Demon's Souls (PPSA01341)

[fisherjoey]

Tested

• Title: Demon's Souls, PPSA01341, v1.07 (Bluepoint engine).
• Host: Linux 7.2 (Nobara 44), AMD Ryzen 5 7600X (6 cores / 12 threads), NVIDIA RTX 5070 Ti (driver 595.91.07), 32 GB, GCC 16.2.
• Build: main at df16c4c2 plus fix(guest-arena): place guest address space in the PS5 window on Linux #861, fix(module): resolve the default dlsym scope against the main program #864, perf(agc): query guest page access with PROCMAP_QUERY on Linux #157, perf(agc): pipeline queue-0 draw commits behind the packet worker #1201/perf(agc): cheaper draw commits; pipelined draws on by default #1202, feat(recompiler): bindless image tables up to the device's sampled-image limit #1184, fix(agc): release the current address space with the image mirrors #1071 and the local changes listed below.
• Last update: 2026-10-08

Status

In game on Linux, but not playable. A new game reaches the tutorial level (Outpost Passage) with the HUD drawn. The level runs at about 2.0-2.2 fps, and the 3D scene renders wrong (blocker B).

Everything before that works: SIE logo, PS Studios movie, language, brightness and subtitle setup, the opening cinematic (33-40 fps), the title screen (50-53 fps), the main menu, "Continue Offline" at the online prompt, character creation (about 10 fps) and the new-game intro. Online fails cleanly and offers offline play. The name entry dialog returns its default text, so the character is named "Slayer".

Shortcut to the level. +ds_enableDebugStartMenu=true in app0/PackageCmdLineArgs.txt (next to the shipped +cp11_callHomeLockDownEnabled=false) opens the game's debug start menu instead of the logos and title. Walk Through, then page 6 of the map list, "(T-1) Tutorial (Fort Entrance)", loads the level with the default loadout about 2 minutes after launch, with no character creation or intro. It needs more bindless slots than the normal route (APS5_BINDLESS_SLOTS=128, see C). It starts at a different spot from a new game, so fps from the two routes isn't comparable.

Settings used: APS5_BINDLESS_SLOTS=48 (128 for the debug start route), APS5_THREAD_PRIORITY=high, APS5_TEXTURE_CACHE_MIB=8192, APS5_STAGING_POOL_MIB=2048, APS5_PIPELINE_CACHE_ENTRIES=4096.

Open blockers

	Stop	Status
A	Level at ~2 fps. A frame is ~480 ms, and ~95% of it is the queue 0 worker on the CPU (capture and validation of ~10k draw packets and ~4k dispatches per frame). The GPU is busy ~205 ms per frame, overlapped.	Pipelined draws (#1201/#1202) took it from 1.4 to 1.9 fps. Next: moving indirect draws and dispatch device calls behind the same committer.
B	3D scene in the level is fog-white (black with #1346), with bright bars and purple bands. The HUD is correct.	Root cause candidate, fix in testing: every pass carries the image faithfully, but the scene gets no direct light. The sun-shadow mask pass reads 16 shadow cascades that are never drawn, because all depth-only draws in the level are rejected (context register 0x2df never written, front/back depth bias differing, color format 0). The mask says "in shadow" everywhere. The bands are a separate aliasing issue.
C	Bindless image tables. Compute shader 0x248946900 selects from a material table with 37-93 keys depending on the scene. #1184 doesn't lift the slot cap for material-selected tables, so APS5_BINDLESS_SLOTS is still needed, and its clamp of 48 is too low for the debug start route.	Local: clamp raised to 1024.

On a cold shader cache the level shows terrain and foliage, because the shadow and fog passes are still compiling and the frames carry only part of their work:

