Refine JIT Runtime/Compilation/Build Capabilities

[hunhoffe]

Description

This issue is used to track progress/plans for changes/extensions to the functionality for JIT Runtime/Compilation/Build Capabilities that are planned

Goals

• Reduce pain points and extend functionality for users with different purposes (users seeking an easy/simple path for education, etc., users seeking to use a piece of the toolstack but not the whole, researchers looking to extend/modify components within the toolstack).
• Sync implementation in this repo and implementation in amd/iron, which includes reducing the amount of duplicated functionality

Specific Tasks and their Related Issues/PRs

Completed

• Loosen XRT ties between the iron.Tensor class and PyXRT (Decouple IRON Tensor from XRT Host Runtime #2692)
• Update documentation on configuration options for the IRON python environment (Add Documentation for IRON Python Configuration Options/Methods #2717)
• Fix amd/IRON so I don't keep breaking it with these changes
  • Streamline versioning of mlir-aie release wheels: Make wheels for tagged releases with the appropriate version #2722 Fix github action syntax error. #2725
  • Use pinned release wheels in amd/IRON: Pin milr-aie version to a release and create action to keep it updated. amd/IRON#27

General Code Cleanup

• Integrate python test lit config with helper classes #2684
• Support for numpy integer type conversion in scf for wrapper #2711 (bug from Jack that never become an issue)
• Refine default Python runtime sequence behavior; remove experimental code. #2687
• Python Device class extension and cleanup #2691
• Disambiguate python buffers #2724
• Tests to compare iron.Tensor and torch.Tensor #2731

In Progress

• Look into PyXRT buffers that do not seem to work as expected (example here: https://github.com/Xilinx/mlir-aie/blob/pyxrt-buffer-bugtest/programming_examples/basic/pyxrt_buffer_bugcheck/passthrough_kernel.py)
• See if we can avoid xrt-smi commands in favor of the PyXRT capabilities. Maybe something like so:

  >>> pyxrt.device(0).get_info(pyxrt.xrt_info_device.name)
  'RyzenAI-npu6'

General Code Cleanup

• Device slicing [WIP] Stateful Destructive Device Slicing #2726
• SequentialPlacer cleanup [WIP] SequentialPlacer #2659 (prep PR: Device and ObjectFifo tests #2733)
• IRON Environment Var Documentation; minor code restructure. #2736

Future Work

• Integrate iron.Tensor into amd/IRON
• Integration of the parallel/algorithms capability: [WIP] Parallel/algorithms #2730

Existing Issues that are Related

• Add trace with JIT example #2541
• Combining multiple object files in JIT #2530
• Add a code object cache for external functions #2509

[andrej]

Is the Compilable/Runnable abstraction you presented previously a part of this or a future stage?

[hunhoffe]

That is part of it, but the first steps are just to capture all the other in-flight efforts. And I've got breaking changes in the pipeline, so I really need to get the wheels pinning in amd/IRON finished before I can make too much progress.