Skip to content

[wip] [vk] Every choice by driver is a line of the extension bans - #2591

Draft
DraChimeraVee wants to merge 10 commits into
mainfrom
dravee/vk/bans-in-table
Draft

DraChimeraVee wants to merge 10 commits into
mainfrom
dravee/vk/bans-in-table

Conversation

@DraChimeraVee

Copy link
Copy Markdown
Contributor

The renderer chose seven ways by comparing a driver or a vendor in the code, where nobody reading the ban table saw them. They are lines of extension_bans.txt now, and a test refuses a new comparison outside the table's reader. No device changes what it does (see Notes for the one exception).

  • A prefer action: a way only the drivers listed take, asked with vk_prefer, with the same source, retest, checked and override fields; vulkan_ban_ignore goes over it.
  • The interlock for feedback draws: prefer lines for NVIDIA's driver, NVK, RADV, AMDVLK and AMD's driver, each with its measure (the integrated GPU's on AMD's own driver is in its line). [vk] Feedback draws through the interlock on AMD: the image was checked at one scene only #2580 stays open for reports.
  • No varying pass on NVIDIA's driver, HMMA by pairs on RADV, cached staging rings on ANV: a prefer line each.
  • Lavapipe's constant buffer windows, the imports that cost every submission on amdgpu's drivers, pipelines made one at a time on NVIDIA's Unix driver: an off line each.
  • runtime_architecture fails on a VK_DRIVER_ID_*, a vendor, driver or device ID compared, or a device name matched in src/, with the file, the line and what to do. DRIVER_CHOICES lists the comparisons that choose no way by driver: a pipeline cache's identity, gpu_pstate's interface by vendor, the GPU family the ASTC encoder and the UI's blend are written for, the GUI's driver packages.

Notes:

  • The interlock was chosen by the GPU's vendor ID and the lines go by driver, so a layered driver (Venus, Dozen, gfxstream) over an NVIDIA or AMD GPU now draws a feedback draw two times.
  • The key names are mine where Vulkan has none for the way (Location for the driver's own interpolation): say if another reads better.

Testing: vk_extension_bans holds each line against the rule the code had, for every driver the table names on every system; the new architecture rule gives 13 errors on main's src/ and none here.

AF52 added 9 commits November 20, 2026 00:44
A choice the renderer made by comparing a vendor or a driver had no place in the table,
whose lines only turn a key off. A prefer line names the drivers that take a way, with the
measure that showed it; vk_prefer asks, and vulkan_ban_ignore goes over it.

Change-Id: Iac8def1ca16ad0176dd48b9877ffc093dfe4015e
Signed-off-by: AF52 <AF52@reden-emu.dev>
device.c took the capture where the GPU's vendor was NVIDIA or AMD. The five drivers of
those vendors are prefer lines now, each with its measure, the integrated GPU's on AMD's
own driver among them; a test holds the lines against the old rule for every driver.

Change-Id: I5c12807c3393e8078dba21ad08e70aec616b464e
Signed-off-by: AF52 <AF52@reden-emu.dev>
render.c skipped the pass by comparing the driver with NVIDIA's. That driver is a prefer
line on Location now, with Eden's rule and the two games' cost as its source.

Change-Id: I7a87a92d06008ca7d17cf06e5732b21196c949f7
Signed-off-by: AF52 <AF52@reden-emu.dev>
device.c took the pairs by comparing the driver with RADV. RADV is a prefer line on
OpGroupNonUniformShuffleXor now, with its bench's 9.7% as the measure.

Change-Id: I33cc29e4b8b3dbb9df261892637b446ba2c98ed7
Signed-off-by: AF52 <AF52@reden-emu.dev>
cmd.c took cached memory by comparing the driver with Intel's Mesa one. ANV is a prefer
line on VK_MEMORY_PROPERTY_HOST_CACHED_BIT now.

Change-Id: I4bdeb875ab0ea84b7c8fbd71cf32cd611a243027
Signed-off-by: AF52 <AF52@reden-emu.dev>
pipeline.c turned the windows off by comparing the driver with lavapipe. The line on
maxPushConstantsSize holds it now, and the log says it with its source as any ban.

Change-Id: Ie03e667aa9ce0ad62dc9dc80764f484d89ac3d7d
Signed-off-by: AF52 <AF52@reden-emu.dev>
uma.c listed RADV, AMDVLK and AMD's Linux driver in a switch. They are off lines on the
host pointer handle type now, which uma_copies goes over as before.

Change-Id: Ibf46ab89883ee192beaad4de34defab6d824f529
Signed-off-by: AF52 <AF52@reden-emu.dev>
device.c compared the driver with NVIDIA's under an #ifdef for Windows. The line on
vkCreateGraphicsPipelines names NVIDIA's driver on every system but Windows.

Change-Id: I04bb29e451cb88b6b2a30e3adf98fba9687e0736
Signed-off-by: AF52 <AF52@reden-emu.dev>
The interlock's choice sat in device.c by vendor, where nobody reading the table saw it.
runtime_architecture now fails on such a comparison in src/ with its file and line, but
for the listed ones that choose no way by driver.

Change-Id: Iba9dc211ea337c71dce505a2b6c1abbe1497125f
Signed-off-by: AF52 <AF52@reden-emu.dev>
@DraChimeraVee DraChimeraVee added kind: enhancement New feature or improvement priority: medium Planned preview-effort: medium A focused change across a few files labels Nov 20, 2026
Four comment lines of this branch were longer than the files' limit.

Change-Id: Iac9278c20f25ae3e3576978c6ac9ae23d6362cb1
Signed-off-by: AF52 <AF52@reden-emu.dev>
@DraChimeraVee
DraChimeraVee force-pushed the dravee/vk/bans-in-table branch from 6341d0b to 6c919d6 Compare November 20, 2026 19:35
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

area: gpu kind: enhancement New feature or improvement preview-effort: medium A focused change across a few files priority: medium Planned

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant