mirror of
https://github.com/tuskyapp/Tusky
synced 2025-02-09 20:20:43 +01:00
125483db61
This pull request aims to dramatically improve the performance of
`BlurHashDecoder` while also reducing its memory allocations.
- Precompute cosines tables before composing the image so each cosine
value is only computed once.
- Compute cosines tables once if both are identical (for square images
with the same number of colors in both dimensions).
- Store colors in a one-dimension array instead of a two-dimension array
to reduce memory allocations.
- Use a simple `String.indexOf()` to find the index of a Base83 char,
which is both faster and needs less memory than a `HashMap` thanks to
better locality and no boxing of chars.
- No cache is used, so computations may be performed in parallel on
background threads without the need for synchronization which limits
throughput.
## Benchmarks
Simple: 4x4 colors, 32x32 pixels output. (This is what Mastodon and
Tusky currently use)
Complex: 9x9 colors, 256x256 pixels output.
**Pixel 7 (Android 14)**
```
365 738 ns 23 allocs Trace BlurHashDecoderBenchmark.tuskySimple
109 577 ns 8 allocs Trace BlurHashDecoderBenchmark.newSimple
108 771 647 ns 88 allocs Trace BlurHashDecoderBenchmark.tuskyComplex
12 932 076 ns 8 allocs Trace BlurHashDecoderBenchmark.newComplex
```
**Nexus 5 (Android 6)**
```
4 600 937 ns 22 allocs Trace BlurHashDecoderBenchmark.tuskySimple
1 391 487 ns 7 allocs Trace BlurHashDecoderBenchmark.newSimple
1 260 644 948 ns 87 allocs Trace BlurHashDecoderBenchmark.tuskyComplex
125 274 063 ns 7 allocs Trace BlurHashDecoderBenchmark.newComplex
```
Conclusion: The new implementation is **3 times faster** than the old
one for the current usage and up to **9 times faster** if we decide to
increase the BlurHash quality in the future.
The source code of the benchmark comparing the original untouched Kotlin
implementation to the new one can be found
[here](https://github.com/cbeyls/BlurHashAndroidBenchmark).