Summary: Initial TOSA support in executorch (pytorch#161)

Summary:
An implementation of a TOSA Partitioner and TOSA backend supporting a small list of operators in fp32 and int8/32. This includes a set of integer networks which will then compile to target Ethos-U NPU using the Vela compiler.

* A small set of test networks in fp32
* A subset of these supporting integer and compiling with Vela
* A test wrapper to register and test networks through the compile stack
  * A flow pt2 -> executorch -> tosa_flatbuffer.tosa file -> tosa_reference_model
  * A flow pt2 -> executorch -> tosa_flatbuffer.tosa file -> vela -> command_stream
* This work depends on a few different python modules
  * https://review.mlplatform.org/plugins/gitiles/tosa/serialization_lib/
  * https://review.mlplatform.org/plugins/gitiles/ml/ethos-u/ethos-u-vela/
* This work uses other projects for validation
  * https://review.mlplatform.org/plugins/gitiles/tosa/reference_model/

Pull Request resolved: pytorch#161

Reviewed By: mergennachin

Differential Revision: D49542254

Pulled By: digantdesai

fbshipit-source-id: bb4074d5b66233c85452c2373a08ddc1cdd0826f

.gitmodules

@@ -28,3 +28,6 @@
 [submodule "backends/xnnpack/third-party/XNNPACK"]
 	path = backends/xnnpack/third-party/XNNPACK
 	url = https://github.com/google/XNNPACK.git
+[submodule "backends/arm/third-party/serialization_lib"]
+	path = backends/arm/third-party/serialization_lib
+	url = https://git.mlplatform.org/tosa/serialization_lib.git

LICENSE

@@ -3,6 +3,7 @@ BSD License
 For "ExecuTorch" software
 
 Copyright (c) Meta Platforms, Inc. and affiliates.
+Copyright 2023 Arm Limited and/or its affiliates.
 
 Redistribution and use in source and binary forms, with or without modification,
 are permitted provided that the following conditions are met:

backends/arm/README.md

@@ -0,0 +1,24 @@
+# Executorch Arm/TOSA Delegate
+
+This subtree contains the Arm Delegate implementation for Executorch.
+
+This delegate is structured to, over time, support a number of different Arm devices
+through an AoT flow which targets multiple Arm IP using the TOSA standard.
+
+The expected flow is:
+ * torch.nn.module -> TOSA -> command_stream for fully AoT flows e.g. embedded.
+ * torch.nn.module -> TOSA for flows supporting a JiT compilation step.
+
+Current backend support is being developed for TOSA to Ethos-U55/65 via the
+ethos-u-vela compilation stack. which follows the fully AoT flow.
+
+## Layout
+- `arm_backend.py` - AoT Partitioner which maps to a subset of Base Inference and Main Inference TOSA profiles, where the subset may be further constrained for early support devices like Ethos-U55. AoT Backend which implements the preprocess step which converts to TOSA and can emit files for ethos-u-vela as shown in `executorch/examples/arm/`
+- `test/` - unit test and test support functions
+- `third-party/` - source dependencies - currently just on TOSA serialization_lib
+- `tosa_mapping.py` - helper functions for mapping edge dialect to TOSA
+
+## Help & Improvements
+If you have problems or questions, or have suggestions for ways to make
+implementation and testing better, please reach out to the Arm team developing this delegate, or
+create an issue on [github](https://www.github.com/pytorch/executorch/issues).