Divinity: Original Sin II Definitive Edition (PPSA06866)

[XatiX]

Existing thread: searched this category for PPSA06866 on 2026-10-31; no thread exists.

Title: Divinity: Original Sin II Definitive Edition (PPSA06866, EUR)

Content version tested: 01.007.000

AnyPS5 commit: XatiX/AnyPS5 test/dos2-retest at 2e499428ca. This is a test branch, not proposed for merging as it stands. It is main f776fddcff plus:

Open pull request	Head tested	Purpose
#9717, carrying #9533 and #6300	a90d0b73bb1d	Stencil read from the 2x D32S8 target
#13021	48b3aa38abef	Blocker 3
#10382	ad201638c144	Speed: a build that never stores through BDA no longer counts as a writer

Own commit	Blockers	What it does
8817987a82	2	Draws DB stencil copies with a driver pixel program
1acc753642	4, 6	Keeps depth publication working over falsely flagged pages
212d8ce0ef	5	Runs the depth publication check before the command buffer is taken
80d7357f0a	7	Forwards socket calls from libkernel to libSceNet
324df3c270	Speed	Publishes only the HTILE plane for a read inside it
5cb8360028	Speed	Uses the collect memo in the native depth check
0d308e9c33	Speed	Raises the draw cache default to 32768 entries
442f617f36	Speed	Keeps a heap run's access check until its page states change
cd373edfe4	Speed	Replaces the slow page hash of 1acc753642
68139cbfac	Speed	Leaves the depth check early when nothing changed in the epoch
bb3d98869e	Speed	Builds a held publication from a queued fill's pattern, without a GPU wait
fb11c2675d	Test	Gives the depth test both constructor flags, which the #9717 and #13021 merge had collided
bc8e544eb0	4, 6	Keeps same-value stores visible to publication
2c0c4f9efd	Test	Covers a queued HTILE fill under a held publication
2e499428ca	Save thumbnails	Stores an occasionally written image before a label

#13734 is merged, so blocker 1 is gone from main. From the order of the stops below, main without these should stop at blocker 2. That was not rerun on this main.

System tested: Windows 11 Pro 10.0.26300, GCC 15.2.0 (WinLibs), RTX 5080 (NVIDIA 617.42), Ryzen 7 9800X3D, 32 GB RAM. Tested 2026-10-31.

Furthest point: In game (host imports on, default settings)

What is shown: in a manual session of 21 minutes on an earlier commit of this work, with no crash and a clean exit, a manual save was written and loaded back, and by the tester's report combat worked and audio appears to work. Saves list correctly with their name, time and version. Before that: boot, title screen, main menu, Campaign, difficulty, character creation, the origin story scene, the intro cinematic, the level load, then the opening in The Hold: Magister Siwan's dialogue, followed by free play with hotbar and minimap. On 456a530efb, before the speed commits, one scripted run of about 570 s and one manual session reached The Hold, with no crash or error in the log. An autosave completed in the manual session (Save to Running on client and server). Two scripted runs on the older base also reached The Hold, and three more reached character creation.

On 2e499428ca, speed depends on the scene. On a later save in The Hold it runs at about 23 fps (224 to 229 presents per 10 s). The opening scene ran at about 49 fps on an earlier commit, and every speed row below up to the rebase is from that opening scene. The main menu last measured about 15.5 fps, on an earlier commit. These come from scripted runs that stand still with the cursor showing. Pad input works for Cross, Options and L3. In play the title reports lower figures in some screens, such as about 17 fps with the Load menu open.

Not tested: everything past The Hold, and combat and audio beyond that one session. Stencil copy values are unverified: the scenes look right, but nothing reads them back and there is no test. Unit test coverage of the publication changes is incomplete, and this branch does not pass every depth publication case. Speed figures are from the run summaries, not tools/trace_digest.py comparisons.

Current blocker

None stops the title in the tested range. Speed is the main problem now, and it is CPU-bound: on the later save the GPU takes about 16 ms per frame, while Driver.Packet takes about 43 ms and the committer thread stays busy. See Speed below. Two picture faults remain, listed under Measured. Each stop below was only reached after the one before it was passed.

#	Stop	Status	Tracked by
1	sceHttp2SslEnableOption on a context handle	Fixed on main by #13734, confirmed on this title	#13715 closed, #13734 merged
2	DB_RENDER_CONTROL 0x88 stencil copy	Open. Local pass, unverified	No issue or PR
3	Native multisample depth publication	Passed with #13021	#13021
4	cannot distinguish an unwritten byte…	Local fix, on the branch	No issue or PR
5	Access violation in nvoglv64.dll, about 30 to 33 s	Cause found. Local fix, on the branch	#10492 describes the same bug
6	guest writes over native multisample depth and publication capture	On the branch, with a known limit	#13023 is related
7	Loop on the load after difficulty selection	Cause found. Local fix, on the branch	No issue or PR

Strong contenders

Open pull requests that bear on these blockers, checked on 2026-10-31. Findings are posted on each one.

