Co-authored-by: vedithal-amd <191402304+vedithal-amd@users.noreply.github.com>
---------
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: vedithal-amd <191402304+vedithal-amd@users.noreply.github.com>
* Optimize analysis performance
* Refactor database schema: separate metric definitions from kernels
Reorganize the database ORM to decouple metric definitions from kernel
objects. This improves the schema design by:
- Rename Metric -> MetricDefinition and Value -> MetricValue for clarity
- Move metric definitions from kernel-level to workload-level, since
metric definitions are shared across kernels
- Update relationships: MetricDefinition belongs to Workload,
MetricValue
references both MetricDefinition and Kernel
- Refactor metric_view to join through the new schema structure
- Update test fixtures to use renamed table and class names
- Update documentation with new example output using nbody workload
- Regenerate database schema and views diagrams
* Add min amd max aggregation in kernel_view
* Add primary key id from tables into the view
---------
Co-authored-by: Copilot <198982749+Copilot@users.noreply.github.com>
Co-authored-by: vedithal-amd <191402304+vedithal-amd@users.noreply.github.com>
* Improve native tool discovery and partition detection
- Enhanced native tool path resolution to support CMAKE_INSTALL_LIBDIR variations
(lib, lib64, lib32, etc.) using glob pattern matching
- Extracted path variables to avoid duplication in error messages
- Improved error message clarity by showing exact paths searched for .so and .cpp files
- Simplified code path construction using consistent Path.resolve().parents[x] syntax
- Fixed redundant partition warnings on pre-MI300 GPUs by adding architecture check
- Only query compute/memory partition on MI300+ series (gfx940+)
- Added proper type hints for gpu_arch parameter
- Moved gpu_info extraction after soc_info to ensure gpu_arch is available
- Improved code comments for MI300 series threshold
* Handle gpu arch like a hex string
* attach: Formalize ROCAttach API
- Make ROCAttach public with public headers
- Change detach to take a PID
- attach and detach are now reentrant
- Cleanup of states and signal handling in ptrace session
- Fixes mixed up definition of ROCPROF_ATTACH_TOOL_LIBRARY
- ROCPROF_ATTACH_TOOL_LIBRARY now always means the tool library loaded by the attachment target
- ROCPROF_ATTACH_LIBRARY refers to the library used to perform attachment
- Add direct call of rocprof-attach
- Fix python library call of rocprof-attach
- Function now named attach(), changed from main()
* attach: rocprof-compute ROCAttach updates
- Update to new library names
- Correct usage of C lib detach
* attach: add test for rocattach
- Disable ASan, TSan, and UBSan for the new parallel-attach test
- Lower log level for LSan tests, existing behavior from other tests
---------
Co-authored-by: Ammar ELWazir <aelwazir@amd.com>
* Replace O(n^2²) nested loop with O(1) dictionary lookup when associating
metric values with metrics. Pre-group values by (metric_id, kernel_name)
to eliminate redundant iteration over entire values dataframe for each
metric-kernel combination.
* This optimization significantly improves database write performance for
workloads with large numbers of metrics and kernels.
* Standalone roofline should create HTML instead of PDF
* Eiminate the dependency on kaleido and plotly_get_chrome by moving
towards plotly native HTML image roofline chart generation
* Address review comments
* 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
* 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
* Test env. vars. in rocprofiler-sdk backend
* Improve rocprofiler-sdk backend test case to check for env. vars. and
ensure we do not overwrite irrelevant env. vars.
* Remove unnecessary usage of ROCPROF_INDIVIDUAL_XCC_MODE env. var.
