Add blur precision

[marcosc90]

This PR adds precision option to .blur().

Currently, using blur on large images is quite slow. In my case, setting precision to approximate significantly improves performance, reducing the processing time from over a minute to just a few seconds.

[kleisauke]

FWIW, I'm uncertain about exposing VIPS_PRECISION_APPROXIMATE because it's not production-quality and remains incomplete, see e.g. issue libvips/libvips#1162.

[marcosc90]

FWIW, I'm uncertain about exposing VIPS_PRECISION_APPROXIMATE because it's not production-quality and remains incomplete, see e.g. issue libvips/libvips#1162.

That's unfortunate, maybe I could remove approximate from the docs and expose it only for advanced users?

[lovell]

Thank you very much for the PR Marcos.

It looks like libvips/libvips#1162 relates to (u)short images processed by vips_conv /vips_conva, but Gaussian blur uses a (faster) separable matrix via vips_convsep / vips_convasep, so might be unaffected.

I guess as a compromise we could limit the use of the approximate value to uchar images, or force-cast to uchar?

[lovell]

Once we're happy with the API for this, please can you update the TypeScript definitions at https://github.com/lovell/sharp/blob/main/lib/index.d.ts

[marcosc90]

Thank you very much for the PR Marcos.

It looks like libvips/libvips#1162 relates to (u)short images processed by vips_conv /vips_conva, but Gaussian blur uses a (faster) separable matrix via vips_convsep / vips_convasep, so might be unaffected.

I guess as a compromise we could limit the use of the approximate value to uchar images, or force-cast to uchar?

You're welcome! Would you prefer to fall back to integer in case approximate is passed for ushort images, or throw an error?

[kleisauke]

It looks like libvips/libvips#1162 relates to (u)short images processed by vips_conv /vips_conva, but Gaussian blur uses a (faster) separable matrix via vips_convsep / vips_convasep, so might be unaffected.

I guess as a compromise we could limit the use of the approximate value to uchar images, or force-cast to uchar?

Good catch, vips_convasep() is indeed not affected.

I opened PR libvips/libvips#4055 to address the crash in vips_conva() that occurs with {u,}{short,int} images, so ignore my previous comment.

[lovell]

Thanks for the updates, I've left a couple of comments inline otherwise this is good to go.

[lovell]

Thank you very much for working on this Marcos.