Uncharted: Legacy of Thieves Collection (PPSA05684) on Linux / NVIDIA

[XairRrR]

Status: boots, plays the studio intro video with sound, reaches the main menu (collection choice, then Story / Extras / Options, navigable) and starts Uncharted 4 chapter 1 (boat intro, Henry Avery text, sound). Not on plain main: the title needs the open image-table PRs #757/#758 and the depth PRs #425/#2076, plus local changes listed below. The 3D menu runs at about 7 fps and its lighting is magenta.

Setup

• Title: Uncharted: Legacy of Thieves Collection, PPSA05684 (EP9000-PPSA05684_00-UNCHARTED4R00000), version 01.002.000, master 01.00.
• Host: Arch Linux, kernel 7.2.6-zen, i9-14900KF, 32 GB RAM, RTX 4070 Ti SUPER 16 GB, NVIDIA 615.71.09 (Vulkan 1.4.351).
• AnyPS5: local test branch on main 197ffac + feat(agc): implement image tables loaded through 64-bit pointers #758 (includes feat(agc): implement image and sampler tables indexed at runtime offsets #757) + feat(agc): depth planes sampled as 2D arrays, cubes and integer views #425 + feat(libSceAgcDriver): take a one-layer 2D array view of a depth surface plane as the plane #2076 + about 40 local commits. Release, GCC.
• Inputs: eboot.bin and sce_module/libc.prx unwrapped from SELF to ELF.
• Run: APS5_HOST_IMPORT_MIB=16384 APS5_NO_SNAPSHOT_CHECK=1 APS5_IMAGE_TABLE_KEY_SCAN=262144, ulimit -n 65536, ulimit -c 0 (a full core dump of the ~20 GB process exhausts RAM). About 20 GB of free host RAM is needed.

Milestones

Date	Milestone	Frame rate
2026-10-08/09	Boots, window opens, first flips	none yet
2026-10-09	Sony studio intro video plays with sound	video pace
2026-10-09	Main menu: collection choice, Story / Extras / Options	not measured
2026-10-09	Chapter 1 boat level, Henry Avery text, sound (textures largely missing, local null-table hack)	~1 fps (window title)
2026-10-09	Image tables from #758 replace the null-table hack; menu still reached	not measured
2026-10-11	Menu performance work starts	4.7 fps
2026-10-11	Current branch	6.4-6.7 fps (one run at 8.4)

Progress, in order

1. Boot to first flip (2026-10-08/09). The video out path needed VRR peg/unpeg (#2918), the voice port mute flag (#2919), command-buffer flip buffers resolved when the GPU reaches the flip (#2204), sceGameLiveStreamingGetCurrentStatus2 (#2205), and marker pops without a pushed marker (#2202). Recompiler fixes found on the way: SDWA destination selects on v_not_b32 (#2200), small return blocks shared by early exits (#2201), extract folding into OpVectorShuffle only between vectors (#2915; NVIDIA crashed in the driver otherwise).

2. Intro video. libSceVdecsw implemented on top of libSceVideodec2 (#2203). The movie never ended at first: the game waits for is_last_frame on the last picture, which is now held until Finalize. Layouts follow shadPS4's vdecsw.h. The Sony studio intro plays with sound.

3. Memory after the intro. Anonymous memory grew from 6 to 10.8 GB in ~100 s. A jemalloc heap profile showed 3.3 GB of SPIR-V decoded from the shader disk cache and kept forever. Local commits reload disk-cached SPIR-V on use and release freshly compiled SPIR-V once the cache holds it: menu RSS 7.9 GB -> 3.9 GB. #3247 (closed) had the same idea. The 1.6 GiB direct range also needs a larger host-import budget (APS5_HOST_IMPORT_MIB=16384), otherwise it goes to heap mirrors and the process runs out of memory.

4. Main menu (2026-10-09). Needed: depth/stencil clear draws through DB_RENDER_CONTROL (#2219), rect-list fault buffer binding in the tessellation control group (#2198), 8_SNORM sampling (#2199), zero-scale viewports (#2197), UINT16 color exports, COLOR_8_8 SWAP_ALT targets (#2916) and blending into them (#3194), shader variants registration did not predict (#2917). Diagnostics merged for the failures seen: #2920, #2921, #2922.

5. Chapter 1. Shadow cascades sample a 2048x2048 D16 depth surface as a 4-layer 2D array T#; that needed depth surfaces read as textures and storage images (#425) and the stencil plane as a storage image (#2076). #425 dropped the first-layer 2D view of a depth plane array, which put a blur over the menu; restored locally. The boat level renders and plays, at about 1 fps at the time.

6. Image tables instead of local hacks. Several compute shaders read T#s from bindless tables with runtime offsets, one with 2809 distinct keys, and the global texture table has 65536 entries. Before, these ran only with a local hack that read oversized or undecodable tables as null (textures missing). The branch now merges #758 (via #757), with its heap sizes aligned to the merged #2347, #2141 and #2282. It needs APS5_NO_SNAPSHOT_CHECK=1: the CPU updates the 65536-entry table while it is captured, which fails the "guest snapshot differs from registered memory" check. That is worth a look on #757.

