Skip to content

b12092

  • b12092
  • 20a51f2
  • Verified

    This commit was created on GitHub.com and signed with GitHub’s verified signature.
  • Choose a tag to compare

  • b12092
  • 20a51f2
  • Choose a tag to compare

  • Verified

    This commit was created on GitHub.com and signed with GitHub’s verified signature.
@tboinovski1 tboinovski1 tagged this 03 Nov 16:26
* hexagon: add support for DSV4_HC_PRE

* hexagon: hsv4-hc-pre plain/gated split kernel and l2 prefetch optimization

* hex-dsv4-hc-pre: double-buffered DMA for prefill

* hex-dsv4-hc-pre: refactor DMA block setup

* hex-dsv4-hc-pre: precompute kernel params

* hex-dsv4-hc-pre: support for n_hc != 4

* hexagon: add support for DSV4_HC_POST

* hex-dsv4-hc-pre: refactor gated PRE prefill DMA to dst writeback after each block

* hex-dsv4-hc-pre: remove in-loop vector spills and redundant device-side checks

* hex-run: add support for test-llama-archs tool

* hex-dsv4: refactoring and optimizations

Remove non-dma paths.
Clean up DMA pipelines and dma_addr_t casts.
Add dims collapsing support.

Tested with test-llama-archs -a qwen4ext and test-backend-ops

---------

Co-authored-by: Max Krasnyansky <maxk@qti.qualcomm.com>
Assets 2
Loading