Blurhash implementation for C#
What is done:
- BlurHash core encoder
- Encoder wrapper for
System.Drawing - Encoder tests for a set of testing images, using all possible component values.
What is TODO:
- Decoder
- Decoder wrapper for
System.Drawing - Decoder tests
Currently, this implementation is ~25x faster than the reference C implementation.
On my reference machine, BlurSharp (Release) can encode all six reference images in 1,1 to 8,8 component forms (64 combinations per image) in under 5 seconds. The C implementation (compiled with -O2) does the same in 125 seconds.
The library has been written with the System.Numerics vector instructions, specifically Vector3 is used heavily to speed up operations. A precalculated lookup table is used to do SRGB to Linear float conversions.
Span<T> is also used extensively in order to reduce object allocations and unnecessary copying of buffers.
For example, Span<T> allows for the encoding of System.Drawing bitmaps without copying the image buffer. Further work needs to be done to allow core support for different pixel formats, so that Bitmap.LockBits doesn't have to do work internally to present the image as PixelFormat.Format24bppRgb.
The library also leverages the new Span<T> based safe stackalloc keyword to avoid heap allocations. Depending on how much stack you allow the encoder to use, an image can be encoded without a single object allocation. This can reduce GC pressure under heavy workloads.
Performance TODO:
- Potentially use parallel processing (questionable benefit for web scenarios)