Display mipmap preview levels as tabs instead of using a spinner

[dgelessus]

This improves a part of the mipmap preview window that I found confusing at first - the "Level" spinner at the top:

When I first used this preview window, I assumed that the level was a property stored in the mipmap object. I expected that changing the spinner would modify the PRP file, so I never touched it. I didn't realize that the spinner has no permanent effect and actually lets you preview other mipmap levels beyond the first one.

This PR replaces the spinner with a tab bar, which IMO communicates the meaning and behavior more clearly. As a bonus, the tab titles also display each level's pixel width/height.

[dpogue]

Plasma has a weird thing where it doesn't actually store data for any levels where the width or height is 2px or smaller.

We can finally account for that here in PrpShop to hide tabs with invalid levels:

if ((tex->getLevelWidth(level) | tex->getLevelHeight(level)) & 0x03)
    continue;

[dgelessus]

Huh, good to know. Is this always the case? I'm trying to make sense of the relevant Plasma code and it seems to be specific to DXT compression. It also reads like small sizes would have some data, just not in the usual format? :/

[dpogue]

It definitely might be specific to DXT compression. I know that OpenGL rejected those smaller levels as invalid data when I was implementing texture loading for plGLPipeline, and there were enough places in Plasma doing tests against & 0x03 to make it apparent that something weird was going on with those.

We also had to workaround this in korman because of similar issues: https://github.com/H-uru/korman/blob/106ef015480fb3dc0ba3cef90491af042458d58b/korlib/texture.cpp#L80-L91

[Hoikas]

Going off of memories of yester-year, the issue I most remember running into is Cyan's level size calculation for DXT5 compression being wrong. IIRC for a 2x2 mip level, it only allocates an alpha block but no color block.

[dgelessus]

Did some quick research (read: punched "DXT" into Wikipedia). Apparently, DXT operates on blocks of 4x4 pixels, so it makes sense that it can't work for textures that are smaller than that along either axis.

I still can't really follow what the Plasma code is doing in that case. But based on some experimentation and staring at hex dumps of a couple of mipmaps, I think any levels narrower than 4 pixels are simply stored uncompressed.

This quick and dirty patch gives me reasonable results for e. g. Ahnonay's mipmaps. (Well, as reasonable as you can get for a texture that's 1 or 2 pixels wide, anyway. I agree with Deledrius' take in the comment that Hoikas linked.)

--- a/src/PrpShop/PRP/Surface/QMipmap.cpp
+++ b/src/PrpShop/PRP/Surface/QMipmap.cpp
@@ -87,7 +87,11 @@
             dp++;
         }
     }
-    fImage = new QImage(fImageData, tex->getLevelWidth(level),
+    const unsigned char* theActualImageData = fImageData;
+    if (tex->getCompressionType() == plMipmap::kDirectXCompression && (tex->getLevelWidth(level) < 4 || tex->getLevelHeight(level) < 4)) {
+        theActualImageData = static_cast<const unsigned char*>(tex->getLevelData(level));
+    }
+    fImage = new QImage(theActualImageData, tex->getLevelWidth(level),
                         tex->getLevelHeight(level), QImage::Format_ARGB32);
     resize(tex->getLevelWidth(level), tex->getLevelHeight(level));
     update();

[dgelessus]

I think it would be better to handle this case correctly in libHSPlasma. AFAICT, the mipmap levels in question do have image data stored, so we shouldn't hide them entirely. The special handling for the different image data format should go into DecompressImage and not here.

[Deledrius]

Great QoL improvement. Thank you!

[dgelessus]

Ah, I wasn't aware of #44 and #69 - I hope this change doesn't conflict too much with those 😅

[zrax]

It may not be too important, but this can also be done by replacing the spin box with QTabBar that still lazy loads the image data, instead of loading all mips at once. It would simplify the diff, as well as avoid populating several mostly-identical widgets with only a different image content.

[dgelessus]

The QTabBar solution works quite well actually. This also avoids adding a small border around the texture like QTabWidget did. Thanks for the suggestion 👍