AnyPS5/docs/dev/debt/libSceOpusSilkEnc.md

libSceOpusSilkEnc technical debt

Unknown function info

• sceOpusSilkEncCancel, sceOpusSilkEncCtl, sceOpusSilkEncEncodeFloatWithPriority, sceOpusSilkEncEncodeFloatWithPriorityAndTimeout, sceOpusSilkEncEncodeWithPriority and sceOpusSilkEncEncodeWithPriorityAndTimeout (libSceOpusSilkEnc) - names and NIDs are present in the PS5 3.20 export inventory. The prototypes mirror the sibling encoder declarations and libopus's own ctl, priority and batch-id entry points, but the console's request codes, priority, timeout and batch semantics were not recovered. They throw through NotImplemented_nid_no_patch; this is not an ABI claim.

Functional

• libSceOpusSilkEnc is an AJM frontend on the console: apart from libc runtime support and sceKernelUsleep the module imports only sceAjmi* calls, Create runs sceAjmiInstanceCreate with codec 17 and the flags 0x9 mono or 0x11 stereo, sceAjmiBatchJobControlBufferRa with request 0x6001 over state +0x10 with size 0x4C and results at state +0xD0, then sceAjmiBatchStartBuffer and sceAjmiBatchWait, Encode hands the samples to sceAjmiBatchJobRunBufferRa, and Destroy calls sceAjmiInstanceDestroy per module runs with stubbed AJM. Our encoders build libopus without AJM, so packets come from libopus and whether they equal the console's is unknown. Closes when: encoder packets are compared against console-produced ones or the library runs through our AJM

• Create and CreateEx zero-fill the 752-byte state and write the word table measured on the modules: the instance id at +0x00, this implementation's id because our AJM does not run, 8 × channels at +0x08, 0x800 at +0x10, rate at +0x14, channels at +0x18 and rate/50 at +0x58 and +0xE4, the bitrate at +0x1C and +0xDC with the per-rate per-channel table 12000/19000 to 51000/99000, the constants 0x44F, 0x5E0, 10, 0x3E8, 0xFFFFFC18 and 0x10, and the 0x17B2 marker at +0xE8. The console's live AJM may write further result bytes into +0xD0 and +0x11C. Closes when: a console run with live AJM settles those result bytes

• Create and CreateEx replacement rule is this implementation's choice: a create over a marked state retires the encoder its instance id names before any state write or allocation, so a later failure cannot strand an entry the state no longer names; explicit libopus error branches leave the state zeroed, while make-unique throwing after setupState or a later registry insert throw leaves the filled state without a live encoder. The console behaviour for its own allocation failure after a create over a live state is unmeasured, and the tests cannot force our failure paths because the library is a shared module against libopus, which has no allocator hook, while a process-wide memory limit fails an unknown allocation first. Closes when: a deterministic allocation-failure injection or a console measurement of the failed replace exists

• Encode and EncodeFloat refuse with the measured codes: null state +0x01, null samples +0x02 before the state marker, a markerless state +0x15, null data +0x04 before the frame check and maxDataBytes 0 +0x05; maxDataBytes below zero is not refused and reaches our libopus backend, where a rejection throws. Frame sizes outside the durations 10, 20, 40 and 60 ms, the measured 480 to 2880 at 48000 with 120 and 240 refused, give +0x03; we scale the counts by rate while the module's rule for other rates is unmeasured. Closes when: the frame rule and the console outcome for a buffer AJM rejects are measured

• sceOpusSilkEncEncodeFloat shares the CELT module's check order: null state, null samples, frame 0 and the marker, then the frame × channels × 4 byte stack reservation for the converted frame before its remaining checks, which kills the console module for negative frames per the measured CELT run; our implementation throws once it exceeds 8 MiB instead, so the failure reports. The bound beyond the measured deaths is our host stack model, not a console measurement. Closes when: the console's stack behaviour for such frames is measured

• Create, CreateEx, Destroy, GetSize, Terminate and Initialize refuse with base 0x81770000, the null pointer +0x01, the sample rate +0x0F, channels +0x0C, application +0x10, the marker in Destroy +0x15 and Initialize(nullptr) +0x18, all from the firmware 13.42 export measurements; Destroy clears the marker and leaves the other bytes, GetSize returns 752 on channels 1 and 2 and refuses other counts with +0x0C. CreateEx context handling, duplicate initialization and termination with live encoders throw as the recovered decoder analogue does. Closes when: those paths are measured on the console