Crimson Spires (PPSA03568)

[JJenkx]

Existing thread

• I searched this category for PPSA03568 and "Crimson Spires"; no thread exists.

Title

Crimson Spires (PPSA03568)

Content version tested

01.003.000 (sce_sys/param.json), Unreal Engine 4. The package has no PlayGo files.

AnyPS5 commit

main b415263451 plus #14298 (285e93acd0), which only adds a Linux crash report. Two open PRs were merged onto the same base for comparison.

System tested

Linux 7.2.11 (Arch), Ryzen 9 7950X, Radeon RX 6650 XT, RADV Mesa 26.2.6, GCC 16.2.1, Release.

Setup

• Relinked with relinker --registry --rpath <build>/core/libs/libs eboot.elf app.elf.
• Run with APS5_PLAYGO_UNLISTED_CHUNKS=absent. Without it, a run on an older main stopped at 20 s: scePlayGo: ... are missing, so chunk ID 1 cannot be checked.
• The Noto system fonts from docs/user/USAGE.md are in anyps5-fonts.
• Input: Cross every 12 s from 30 s, then left-stick holds.

Furthest point

The opening story. A car-dashboard scene at night ("It all seems so perfect and peaceful. Until..."), then Erika's dialogue with character portraits ("I understand. Who is the prisoner in transport?"). The dialogue advances with Cross. The run ends at the scheduled close at 293 s, exit 0, with no NotImplemented throws.

Performance

About 8 to 18 FPS in these mostly static scenes, median 10.7 after 60 s.

Host imports: 1,831 fail per run at vkAllocateMemory (VK_ERROR_INVALID_EXTERNAL_HANDLE) and fall back to copies. This is the scattered direct-memory case. With #4449 (2f225f0ff2) merged:

• 1 failure;
• median 12.0 FPS after 60 s.

That is one run per arm, so it does not establish a difference.

perf on main, sampled for 15 s at 150 s:

• 43% of CPU is one game thread spinning in sched_yield and clock_gettime, apparently waiting for the frame.
• 33% is the driver's draw-pipeline thread building per-draw resources (ShaderResources::buildComplete, BdaResources) and copying uploads (GuestBufferMemory::UploadFinish, copyRegion, GuestMemory::Read).
• 20% is the queue thread in waitMemory and DrawPipeline::Drain.

So the frame rate is set by per-draw CPU work in the driver, not by the GPU.

Not verified

Gameplay past the opening dialogue, audio, saving and other GPUs.

AI-assisted analysis: yes.

[JJenkx]

Windows result, on plain main 7feb37ee3e (Release, no other PRs). Windows 11 22621, Ryzen 9 5950X, Radeon RX 6700 XT with the AMD proprietary driver 26.10.1. Same setup as above: no --module-dir (the relinker picks up prx/ itself), APS5_PLAYGO_UNLISTED_CHUNKS=absent, Cross every 12 s from 30 s. The Latin Noto fonts were not installed.

Furthest point: the same as Linux. The night car-dashboard opening ("On a brisk autumn evening, the drive from Cape Girardeau to St Louis…"), then Erika's dialogue with portraits ("Do you mean… the serial killer? The Heartbreaker?"). Text draws without the Noto fonts. Two runs of 300 s each ran to the scheduled close with no FATAL.

Performance (APS5_TRACE_SECONDS=200:10, one run each):

Run	Presents in 10 s	Frame p50	GPU busy per present
1	74	106 ms	111 ms
2 (APS5_OUTPUT_RESOLUTION=1920x1080)	89	80 ms	86 ms

• Resolution: the game draws a 3840x2160 display buffer in both runs, so the 1080p class does not change its render size. The difference between the two rows is run-to-run variation, not resolution.
• Host imports: no VK_ERROR_INVALID_EXTERNAL_HANDLE failures on Windows, against 1,831 per run on Linux.
• Where the time goes: the trace's top phases per frame are Driver.Packet (195–242 ms) with memory_wait (84–109 ms) and Recorder.Sync (42–60 ms).
• The p90 frames: in the second run, every 500 ms one frame takes about 350 ms. It holds about 730 draws and 190 dispatches, against about 310 and 30 in the other frames. The per-draw cost is the same in both (median 0.031 against 0.029 ms on the queue worker), so it is a periodic heavy pass of the title's with the GPU already near saturation, not added per-draw overhead. The memory_wait time is mostly the second queue worker waiting for the first.

This is a different machine and GPU from the Linux run above, so the FPS is not a like-for-like comparison.

AI-assisted: yes.