Add 16bit xtensa depthwise conv kernel support

[narrietal]

This PR adds support for the optimized Xtensa depthwise convolution kernel when using 16-bit activations and 8-bit weights. Previously, this configuration would fall back to the reference implementation.

Changes:

• Removed hardcoded if-else logic in the Prepare function that restricted inputs to int8 activations only
• Removed TF_LITE_ENSURE_EQ assertion enforcing int8-only inputs
• Renamed the existing int8 evaluation function for clarity
• Added a new evaluation function to support int16 activations with int8 weights

bug=fixes #3484

[rameshkunasi]

Hi @narrietal,

Thank you for this PR. Can you please resolve the failed test cases for Hifi3z and Fusion F1 platforms and update the PR?

[narrietal]

@rameshkunasi I pushed a new commit which should solve the previous conflicts. Could you approve the execution of the automated test suite to verify it?

[narrietal]

@rameshkunasi I just pushed a commit with the formatted code. Could you approve the execution of the automated test?

Thanks.

[narrietal]

@rameshkunasi the CI/CD test pipeline went through 👍 It seems ready to be merged

[rameshkunasi]

@unmeshna017 Can you please have a look into these changes?

[unmeshna017]

Hi,

1. In depthwise_conv_hifi.cc, the fused activation variant of the Conv2D depthwise kernel (xa_nn_conv2d_depthwise_v2_per_chan_sym8sxsym16s) can be used.
2. In xtensa_depthwise_conv.h, the REF eval function declaration for INT16 precision (DepthwiseConvReferenceEvalInt16) can be added.

Adding @vp-cad and @joshih-cad as watchers.

[rameshkunasi]

Hi @narrietal,

Can you please create a new PR with the suggested changes?

[narrietal]

Hi @unmeshna017 and @rameshkunasi,

1. I believe the current implementation is correct. The function xa_nn_conv2d_depthwise_v2_per_chan_sym8sxsym16s is called within xa_nn_conv2d_depthwise_per_chan_sym8sxsym16s. Additionally, the code includes an if–else check on the data format to ensure the appropriate function is called for each case.

2. The file xtensa_depthwise_conv.h corresponds to the HiFi kernels, not the reference implementations. For reference kernels, the appropriate header is tensorflow/lite/micro/kernels/depthwise_conv.h. I noticed that xtensa_depthwise_conv.h declares a DepthwiseConvReferenceEvalInt8 function, which may be the source of the confusion. However, there does not appear to be any implementation of this function, suggesting it may be leftover or legacy code.

Given this, I would suggest removing the DepthwiseConvReferenceEvalInt8 declaration from xtensa_depthwise_conv.h to avoid any future misunderstanding.

[unmeshna017]

Hi @narrietal, apologies for the delayed response.

1. In the mapping layer, we are always passing the input data format as '0' which internally calls v2 API. Additionally, replacing current API call with the v2 version, the subsequent activation function call won't be required, as "v2" is a fused activation variant of the kernel.

2. In xtensa_depthwise_conv.h, your observation is correct; it is okay if we don't add the REF eval function declaration for INT16 precision.