[ntamas94]

What

Knowledge is Power: Decades (CUSA12435, Wwise), the one PS4 title observed, builds its AJM batches with the PS4 SDK's inline batch helpers, which end in four libSceAjm exports: sceAjmBatchJobControlBufferRa, sceAjmBatchJobRunBufferRa, sceAjmBatchJobRunSplitBufferRa and sceAjmBatchStartBuffer. Unlike the PS5 builders, whose batch records are private to the library, here the batch buffer belongs to the title: the SDK code reads the position each builder returns, sizes the buffer for the console's jobs and may write chunks itself. So the buffer has to use the console's layout. On main, boykopovar#9433 added the PS4 libSceAjm with three of the four as throwing stubs, and sceAjmBatchJobRunBufferRa (ElslOCpOIns) is only a throwing stub in libSceAjmi.

This implements the layout in libSceAjm.native/src/Ajm.cpp, which libSceAjm is built from:

• The builders write one job each: a job chunk (instance as payload, byte count of the job's chunks), an optional return-address chunk, the input buffers, a flags chunk (control or run, flags bits 32-51 as payload), the output buffers and the output control (sideband) buffer. They return the position after the job.
• sceAjmBatchStartBuffer walks the chunks, skips inline-data chunks, turns each job into the library's existing control, run or statistics job and runs it with the existing decoders (ATRAC9, MP3, M4AAC, Opus), so the sideband results are the ones the PS5 path writes. Like sceAjmBatchStart, it rejects a context that initialization did not allocate, and the batch belongs to its context.
• A chunk header is read as the console's builders write it: bits 0-3 the kind, bit 4 the output direction of a buffer chunk (0x11 output run, 0x12 output control), bits 6-25 the payload. The builders leave bits 5, 26, 27 and 31 as they were, and a title that reuses its batch memory carries the previous batch's bytes there, so those bits are ignored.
• An ATRAC9 run with RUN_MULTIPLE_FRAMES reports which side ran out, as the console does: PARTIAL_INPUT (0x8) when the input ran out with room left, on a superframe boundary or not; NOT_ENOUGH_ROOM (0x10) when the room ran out with input left, after any number of frames, and with nothing consumed for a zero room; both (0x18) when the input ended just as the output filled, which is every Wwise run job. The ATRAC9 format reports the frame's bitrate (frame bytes x 8 x sample rate / 256: 168000, 84000 and 384000 measured). Resampled runs keep the earlier rule, as they were not measured.
• A run job read from a PS4 batch buffer without RUN_MULTIPLE_FRAMES decodes exactly one ATRAC9 frame and reports 0, as the probe measured on the PS4. A job from the PS5 builders still decodes its whole input, which FMOD in PPSA29141 relies on; the PS5 console's behaviour there is not measured.
• Buffer chunks follow what probe batches on the PS4 showed: a chunk of size zero is dropped whatever its address, so the job runs as without it, and a null address with a nonzero size refuses the whole batch before any of its jobs runs: sceAjmBatchStartBuffer returns 0x80930011 (MALFORMED_BATCH) and fills the error record with 0x80930016 (INVALID_ADDRESS), the job chunk's address, the chunk's offset in the job and the job's return address. A size that is not a multiple of 8, a chunk that overruns its job or the batch, and a job without a flags chunk are refused the same way, not measured; the record then holds 0x80930011. Unknown chunk kinds and control flags other than RESET, INITIALIZE and SIDEBAND_GAPLESS_DECODE throw.
• The three stubs in libSceAjm/Export.cpp are dropped. The throwing sceAjmBatchJobRunBufferRa of libSceAjmi stays, declared as sceAjmBatchJobRunBufferRa_nid_disambig1: the NID patcher exports it under the name's NID, so a title importing it from libSceAjmi still gets the not-implemented error at call time, and the two libraries no longer export one name.

The layout (chunk kinds 0, 1, 2, 3, 4, 6, 7, 17, 18, the builders' chunk order, statistics instance 0x80000) follows shadPS4's PS4 library reimplementation. A GoldHEN user-mode plugin dumped 32 545 batches that Knowledge is Power: Decades (CUSA12435, Wwise) started on a PS4. They confirmed the 64-byte control and 80-byte run jobs byte for byte, the header bits above, and the 0x18 result and bitrate of all 14 680 traced run jobs, which each decode 4 frames into exactly 4 frames of room. Probe runs of one of the title's ATRAC9 streams through the same builders on the PS4 (fresh instance per case: input ending with room left, room ending with input left, half a superframe of room, zero room, exact fit) gave the result rules above, showed that a run without RUN_MULTIPLE_FRAMES decodes exactly one frame on the console (kept as the whole input here, which a PS5 title needs; noted in the debt), that the console rejects the tests' synthetic silent superframe after three frames where LibAtrac9 decodes four, and that LibAtrac9 matches the console's PCM of that stream within 4 LSB (83.5% of samples exactly). The statistics instance and the inline chunk come from shadPS4 only; the debt entries say what is measured and what is not. The PS5 sceAjmBatchStart path and its records are unchanged, apart from the ATRAC9 exact-fit result and bitrate, which both paths share.

Tested

Windows 11 Pro 26200, WinLibs GCC 15.2.0 r7, NVIDIA RTX 2080, on ps4/limbo-gpu at 45625370e4, merged into this branch. The same change is open upstream as boykopovar#9637; this branch carries the identical Ajm content for the integration branch.

