Performance

Tips that matter

• Load in the type you will work in. Raster.Load<Rgba8> converts row by row while decoding; loading the natural type and converting afterwards touches every pixel twice.
• Use Derive for thumbnails. It reads the source once and writes only the small result; Copy then Apply copies the full image first.
• Decode small. ReaderOptions.MinimumSize decodes JPEG at 1/2, 1/4 or 1/8 scale straight from the DCT, and averages blocks of rows as they are decoded for every other format, so a thumbnail never holds the full image.
• Stream large files. Raster.Transcode and RasterDecoder hold a few rows instead of the whole image; a leading crop or resize in a transcode chain becomes a decode region or reduction.
• Chain operations. A chain is planned as a whole: a Crop before a resize reads only the crop, adjacent color steps run as one pass, and AutoOrient before a resize turns only the result.
• Probe before you load when you might reject the file.
• Batch work on one thread per image (MaxParallelism = 1) and run images side by side; for a single large image, leave MaxParallelism at the processor count.
• Dispose rasters. Pixel memory is pooled; disposing returns it at once for the next image.

Running the benchmarks

dotnet run -c Release --project benchmarks/Egami.Benchmarks -- --filter "*"
dotnet run -c Release --project benchmarks/Egami.Benchmarks -- --filter "*SimpleDecode*"

The benchmarks use BenchmarkDotNet and generated test images (a 1920 × 1080 photo-like image and a screenshot-like one), so anyone can reproduce them.

Egami's numbers

Benchmarks compares Egami with other .NET imaging libraries on the same tasks.

Median times on one machine (24 cores, .NET 10, x64 with AVX2), decoding from memory into the file's natural pixel type:

Scenario	1 thread	All cores
BMP 24-bit decode, 1920 × 1080	0.31 ms	0.37 ms
BMP 8-bit palette decode	1.6 ms	1.7 ms
TGA 32-bit RLE decode	3.1 ms	2.9 ms
QOI decode, photo	15.3 ms	15.6 ms
PPM decode	0.32 ms	0.39 ms
QOI encode, photo	14.3 ms	14.4 ms
TGA 32-bit RLE encode	3.6 ms	3.6 ms
BMP 24-bit encode	0.46 ms	0.46 ms

Steady-state decoding and encoding allocate only the result raster: under 2.5 KB of managed memory a call for these formats.

Native AOT

Egami is trimming- and AOT-clean: no reflection, no runtime code generation. Publish as usual:

dotnet publish -c Release -r linux-x64 -p:PublishAot=true

Egami's own build publishes a small console app this way on every change and checks that it runs with no warnings.