Reverse-engineering the Apple Video Decoder (AVD) found on Apple Silicon, primarily the custom instruction streams driving each of the 3 (or 4) programmable per-codec processors of the first stage of the decoding pipeline. Given a stream of little-endian u32 words and the slice data, if all goes well, the processors will emit a codec-agnostic residual IR to be transformed by the shared stage two prediction block.

• avid/: AVD Instruction model + Python stateless decoder glue. The "i" is to make Python happy
• codecs/: In-house C bitstream parsers
• tools/: Nifty tools/wrapper scripts to assist RE
• avd_emu.py: Coprocessor firmware emulator repurposed to spit out the instruction stream FIFO

All for my J293AP only, but different hardware revisions should only require minor encoding changes at the instruction HAL. The processors are really glue for a unified IR (this helps reuse silicon), and the hard part is figuring out the codec-specific quirks.

Where

• Tracer: m1n1-side hypervisor hooks for tracing and dumping reference macOS source data.
• Emulator: avd_emu.py support for extracting the instruction stream out of trace dumps.
• Parser: Bitstream parser, demuxer if needed. "Syntax" in Apple driver terms.
• FParser: macOS frame_params struct + other blobs (e.g. VP9 probabilities blob) documentation.
• Decoder: Reference List Management (RLM) / DPB management logic. Includes memory allocation calculations.
• Matrixbench: Valid instruction stream generation for

   ffmpeg -i matrixbench_mpeg2.mpg -s 1024x512 -pix_fmt yuv420p -c:v libx264 matrixbench.h264
  

  Basically a POC and that it'll happen soon™.
• Works?: Stuff's never "done", but it's reasonably feature-complete in my view, e.g. supports all resolutions up to hardware cap (usually 4K), supports all notable features the macOS driver does, has been battle-tested with gnarly samples, etc. But all of the following asterisks apply.

• What works:

  • High Profile i.e. 8-bit 4:2:0 I/P/B (i.e. normal x264 output)
  • High 4:2:2 Profile. Tested less than 4:2:0 but I haven't seen it not work
  • High 10 Profile (10-bit)
  • High 10 4:2:2 Profile
  • Up to hardware cap of 4096x4096 (that includes 4K)

• Works in macOS but not in mine:

  • frame_num gaps. This is a 10-line fix if I can see a sample.
  • Multiple slice groups. Ditto but 20

• Unsupported by hardware:

  • MBAFF/PAFF interlaced. Though you will not see this unless you work in broadcast.
  • Top/bottom field coding (x264 doesn't support)

• H.264 but worse. Doable

• 12/16/2023: It works for normal videos encoded with normal parameters, e.g. a youtube video, but it's not passing all of the ITU-T conformance suite, which are samples designed to stress the decoder. We have most features down e.g. I/P/B, dequant/deblock, scaling lists, tiling/WPP, multiple slices, PCM, etc. Motion vector merging needs work. There's also some PPS extensions supported in hardware. Also 4:2:2 and 10-bit, but those shouldn't take long.

• Looks easier than H.264. Up next

• Would you like to buy me an M3? :D

Other than RE-ing

• Add (more) tests
• Add SoC/Revision/Codenames/Caps matrix