Forgejo Fairy reads this file from the default branch and includes it in the prompts of its reviews.
filter_plane() declares coarse[BINS], fine[BINS][BINS] and luc[BINS] as on-stack arrays reset on each row.
Use moof-relative addressing when global sidx shifts fragmented data, unless absolute offsets are already avoided or no sidx is inserted, and test it in api-movenc-test.
AVX-VNNI is the VEX-coded form of the VNNI dot products (vpdpbusd(s), vpdpwssd(s)) on xmm and ymm registers.
The SILK resampler is primed with 4, 8 or 11 input samples of silence for NB, MB and WB, which sets the decoder delay.
A chained Ogg Theora link has its own headers and starts a new logical stream, which the decoder only takes up from header packets and without frame threading.
When a chained stream moves to a new link, its headers are passed on as packets, which the decoder refuses with frame threading.
A stream copied chained Theora file keeps the headers of each later link, so the copy decodes as the original.
Replace the hardcoded preanalysis lookahead maximum with the maximum reported by AMF at runtime.
prores_ks wrote reserved codes as-is and prores_aw dropped equivalent ones; map bt470bg matrix and bt601/bt2020 transfers to their codes, others to unspecified.
5a75d905cb24 added stream_groups[MAX_STREAMS_PER_PROGRAM] to struct Program, making it 134688 bytes on 64-bit targets.
An SPS change at an IDR can increase the decoder's reorder delay from zero.
The decoder had no flush callback, so after a seek it kept the LSP, excitation, filter, postfilter and comfort noise state from before the seek.
An untransmitted frame right after an active frame means the SID frame that should have started the silence was lost.
On the third lost active frame in a row the ITU reference sets the excitation to zero, so the synthesis filter only rings out its memory (DS2-Anywhere SPEC.md section 4.9).
The ITU reference starts with the previous frame type set to active (DS2-Anywhere SPEC.md section 3).
The ITU reference computes each LPC coefficient as neg16(round16(x)), and neg16(-32768) saturates to 32767 (DS2-Anywhere SPEC.md section 4.3.2, step 4).
The ITU reference runs the LPC synthesis filter in 32 bits and saturates the sum after every product (DS2-Anywhere SPEC.md section 4.7).
The formant postfilter's tilt follows the first reflection coefficient k of each subframe (DS2-Anywhere SPEC.md section 4.8.1, step 5).
The previous commits saturate the synthesis and formant filters after every product, as the reference does.
The reader moves from oggparsevorbis.c to vorbis/comment/reader.c and its declarations to vorbis/comment.h, next to the writer.
Add initial support for Video Processing Engine (VPE) as a hardware accelerator option for AMF VPP.
seek2any adds one flag to every seek, so each other flag would need an option of its own.
GCC 16.2.1 reports -Wstringop-overflow in frame encoding because the psychoacoustic window count, which comes from an indirect model callback, is not visibly bounded before it indexes the per-window arrays.
The C YUV420P/YUV422P/YUVA420P to RGB converters process each row in blocks of 8, 4 and 2 pixels, so for an odd output width the final pixel of every row was never written and kept whatever was in the destination buffer. The unscaled path selection only rejects odd heights, and the ARM and AArch64 NEON versions require widths divisible by 16, so on those platforms odd widths always reach this code.
A corrupt packet may hold only part of a frame, which skews the average frame rate estimated from the few probed packets.
HAVE_FAST_CMOV is only disabled on x86-64 when configuring with--cpu=pentium4, pentium4m, prescott or nocona (or --cpu=host on these CPUs), on which cmov is supported but slow.
Each decision computes its range from the range of the previous one, so the instructions that go from one range to the next are the critical path of get_cabac, and their latency limits decoding speed. The LPS range lookup in the x86-64 PIC version builds its table address from the state and two bits of the range.
Each decision computes its range from the range of the previous one, so the instructions that go from one range to the next are the critical path of get_cabac, and their latency limits decoding speed. The renormalisation shift in the x86-64 PIC version was looked up from the new range, so the lookup had to wait for the MPS/LPS decision.
The x86-64 version of the get_cabac inline asm keeps the address of the tables in a register.
The refill shift is the number of trailing zeros in low minus 16.
Both versions of get_cabac now use the same tzcnt refill, so define it only once.
Drivers without it, like ANV, rejected every AV1 stream by default; sequences with film grain still fail there unless it is exported.
There are a number of cases in which deriving the filename from a hash of the URL is not ideal: For example, when the URL embeds parameters that don't affect the underlying file data, or when the user wants to be able to clean up or access the cache files directly, e.g.