3DMiniGolf (PPSA03647) on Windows with NVIDIA

[authurtabo-byte]

Tested

• Title: 3DMiniGolf, PPSA03647, v01.000.000. Title ID and version checked against sce_sys/param.json.
• Host: Windows 11 Pro, Ryzen 7 8845HS, NVIDIA RTX 3050 6 GB Laptop GPU, WinLibs GCC 15.2.0, Release.
• Development base: main 8b393b6, with the integrations below. The first hole was completed on this updated base with the original quality asset and a 1920 x 1080 render configuration. Earlier profiling used main ae3d5f9 plus the local integration.
• Last update: 2026-10-09.
• Investigated with Codex assistance; observations and tests were collected on this machine.

Status

In-game progress on an experimental local integration; performance, audio quality and rendering correctness remain open. Menus and the first course render, input works, and the first hole has been completed. This is still short of a fully playable compatibility result.

The integration includes an earlier sampler revision. Its in-game results do not establish compatibility with the current head of #2092 or stock main. Current #2092 correctly retains rejection of unsupported anisotropic encodings. A supported stack using validated sampler behavior still needs to be established.

Integrations being tested

The table records relevant patch areas. The minimal dependency set remains unverified; existing contributors retain credit for their changes.

Area	References
Relinking and runtime-loaded Unity modules	#2088, #846
Audio3d imports used during startup	#1849
HTILE/depth maintenance	#2053
Null/unwritten pixel programs and rect-list preparation	#1902, #1928, #2022
Multisampled targets, resolves and custom sample positions	#419, #2121, #2169
Sampler allocation and color-comparison descriptors	#2141, #2092; sampler caveat above
Save-memory sizing and local-time conversion	#2131, #2157
Audio state queries during blocking output	#2130
Signal delivery while waiting for mutexes/rwlocks	#2051, by esatturan
Windows audio pacing	#1486, by popcoder565
Vertex layouts and imported vertex/index buffers	#1316, by TibixDev; #2253, by Raghav314

All eight converted guest modules loaded in the startup investigation, including Il2CppUserAssemblies.prx from Media/Modules. The earlier missing-module suspicion was incorrect. Later blockers occurred after entry into Unity and graphics initialization.

Putting freeze

The reported putting crash was reproduced as a live-process stall. The main thread had a redirected Unity signal pending while blocked in a native mutex wait. Graphics workers were waiting rather than reporting a device loss.

The mutex/rwlock interruption change from #2051 was integrated locally. Its signal-delivery regression timed out before the fix and passed afterward; five related kernel tests passed. A subsequent run completed the first hole in two shots. Longer sessions and other courses still need testing.

Frame rate

Low-angle putting is the main performance problem. An original-quality asset check on the preceding integration showed roughly 3–5 FPS in a demanding putting view. Other camera views run substantially faster; those should be reported separately from putting performance. Stable 30 FPS while putting has not been achieved.

Profiling uses APS5_PROFILE_DRAW=1 and APS5_PROFILE_GPU=1, with APS5_DRAW_CACHE_ENTRIES=32768. The larger draw cache avoids severe default-cache churn. Thousands of scene draws, CPU preparation and GPU work still contribute to frame time.

#1316 and #2253 are under title testing. Representative vertex-input preparation samples fell from about 8–10 to 4–5 microseconds per draw after integration, but camera/ball position and settings differ between some runs, so this is not a controlled FPS comparison.

Audio

Crackling/flickering audio was reported during gameplay. One Windows pacing defect was reproduced independently: with 256-frame, 48 kHz output, 192 of 300 calls returned in under 1 ms, with waits up to 17.5 ms. After #1486, none of 300 calls returned in under 1 ms or took over 10 ms; intervals were 3.6–8.2 ms. Eight focused audio/graphics tests passed before the subsequent main update.

A paused WASAPI device query reported 48 kHz float stereo, matching the guest's main port. A 20-second SDL disk capture after the timing fix contained no non-finite samples or exact silence gaps, and 11 samples at or above full scale. This covers the main port only and does not verify audible WASAPI playback. Audio-quality confirmation remains open.

Validation on the updated base

On main 8b393b6 plus the local integration, 15 focused tests passed: vertex inputs, shader capture, graphics, PM4, BDA resources, staging-buffer pooling, guarded null pointers, audio state/mixing/timing, Audio3d, recursive condition waits, redirected signal delivery and CPU affinity. The isolated imported-draw-input GPU checks also passed on NVIDIA. This is targeted validation; the full test suite was not run.