PR	Blocker	Tested on this title	Notes
#13734	1	Merged. No sceHttp2 failure with main's copy (findings)	Fixes #13715
#9717, #9533, #6300	2	Yes. No change needed for what they claim (findings)	Required: without #9717 the title stops at about 45 s on native multisample depth sampled identity or format mismatch. The stencil copy pass itself is still an own commit
#13021	3	Yes. Needed two own commits to keep running (findings)	Exposes blockers 4, 5 and 6 on this title
#10492	5	No (note)	Same mechanism, written for D32 on #9950. It needs a rebase, and it may not cover the D32S8 path through #13021
#13023	6	No	Guest stores of the clear value over the stencil plane during import. Related, but a different case: blocker 6 hits the depth plane after import

No open pull request found for blocker 2 (stencil copy) or blocker 7 (socket forwards). Merged #9515 fixed the same kind of gap for getsockopt on sceNetSocket sockets, so blocker 7 could follow that pattern.

Measured

• Blocker 1: the handle is a context: id=1 kind=0 owner=0 transport=0. No sceHttp2 failure with fix(libSceHttp2): accept SSL options on a context #13734.
• Blocker 2: registers of the draw: DB_RENDER_CONTROL 0x88, DB_DEPTH_CONTROL 0, DB_Z_INFO 0xa0000187, DB_STENCIL_INFO 0x20000181, CB_COLOR0_INFO 0x00050c04, CB_TARGET_MASK 0x2, CB_SHADER_MASK 0x3, SPI_SHADER_COL_FORMAT 0x2. The surface is 1920x2160 with 2 samples. Same value on PPSA05666 (feat(agc): run DB depth copies into a single-sample R32 float target #5499, Vr compatibity? #7649).
• Blocker 4: pages under a host import carry ImportWindow stamps, which set the CPU flag but not the stored flag. Every imported page traced kept its bytes. Publication now checks stores only, through a new NoStoreSinceLastCollect.
• Blocker 5: a command buffer use after free, seen directly with VK_LAYER_LUNARG_api_dump. On one command buffer: vkCmdBindPipeline, vkEndCommandBuffer, vkQueueSubmit, vkFreeCommandBuffers, then vkCmdBindDescriptorSets, vkCmdPushConstants, vkCmdDispatch and vkCmdPipelineBarrier, which crashes. recordDispatch caches the handle, then ShaderResources::Bind calls requireGuestBuffers, which reaches the feat(agc): publish a written native multisample D32S8 surface to its guest planes on a guest read #13021 publication, and that flushes the recorder. The local fix runs the check before the handle is taken, in recordDispatch and before StartDrawPass. The crash does not need feat(agc): sample the stencil plane of a native multisampled D32S8 target #9717: with feat(agc): sample the stencil plane of a native multisampled D32S8 target #9717 out, the stencil copy draw skipped and this fix removed, it recurred at the same address, nvoglv64.dll+0xe35467, about 31 s in. That build still held ed82dbdccf and 456a530efb. Not the cause, each tested: VRAM pressure (it crashed with 971 MiB in use), descriptor reuse and retired imports.
• Blocker 6: the whole depth plane gets CPU stamps with no driver store, also with imports off. Local change: keep per-page hashes at import, capture and publication, and accept flagged pages only if the bytes match.
• Blocker 7: a true loop, not a slow load: of 23,432 file opens, nearly all were six identical read patterns over the module .pak files. The cause was in the game's own log: AbstractPeer::Init: Failed to bind socket to port in range 0 - 0, then GameServer::Init: Initializing server failed, so the local server never started and the client waited forever. RakNet creates its socket with sceNetSocket, then calls bind, getsockname and sendto from libkernel, which does not know that socket. The local fix is a table of forwards that libSceNet registers with libkernel for bind, getsockname, sendto and recvfrom. setsockopt still refuses SO_LINGER and level 0 option 2, and the title carries on regardless.
• Save thumbnails, fixed in 2e499428ca: every save stored noise as its PNG. The file was valid, and the bytes were BC1 blocks of a ground texture with the alpha byte forced to 255. The title renders the thumbnail by compute into a 768 KiB tiled 480x270 image at 0x3a8780000, then reads it on the CPU after a label. StoreReadbackBeforeLabel stores only linear images and those up to 64 KiB, so the CPU read the old bytes at that address. Images up to 1 MiB not written in the last 8 flips are now stored too. Two saves produced correct pictures: one on this fix's final form, one on a wider test version. It was not repeated after the last rebase, and there is no automated test. The 1 MiB and 8 flip limits are chosen from this one case.
• Container panel: one container's panel was garbled every time it opened during one session, and rendered correctly in the next. That points away from the asset. The cause is unknown.
• First loading screen: the top third was garbled once around present 779, and in a later manual session. It did not appear in two runs dumping frames across that point. It is intermittent and not explained, and it is not known to share a cause with the panel.

