FAR: Changing Tides (PPSA03766)

[JJenkx]

Existing thread

• Searched the title ID and “Changing Tides”; this is the existing thread.

Title

FAR: Changing Tides (PPSA03766)

Content version tested

01.002.000 (master01.00), from sce_sys/param.json.

AnyPS5 commit

Latest reviewed run: private integration 032ff3f70bd002033cb37e95244924b84d3cd64d, main a5a2c946a3a98bd39532d284b1283d78415bb124 plus original #6782 (4433a62b3bf6248df830994a255701c72581cc95, XairRrR) and #6913 (d89f7050b7de6e7ec337f8edfcc8e41feb644b7b, JJenkx). No other production or title-specific changes. This is not a vanilla-main result.

The earlier report used ec4ad3f67f0cfd01ab4d2f15148122caac543eba with #8091 as well. The new run reaches the surface without #8091; this does not establish whether other scenes need it.

System tested

Arch Linux 7.2.9-arch1-1, GCC 16.2.1, Ryzen 9 7950X, AMD Radeon RX 6650 XT / RADV Mesa 26.2.4. Game window on DP-2, the 24-inch display; captures are 1536×864.

Furthest point

In game: from the opening underwater movement tutorial to swimming at the surface of the flooded village.

What is shown

Directly reviewed captures at 230/280/370 seconds show the underwater character with the left-stick “TO MOVE” prompt, the character rising toward the surface, then swimming at the surface with buildings, mountains and water reflections visible. Scripted left-stick up/right inputs were sent between 235 and 313 seconds. No separate guest input-acknowledgment log was found; the visible progression is established by the captures. Audio, extended gameplay and PS5 same-frame lighting comparison remain unverified.

The bounded run closed normally near 380 seconds, native exit 0, with no remaining runtime process. Private configuration/save seeds and the kernel import limit were restored. Actual Khronos validation insertion appears in the native log; zero VUID errors were recorded. Game output was redirected and no performance trace was captured, so this run makes no FPS-improvement claim.

Current blocker

No fatal stop occurred in this bounded run. Performance, audio and later gameplay still need qualification. The tested open-PR composition is #6782 plus #6913; their full title dependency scope has not been isolated.

Before the native run, a matched same-base focused check used main plus #6782 with the identical #6913 regression fixture in both arms. The control mock test fails specifically on chunked vertex-buffer usage; the candidate passes. The other mock check and four physical Vulkan checks pass in both arms, with actual Khronos insertion and zero VUID errors in every physical check. Both full builds and relinker_conversion suites pass. This is focused correctness evidence, not a matched native performance comparison. The older native control was on another main commit and must not be treated as such a comparison.

How it was run

Converted existing extracted assets into a private runtime: app plus 14 bundled modules; the 15 original inputs remain unchanged. Relinker arguments:

relinker --registry --no-patches --rpath <build>/core/libs/libs --sce-module-path <private-ELF-root> --module-dir Media/Modules --module-dir Media/Plugins <private-ELF-root>/eboot.elf <private-runtime>/app.elf

Ran with APS5_PIPELINED_DRAWS=1, APS5_HOST_IMPORT_MIB=16384, APS5_DRAW_CACHE_ENTRIES=4096, APS5_PLAYGO_UNLISTED_CHUNKS=absent, ANYPS5_NO_SHADER_CACHE=1, MESA_SHADER_CACHE_DISABLE=true and VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation; RADV ICD and the installed layer path selected explicitly. The diagnostic harness temporarily set udmabuf's size_limit_mb from 64 to 8192 and restored 64 afterward; the default limit has not been qualified. The entire GPU qualification/native harness used the shared launch lock (ticket 542).

Exact pins, integration patch, focused test logs, conversion metadata, native log/report, harness and reviewed captures. SHA-256: f8246a7bd2f814ca530caaf74a789cb43bda993a290cff7800d9d39e3a0cf174. No game binaries, SDK libraries or extracted asset files are included.

Earlier opening-only evidence is retained as a separate historical build.

AI-assisted: yes.

[JJenkx]

Warm-cache baseline completed on Linux, head 3dee9cc73cf71e02894998214158d655e5912e86: main 5d3ab7791dc444be2996a0bb730e89c78ecfa8d7 plus original #6782/#6913, no #8091. Same content 01.002.000; Ryzen 9 7950X, RX 6650 XT, RADV Mesa 26.2.4, Linux 7.2.9. VideoOut is 1920×1080; captured client is 1536×864 on DP-2.

The scene ran once to warm both private shader caches, then the title restarted with restored private writable seeds. Directly reviewed warm/capture screenshots at 230 seconds and capture at 270 seconds show the stationary underwater character and “TO MOVE” tutorial. This run intentionally sends no movement after reaching that scene; the earlier surface-progression result remains separate. Both runs closed normally near 280 seconds, native exit 0, no remaining runtime; private writable state and the kernel import limit were restored.

