* Show description of metrics during analysis
* Use --include-cols Description show the Description column in analyze mode (this is hidden by default)
* Remove tips field from analysis config
* Align metric names in analysis config and documentation
* Add unified config utils/unified_config.yaml
* Add python script utils/split_config.py to auto generate analysis configuration and documentation metrics description
* Add test case to ensure unified config is older than auto-generated config
* Auto generate analysis config and documentation metrics description
* Update CONTRIBUTING.md to add instructions to build documentation assets
* Add docker image and compose file to build documentation
* Update CHANGELOG and Documentation
* Use jinja template instead of hardcoding metric tables in documentation
[ROCm/rocprofiler-compute commit: bb44e90b2d]
* Show description of metrics during analysis
* Use --include-cols Description show the Description column in analyze mode (this is hidden by default)
* Remove tips field from analysis config
* Align metric names in analysis config and documentation
* Add unified config utils/unified_config.yaml
* Add python script utils/split_config.py to auto generate analysis configuration and documentation metrics description
* Add test case to ensure unified config is older than auto-generated config
* Auto generate analysis config and documentation metrics description
* Update CONTRIBUTING.md to add instructions to build documentation assets
* Add docker image and compose file to build documentation
* Update CHANGELOG and Documentation
* Use jinja template instead of hardcoding metric tables in documentation
- Introduced double-buffering to reduce copy overhead and overlap BF16 arithmetic with data prefetching.
- Aimed to improve performance of reduction-based collectives by up to 10%.
- Implemented based on recommendations from Guennadi Riguer (AMD)
- Added --force-reduce-pipeline option to install.sh to activate this optimization for BF16 reductions.
- Feature is disabled by default to prevent regressions with large messages until auto-tuning logic is upstreamed.
---------
Co-authored-by: Jeffrey Novotny <jnovotny@amd.com>
Co-authored-by: Pedram Alizadeh <pmohamma@amd.com>
[ROCm/rccl commit: 0ce20e7e07]
- Introduced double-buffering to reduce copy overhead and overlap BF16 arithmetic with data prefetching.
- Aimed to improve performance of reduction-based collectives by up to 10%.
- Implemented based on recommendations from Guennadi Riguer (AMD)
- Added --force-reduce-pipeline option to install.sh to activate this optimization for BF16 reductions.
- Feature is disabled by default to prevent regressions with large messages until auto-tuning logic is upstreamed.
---------
Co-authored-by: Jeffrey Novotny <jnovotny@amd.com>
Co-authored-by: Pedram Alizadeh <pmohamma@amd.com>
The ape1_size_ member was leftover after the removal
of KV and is no longer used.
Remove it to remove some compiler warnings.
Signed-off-by: Tony Gutierrez <anthony.gutierrez@amd.com>
The ape1_size_ member was leftover after the removal
of KV and is no longer used.
Remove it to remove some compiler warnings.
Signed-off-by: Tony Gutierrez <anthony.gutierrez@amd.com>
[ROCm/ROCR-Runtime commit: 5285c24657]
* adding summary.py to generate tmp <category_region>_summary views
* migrating CSV summary to SDK method of writing CSVs
- Add domain_view to summary.py
- omit the C++ code of writing CSV because it gets revered later anyway
* Add summary subparser and write_sql_view_to_csv function
* adding all <>_summary views generation to summary.py
* add summary_per_rank feature
* add --summary-per-rank
* reconstruct generate_summary_view and create_domain_view
-introduce by_rank
* remove sqr and variance in summary views
* use RocpdImportData instead of connection
* two fixes on summary.py
--modify the generate_summary_view function to return a tuple with view name and sql code
add if_not_exits parameter to generete_summary_view
* Refactor summary.py to allow output path and filename args, and apply time_window
- clean up summary table column headers
- only generate by-rank views if that param is specified
* Add ProcessID to Hostname output and csv, so users can identify the system in the by-rank summaries
* Summary.py, just add hostname to by-rank summaries, instead of creating mapping table
* Summary - migrate csv writer to pandas, for more future flexibility
* Adding a few simple tests for summary.py
* Linting fixes
* add region_categories to summary options
- Automatically retrieve region categories from the database if argument is None
* add backticks for view_names
* fix tests after rebase
* Made code review changes
- fixed whitespace in CMakelists.txt
- adding query.py module & subparser in __main__.py
- refactor summary function to return query
- used query.py to output csv
- used query.py to also output summary to console
- provided new command line options to select summary output to csv or console
* Made fix to jinja template in query.py, as suggested by copilot
* Consolidated output calls to query in export_view function based on feedback
- refactored: helpers, query functions, create view functions
- extended formats to include what query supports (md, html, pdf, json)
- added json format to query, and changed orient=records
- adding jinja2 and reportlab to requirements.txt
* Add version_info for rocpd and roctx
* Add rocpd commandline tool
* Add executable permissions to source/bin/rocpd.py
* Removed rocpd2query, and cleaned up --help examples