Speed

The opening scene of The Hold, idle, on the Continue route, 10 s windows. Each row adds to the one before it, and the first one or two windows after loading are left out.

Change	fps
Before speed work	about 5
Publish HTILE alone	about 20
Memoized collect in the depth check	about 34
Draw cache of 32768 entries	about 35 (no effect in play, menu 11.5 to 15.5 fps)
#10382	about 38
Access check held for the heap run	about 42
Faster page hash and depth check early-out	about 44
Publication built from a queued fill	about 50 (499 to 507 presents per 10 s, against 451 to 460 before it, from tools/trace_digest.py summaries without samplers)
Rebased onto main 9fe758d752	about 49, one run

The same route on a later save in The Hold, which is the fairer figure for play:

Build	Presents per 10 s	Frame p50
Without the thumbnail fix	223 to 230	43.4 ms
With the thumbnail fix	228 to 233	43.1 ms
With it, on main f776fddcff	224 to 229	not captured

This is the bottleneck tracked in #7895: one draw committer. On the later save, with draw profiling on, about 2,070 draws per frame reach it at about 18 µs each, and it is idle about 7% of the time. In the opening scene the figures are about 672 draws at about 29 µs. The cost is spread thinly: recording about 24%, the write watch walk about 12%, key and lookup about 10%. The GPU alone takes about 16.2 ms, so 60 fps would leave no margin here even with the CPU cost removed.

Remaining committer cost in the opening scene, and why:

• Budget. One submission per frame, and the game waits in waitForFlipRoom, so the emulated GPU work sets the pace. About 672 draws per frame reach the committer, which stays nearly full. The frame takes about 19.7 ms and 60 fps needs 16.6 ms.
• Write watch walk, about 3 ms per frame. A 5 GiB allocation at 0x376200000 does not get a host import, so a 3,008 MiB heap mirror is checked by address-based builds (about 1,800 per second, about 41 per collect epoch). Each epoch walks the whole range through GetWriteWatch: about 1,500 runs of 2 MiB at about 2 µs each, with no dirty pages seen. A standalone benchmark puts one call over 3 GiB at about 1.6 ms, and the reset flag adds nothing, so merging the runs would save about 1.4 ms at most. Moving this walk to the queue worker worked (AcquireRegistered fell from 17% to 1.8% of the committer) but gained no fps, also with a pipeline depth of 128, so it was not kept. The committer went idle instead, probably because the walk follows a drain and leaves nothing queued to overlap with. The walk has to get shorter or rarer, not move.
• Draw recipes: 0 hits. 85 to 90% of draws fail on resolved.moved, read-only buffers at new addresses each frame, and only 0.1% would qualify. fix(agc): look up and attach draw recipes on the relocated donor entry #7369 cannot help this title, so it was not resolved or tested.
• Depth publication, about 1 ms per frame. A held publication used to wait for each frame's queued clear fill so it could capture the plane. The whole-plane fill already fixes the bytes, so the snapshot is now built from its pattern. That took CheckCurrent from 11.8% to 2.0% of the committer. It gained more than the wait's sampled cost, probably because the wait also split GPU work.
• HTILE write-back, about 0.5 ms per frame. A shader binds the 320 KiB HTILE at 0x383d00000 as a written buffer, and storeBackClearedHtile waits for it. It always reads back uniform 0xfffc000f, but it is a real GPU write, so the wait stays.
• The rest is spread across draws with no single large item.

Tested with no gain on this title: #13858, #9899, #9858, #9898, #12070, #10886. A larger APS5_HOST_IMPORT_MIB made it worse: with 13000, vkAllocateMemory failed at about 8.5 GiB of imports, the budget was lowered to that, fps fell to about 17 and loads stalled.

Analysis (AI-generated, not measured)

• Blocker 2: 0x88 reads as STENCIL_COPY plus COPY_CENTROID, and the 1920x2160 2x target looks like checkerboard rendering. The local pass assumes the DB exports stencil on both of the first two channels. That has not been measured on a console.
• Blocker 6: the hash check cannot see a guest store of the same value, the same limit as the debt entry on import windows, and it loosens a check in feat(agc): publish a written native multisample D32S8 surface to its guest planes on a guest read #13021. What stamps the plane with imports off is still unknown.

