Skip to content

Add HiFi int8/int32 optimized paths for LeakyRelu, Pad, and SVDF on Xtensa - #3644

Draft
unmeshna017 wants to merge 3 commits into
tensorflow:mainfrom
unmeshna017:xtensa-hifi-int8-leaky-relu-pad-svdf
Draft

Add HiFi int8/int32 optimized paths for LeakyRelu, Pad, and SVDF on Xtensa#3644
unmeshna017 wants to merge 3 commits into
tensorflow:mainfrom
unmeshna017:xtensa-hifi-int8-leaky-relu-pad-svdf

Conversation

@unmeshna017

Copy link
Copy Markdown
Contributor

Enable HiFi-accelerated kernels for three Xtensa ops

  • LeakyRelu: adds int8 (asym8s) accelerated path via xa_nn_vec_leaky_relu_asym8s_asym8s, mirroring the existing int16 path.
  • Pad: adds int8 (xa_nn_pad_8_8) and int32 (xa_nn_pad_32_32) paths alongside the existing int16 path; >4D inputs fall back to reference_ops.
  • SVDF: extends existing HIFI guards to HIFI_IQ so the optimized paths compile for that target.

Adding @cad-audio and @joshih-cad for review.

Add an int8 (asym8s) accelerated path using xa_nn_vec_leaky_relu_asym8s_asym8s
for HIFI3/HIFI4/HIFI5/HIFI_IQ targets, mirroring the existing int16 optimized
path. Falls back to the reference QuantizeLeakyRelu on other targets.
Add int8 (xa_nn_pad_8_8) and int32 (xa_nn_pad_32_32) accelerated Pad paths for
HIFI targets, complementing the existing int16 path. Inputs with more than 4
dimensions fall back to reference_ops. The int32 path is gated on
HIFI3/HIFI4/HIFI5.
Extend existing HIFI kernel guards to include HIFI_IQ so the optimized SVDF
paths are compiled for the HIFI_IQ target. No functional change on other
targets.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant