Skip to content

Conversation

CyrilKoe
Copy link
Contributor

@CyrilKoe CyrilKoe commented Jan 5, 2025

The region at [0x2_8000_0000 ; 0x3_0000_0000[ can be used to bypass the LLC when the LLC is enabled.
This is useful to speedup device DMA when LLC is enabled (of course requires managing LLC flushes in software).
Note that the LLC already has an AXI demux - mux which bypasses the LLC by default. Thus there is no need to change anything in the LLC or its internal address map.

Maintainer's Note: TODO:

@CyrilKoe CyrilKoe requested a review from alex96295 January 5, 2025 13:43
@CyrilKoe CyrilKoe self-assigned this Jan 5, 2025
@CyrilKoe CyrilKoe changed the title Add LLC bypass region for Carfield Draft: Add LLC bypass region for Carfield Jan 5, 2025
@CyrilKoe CyrilKoe marked this pull request as draft January 5, 2025 14:17
@paulsc96
Copy link
Member

Would you be interested in retargeting this to main? I cannot adopt the LLC partitioning change as is (#74 seems to be stuck in serious limbo) but I am generally interested in this feature for mainline, e.g. for efficient VGA frame buffer accesses.

What I am not sure about is whether it is flexible enough. A configurable range would be nice.

@alex96295
Copy link
Member

alex96295 commented Oct 10, 2025

For me, yes; it was not merged in Carfield as I was in the process of bumping Cheshire first (PR308, which is pending as it waits for this PR to be merged)

I agree that the LLC partitioning as-is cannot be merged, but this features could be placed on top of LLC main

@paulsc96
Copy link
Member

paulsc96 commented Oct 10, 2025

Then that sounds like a path forward. Kinda. Updated the PR message.

There is an LLC PR for partitioning (pulp-platform/axi_llc#10), but either something is bizarrely wrong with it (5k+ changes) or that is just not getting merged. Either way, it was never undrafted.

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.

3 participants