Grand Theft Auto V (PPSA04264)

[MichonGoddijn231849]

Tested

• Title: Grand Theft Auto V, PPSA04264, v01.005. Relinked with its own sce_module.
• Host: Windows 11 Pro, AMD Ryzen 5 7600, NVIDIA GeForce RTX 5080 (driver 32.0.16.1074), 32 GB DDR5-5600, MinGW-w64 GCC 15.2.0.
• Last update: 2026-10-09

Status

On the gta-v/testing branch (based on main from before #639): in game on Windows. "Resume Story" starts the North Yankton prologue, which plays with rendering, reflections, music, speech, sound effects and the HUD working.

On current main (d70b899, 2026-10-09): stops at boot. With the open pull requests under "Needed on current main" below it presents its first frames (the start screens), then stops on a T# read from a material table (bindless image tables, #1184). Getting it in game on main is what I am working on now.

Graphics mode (gta-v/testing)	State
Performance	Playable, around 27-28 fps in the bank (was 20-23)
Fidelity, Performance RT	Renders correctly since #919, but only a few fps. The ray tracing passes are being worked on.

Only tested in the prologue, in the bank. The frame rate drops a bit depending on where you are and where you look.

Story Mode is an add-on entitlement. Listing GTAVSTORYMODE001 in anyps5-entitlements.ini (#153) enables "Resume Story".

Needed on current main

Since #639, main needs these open pull requests to get this far, in the order of the stops they fix:

• fix(agc): skip prim state preparation for a patch/hull mismatch #2293 prim states whose patch topology and HS_EN disagree (fix by @CouchRiv)
• the image heap capacity of Sampled image heap of 16 rejects a retail shader before the metadata limit #1789 (candidate by @CouchRiv on fix-shader-sampled-heap-capacity)
• fix(agc): defer shaders that cannot be prepared at registration to their first use #1836 shaders that cannot be prepared at registration deferred to their first use, and fix(agc): accept bound copies of a registered shader header that differ only in user_data #1837 bound copies of a registered shader header
• fix(shader): keep runtime image mode switches inside continue targets #1964 runtime image mode switches inside continue targets
• fix(shader): lower the packed pixel ancillary when its extract operands fold after it #1904 and fix(shader): carry a packed pixel ancillary through phis when every reader extracts a modelled field #1934 the packed pixel ancillary word (or feat(recompiler): pixel ancillary word used whole #1179)
• next: bindless image tables beyond 16 slots on the new runtime ABI (feat(recompiler): bindless image tables up to the device's sampled-image limit #1184, being reworked)

How to run it

The gta-v/testing branch is the pull requests in the "Pull requests" section below merged together, plus the 2026-10-08 Performance work that is not split into pull requests yet. It is a testing branch, not for merging: each change is, or will be, its own pull request. The first launch compiles the shader cache and can sit on a black screen for a few minutes before the first frame.

Performance

Performance mode is CPU-bound, not GPU-bound. During gameplay the GPU is about 48% busy and the CPU about 25% (3 busy threads). A frame has about 1,800 draws, and one thread records them into Vulkan command buffers at about 23 µs each, which caps the frame rate. Going from 11 to around 20-23 fps came from pipelining that recording behind the packet decoding and making each draw cheaper. On 2026-10-08 a further set of draw-path changes took the bank from 23.5 to 27.8 fps (measured with the same build, switches off and on): finished draws freed by the decoding thread, fewer and batched GPU submits on a submit thread, render-target checks and draw keys reused while nothing changed, and wider write-watch reads. They are on by default on the testing branch and reach the pull requests separately. Next: moving more of the per-draw work off the recording thread; compute dispatches that no longer drain the pipeline are in #1747 (no change in GTA V, which has few dispatches between draws).

Open issues

	Issue	Status
A	Ray tracing modes run at a few fps	Being worked on: the ray tracing passes take most of the GPU time
B	Performance mode below 60 fps	Draw recording on one thread is the limit (see Performance)
C	First launch after a shader recompiler change waits on the shader cache	Known, not specific to this title

Pull requests

Every change on the testing branch that the title needs is in one of these. Stacked pull requests say so with "Depends on".

