[Toolkit bug]: Texture ingest spins up multiple instances of nvtt_export, saturating vram.

[BinqAdams]

Describe the bug

It seems to run out of VRAM and then grind to a halt.

How do you reproduce the bug?

Attempt to ingest a high resolution texture set (8k) of normal, albedo, height, roughness on a system with limited VRAM.

Version

1.0.0 | 106.5.0+release.1062521.d02c707b.gl

[nxkb]

Hi @BinqAdams, Thanks for reporting. Can you give a bit more detail on the repro steps just to narrow down the exact mode you are running in? One of those files may be helpful too if it's not too large.

Looks like we're allowing 4 procs here so one potential fix is to make that configurable. 8k textures may indeed just demand a lot of resources.

[BinqAdams]

Hi @BinqAdams, Thanks for reporting. Can you give a bit more detail on the repro steps just to narrow down the exact mode you are running in? One of those files may be helpful too if it's not too large.

Looks like we're allowing 4 procs here so one potential fix is to make that configurable. 8k textures may indeed just demand a lot of resources.

Repro is quite simply to load lightspeed.app.trex.ingestcraft.bat (or the full toolkit, same behaviour), and attempt to ingest an 8k texture set.

I changed max_workers to 2, and the ingest completed successfully for the first time.

[BinqAdams]

The textures are too large to upload here, but you get the same issue with (for example) any large 8k texture set.

For example, png's of the albedo, height, roughness, normals here:

https://polyhaven.com/a/climbing_wall_02

[nxkb]

We've created an internal ticket for tracking: https://jirasw.nvidia.com/browse/REMIX-4404

[nxkb]

Nice, I'm glad the setting work for ya.