7. Menu performance (2026-10-10/11), measured on the menu with APS5_PROFILE_DRAW=1 and A/B runs of 240 s:

Change	Where
Import adjacent registered ranges as one host-import span	#3158, merged
Prepare shader ABI variants before taking the preparation lock (Shader.Register 169 s -> 15.5 s per run)	#3159, open
Insert a draw into the draw cache on the second sighting of its key	#3985, open
Bulk image-table scans: ShaderMemory::readWords bulk copy, single-pass key narrowing without divisions, one read per address-mode record, last-page reuse	local; depends on #757/#758, last-page reuse is in #1193
Linux readable-page cache in GuestMemory::describePages (one PROCMAP_QUERY ioctl per check before)	local; open upstream as #1810 and #1330
Single-layer 2D and 2D-array views share one storage image (stopped ~7 GB per 10 s of write-back and re-upload of 2560x1440 images)	local
Resident single-slice linear color targets; small storage image baseline without a GPU wait	local; open upstream as #1191, merged as #2791

Menu measurements. Each run boots the title and reaches the menu with scripted Cross presses, 240 s in total. Frame rate and frame times are taken over the last 120 s from frame-timing.log. "1 s frames" counts frames 990-1030 ms after the previous one (the stalls listed under open problems). RAM is the peak anonymous memory of the process.

Build	fps	Median frame	p90 frame	1 s frames	Peak RAM
Baseline, start of the perf work	4.73	198.6 ms	226.8 ms	0	11.0 GB
+ bulk image-table reads and key narrowing	6.55	140.8 ms	172.5 ms	0	10.2 GB
A/B: before shared single-layer storage image + page cache	4.71	196.9 ms	217.2 ms	1	9.1 GB
A/B: after	6.55	140.1 ms	167.2 ms	0	9.1 GB
Layer aliases, off / on	6.03 / 5.90	152.7 / 153.8 ms	171.7 / 179.8 ms	0 / 0	9.3 / 9.0 GB
Layer borrowing, off / on (2 runs each)	6.09, 6.12 / 6.12, 5.94	~152 ms	~174 ms	1, 1 / 0, 1	~8.9 GB
Resident linear targets (2 runs)	6.18, 6.31	143.8, 146.0 ms	163.7, 164.2 ms	2, 0	7.8, 8.3 GB
+ deferred small-image baseline (stall-free run)	6.80	134.9 ms	156.8 ms	0	9.6 GB
Final A/B: before the last two commits (2 runs)	5.73, 6.26	147.9, 137.2 ms	173.8, 171.1 ms	0, 58	10.6, 11.1 GB
Final A/B: after (2 stall-free runs)	6.55, 8.40	135.9, 119.8 ms	161.0, 144.8 ms	1, 0	10.8, 11.7 GB
Current branch (3 runs)	6.69, 6.63, 6.42	133.8, 133.2, 134.3 ms	161.0, 159.3, 157.0 ms	1, 2, 0	9.4, 9.6, 9.4 GB

8 of the 38 menu runs that day had 20 or more 1 s frames, with and without each change; the worst dropped to 2.6-3.6 fps over the window. Those runs are left out of the table except where noted.

Result: 4.7 -> 6.4-6.7 fps in the menu, median frame 198 ms -> 134 ms. GPU time is only ~28 ms per frame (APS5_PROFILE_GPU=1); the CPU queue thread is the limit. Per 10 s on that thread: draw capture walk ~2 s, Graphics::Draw ~2.4 s, image tables ~1.1 s, WAIT_REG_MEM ~1.5 s.

Tried with no measurable gain: layer aliases of 2D arrays and GPU copies between layer images and the array (the 4-layer R8 array round trips moved to a later buffer copy that reads the same memory), and larger result memos (snapshots never repeat).

Fixes from this bring-up that are open upstream

• perf(agc-driver): prepare shader ABI variants before taking the preparation lock #3159: shader preparation lock (above).
• fix(agc): read zero for pixel inputs of parameters the vertex stage does not export #3523: pixel inputs of parameters the vertex stage does not export read zero (Mesa SI_PS_INPUT_CNTL_UNUSED), fixes Pixel shader reading a parameter the vertex shader never exports: rect-list draws zero-fill it, ordinary draws fail validation #2975.
• fix(agc): render a slice of a 2D array color target into that layer of the array #3524: a slice of a 2D array color target renders into that layer. The detiler's XOR swizzle takes the slice index, so a 2D target at the slice address scrambles slices. Uncharted itself does not use slice views.
• fix(agc): reject constant color factors and memory-writing stages in split SWAP_ALT blends #3561: follow-up to feat(agc-driver): blend into COLOR_8_8 SWAP_ALT targets with a separate alpha equation #3194, rejects constant color factors and memory-writing stages in split SWAP_ALT blends.
• perf(agc): insert a draw into the draw cache on the second sighting of its key #3985: draw cache admission (above).

Open problems

• Magenta lighting in the menu: the G channel is close to zero over the lit scene. Not found yet.
• Stalls at exactly 1 frame per second, often from boot, lasting tens of seconds, in some runs and not others with the same build. Not attributed.
• Intermittent NVIDIA Xid 13 (compute device loss).
• Fidelity mode crashes on MSAA targets; feat(agc): multisampled color and depth targets with CB resolves #419 covers multisampling.
• A bug on main seen during this work: creating a storage image of a multi-layer 2D array while a 2D image over one of its layers has pending results loses those results (flush with PublishScope::None, the CPU upload reads stale bytes). No PR yet.
• agc_driver_emulated_color_compare fails on plain main (a11c8f7).

Not proposed upstream

The branch also carries local test commits that are not meant for upstream: hacks that read undecodable descriptors as null, skipping draws whose stage interfaces disagree, the #758 adaptation, larger image capacities (32 slots per bindless table, 128 images, 4096 cached pipelines), and auto-input / debug traces used for these runs.

Help welcome on the magenta lighting and the 1 fps stalls. If someone has Legacy of Thieves on AMD, a comparison of the menu colors would tell whether the lighting is NVIDIA-specific.

AI-assisted: yes.