[BUG] ImageBuf::nmiplevels() returns 0

[AlexeySmolenchuk]

OpenImageIO 3.1.14.0

Only cached version of ImageBuf object returns correct number of mipmaps.

buffer = ImageBuf(imageFileName, subimage, mip, cache);
nmips = buffer.nmiplevels(); // Works as expected

buffer = ImageBuf(imageFileName, subimage, mip);
nmips = buffer.nmiplevels(); // Always returns zero

Same for both exr and tiff backends.

Could be the same issue: #4984

[luna-y-kim]

The ImageBuf constructor path without an ImageCache ends in ImageBufImpl::init_spec(). There, the non-ImageCache branch executes m_nmiplevels = 0; and never touches it again. Confirmed by changing that line to assign an arbitrary number, which is exactly what ImageBuf::nmiplevels() then returns.

I'm going to look into how to properly assign the value.

OpenImageIO/src/libOpenImageIO/imagebuf.cpp

Line 1258 in a1045a4

m_nmiplevels = 0;

[AlexeySmolenchuk]

The cached version of ImageBuf also returns wrong values, e.g.
ImageBuf("kitchen_probe.hdr", 0, 0, cache); returns 10, whereas no mipmaps could be stored in the .hdr file.
The result is more like "how many mipmaps are suitable for this resolution" rather than the number of existing mipmaps.

[luna-y-kim]

The cached version of ImageBuf also returns wrong values, e.g.
ImageBuf("kitchen_probe.hdr", 0, 0, cache); returns 10, whereas no mipmaps could be stored in the .hdr file.

I can't reproduce this with a plain ImageCache on a single-image file (I tested with a .hdr, which returns 1 as expected). One way I could get a larger number on my end was by setting the automip attribute on the ImageCache. Did you happen to have automip enabled?

OpenImageIO/src/include/OpenImageIO/imagecache.h

Lines 134 to 142 in a1045a4

	/// - `int automip` :
	/// If 0 (the default), an untiled single-subimage file will
	/// only be able to utilize that single subimage.
	/// If nonzero, any untiled, single-subimage (un-MIP-mapped)
	/// images will have lower-resolution MIP-map levels generated
	/// on-demand if pixels are requested from the lower-res
	/// subimages (that don't really exist). Essentially this
	/// makes the ImageCache pretend that the file is MIP-mapped
	/// even if it isn't.

[lgritz]

That's a good hypothesis, @luna-y-kim. Automip would make the IC behave as if there are MIP levels, but a direct ImageBuf access not going through the cache would fail to find those virtual levels.

[AlexeySmolenchuk]

Hi @luna-y-kim
If it helps:
Here is how it was discovered:
AcademySoftwareFoundation/OpenShadingLanguage@034df86

And here is the complete log but with OIIO 3.1.6.2 :
https://github.com/AcademySoftwareFoundation/OpenShadingLanguage/actions/runs/33069945076/job/98509300701#logs