GS/DX11/GL: Expand the fb optimization hit detection. by lightningterror · Pull Request #14567 · PCSX2/pcsx2

lightningterror · 2026-06-09T08:09:48Z

Description of Changes

GS/GL: Expand the fb optimization hit detection.

GL: allows the same texture to be bound as an attachment or shader resource under the condition that only one operations is performed, we either only read from it or only write to it, so we can just check if we are actually writing to it which is enough.

Both DX11/GL: we can actually expand the hit detection even if either color or depth trigger a state change. This allows not the current but subsequent potential state changes to be avoided.
Also clean up the checks on dx11, only keep the texture comparisons themselves as they should be enough.

Rationale behind Changes

Optimizations.

Suggested Testing Steps

Test various games as a lot of games were hit, dump run only done on dx11, need to do gl.
dx11: Most see a single digit reduction, there are a few in double digits.

Did you use AI to help find, test, or implement this issue or feature?

No.

TJnotJT

Looks fine as far as correctness

SternXD

Tested in NFS MW 2005 and KH1 didn't see any issues

----------------------------------------------------------------------------------- eden.mk 4edb1ac6c5de888c83d0c838f30a44e929742979 # Version: Commits on Jun 12, 2026 ----------------------------------------------------------------------------------- [common/fs] create XDG_DATA_{HOME,CACHE,CONFIG}_HOME if we are the first application to ever be ran or they dont exist for a reason (#4087) ------------------------------------------------------------------------------------- hatari.mk c9906f10bc88b2ed57d5b01392a0c0c950a5b86b # Version: Commits on Jun 12, 2026 ------------------------------------------------------------------------------------- Rename JIT file to .cpp (not used in Hatari) (from WinUAE 6.1.0 beta8 2026/06/11), ---------------------------------------------------- pcsx2.mk v2.7.409 # Version: Commits on Jun 12, 2026 ---------------------------------------------------- - [GS/DX11/GL: Expand the fb optimization hit detection.](PCSX2/pcsx2#14567) ------------------------------------------------------------------------------------- ppsspp.mk 5d325290403006fc2b57f2488881f1cfd1463166 # Version: Commits on Jun 12, 2026 ------------------------------------------------------------------------------------- Merge pull request #21820 from fp64/vdot-fix Fix VFPU dot bug, ------------------------------------------------------------------------------------- snes9x.mk 2462dc98b288b66082ad57f8429cab2073f4108f # Version: Commits on Jun 12, 2026 ------------------------------------------------------------------------------------- Qt: Poll joysticks more often., ------------------------------------------------------------------------------------- vita3k.mk 985ac699c600277d5012ac06caa8212e6f5bf230 # Version: Commits on Jun 12, 2026 ------------------------------------------------------------------------------------- emuenv/app: Remove base_path (#4018), --------------------------------------------------- nblood.mk r14355 # Version: Commits on Jun 12, 2026 --------------------------------------------------- - ---------------------------------------------------------------------------------------- openjkdf2.mk 02bc22e8d6104d744825fa0e448b340a5a46b62a # Version: Commits on Jun 12, 2026 ---------------------------------------------------------------------------------------- Merge pull request #422 from bobberb/nix-flake-pr Add nix flake for Linux build and dev shell, ---------------------------------------------------------------- wine-mono.mk wine-mono-11.2.0 # Version: Commits on Jun 12, 2026 ---------------------------------------------------------------- Release 11.2.0., ------------------------------------------------------------------------------------------ openbor7530.mk 8a36651101b2051392845fa6b5bcd3a5ccb59702 # Version: Commits on Jun 12, 2026 ------------------------------------------------------------------------------------------ Add PAK64 support to borpak - Add PAK64 archive format with 64-bit offsets and sizes - Preserve legacy PACK output as the default format - Add -6 and --pak64 options for large archive creation - Auto-detect PACK and PAK64 during list/extract operations - Improve malformed archive checks and path traversal protection - Add clearer errors when legacy PACK exceeds 32-bit limits - Remove mandatory Windows pause from normal terminal builds - Add helper batch files for PACK and PAK64 creation, --------------------------------------------------------------------------------------- vpinball.mk 8be49665164d9e80479df9349151580370607f48 # Version: Commits on Jun 12, 2026 --------------------------------------------------------------------------------------- ui: let floating ancillary windows be moved by dragging on Wayland (#2772) On Wayland applications may not position windows themselves, so the drag support of the display settings UI cannot work and the borderless windows could not be moved at all. Declare them draggable there so the window manager moves them when dragged anywhere. Other platforms keep the settings