Aliasing issue when resizing on 32-bit systems

First, thanks for the awesome work you did with Sharp ! 👍

I have a question (I don’t think it is a bug) regarding the resize function. My code looks like this :

const outputImage = sharp(imagePath);
outputImage
        .resize(Number(width), Number(height))
        .max();

When I run it on my Windows machine, the resized image is perfectly fine, without aliasing.

When I run it on a Debian 8 machine, it generates the following image : Source image :

Output with size 350x350 (the “bug” happens) :

Output with size 351x351 (perfectly fine) :

It seems to have this behaviour only with some “round” numbers and only on my Debian server, not on Windows.

Any help would be much appreciated. Thanks !

[PierreM-Dev]

Thanks a lot for these tips. You were almost right and pointed me in the right direction. I had to follow libvips docs to “build with the GCC auto-vectorizer and diagnostics (or just -O3)” (from here : https://github.com/jcupitt/libvips) :

$ FLAGS="-O2 -msse4.2 -ffast-math"
$ FLAGS="$FLAGS -ftree-vectorize -fdump-tree-vect-details"
$ CFLAGS="$FLAGS" CXXFLAGS="$FLAGS" \
	./configure --prefix=/home/john/vips 

It worked with these params. Maybe you could add this tip in Sharp docs if someone faces the same issue.

For information I tried enabling “simd” first but it didn’t work.

Again, thanks a lot for the time you give to the community, you are awesome and Sharp is amazing 😃

[lovell]

Enabling simd might help too as that will keep the floating-point reduction away from the x87 FPU.