Video codec

[igankevich]

This PR introduces a fast and reasonably efficient video codec optimized for RISCV. This codec uses YUV color model, Haar wavelet transform, scalar quantization, rANS encoder, and simple inter-frame delta encoding. To the best of my knowledge this is the smallest number of stages of a video encoding pipeline that are required to get decent compression ratio without significant performance degradation.

There are many royalty-free and open source video codecs but they tend to trade encoding speed for better compression ratio. This is acceptable tradeoff for a codec that is made for the web, but unacceptable in CoreVM context where a guest has only 6 seconds to generate and encode all video frames for the next time slot (i.e. next 6 seconds). Another problem is that a sizable portion of the source code in such codecs is written in x86/ARM assembly (example), and fallback code doesn't perform well on RISCV. Hence a custom codec.

Benchmarks

Below are the results of a benchmark for AV1, CoreVM and QOI+rANS codecs. The benchmark encodes 6 seconds worth of Quake frames. The benchmark was run inside CoreVM. Original video size is 9327 KiB.

Codec	Compressed size	Encoding time	Compression ratio	Crate
CoreVM	2283 KiB	308 ms	76%
QOI+rANS	5142 KiB	192 ms	45%	qoi
AV1	275 KiB	22403 ms	97%	rav1e

Same benchmark on x86_64 without CoreVM.

Codec	Encoding time
CoreVM	172 ms
QOI+rANS	88 ms
AV1	1109 ms

Demo

Below are two animated PNG images generated with ffmpeg: the left image was generated from raw Quake frames (convert Quake frames RGB, then to PNG), the right image was generated from the CoreVM video stream (convert CoreVM frames to raw RGB, then to PNG).

Original	CoreVM