- Also, fixes to rocprofiler-sdk when incorporating recent features.
- URL encoding algorithm converts trailing '=' in the base64 string to an integer representing the number of those trailing '=' characters.
- pip update click module for test failures.
- Test results are at 99.8% with these fixes.
- Missing cmake dependency from last PR for ROCR-Runtime
- Missing pkg-config dependency for amdsmi
- Modify PATH to find pip's cmake for rocprofiler-sdk
- Dynamically write a Dockerfile based on the environment for the failing job.
- Account for additional dependencies that need to be installed and setup.
- Build and push a custom container based on that dynamic Dockerfile to capture that failing environment.
- Documenting additional setup to install Docker on VMSS during provisioning.
- Recent vision compilation has been failing, and debugging hasn't been fruitful in finding cause.
- Should unblock nightly job to at least build and test pytorch while debug effort continues after the holidays.
- pytorch build and test is unblocked by temporarily patching the composable_kernel submodule on upstream pytorch to latest develop, until that submodule is updated to have explicit cast for hneg.
- Exclude lone, consistently failing MIOpen test.
- test_rnn_seq_api is the only ctest failure, so let's filter it out for now to easily identify new failures.
- aomp: Account for path changes due to LLVM_INSTALL_LOC from aomp PR #1012
- aomp: Add llvm-legacy build script step for aomp PR #1062
- rocWMMA: Fix rpath issue when using ninja.
- Swap to default gcc on OS to resolve build errors from recent commits.
- Added libdnnl-dev dependency from iterative attempts with compiler change.
- Referred to the passing GitHub checks to observe the compilers that was used.
- Build CK jit lib and include in AMDMIGraphX build.
* External CI: omnitrace build pipeline starter
- Adding initial set of dependencies and build flags.
* External CI: omnitrace build pipeline
- Add bison, rccl, texinfo dependencies based on build failures.
- Add AMDGPU_TARGETS flag
- Add ROCm binaries to PATH for clang-format and other tools used.
* Fix indentation
---------
Co-authored-by: Daniel Su <[email protected]>
* External CI: torch and torchvision smoke tests
- Fixed issues with package name and version for the vision wheel that prevented it from installing. A patch is used until my pull request in vision repo is merged.
- Referred to rocAutomation scripts to pick which test scripts to run out of the many in the torch and vision repo, and iteratively tested suggested scripts to see which ones completed in a timely manner.
- Leveraging pytest-azurepipelines module to automatically parse and graph results from these tests.
- Referred to public documentation, source, and iterative attempts to create and improve build and test pipeline.
- ctest failures are due to the test node not having expected marketing name string and override not working.
- The fix should be on the omniperf repo side of things, so this pull request should be fine as is.
- Since this is an autotools project and not cmake, build and test on gfx942 system instead of separating into two jobs. Pipeline time is short anyway.
- Follow build instructions to update build flags and to incorporate the ROCdbgapi.
- Results are not parsed and graphed, but the log contents are printed at the end. This was helpful for debugging and will be kept in the pipeline, as the make check-gdb command's output was not helpful on its own.
- Add roctracer dependency to hipBLASLt build to address recent failures.
- Change build pool to ultra due to increased build times.
- Enable ccache to help with build times.
- Referred to public documentation, build instructions, source code in tests directory, and iterative runs to modify build flags.
- rdci test failures are known due to singleton nature of rocprofiler, but gtest attempting to spawn multiple instances. There is an internal ticket to track the issue.
Referred to public documentation, build instructions, and iterative debug runs to update build flags, publish new artifacts, and run tests. Test results are not parsed and graphed in Azure.
40% pass rate for this initial pass. Would like to push this through to at least change the build process and then defer fixing the remaining test failures.
- Test results are not parsed to be graphed in Azure reports.
- Added ccache to potentially improve build times, keyed against the date and hash based on amdclang++ binary.
Replace cmake calls with bash script calls to compile the components comprising openmp-extras.
Added inline comments to describe the bash scripts from aomp repo being executed.
- Added steps for creating wheel file for torchvision.
- Tried to add torchaudio as well, but it was not reading in AMDGPU_TARGETS value in the nested cmake calls from the python setup.py execution.
- Upstream pytorch builder scripts were updated, so it broke the patching step in the job. Removed the need to patch by using a flag to skip the tests.
- Will work on adding smoke tests of pytorch and torchvision later, just getting this out to fix the nightly build.
- Updating pipelines to account for combined repo changes of ROCR-Runtime and ROCT-Thunk-Interface.
- Removed dependencies referring ROCT-Thunk-Interface since it is now in the ROCR-Runtime repo.
- Changed ROCR-Runtime build command to account for directory changes.
Extension of PR #3544 and additional logic for ROCm dependency downloads to account for gpu target for components that can specify GPU target when building or have direct dependencies of these components. Plus, refactoring if statements to reduce lines of code.
Adding support for parallel build jobs where the only difference is the singular GPU target. This allows nightly packaging jobs to pick and choose based on GPU target to reduce download size.
To accommodate this new feature producing multiple artifacts for a component, added support for a file filter when downloading a ROCm component using the format "componentName:fileFilter".
With MIOpen now building with latest source on External CI, this unblocked AMDMIGraphX from building with latest source.
Determined rocMLIR also needed to be built with latest source as a dependency.
Template with bash commands to update cmake with snap.
Use template for two components that want updated cmake with latest source on their default branches.
* Disable aomp offload build for initial ci-build work
* Missing dependency for medium pool use of rocPRIM
* Latest rocBLAS source needs added ROCm dependencies
aomp build is not triggered by changes to aomp repo, but by updates to llvm-project and ROCR-Runtime, so trigger definition can remain this ROCm/ROCm repo.
Instead of using docker and apt install of ROCm component dependencies, use tarballs from Azure Pipeline builds to enable updates of ROCm interdependencies without waiting for releases..
* Update External CI Interdependencies for more repos
- composable_kernel
- hipBLAS
- rocBLAS
- rocSOLVER
Cleaned up unused flags from llvm-project
* Remove LD_LIBRARY_PATH change. Should not be needed.
- Fixed compilers to pick amdclang.
- Added ldconfig step for setting up linking of shared libraries.
- Set Azure VMs to medium only.
- Remove empty directories in published tarballs.