---------
Co-authored-by: acanadas <acanadas@amd.com>
Co-authored-by: Jin Tao <jintao12@amd.com>
Co-authored-by: a-canadasruiz <Araceli.CanadasRuiz@amd.com>
Co-authored-by: Jonathan R. Madsen <Jonathan.Madsen@amd.com>
[ROCm/rocprofiler-sdk commit: 3954cedd25]
* adding summary.py to generate tmp <category_region>_summary views
* migrating CSV summary to SDK method of writing CSVs
- Add domain_view to summary.py
- omit the C++ code of writing CSV because it gets revered later anyway
* Add summary subparser and write_sql_view_to_csv function
* adding all <>_summary views generation to summary.py
* add summary_per_rank feature
* add --summary-per-rank
* reconstruct generate_summary_view and create_domain_view
-introduce by_rank
* remove sqr and variance in summary views
* use RocpdImportData instead of connection
* two fixes on summary.py
--modify the generate_summary_view function to return a tuple with view name and sql code
add if_not_exits parameter to generete_summary_view
* Refactor summary.py to allow output path and filename args, and apply time_window
- clean up summary table column headers
- only generate by-rank views if that param is specified
* Add ProcessID to Hostname output and csv, so users can identify the system in the by-rank summaries
* Summary.py, just add hostname to by-rank summaries, instead of creating mapping table
* Summary - migrate csv writer to pandas, for more future flexibility
* Adding a few simple tests for summary.py
* Linting fixes
* add region_categories to summary options
- Automatically retrieve region categories from the database if argument is None
* add backticks for view_names
* fix tests after rebase
* Made code review changes
- fixed whitespace in CMakelists.txt
- adding query.py module & subparser in __main__.py
- refactor summary function to return query
- used query.py to output csv
- used query.py to also output summary to console
- provided new command line options to select summary output to csv or console
* Made fix to jinja template in query.py, as suggested by copilot
* Consolidated output calls to query in export_view function based on feedback
- refactored: helpers, query functions, create view functions
- extended formats to include what query supports (md, html, pdf, json)
- added json format to query, and changed orient=records
- adding jinja2 and reportlab to requirements.txt
* Add version_info for rocpd and roctx
* Add rocpd commandline tool
* Add executable permissions to source/bin/rocpd.py
* Removed rocpd2query, and cleaned up --help examples
---------
Co-authored-by: acanadas <acanadas@amd.com>
Co-authored-by: Jin Tao <jintao12@amd.com>
Co-authored-by: a-canadasruiz <Araceli.CanadasRuiz@amd.com>
Co-authored-by: Jonathan R. Madsen <Jonathan.Madsen@amd.com>
* Add host APIs for querying device ctx and remote heap pointer
* Host API to query device pointer for ROCSHMEM_DEFAULT_CONTEXT,
this is needed to support dynamic module initialization via device kernel
library bitcode.
* Host API to query remote symmetric heap pointer that can be used in
custom device kernel for RMA operations.
* Added rocshmem_ptr implementation within the Host Context class
* Enables pointer retrieval functionality for symmetric data objects
* Copy IPC pointers to host memory in RO host context
---------
Co-authored-by: avinashkethineedi <avinash.kethineedi@amd.com>
[ROCm/rocshmem commit: 87f99e7ec6]
* Add host APIs for querying device ctx and remote heap pointer
* Host API to query device pointer for ROCSHMEM_DEFAULT_CONTEXT,
this is needed to support dynamic module initialization via device kernel
library bitcode.
* Host API to query remote symmetric heap pointer that can be used in
custom device kernel for RMA operations.
* Added rocshmem_ptr implementation within the Host Context class
* Enables pointer retrieval functionality for symmetric data objects
* Copy IPC pointers to host memory in RO host context
---------
Co-authored-by: avinashkethineedi <avinash.kethineedi@amd.com>
Prevent initialization failures in certain configurations when attempting
to load fp8-specific symmetric multicast kernels on GPUs older than
Blackwell.
[ROCm/rccl commit: 593de54e52]
Prevent initialization failures in certain configurations when attempting
to load fp8-specific symmetric multicast kernels on GPUs older than
Blackwell.
* Do not force unsupported metrics to be specified in older gpu
architectures as None
* Remove metrics which are explicitly set to None
* Update CHANGELOG
* Fix analysis configuration to fix baseline comparisons across all gpu
architectures
* Add missing 1812 section for gfx908
* Add missing 1812 section for gfx90a
* Baseline comparision will only show common metrics
* First workload will be used to set Metric ID index column
[ROCm/rocprofiler-compute commit: d4c316a730]
* Do not force unsupported metrics to be specified in older gpu
architectures as None
* Remove metrics which are explicitly set to None
* Update CHANGELOG
* Fix analysis configuration to fix baseline comparisons across all gpu
architectures
* Add missing 1812 section for gfx908
* Add missing 1812 section for gfx90a
* Baseline comparision will only show common metrics
* First workload will be used to set Metric ID index column
Yiannis Papadopoulos
Yat Sin, David
Kanangot Balakrishnan, Bindhiya
vedithal-amd
Atul Kulkarni
Williams, Justin
Sarat Kamisetty
Mustafa Abduljabbar
Tony Gutierrez
Roy, Rakesh
Dittakavi, Satyanvesh
Vaddireddy, Sushma
David Addison
Hui, Young
Madsen, Jonathan
Dimple Prajapati
Baraldi, Giovanni
Kamil Iskra
xuchen-amd