Skip to content

Add Xtensa fully connected kernel registrations - #3730

Open
unmeshna017 wants to merge 3 commits into
tensorflow:mainfrom
unmeshna017:xtensa-fully-connected-hifi
Open

unmeshna017 wants to merge 3 commits into
tensorflow:mainfrom
unmeshna017:xtensa-fully-connected-hifi

Conversation

@unmeshna017

Copy link
Copy Markdown
Contributor

Add HiFi-optimized Xtensa fully connected kernels, including int8, int16, and float32 support paths, shared Xtensa helpers, and the required fully_connected.h registration updates for the Xtensa-specific variants.

@cad-audio @joshih-cad

Optimize the Xtensa fully_connected kernels using the HiFi NNLib with per-channel and Int4 support, and add float32/int16 eval paths (fully_connected_float32.cc, fully_connected_int16.cc). Registers the two new sources in xtensa_makefile.inc. Depends on the 09_03_2026 NNLib release (PR#3703) for the v2 fully_connected/matmul APIs (e.g. xa_nn_fully_connected_v2_asym4sxasym8s_asym8s).
Add the updated fully_connected.h declaration needed by the Xtensa fully connected kernels.
@unmeshna017
unmeshna017 requested a review from a team as a code owner September 17, 2026 08:40
Add HIFI_IQ to the include guard so xa_nnlib_api.h and xa_nnlib_standards.h are pulled in for the hifi_iq build.
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