Resident Evil 2 (PPSA04288)

[AnthonyClemens]

Existing thread

• I searched this category for the title ID and no thread exists for this title on any platform.

Title

Resident Evil 2 (PPSA04288)

Content version tested

01.000.003

AnyPS5 commit

main 0214069 (2026-10-13) plus the pull requests and local changes listed below.

System tested

Windows 11 Pro (26200), AMD Ryzen 5 5600X, AMD Radeon RX 9070 XT (Adrenalin 26.9.1), 32 GB of RAM with a 19 GB page file, WinLibs GCC 15.2.0, Release.

Furthest point

Menu

What is shown

• On main before fix(agc): start the sample location registers at their clear-state value #4949, it stopped about 21 s in, at the first presented frame, with missing register in context bank at DWORD 0x2fe (766). This is the same stop as Resident Evil 4 in Resident Evil 4 (PPSA07411): start-up needs a library named Party.prx; with the missing exports added it stops at sce::Json::InitParameter2::setSpecialFloatFormatType #2776 and Resident Evil Village in Resident Evil Village (PPSA01556) #5169.
• With the changes below and a warm shader cache, the screen stays black for about 3 minutes while the game registers its shaders (about 7,800 variants). Then it shows:
  • the Capcom logo;
  • the PS4 save-conversion question (Yes/No), by about 190 s;
  • the autosave notice, Initial Settings, and the max/min brightness and brightness screens;
  • the Graphics setup screen, with its 3D corridor preview rendered. Ray Tracing was switched to Off for these runs;
  • the RE NET question and the "new game mode added" notice;
  • the main menu, about 4.5 to 6.5 minutes in: Leon S. Kennedy / Claire Redfield / Options over the rendered rainy 3D background, and then "Select a game mode".
• It stayed on the main menu, with no crashing.
• Frame rate: about 3.8 frames per second on the setup screens and the main menu.
• Memory: 13.9 GB peak private memory with APS5_TEXTURE_CACHE_MIB=1024 and the memory changes below. Without them, it went past 28 GB.
• New Game has not been tried yet.

Current blocker

• Retest on current main. The waterfall planning in feat(recompiler): plan waterfall loops over per-lane T#s as non-uniform tables #5089 merged and was then reverted (Revert #5089: agc_driver_waterfall_images fails SPIR-V validation on main #5824), so main again needs feat(agc): implement image and sampler tables indexed at runtime offsets #757/feat(agc): implement image tables loaded through 64-bit pointers #758 and the local per-lane changes for the waterfall image read below.
• Memory: preparing all the shaders at registration held 16.5 GB of private memory without the memory changes listed at the end. perf(shader): keep only evaluable branch conditions in resource plans #5181 (plans keep only evaluable branch conditions) covers most of it. After that, the biggest owners are texture uploads (about 3.8 GB, capped with APS5_TEXTURE_CACHE_MIB) and the dispatch cache (about 2.5 GB).
• Boot time (about 3 minutes of black screen with a warm cache): in a profile of the boot, the game's loader threads spent about 80% of their time waiting on the process-wide registration lock (Shader registration runs the recompiler front end under a process-wide lock even on disk-cache hits #2292). fix(agc): settle registration preparations before a preparation scope locks #5023 works in this area.
• Menu speed (about 3.8 FPS): not profiled yet on this title. perf(agc): pipeline queue-0 draw commits behind the packet worker #1201 and perf(agc): dispatch-cache recipe hits as ordered commit items #1747 work on the single queue worker.
• Unproposed stops, each passed locally with an experiment:
  • domain programs that read the ring at a control point computed from data (the domain counterpart of feat(shader): read hull inputs at a control-point index computed from data #4727);
  • two-sample rasterization into one-sample color targets, for the checkerboard rendering (approximated locally by drawing the first sample). simonbronk mentions a mixed-samples approach in Resident Evil Village (PPSA01556) #5169;
  • wave32 compute programs whose v_movreld window hid descriptor dwords (six programs);
  • guest snapshot differs from registered memory for image-table records. APS5_NO_SNAPSHOT_CHECK=1 also passes it; locally the dispatch is recaptured instead.

Pull requests

PR	Fixes
#4949 (merged)	First-frame stop on the unwritten sample-location registers (0x2fe)
#5234 (merged)	Integer tessellation partitioning
#5457 (merged)	Hull programs with one input or one output control point
#4172 (merged)	Hull exec masks restored through phis (branch condition Unreachable in a domain program)
#757, #758	Image and sampler tables at runtime offsets and through 64-bit pointers; the waterfall image read needs them
#5089 (merged, then reverted in #5824)	Waterfall loops over per-lane T#s. Locally, per-lane table planning on top of #758 was used instead
#5188 (merged)	Image descriptors read from constant buffers at uniform offsets
#5861 (merged)	Sample locations that vary between pixels (2 samples, swapped on odd rows). The local branch had its own approximation
#5884	Line-strip output of NGG geometry programs (VGT_SHADER_STAGES_EN=0x2030: unsupported geometry input or output assembly)
#5005 (merged)	104 compute programs that failed registration on scalar aperture source ... not modelled
#5181	Host memory of resource plans
#4806 (merged)	Prepared SPIR-V mapped from the shader cache entries. Locally, an older port of #3247 was used

Local changes not proposed

• The four stops listed under Current blockers.
• Per-lane image table planning for the waterfall loop, on top of feat(agc): implement image tables loaded through 64-bit pointers #758.
• A memory arena for resource plan values (taken from perf(shader): keep only evaluable branch conditions in resource plans #5181) and a filter that drops plan conditions the walker can never evaluate.
• SPIR-V eviction to the disk cache entry (a port of perf(shader): evict prepared SPIR-V to its disk cache entry #3247).
• Diagnostic switches: an allocation tracker and plan statistics.

How it was run

• Relinked with relinker --windows --registry from the dump's eboot.bin.
• APS5_TEXTURE_CACHE_MIB=1024 and a separate ANYPS5_SHADER_CACHE_DIR.
• Menus driven with a pad command file (APS5_PAD_INPUT_FILE, tools/pad_input.py). At about 4 FPS, 400 ms d-pad presses are sometimes missed.
• Every run had a private memory cap of 20 GB and was stopped by closing its window. An earlier run that was force-killed at about 28 GB was followed by a bluescreen (DPC_WATCHDOG_VIOLATION) on this PC.