Stops on main, in the order they are hit

1. Black screen before the first logo. The title opens /app0/packagecmdlineargs.txt; the file is PackageCmdLineArgs.txt. Linux: guest paths are resolved case-sensitively, so titles miss files the console would find #1047; fixed by fix(libc): resolve case differences in guest paths and mounts #1321 (merged).
2. Every allocation fails after one refused host import (NVIDIA on Linux, read-only ranges). Linux/NVIDIA: a refused host import of a read-only range makes every later vkAllocateMemory fail #1022; fix(agc): skip Linux host imports of ranges not mapped read-write #1051 (merged).
3. 0.65 fps in the intro movie. Dispatch validation re-read /proc/self/maps about 48,000 times in 25 s. perf(agc): query guest page access with PROCMAP_QUERY on Linux #157 (merged).
4. Title screen at ~6 fps. The 6 GiB host import budget runs out and a 64 MiB render target falls back to CPU write-backs. feat(agc): host import windows, least-recent eviction under the budget, recoverable GPU out-of-memory #290's default (half of RAM) fixes it.
5. Sets that sample a depth surface always take the full proof walk. AGC: sets that sample a depth surface never pass the fast proof (no record) #1132; PR perf(agc): prove sampled depth surfaces in the fast resource proof #1430.
6. A comparison-sampled shadow map is compared against guest memory under the GPU mutex. AGC: a comparison-sampled R32F surface stays a CPU snapshot, compared against guest memory under the GPU mutex after every GPU write #1133; fix on fix/depth-compare-copy.
7. The title's job threads spin and starve the queue workers on a 12-thread CPU. Opt-in priority for queue workers and the presenter: feat/queue-thread-priority. Locally, a port of KytyPS5 test(shader): execute tbuffer_load_format_d16_xyzw against hardware results #599's idle backoff for the job-worker poll loop (title-specific, patched by signature) cuts the title's CPU use from ~920% to ~300% with no fps loss.
8. Garbage descriptors about a minute into boot (compute shaders 0x248a1dd00 and 0x2489e6d00). Compute buffers wait on a label written at the end of the graphics buffer, and the title fills the dispatch tables after submitting; the wait passed when queue 0 recorded the label. 15 of 15 clean runs with the fix. Branch fix/agc-compute-gate-capture; issue to follow.
9. Crash while loading the tutorial level. A cutscene uses a 16-channel ATRAC9 stream (config 30 73 c1 7e) that libSceAjm rejected. Fixed locally (one LibAtrac9 decoder per channel, interleaved). Issue to follow.
10. Level: every guest-memory check is a system call on Linux. describePages stores PROCMAP_QUERY results only on Windows: 2.1M calls per 20 s in the level. Caching them in the page table on Linux takes that to ~2.7k. Fixed locally; perf(agc): cache page permissions within registered allocations #1330 covers part of it (it leaves ~500k queries per 10 s here).
11. Level: texture cache churn. The texture key includes the T# streaming-feedback fields (MIP_STATS_COUNTER_EN/ID), which the title rotates, so each texture was cached under about six keys and evicted live twins. Masking them took evictions from ~800 per 10 s to 0. Fixed locally.
12. Level: the pipeline store is bounded at 256 entries; the level uses ~317, so ~1,600 pipelines were recreated per 10 s. Fixed locally with a configurable bound.

Also fixed locally and not filed yet: rejected draws memoized instead of rethrown (~36k per 10 s in the level), indirect-dispatch recipe rebuilds on the title screen, and index-based overlap queries for pending storage images.

Not established

• Whether blockers A and B also happen on Windows or on AMD GPUs.
• Whether the window-close crash (exit -11) is the title's own audio shutdown: with fix(agc): release the current address space with the image mirrors #1071 the driver-side fault is gone, but the exit still faults in a job-worker fiber after sceAjmBatchWait fails.

[fisherjoey]

Update, top post refreshed.

• The tutorial level is at about 2.0-2.2 fps, up from 0.2. Most of it came from local fixes: the Linux PROCMAP_QUERY cache, a texture key that ignores the T# streaming-feedback fields, and a larger pipeline store. perf(agc): pipeline queue-0 draw commits behind the packet worker #1201/perf(agc): cheaper draw commits; pipelined draws on by default #1202 added 1.4 to 1.9 fps on top. The rebase is now on main df16c4c.
• The debug start menu (+ds_enableDebugStartMenu=true) loads the tutorial about 2 minutes after launch, without going through the title or character creation.
• Blocker B is narrowed to the sun shadows. Lighting, TAA and tonemap all carry the image correctly, but the shadow mask is computed from cascades that are never drawn, because the level's depth-only draws are rejected. A fix for those rejections is being tested. With a cold shader cache the scene is partly visible (new screenshot), because the shadow and fog passes haven't compiled yet.
• A port of KytyPS5 test(shader): execute tbuffer_load_format_d16_xyzw against hardware results #599's job-worker idle backoff takes the title from ~920% CPU to ~300% with no fps loss.

AI-assisted: yes