Actual game output stayed on a PTY, caches were enabled, validation disabled, and the capture used only APS5_TRACE_SECONDS=250:10. Official summary: 138 presents / 10,023.525 ms; frame p50/p90/p99/max 70.081 / 85.351 / 97.432 / 99.748 ms. This is a single-build baseline, not an improvement comparison.

The raw queue-0 slice buffer filled at 5.86 seconds, losing 184,191 events. totals_full=false: the documented summary aggregates remain complete, while raw slice timelines/rankings after overflow are incomplete. The summary reports 38.203 ms/frame tagged Driver.Packet.memory_wait, 26.115 tagged Graphics.Draw.draw_and_resource_release, and 21.941 tagged Driver.Draw.shader_compile_and_link. These are inclusive and overlap across threads; they are not additive, and the last tag alone does not establish actual fresh shader compilation. All seven sampled focus checks around the window match. Read-only CPU sampling estimates 2.42 application cores and 13.23% of machine CPU busy outside the application; this was not an exclusive-host run.

Exact source pins/patch, official summary and trace, digest, reviewed captures, cache/protocol metadata and CPU observations, SHA-256 8bd1dfeaf06e56b00e2c5cdbdb2d9be3ea73c44a09c7e9ca0073ad38b844c256. No game binaries, SDK code bytes or extracted asset files are included.

A same-build repeat using the documented APS5_TRACE_EVENTS=1048576 buffer control is queued for the full timeline. Audio, extended gameplay and a PS5 same-frame visual comparison remain unverified here.

[JJenkx]

Follow-up diagnostic on the same pinned build 3dee9cc73cf71e02894998214158d655e5912e86 (main 5d3ab7791d + original #6782 + #6913, no #8091): the repeated warm-cache 250:10 capture with documented APS5_TRACE_EVENTS=1048576 has no dropped trace events. The earlier default-capacity capture remains useful for aggregate totals but has an incomplete raw timeline.

The new window contains 144 presents / 10046.011 ms, frame p50/p90/p99 69.279/75.904/85.533 ms. Both 230/270-second screenshots show the same stationary underwater character and movement prompt. This is one build, not an improvement comparison. One of seven focus samples departed at 254.62 s, and the separate host CPU observer measured about 12.08% machine load outside the title. Inclusive/cross-thread phase times are not additive; shader_compile_and_link does not by itself establish a fresh shader compilation. The full trace is available for examining the repeated stage preparation and draw work.

Reusable game checkpoint: a separate run copied the earlier game-generated Zone1_GameStarted.save and its companion files into private writable storage. On this same head, directly reviewed screenshots at 80/140/230 seconds show the character at the water surface amid the submerged village, so that checkpoint can avoid repeating the underwater opening in later tests. This is one successful load, not a process save state, a loading-time benchmark or new furthest progression. Only menu-button inputs were scripted; no new analog progression or audio result is claimed. Both runs closed normally and restored their configuration/private files; the original checkpoint hashes are unchanged.

Complete trace, summaries, native report, screenshots and checkpoint hashes/results, SHA256 076c4afac6aab9befb707fb4d60338635a3f43adc5720a1ef3a687614ee9a61f. No game binaries, assets or checkpoint contents are included.

[JJenkx]

Current-main recheck of FAR: Changing Tides (PPSA03766). Linux 7.2.11, RX 6650 XT, RADV Mesa 26.2.6, Release.

Setup:

• Relinked with relinker --registry --rpath <build>/core/libs/libs --module-dir Media/Modules --module-dir Media/Plugins eboot.elf app.elf.
• Run with APS5_PLAYGO_UNLISTED_CHUNKS=installed.
• Input: Cross every 12 s from 30 s, then left stick up.

On main b415263451 (plus #14298, which adds a Linux crash report), three runs stop in start-up:

• Two stop at 18 and 24 s with sceHttp2SetRedirectCallback not implemented.
• One stops at 49 s on a SIGSEGV. The new report shows a general-protection fault (code 128) at libc.prx.guest.prx+0x3e57a. That is strcmp's first byte read, through a garbage pointer rdi = 0x6cdb3b69edcb2d76 passed from the title's code.

The HTTP PRs from this account, merged onto that main in order:

Added	Result
#11575 (6ae1f9381b)	stops at 16 s: sceHttp2SetSslCallback not implemented
+ #11586 (7e4805822f)	stops at 16 to 18 s: sceHttp2SendRequest: HTTP/1.x transport is not implemented
+ #11611 (717e3e78c2)	two runs of 294 s, exit 0, reaching gameplay

With all three, it reaches the underwater opening ("L to move"), then the swim up to the surface of the flooded world with its half-sunk houses, at 8 to 14 FPS. #11586 and #11611 currently conflict with main only in docs/dev/debt/libSceHttp2.md; I kept main's text for this test.

Not verified: the boat, audio, saving and other GPUs.

AI-assisted analysis: yes.