commit 65690a1e175f9e81e815042775ec4628dbc7f548
author Stella Laurenzo <stellaraccident@gmail.com> Mon Jan 31 12:43:00 2022 -0800
committer GitHub <noreply@github.com> Mon Jan 31 12:43:00 2022 -0800
tree 39d91b278f94bb4613002a0fbc3a907c79a3cbbc
parent caedcd7a8fd37750e1302cb872edefec655e8303

Expose HAL low levels to Python and factor out a DeviceArray. (#8206)

* Expose the HAL Allocator and Buffer directly in the Python API.
* Supports explicit management of device memory and transfers.
* Reworks function dispatch to be based on this.
* Allows raw BufferViews to be passed into functions (vs an array) for expert control if needed.
* Exposes allocator statistics to python.
* Add a real DeviceArray class, implementing the numpy array and ufunc protocols.
* Rewire the VM layer to always work on BufferView and the Python layer to coerce in/out of DeviceArray.
  * This eliminates the implicit map to host memory that was happening on function return. Instead this will be done only if accesses are done to the DeviceArray that need it on the host.
* Generalize array input handling: DeviceArrays are not transferred back to the host to be passed to the next function and other framework __array__ implementations are coerced.
* Add iree.runtime.asdevicearray (casing matches numpy) as a quick shortcut for getting a device array when you aren't *that* concerned about the details.
* Support deepcopy/pickle (the TF tests use this).
* Implement DeviceArray.astype (needed for two tflite tests).