• guest_ajm passes, with seven new cases:
  • TestConsoleBufferJobs builds a control job (with a return address), a run job and a split run job with the three builders, checks the returned positions and the chunk words, starts the batch with sceAjmBatchStartBuffer and checks the control result and both run sidebands (result, consumed, written, total samples, frames) and the decoded PCM.
  • TestConsoleHandWrittenJobs writes a batch word by word: an inline chunk holding ATRAC9 initialize parameters, a control job whose input points into it, a run job and a statistics job. It checks their results, then that a size not a multiple of 8, a job without flags and an overrunning job return 0x80930011, and that an unknown chunk kind throws.
  • TestConsoleStaleHeaderBits runs a hand-written batch with every bit the console's builders never write set, and checks it runs exactly as the clean batch.
  • TestAt9FlagRules runs the five fits above and checks the result, the bytes consumed, the output written and the frame count of each.
  • TestConsoleBufferEdges runs a job with zero-size input and output chunks beside real ones (decodes as without them), a job whose only output has size zero (0x10, nothing consumed), and two batches with a null input or output address, which are refused with the measured error record while the control job before them stays unrun.
  • TestConsoleEmptyBatch checks the empty-batch contract: a null batch address is 0x80930005 with size 0 or 16, a null batch id is 0x80930005, and a batch of zero bytes starts, returns an id and completes at once.
  • TestConsoleSingleFrameRun runs a console job without RUN_MULTIPLE_FRAMES on a whole superframe with room for four frames and checks that one frame is decoded with result 0, the PCM after it untouched.
  • The existing ATRAC9 cases whose input exactly fills the output now expect 0x18; the helpers that stream a decode accept 0x8 and 0x10 where one side runs out.
• guest_ajm_vibration passes; its exact-fit run now expects 0x18 (it expected 0 before). Vibration streams were not in the trace.
• check_conventions passes; the patched libSceAjmi.prx still exports ElslOCpOIns (objdump).
• PS4 probe batches (FW 10.71, ps4probe, the console's builders, a control job before each run job): zero-size input and output chunks alone, beside a real chunk and with a null address; null addresses with a nonzero size as the first or second input chunk and as the output chunk. They gave the rules above; the refused batches ran nothing.
• Knowledge is Power: Decades, locally on my bring-up branch, with the first version of this change (before the header-bit and result rules above): Wwise builds every batch with these builders and a return address (80-byte run jobs, 64-byte control jobs, up to 6 jobs per batch). In four 150 s runs it ran about 30,800 jobs per run on five stereo ATRAC9 instances (control flags 0x6001, run flags 0xc00000001001, 56-byte sidebands), every one with result 0 under that commit's rules, about 650 s of decoded PCM per run, and the title reached its animated title screen. One of the four runs stopped before that on a UnityGfxDeviceWorker fault in the guest heap that runs with every AJM job skipped also hit; it is not in AJM. Without this PR the title aborts at its first sceAjmBatchJobControlBufferRa (not implemented).

boykopovar#9483 changes the last lines of sceAjmBatchStart's job loop, next to where this moves the batch registration into RegisterBatch, so the two conflict in three lines; boykopovar#9481 and boykopovar#9034 also add debt entries to libSceAjm.native.md. I can rebase onto whichever lands first.

Checklist

• Based on current ps4/limbo-gpu (45625370e4, merged in); no other open PR implements the same functions (searched boykopovar/AnyPS5 and Get-fix/AnyPS5 for Ajm, BufferRa and StartBuffer)
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: feat: run PS4 games on the AGC driver — Limbo (CUSA01664) in game (integration branch) boykopovar/AnyPS5#6977 (this targets its branch); upstream twin feat(libSceAjm): run PS4 batch buffer jobs in the console's chunk layout boykopovar/AnyPS5#9637
• Fixes: none
• Unblocks: Knowledge is Power: Decades (CUSA12435) Knowledge is Power: Decades (CUSA12435) boykopovar/AnyPS5#9718, which stops at its first sceAjmBatchJobControlBufferRa without this
• GPUs tested: n/a (no AGC change)
• AI-assisted: yes (Claude Code)

🤖 Generated with Claude Code

[Get-fix]

Thanks @ntamas94. boykopovar#6977 is now a draft integration branch, and its own work is split into boykopovar#9429 to boykopovar#9433 (details). The maintainers asked that the branch not land anything without its own review, so could you open this upstream against main, on top of boykopovar#9433, with Depends on? Once it's open, I'll merge it into ps4/limbo-gpu, list it as a member and run Limbo with it. ps4/limbo-gpu now moves to the split branches, so this pull request may show conflicts here.

AI-assisted: yes (Claude Code)

🤖 Generated with Claude Code

[ntamas94]

Thanks @Get-fix. Opened upstream against main as boykopovar#9637: one commit with this PR's net change, on top of boykopovar#9433, which is already merged, so Depends on: none. guest_ajm, guest_ajm_vibration and check_conventions pass on main at da54c652e. I'll leave this one open until you've merged boykopovar#9637 into ps4/limbo-gpu.

[ntamas94]

Closing: the upstream twin boykopovar#9637 was merged into main on 2026-10-25 (ae7d59258e), so ps4/limbo-gpu can take the PS4 batch buffer change from there. This branch also picked up an accidental merge of main on 2026-10-29, which is why it now shows main's whole diff and fails conventions; nothing in it needs to land here.

AI-assisted: yes (Claude Code)