Next checks

• Establish a clean, supported integration using current PR heads, including validated sampler behavior.
• Measure the same stationary putting view before and after optimizations, and separate CPU draw preparation from GPU time.
• Verify native-device audio after the pacing fix and check longer sessions, further holes and map transitions.

Reports from other NVIDIA or AMD systems are welcome; please include the exact base/patch revisions, camera view, settings and how far the title gets.

[authurtabo-byte]

Update: Dynamic Buffer Decoupling, Pipeline Cache Persistence & 35 FPS Putting Benchmark

1. Root Cause Analysis: Shot Trajectory Stalling

In detailed profiling (APS5_PROFILE_DRAW=1), stationary putting held 35 FPS, but tracking a moving ball across a shot previously experienced severe framerate degradation down to ~15 FPS.

Telemetry isolated the primary stall to Recorder.cpp's FlushForAccess memory hook during dynamic draw input preparation (DRAW_INDEX_AUTO / DRAW_INDEX_2):

[hooksync] 740 pending-write syncs waited 5166 ms (10 s);
targets: 727 open, 0 already signaled, 2390 batches finished;
top by wait: [0x0 DRAW_INDEX_AUTO unknown: 740/5166ms avg range/access 192K/127K]

Over 51.6% of total execution time was lost to synchronous GPU timeline stalls (WaitTimeline on semaphore). When Unity updated dynamic transform/vertex buffers each frame, InPlaceDrawInput and CopyDrawInput treated vertex attribute reads as requiring immediate GPU flushes if overlapping the open command buffer batch.

2. Implementation: Dynamic Buffer Decoupling

We implemented decoupling in core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp:

• Introduced explicit GuestMemory::ReadSiteScope tags for ReadSite::IndexBuffer and ReadSite::VertexBuffer during draw input validation.
• In InPlaceDrawInput and CopyDrawInput, when a draw input range overlaps an unsubmitted write in the open batch (recorder->OpenWriteOverlaps(address, bytes)):
  • InPlaceDrawInput immediately yields to CopyDrawInput without invoking FlushGpuWrites, preventing mid-frame command buffer fragmentation.
  • CopyDrawInput skips FlushGpuWrites and directly snapshots the host-visible memory into pooled staging buffers via std::memcpy.
  • Intra-batch hazards are ordered via Vulkan pipeline barriers (VkMemoryBarrier in NoteAccess), eliminating CPU thread blocking.
• Controllable via APS5_NO_DECOUPLE_DYNAMIC_BUFFERS environment switch.

Results:

• Pending-write sync wait time plummeted from 5,166 ms down to 646 ms (an 87.5% reduction in sync wait latency).
• DRAW_INDEX_2 sync waits dropped to 0 ms (0 stalls).
• Full AGC driver CTest suite: 259/260 tests passed (1 skipped, remaining test failure is an unrelated AMD iGPU NaN comparison quirk in vopc_compare_table).

3. Persistent Pipeline Cache

Verified the Vulkan pipeline cache serialization:

• Cache binary: shader_cache/vk-pipelines-10de-25ac-9a100000-6dcaa8fd5d42c5f1fd532c2a423d6fe9.bin (46.1 MB).
• 727+ pipeline hits loaded at boot in ~573 ms with 0 runtime misses, preventing mid-shot stutter when newly revealed course assets enter the camera frustum.

4. Upstream PR Status & Candidates

Analyzing generic fixes for upstream contribution:

1. PR Candidate 1 (Clean / Ready): Pipeline Cache Bound Expansion (Pipeline.cpp)
   • Expands the hardcoded 256 entry pipeline store capacity to 4096 (configurable via APS5_PIPELINE_CACHE_ENTRIES). Prevents eviction thrashing in titles with diverse materials.
2. PR Candidate 2 (Clean / Ready): Rect-List Tessellation Qualification (Draw.cpp)
   • Qualifies generated tessellation stages (rectList) in resolveDrawResources to prevent cache miss churn on rect primitives where PipelineVariantId() == 0.
3. PR Candidate 3 (Under Review): Dynamic Buffer Decoupling (Draw.cpp)
   • Eliminates synchronous timeline wait locksteps on CPU ring-buffered vertex/index data.