UI drag (a full surface hit test would swallow the mouse events it relies on and make the windows movable during play)., -------------------------------------------------------------------------------------- yquake2.mk 4857b654e57154526e1471381d25bde7fec98899 # Version: Commits on Jun 12, 2026 -------------------------------------------------------------------------------------- game: make monster internal functions static, --------------------------------------------------------------------------------------------- libretro-fbneo.mk 9723c47c2d782ccb8cf48f96568a89a44631e6ba # Version: Commits on Jun 12, 2026 --------------------------------------------------------------------------------------------- (libretro) update files, -------------------------------------------------------------------------------------------------- libretro-gearcoleco.mk 09ab31b7b939e5fe32771d29a1a2a170f05b5887 # Version: Commits on Jun 12, 2026 -------------------------------------------------------------------------------------------------- Harden MCP HTTP transport configuration and authentication, ------------------------------------------------------------------------------------------------- libretro-geargrafx.mk 7c51cfa4b4654e823da7f5a6be7e489979bce17d # Version: Commits on Jun 12, 2026 ------------------------------------------------------------------------------------------------- [debugger] Add option to pause on BRK #n, -------------------------------------------------------------------------------------------------- libretro-gearsystem.mk ed2949d7e39d6085cfe5de9d3e32f70e354eee94 # Version: Commits on Jun 12, 2026 -------------------------------------------------------------------------------------------------- Harden MCP HTTP transport configuration and authentication, ---------------------------------------------------------------------------------------------------- libretro-parallel-n64.mk 324974795b925fea4b13fea96c446615d34f1868 # Version: Commits on Jun 12, 2026 ---------------------------------------------------------------------------------------------------- rdp_emit: scale gSPModifyVertex XYSCREEN writes to the pipeline's form The G_MWO_POINT_XYSCREEN case stored the command's two raw s10.2 halves into the vertex record's screen x/y, but this pipeline carries screen coordinates as s10.2 << 14 (the triangle bridge reads them back with >> 14 for the cross-product cull and quantizes to the quarter- pixel grid from the same scale). Storing the halves unscaled collapsed every patched vertex to screen (0, 0), so triangles built from them were rejected as zero-area degenerates. The New Tetris is the canonical victim: it draws each tetromino sprite (the falling piece, the next queue, the hold box, the ghost, and most HUD elements) as quads of four zero-position vertices whose screen placement comes entirely from gSPModifyVertex, so all of them were invisible. With the halves scaled into the record's form, the game's boot, menus, and marathon gameplay verify pixel-exact against the cxd4 LLE stream, including the previously divergent shade alphas on patched background geometry., ---------------------------------------------------------------------------------------------- libretro-ppsspp.mk 5d325290403006fc2b57f2488881f1cfd1463166 # Version: Commits on Jun 12, 2026 ---------------------------------------------------------------------------------------------- Merge pull request #21820 from fp64/vdot-fix Fix VFPU dot bug, ---------------------------------------------------------------------------------------------- libretro-prboom.mk 061f9510746e67e8fc6cc66e6ac5b274bb89c489 # Version: Commits on Jun 12, 2026 ---------------------------------------------------------------------------------------------- decorate: recolour custom projectiles via their Translation property The female imps fire A_CustomMissile(\LowerImpBall\, ...), and LowerImpBall is a custom projectile whose DECORATE Translation remaps the imp fireball's palette (and which the engine previously ignored, firing the stock orange ball). Parse the Translation property into a 256-byte palette table, capture the projectile class each A_CustomMissile names, resolve it to its table, and carry that onto the fired ball so the sprite renderer recolours it. Mechanically: a mobj gains a translation-table pointer and a vissprite an xlat pointer; the sprite column path uses an arbitrary remap table when present, beyond the handful of built-in player-colour translations. The renderer validates the pointer against the built-table storage so a stale vissprite slot can never feed a wild table to the column drawer. The mobj field is appended at the end of the struct and zeroed across the savegame load path alongside the other appended pointers, so a recycled heap block cannot leave it poisoned. Vanilla behaviour and the demo hash are unchanged., ---------------------------------------------------------------------------------------------- libretro-stella.mk 98cf11a548e26bcbdf54663243a79442b17bc42c # Version: Commits on Jun 12, 2026 ---------------------------------------------------------------------------------------------- Fixes for minor warnings from clang-tidy.,