Rendering:

• fix(agc): depth fast clears through the HTILE metadata #374 depth fast clears through HTILE
• feat(agc): multisampled color and depth targets with CB resolves #419 multisampled targets with CB resolves (the reflection cube); fix(agc): CB resolve into any tiled color target #1189 resolves into any tiled target
• feat(agc): vertex clip and cull distances #423 vertex clip and cull distances (reflections, glass)
• feat(agc): depth surfaces read as textures and storage images #425 depth surfaces read as textures and storage images; feat(agc): sample depth plane arrays from a later base slice #1186 depth plane arrays from a later base slice
• feat(agc): single-sample sample masks as a per-quad-pixel discard #448 single-sample sample masks as a per-quad-pixel discard
• fix(agc): a fill of a surface's DCC keys is a key fill over other images' memory #1170 DCC key fills inside the memory of other images (the bloom stacked up into blue blobs)
• feat(recompiler): pixel ancillary word used whole #1179 the pixel ancillary word used whole

Shaders:

• feat(recompiler): bindless image tables up to the device's sampled-image limit #1184 bindless image tables up to the device's sampled-image limit
• feat(recompiler): FLAT addresses in the shared and private apertures reach LDS and scratch #1190 FLAT addresses in the shared and private apertures (the ray traversal stack)

Memory:

• feat(agc): host import windows, least-recent eviction under the budget #290 host import windows, least-recent eviction under the budget

Performance (11 to 23 fps):

• perf(agc): draws that differ only in user data reuse their shape's decode and stages #494 data-only draw hits
• perf(agc): linear color targets in registered guest memory get a resident image #1191 resident linear color targets
• perf(agc): cheaper shader-memory captures per draw #1193 cheaper shader-memory captures per draw
• perf(agc): pipeline queue-0 draw commits behind the packet worker #1201 pipelined draw recording
• perf(agc): cheaper draw commits; pipelined draws on by default #1202 cheaper draw commits, pipelined draws on by default
• perf(agc): dispatch-cache recipe hits as ordered commit items #1747 compute dispatches as ordered commit items (off by default)

[MichonGoddijn231849]

Update 2026-10-08: Performance mode in the prologue bank is now around 27-28 fps (23.5 before, same build with the new changes switched off). The changes are on by default on the gta-v/testing branch, which I updated today; the first launch after updating recompiles the shader cache. Details are in the Performance section above.

[MichonGoddijn231849]

@CouchRiv thanks a lot, both very helpful! The strcmp fix and main's 2093e74 are now on gta-v/testing (57c8a97). I took your rebases for #290 and #494 (checked and built here on Windows too, with you as co-author), and both are mergeable again. The #1191–#1202 stack stays parked for now: I'm re-cutting it from my main-based branch and will compare against yours.

[MichonGoddijn231849]

@CouchRiv thanks! I opened your prim state fix as #2293 (your commit, you as author). Your #1789 candidate also works here on Windows: agc_driver_typed_heaps passes and GTA V gets past the image heap stop to the s_swappc_b64 one, as you saw. With #1836, #1837, #1964, #1904 and #1934 on top it reaches its first frames, then stops on a T# read from a material table (the full table is in my reply to boykopovar's question).

[boykopovar]

Do you know what stage the game has reached on the current main branch?

[MichonGoddijn231849]

On current main (d70b899) it doesn't get past boot: it aborts a few seconds in, during shader registration, with VGT_SHADER_STAGES_EN=0x2002000: Patch topology and HS_EN disagree (GTA registers a tessellation stage set before its primitive state is set). On my branch I drop that failing hint and prepare the shader at first use instead. With that and a few other fixes it reaches story mode and renders correctly at ~24 fps. I can PR the hint fix if you want.

[MichonGoddijn231849]

Follow-up with a stack of open PRs on top of main d70b899 (Windows, RTX 5080):

On top of main	Result
nothing	aborts at boot: patch topology and HS_EN disagree
+ #2293	aborts at boot: shader image heap capacity exceeded (#1789)
+ CouchRiv's #1789 candidate	aborts at boot: data-dependent s_swappc_b64 (#1817)
+ #1836, #1837, #1964	~1 min in: packed pixel ancillary input has an unsupported live use
+ #1904 and #1934	first frames (start screens, ~29 fps), then ~2 min in: GetImageResource dword 0 is not a valid runtime value (a T# read from a material table, the #1184 area)

So with those PRs it presents its first frames on main; the next stop is the bindless material tables.