* Formatting fixes
* Test fixes
* Remove redundant code in tests
* Remove usage of utils_mod and use utils instead, this prevents
duplicate imports
* Fix for multi process workload profiling
Native counter collection tool updates:
* Do not dump empty counter data for a process
* Use PID instead of UUID for dumped csv files to facilitate correlation
* Handle merging multiple pairs of rocpd (from sdk tool) and csv (from
native tool) files
* Handle merging multiple pairs of csv (from sdk tool) and csv (from
native tool) files
Rocpd output format updates:
* Merge multiple rocpd databases into a single csv
* Reset dispatch id and kernel id for unique dispatches and unique
kernels respectively
* Retain multiple rocpd databases per run for multi process workloads
* Add test case for multiprocess profiling using rocflop workload
* Add rocflop
* Fix native counter csv to rocprofv3 csv conversion
* Use kernel_id instead of dispatch_id to correlate native counter csv
and kernel trace csv
* python formatting using ruff 0.14 instead of 0.13
* Improve Iteration multiplexing
* Improve iteration multiplexing documentation by adding usage note and
listing caveats
* Bugfixes for iteration mulitplexing
* Use merge iteration multiplexing in analysis webui and db mode
* Do not remove Dispatch_ID column in merge iteration multiplexing
since it is needed for analysis of top dispatches based on
duration
* Bugfixes for analysis logic
* Graceful handling of missing counters in case of iteration
multiplexing
* Improved warnings when metrics could not be calculated due to
missing counter data
* Fix the check to prevent showing table when a column is full of
N/A
* Improve detection of empty values when metric evaludation fails
due to missing counter data
* Bugfixes for profile logic
* Fix kernel filtering during roofline benchmark phase
* Update changelog for bugfixes
* Remove unnecessary columns when merging dispatches for iteration multiplexing
* bugfix
* Better analysis warnings
* fix to_std() in parser
* Use median in merge iteration multiplex
* Address review comments
* Fix cmake formatting
* fix None handling of parser util functions
* Enable stochastic counter accuracy test
* fix cmake formatting
* added graceful errors/exit in profile/analyze roofline.csv
* edit if statement truth
* restore if statement truth (roofline_csv needs at least 2 rows)
* addressed comments and skipped showing roof metrics when data invalid
* fix workload merge
* changed warning to error
* removed redundant variable definition
* added roofline csv validate check in TUI
* add test cases to test validation function
* ruff format
* simplified TUI roofline handling
* Improve amdsmi interface
* Fix issue where max mem clock was being set as max gfx clock
* Handle the case when all device handles might not be usable due to
devices being hidden by ROCR and HIP environment variables
* Fix get gpu vram size to return str in KB
* Improve testing of amdsmi interface functions
* Remove TCP_TCP_LATENCY_sum counter for MI300
* Remove TCP_TCP_LATENCY_sum counter which is unsupported for MI300 per register specification
* Remove VL1 Lat metric from memory chart section (block 3) for MI 300
since it uses TCP_TCP_LATENCY_sum counter which is unsupported
* Remove references to TCP_TCP_LATENCY_sum
* Update CHANGELOG
* reword changelog
* Add exception handling for native tool path search
* Fix formatting in roofline benchmark code
* Fix detection of .so files
* include hip code and native tool code in standalone binary
* add fallback path for ROCM_PATH
* Implement AMDGPU driver info and GPU VRAM attributes in system info.
section of analysis report.
* Backward compatibility for rocprofiler-sdk avail module path migration
* Fix roofline calculation where AI data points are N/A
* Removing FP8 from peak VALU datatypes - PEAK_OPS_DATATYPES.
* Similar change for BF16.
* Roofline binaries from rocm-amdgpu-bench generated 10/22.
https://github.com/ROCm/rocm-amdgpu-bench/commit/2113ef1f5eada8a4a6e44e6d07fd6abac9b0a3f8
Bins include change that removes FP8 and BF16 peak VALU benchmarks.
Built and tested on rhel8, azl3, ubuntu22.04, sles15sp6.
Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>
* Re-committing the bins
accidentally copied over bins from the wrong folder earlier, caught by Gowthami during testing.
Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>
* Updated changelog
* gersemi fix
* Changelog corrected.
* Changelog fix.
* Adding this to the 7.2.0 section to be picked up in an RC build.
* Moving changelog entry into unreleasesd section - team reconfirmed cutoff date after I requested this change so I am just quickly correcting my mistake in my ask.
Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>
---------
Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>
Co-authored-by: Carrie Fallows <Carrie.Fallows@amd.com>
Co-authored-by: JeniferC99 <150404595+JeniferC99@users.noreply.github.com>
* Split roofline tests
* Use N/A for missing values
* Test eval_expression for no valid data
* Fixed tests
* Updated Changelog for N/A
* Fixed platform specific test failure
* Use native tool for counter collection
* Add native counter collection tool which uses rocprofiler-sdk C++
library public API to get counter collection data
* This is enabled by default, unless --no-native-tool option is
provided or ROCPROF=rocprofv3 env. var. is provided
* This tool is only supported for ROCm version >=7.x.x
* This tool is not supported for attach/detach scenario
* Build native tool shared object during build time
* If using rocprof-compute without building then runtime compilation of
t push native tool shared object is performed
* rocprofiler-sdk tools is still used for services other than counter
collection and data collected by native tool is merged into the
rocpd/csv output of rocprofiler-sdk tool
* Make `rocpd` choice the default choice for `--format-rocprof-output`
option
* If `rocpd` public API from rocprofiler-sdk library is not present,
then fallback to `csv` choice
* In this case only `pmc_perf.csv` is written in workload folder
instead of multiple `csv` files for each profiling run
* Remove `json` choice from `--format-rocprof-output` option since it
functions identical to `csv` option
* Rename option `--rocprofiler-sdk-library-path` to
`--rocprofiler-sdk-tool-path` since we LD_PRELOAD the
rocprofiler-sdk tool shared object and not the rocprofiler-sdk library
shared object
* Fix the meaning of `--dispatch` option in `profile` mode to mention
dispatch iteration filtering instead of dispatch id filtering
* --dispatch option in analyze mode does dispatch id filtering
* Move standalone binary creation logic from cmake file to docker file
* fix native counter collection tool during attach/detach
* improve logging
* fix attach detach with native tool
* fix attach detach with native tool
* do not support attach/detach in native tool
* Update changelog
* add standalone binary creation functionality in cmake
* address review comments
* address review comments
* fix formatting
* address review comments
* Adding paths for cmake to search. Also updated min. cmake requirement to 3.21 as this was when hip was supported.
Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>
* Update hip compiler ID check, sometimes comes up as Clang, sometimes ROCMClang- depends on setup.
Updated formatting.
Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>
* RHEL8.10 unable to compile due to defaulting to old c++ version, need to force c++17
Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>
* Updating changelog per docs team recommendations
Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>
* Apply suggestions from code review to changelog
Co-authored-by: Pratik Basyal <pratik.basyal@amd.com>
* Do not required HIP complier to build native counter collection tool
* fix cmake
* gersemi formatting on latest cmake change
Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>
* ex ci updated dependencies to include rocprofiler-sdk, but cmake was still not capturing the path- there was a commit that added to the cmake_prefix_path entry that specified rocprof-sdk's cmake location ut was too specific for the search paths in find_package's config mode.
removing the cmake_prefix_path var and adding hints to find_package call instead, and specifying config mode so it knows how to construct the search paths
Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>
* gersemi run for formatting
Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>
* Still need prefix path, should not have been removed in last commit but does need to be shortened to just the rocm path to allow for find_package config mode to do the job
Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>
* include cstdint for uint32_t
* Run formatting on helper.cpp
Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>
* Remove rocm 7.2 release stuff from version and changelog and handle it in separate pr
* fix version
* fix changelog
* fix changelog
* run ruff formatter
Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>
* fix rocprofiler-sdk attach so path
---------
Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>
Co-authored-by: Carrie Fallows <Carrie.Fallows@amd.com>
Co-authored-by: Pratik Basyal <pratik.basyal@amd.com>
Force tencentos to use rhel-based bin since tencent is branched off of centos, which is branch of fedora.
Verified rocprof-compute run correctly selects bin to use, and the roofline benchmark values look similar between runs on rhel vs tencentos4 docker images on same system.
Signed-off-by: Carrie Fallows <Carrie.Fallows@amd.com>
* attach: rename librocprofv3-attach
- Renames library to librocprofiler-sdk-rocattach
- ROCAttach library will be formalized and documented in future commit
* Address review comments
- Rename rocprofv3-attach.py to rocprof-attach.py
- Use common filesystem.hpp in rocattach
* Fix component name typo
* Doc fixup
---------
Co-authored-by: JeniferC99 <150404595+JeniferC99@users.noreply.github.com>
* roll back json file processing logic for pc sampling
* format cmake files
* Revert "format cmake files"
This reverts commit e64df65a8f30abcb6738e3a0d7ffd4270bd1d302.
Were not handling the case where the eval result is None e.g. some
columns have a peak value, but it is unused, so we use 'None', which
evaluates to the None object.
Return empty string in this case.
* Analysis db changes for visualizer
* Add support for per kernel analysis metrics
* Add support for dispatch timeline visualiztion
* Show median instead of mean of dispatch duration in kernel view
* Add test case to validate analysis db schema
* Analysis db schema updte
* Add Kernel table and make Metric and Dispatch table its children
* Kernel table is a child of Workload table
* Update metric_view to show kernel_name column
* Add disptach timestamps to Dispatch table for dispatch timeline
visualization
* Update kernel_view to show duration_ns_median instead of mean
duration
* Add mean duation in kernel view
* update changelog
---------
Co-authored-by: Fei Zheng <44449748+feizheng10@users.noreply.github.com>
* fix crash created by path and arg for pc_sampling and add copyright for mat_mul
* resolve fomat issue of line too long
* bugfixes
* copy gfx9 config template to analysis config in src
---------
Co-authored-by: Wang <ywang103@ctr2-alola-login-01.amd.com>
Co-authored-by: Vignesh Edithal <Vignesh.Edithal@amd.com>
ggottipa-amd
vedithal-amd
cfallows-amd
jamessiddeley-amd
abchoudh-amd
Mark Meserve
xuchen-amd
Ben Richard
Kunal Malviya
ywang103-amd