I don't really have much to add other than on a Raspberry pi 5, the image crate outperforms this one. While not by much, and they remain close at lower qualities, the fact that x86 is like twice as fast as image seems to indicate that something is wrong on arm. I realize this might not have been a design priority, and sometimes optimizing for one is worse for the other. But when vectorization is also not supported for that architecture, it seems to leave a gap in performance.
Not really an issue per say, feel free to close if this isn't part of the goal. But I thought it was interesting, and maybe you just never tested on that architecture.
Edit: Forgot to mention, std_simd branch was also about 5% slower. Just a rough observation in a test as part of my app, not a proper cargo bench result, but across multiple results it's almost certainly not faster.
I don't really have much to add other than on a Raspberry pi 5, the image crate outperforms this one. While not by much, and they remain close at lower qualities, the fact that x86 is like twice as fast as
imageseems to indicate that something is wrong on arm. I realize this might not have been a design priority, and sometimes optimizing for one is worse for the other. But when vectorization is also not supported for that architecture, it seems to leave a gap in performance.Not really an issue per say, feel free to close if this isn't part of the goal. But I thought it was interesting, and maybe you just never tested on that architecture.
Edit: Forgot to mention,
std_simdbranch was also about 5% slower. Just a rough observation in a test as part of my app, not a proper cargo bench result, but across multiple results it's almost certainly not faster.