MagicScaler v0.10.2.0
MagicScaler v0.10.2
Fixes
- Fixed an issue where column order was swapped in 64bpp CMYK TIFF images when running on .NET Core 3.x with an AVX2-capable processor.
Performance Improvements
- Added SIMD implementations to the hybrid box scaler, improving on speed of the previous SWAR implementations by another 3-4x.
- Aligned pixel buffers where possible, reducing cache-line splits and improving overall SIMD throughput.
- Changed convolution kernel layout to align for SIMD reads, and deduplicated weights to improve cache hit ratio.
- Added small convolution kernel cache to allow weights to be re-used between pipeline stages, such as for identical transforms of Cb and Cr planes.
WebRSize 0.5.0
The algorithm used for file cache names has changed. You should clear your disk cache when upgrading.
Changes
- Updated to use MagicScaler v0.10.2.
- Lighter-weight cache hash calculation.