swift-gif

swift-gif copied to clipboard

This repository could be much faster by encoding each frame in parallel. This is necessary to achieve acceptable latency in a production renderer, where videos are currently encoded in the GIF (640x640, 20 FPS). I am forking swift-gif and will update this thread when it has been optimized.

Using the GIF in the tests as a benchmark, I first decreased a bottleneck in the single-core execution context. Replace the indices of [Int] with [Int32], reducing execution time by 20%.

Before	After
17.1 s	13.7 s

Next, I parallelized across all CPU cores, with 8 performance and finally 2 efficiency cores. Then, re-ran the benchmark with a work redistribution algorithm to balance work across cores. Parallel efficiency is a metric usually used to measure strong scaling on supercomputers. Efficiency cores count as half a core for this metric.

Cores	Original	Load Balanced	Speedup	Parallel Efficiency
2	7.91 s	7.12 s	$1.92\times$	96%
3	5.35 s	4.95 s	$2.77\times$	92%
4	4.21 s	3.74 s	$3.66\times$	92%
5	3.35 s	3.07 s	$4.46\times$	89%
6	3.00 s	2.62 s	$5.24\times$	87%
7	2.72 s	2.34 s	$5.85\times$	84%
8	2.40 s	2.06 s	$6.65\times$	83%
9	2.34 s	1.99 s	$6.88\times$	81%
10	2.15 s	1.90 s	$7.21\times$	80%

Overall, this was a +801% speedup.

Pretty cool, feel free to open a PR. I'm actually surprised the Int -> Int32 made a big difference, I would not have expected that, at least not on modern 64-bit CPUs.

It's because the compiler can vectorize the instructions, and perform more comparisons in one instruction (theory 1). Another theory is slightly reduced memory bandwidth.