Cuda Memory Exception

[xohm]

Hi,

First of all, thank you so much for this fantastic library!

I encountered CUDA memory exceptions when using free camera views that were different from the original cameras in the test dataset (e.g., counter).

My setup is as follows:

• CUDA version: 12.4
• GPU: NVIDIA RTX A2000 Laptop GPU with 4GB memory
• OS: KDE Neon 22.04
• Development environment: C++ project set up with CMake, using LibTorch 2.5.1 with CUDA and the cxx11 ABI.

The crashes occurred when the camera viewport was partially or fully outside the bounds of the Gaussian splats.

I managed to resolve the issue, but I must admit that I don’t fully understand your code yet. Here are the changes I made:

1. After Preprocess

   • If there are no Gaussian splats to draw, abort the sorting and rendering processes. Otherwise, memory exceptions occur.

2. renderCUDA

   • I identified an uninitialized variable in certain cases and an array out-of-bounds access when a Gaussian splat was skipped. I did attach a patch. crashFixes.zip

[chensiyan96]

Your GPU memory is small, so maybe there's an oom? Set MAX_NUM_RENDERED = 2 ** 26 and try a small scene.

[xohm]

Yes, your right, i already did this: _rasterizer = std::make_shared(*_scene, 1 << 24, 1 << 18);

Otherwise it would have right away oom. With these settings i do see the splats, but as mentioned above, if i move into an area where the splats are clipped i see the crashes.