`ktx create` hangs on image

[BenLubar]

I'm converting user-uploaded images from my game to ktx2 for faster loading. Most take a few milliseconds, but this one has been running for several minutes (12 minutes so far at the time of writing).

ktx create --format R8G8B8A8_UNORM --generate-mipmap --encode basis-lz --testrun --clevel 5 1W0CCXARX53PKQ0D.png 1W0CCXARX53PKQ0D.ktx2

The image itself doesn't seem to be anything out of the ordinary; it's just a 300x300 8-bit PNG file with a drawing of a dinosaur.

Original image is at https://spy-cards.lubar.me/spy-cards/user-img/1W0CCXARX53PKQ0D.png in case uploading the file to GitHub changes it.

Edit: More images that hang:

• https://spy-cards.lubar.me/spy-cards/user-img/2VTP0T6Y0R6Y3387.png (290x340, 8-bit)
• https://spy-cards.lubar.me/spy-cards/user-img/3QNXJB43Z1RESN07.png (330x410, 8-bit)
• https://spy-cards.lubar.me/spy-cards/user-img/68M79RS2HPH5HQKD.png (79x136, 8-bit but no warning about a transfer function is printed)

[MarkCallow]

Thank you for the report. Investigating. The issue appears to be related to the input PNG being RGB while you are specifying --format R8G8B8A8_UNORM. The Debug configuration hits an assert in the PNG loader. Without the assert the loader goes into an infinite loop. As a workaround try --format R8G8B8_UNORM.

Please do not use --testrun. That is for internal use only while running our test suite. It causes a default version number to be set in the KTXwriter metadata in the generated file. We need all files created by users to properly identify the version of the writer that was used.

[MarkCallow]

May I have your permission to add these 4 files to our test suite. They go through an infrequently used code path in our loader due to having an sBIT chunk. That in combination with --format R8G8B8A8_UNORM adding an alpha channel is the trigger for the hang. I want to add tests to make sure there will be no regression.

[MarkCallow]

P.S. Use an _SRGB format for the texture as PNG files without color space chunks are always, per W3C, treated as sRGB and nearly always are.

[BenLubar]

The command line arguments I ended up using after some testing are:

timeout 60 ktx create --format R8G8B8A8_SRGB --generate-mipmap --encode uastc --uastc-quality 4 --uastc-rdo --assign-oetf linear --convert-oetf srgb --assign-primaries none [...].png [...].ktx2

You have my permission to use any of the failing images in your test suite. I think two of these failed with actual error messages and the rest hung. List generated by searching for PNG files and printing the names of the ones that didn't have matching ktx2 files:

1W0CCXARX53PKQ0D
2VTP0T6Y0R6Y3387
3QNXJB43Z1RESN07
65V3B6AED6GPNZ8Q
68M79RS2HPH5HQKD
6F2EB6RGA06WHS8D
7DB1WWRW22A3HRGD
BN8C0MR852ZR5N87
C2XEZPZS5P49NK8J
CFFAX3Z0829XQ08J
DRHEE8AP4KSF3M18
EFSTPKJVGMSVH08T
EK60NKGYD7S1B707
FMF891YMGW11Q58S
G58YQ3WTQN005S8S
GCMEJEDH8RF0KG8H
GFX86HWH6DVZKTRD
H459KFCA2VZTHKGJ
HP0T7GB5TAG4BK0J
JMY766QHVZH0NM8D
JQPXFRQBTY5PDDM904
KB8BT9F881BN1G0H
MTVZPSP2GGH2ZZRG
NSZ8D7S45TKKVP0D
SFW3H7RP56F2K3AB
TQG979N5H9XXHNGJ
TX1KMZTWD2XX7GGH
W5VG34WYNE5RXXGS
XZB1HP70ZPAHVKRJ

[MarkCallow]

What licence do you have on the images? I need to make sure I label them properly for the reuse tool.

I notice you switched to UASTC. Do you see a significant quality difference between that and ETC1S for these images?

You should not need --assign-oetf linear --convert-oetf srgb. I don't know how your game is saving these PNG images but if they are screenshots they are almost certainly sRGB already, as strongly implied by the lack of any color space info in the file. Therefore you should omit these 2 options. If you want to get rid of the warning about sRGB being assumed, then use just --assign-tf srgb. As is, you are potentially giving an incorrect tf to ktx create and having it convert from that to sRGB which will undoubtedly mess up the shades in the images.

[MarkCallow]

https://spy-cards.lubar.me/spy-cards/user-img/68M79RS2HPH5HQKD.png has no warning about transfer function because it has an sRGB chunk. In that case using --assign-oetf linear is wrong.