SaoirseARM

Fix accumulator and bias types coherence for int16x8 FC/Conv2d/TransposeConv2d operators

13

This PR implements a fix for this issue: https://github.com/tensorflow/tensorflow/issues/53763 This adds a quantized_bias_type parameter to the Conv2d, TransposeConv2d and FullyConnected operators in the flatbuffer schema. If _experimental_full_integer_quantization_bias_type is set in...

When using the Arm NN Tensorflow Delegate or Arm NN TfLite parser an abort is caused by the destruction of flatbuffers::ClassicLocale::~ClassicLocale() before the Arm NN ExecuteNetwork process exits. Setting FLATBUFFERS_LOCALE_INDEPENDENT=0...

Use single rounding as default for TOSA lowering

4

- Use single rounding at rescale to int8 - Minor refactor of op_addmm to use shared functions

[TOSA ] Add GELU approximate attr and transform to Tosa

9

Hi, This PR adds check for getApproximateAttr when converting TFL -> TOSA legalization and adds GELU approximation transform. Thanks, Saoirse