Small, but yet reasonably fast H264/SVC encoder single-header library with SSE/NEON optimizations. Decoder can be popped up in future.
Disclaimer: code highly experimental.
Rough comparison with x264 on an i7-6700K:
x264 -I 30 --profile baseline --preset veryfast --tune zerolatency -b 0 -r 1 --qp 33 --ipratio 1.0 --qcomp 1.0 -o x264.264 --fps 30 vectors/foreman.cif --input-res 352x288 --slices 1 --threads 1
|out size: 320kb||391kb|
x264: PSNR y:32.774824 u:38.874450 v:39.926132 average:34.084281 min:31.842667 max:36.630286 minih264: PSNR y:33.321686 u:38.858879 v:39.955914 average:34.574459 min:32.389171 max:37.174073
First intra frame screenshot (left-to-right: original 152064, minih264 5067, x264 5297 bytes):
You can compare results in motion using ffplay/mpv players on vectors/out_ref.264 and vectors/x264.264 .
Minih264 supports both spatial and temporal layers. Spatial layers are almost same as encode 2 independent AVC streams except for Intra frames prediction. Following diagram shows minih264 SVC scheme for two spatial layers:
That's because P frames spatial prediction are almost useless in practice. But for Intra frames there is a ~20% benefit in full resolution frame size. Note that decoder must have both base layer I frame and full resolution SVC I frame to decode whole sequence of next P frames in full resolution.
The following major features are not supported compared to x264 (baseline):
- Trellis quantization.
- Select prediction mode using Sum of Absolute Transform Differences (SATD).
- 4x4 motion compensation.