Next steps

1. Add tests for the stencil copy, the publication changes and the socket forwards, then propose them.
2. Follow perf(agc): the queue-0 draw committer bounds draw-heavy titles on the CPU — is parallel draw recording planned? #7895, which 60 fps on this title depends on. refactor(agc): split prepareDraw's read-only plan from its record-time work #14051 is a preparation step and does not speed anything up by itself. Smaller local candidates are readTarget, markLayersPending, perf(agc): build a pipelined draw's resource-cache key before it reaches the committer #10678 and merging the 2 MiB runs. They help, but they will not reach 60. Recipes that accept moved buffers look worth about 2 ms at most in the opening scene, an estimate from samples.
3. Find what garbles the container panel and the first loading screen.
4. Test past The Hold.

How it was run

Dump content in app0\, with eboot.bin and the three sce_module PRX unwrapped. Noto fonts in anyps5-fonts\. No PlayGo or other title-specific workarounds.

relinker app0\eboot.bin game.exe --windows --registry --sce-module-path app0
game.exe

APS5_LOG_TIMESTAMPS=1, APS5_TRACE_SUMMARY_EVERY=10, APS5_DUMP_FRAMES=400, APS5_DUMP_FRAMES_EVERY=40, APS5_PAD_INPUT_FILE=<path>. Input was scripted: Cross through the menus from 45 s, Cross at 150 s to close the tutorial popup, Options at 170 s, Cross at 190 s to confirm, then Cross every 12 s from 300 s in the longest runs. Each lasted 700 ms, because shorter presses can fall between frames at these rates. Shader caches were warm except for the first run of a build. Times are from launch.

Creating temp0/Logs before launch turns on the title's own network.log and gold.log, which show the client and server state machines. It is not needed to run the title.

Host report (integration build)

AnyPS5 host report (format 1)
commit: b0c711fa2620
[system]
os: Windows 10.0.26300
cpu: AMD Ryzen 7 9800X3D 8-Core Processor
cpu_threads: 16
memory_mib: 32404
cpu_gate: passed
[sdl]
linked: 2.33.0
compiled: 2.33.0
[gpu]
device: NVIDIA GeForce RTX 5080
type: discrete
vendor_id: 0x10de
device_id: 0x2c02
vulkan_api: 1.4.351
driver: NVIDIA
driver_info: 617.42
driver_version_raw: 0x9a4a8000
subgroup_size: 32
device_local_mib: 15977
candidates: NVIDIA GeForce RTX 5080
present_modes: unknown
present_mode_requested: default
[vulkan]
VK_KHR_8bit_storage: extension required present
VK_KHR_buffer_device_address: extension required present
VK_KHR_shader_float_controls: extension required present
bufferDeviceAddress: feature required present
fragmentStoresAndAtomics: feature required present
samplerAnisotropy: feature required present
shaderInt64: feature required present
shaderSignedZeroInfNanPreserveFloat32: feature required present
storageBuffer8BitAccess: feature required present
textureCompressionBC: feature required present
vertexPipelineStoresAndAtomics: feature required present
VK_EXT_external_memory_host: extension optional present
VK_EXT_fragment_shader_interlock: extension optional present
VK_EXT_hdr_metadata: extension optional present
VK_EXT_memory_budget: extension optional present
VK_EXT_mesh_shader: extension optional present
VK_EXT_present_mode_fifo_latest_ready: extension optional present
VK_EXT_provoking_vertex: extension optional present
VK_EXT_sample_locations: extension optional present
VK_EXT_sampler_filter_minmax: extension optional present
VK_EXT_shader_demote_to_helper_invocation: extension optional present
VK_EXT_shader_image_atomic_int64: extension optional present
VK_KHR_draw_indirect_count: extension optional present
VK_KHR_driver_properties: extension optional present
VK_KHR_fragment_shader_barycentric: extension optional present
VK_KHR_maintenance5: extension optional present
VK_KHR_present_mode_fifo_latest_ready: extension optional present
VK_KHR_shader_atomic_int64: extension optional present
VK_KHR_shader_clock: extension optional present
VK_KHR_shader_terminate_invocation: extension optional present
VK_KHR_spirv_1_4: extension optional present
[shader_cache]
state: present
size_mib: 112
[environment]
ANYPS5_HOST_REPORT: 1
APS5_DUMP_FRAMES: 400
APS5_DUMP_FRAMES_EVERY: 300
APS5_LOG_TIMESTAMPS: 1
APS5_PAD_INPUT_FILE: <path> (v01.007.000) (Backport <path>
APS5_TRACE_SUMMARY_EVERY: 10
VK_SDK_PATH: <path>

This report is from an earlier build of this work on main d30e04fb8a, so its commit: line differs from the branch above. The system is the same.

AI-assisted: yes