Skip to content

[vk] Take a range's first page from its first word in the blk question - #2574

Draft
DraChimeraVee wants to merge 1 commit into
mainfrom
dravee/vk/uma-blk-first-word
Draft

DraChimeraVee wants to merge 1 commit into
mainfrom
dravee/vk/uma-blk-first-word

Conversation

@DraChimeraVee

Copy link
Copy Markdown
Contributor

rg_blk masked the bits before a range's first page in every word but the first: a range that starts inside a word was answered by pages before it, and missed a blk page in the first bits of its later words, so vk_uma_read could take in place a range whose newer bytes are elsewhere. From an outside static audit of the tree (its F171). The page model is #1714's; this changes the one mask.

The mask moves to gpu/vk/uma_pages.h, as uma_budget.h holds the budget's arithmetic, so a test reaches it without a device.

Testing, on main 3abdb88 with the other branches of this audit applied:

  • The new vk_uma_pages passes: every range of the first 300 pages against its pages one by one. With main's expression it fails: pages 70 to 200 took bits 0 to 5 of word 1 and missed them in words 2 and 3.
  • vulkan_headless_uma was not run here.

rg_blk masked the bits before the first page in every word but the first: a range from
inside a word was answered by pages before it and missed blk pages in its later words.

Change-Id: Ifc62e06bc5da427e4c41f38aa02fee6af9020bdf
Signed-off-by: AF52 <AF52@reden-emu.dev>
@github-actions

Copy link
Copy Markdown

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

area: gpu kind: bug Something does not work preview-effort: small A contained change priority: high Next in line

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant