Review: maketx --alpha1-detect

[lgritz]

maketx --alpha1-detect will omit alpha from textures whose input images had designated alpha channels that were 1.0 for all pixels. Somewhat similar in operation to --monochrome-detect and --constant-detect.

Also adds ImageBufAlgo::isConstantChannel() utility and does some minor cleanup/simplification to isConstantColor.

(Yes, this is because we have textures with pointless alpha=1 everywhere, and it's easier to fix in image-to-texture conversion than it is to actually fix the problem upstream.)

[lgritz]

Also note that I put the test for alpha=1 before the "is it monochrome" test. So it will correctly collapse CCC1 textures to C, if you know what I mean, whereas previously CCC1 textures would not only keep the pointless alpha, they also would fail to recognize the texture as monochrome (except for the very rare case of C==1).

[jeremyselan]

I like the refactor of this loop, much cleaner. Are there any issues with src.getpixel if the src image is null or doesnt have any pixels? (what does getpixel do?)

[lgritz]

src is a ref, not a pointer, so it can't be null. I'll put in a check to immediately return 'true' if there are no valid pixels at all.

[jeremyselan]

Looks good to me. I'm fine with this name (maketx --alpha1-detect) or any other one.

[fpsunflower]

LGTM

+1 for Solomon's suggestion of --opaque-detect though

[jeremyselan]

Someday we may want to tag what the original alpha value was in the metadata. (Particularly if we ever introduce an option to drop alphas with a constant 0.0 value). But until we cross that line, I'm all for not adding extra complexity until needed.

[lgritz]

At SPI, we checked our shader library and without exception, RGB textures are treated the same as RGBA textures where alpha=1, and I suspect that this is probably a widespread convention.

Dropping constant alpha that != 1 seems inherently dangerous, and hard to imagine that the extra logic in the texture system or shaders would be worth the storage savings, for which is surely a rare case of completely constant but non-1.0 alpha.