lightningterror added this to the Release 2.8 milestone Jun 9, 2026

github-actions Bot added GS GS: Direct3D 11 GS: OpenGL labels Jun 9, 2026

lightningterror commented Jun 9, 2026

View reviewed changes

Comment thread pcsx2/GS/Renderers/OpenGL/GSDeviceOGL.cpp Outdated

lightningterror force-pushed the gs_aggressive_fb_opt branch from e3d6ae5 to eeb7acc Compare June 9, 2026 08:16

lightningterror changed the title ~~GS/GL: Expand the fb optimization hit detection.~~ GS/DX11/GL: Expand the fb optimization hit detection. Jun 9, 2026

lightningterror requested a review from TJnotJT June 9, 2026 09:24

TJnotJT reviewed Jun 9, 2026

View reviewed changes

Comment thread pcsx2/GS/Renderers/DX11/GSDevice11.cpp

TJnotJT approved these changes Jun 9, 2026

View reviewed changes

lightningterror force-pushed the gs_aggressive_fb_opt branch from eeb7acc to f41660f Compare June 9, 2026 23:45

GS/DX11: Expand the fb optimization hit detection.

384d2d3

lightningterror force-pushed the gs_aggressive_fb_opt branch 2 times, most recently from 2438cf2 to 7491cd1 Compare June 10, 2026 10:11

GS/GL: Expand the fb optimization hit detection.

c23c06a

lightningterror force-pushed the gs_aggressive_fb_opt branch from 7491cd1 to c23c06a Compare June 10, 2026 10:15

SternXD approved these changes Jun 12, 2026

View reviewed changes

JordanTheToaster approved these changes Jun 12, 2026

View reviewed changes

lightningterror merged commit 16592f3 into master Jun 12, 2026
28 checks passed

lightningterror deleted the gs_aggressive_fb_opt branch June 12, 2026 20:29

lightningterror mentioned this pull request Jun 16, 2026

GS/DX12: Expand the fb optimization hit detection backport from GL/DX11. #14595

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

GS/DX11/GL: Expand the fb optimization hit detection.#14567

GS/DX11/GL: Expand the fb optimization hit detection.#14567
lightningterror merged 2 commits into
masterfrom
gs_aggressive_fb_opt

lightningterror commented Jun 9, 2026 •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

TJnotJT left a comment

Uh oh!

SternXD left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Uh oh!

Conversation

lightningterror commented Jun 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description of Changes

Rationale behind Changes

Suggested Testing Steps

Did you use AI to help find, test, or implement this issue or feature?

Uh oh!

Uh oh!

Uh oh!

TJnotJT left a comment

Choose a reason for hiding this comment

Uh oh!

SternXD left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

lightningterror commented Jun 9, 2026 •

edited

Loading