CUDA illegal memory access when Gaussian opacity < 1/255 in duplicateToTilesTouched (compact box)

[LiSaiCSU]

Summary

Training crashes with CUDA error: illegal memory access during _C.rasterize_gaussians (often reported asynchronously at (radii > 0).nonzero() in Python). Root cause: compact-box tile culling uses log(con_o.w * 255.0f) without guarding con_o.w <= 1/255, while the rasterizer already skips splats with effective alpha below 1/255.

Environment

• OS: Windows 11
• GPU: NVIDIA RTX 5090 (sm_120)
• PyTorch: 2.11 + CUDA 12.8
• Dataset: COLMAP, ~100 views, images rescaled to ~1600×900
• Command: python train.py -s data -m data

Reproduction

1. Train until ~4k–7k iterations (densify continues; opacities drift down).
2. Crash stack (representative):
   • gaussian_renderer → diff_gaussian_rasterization_fastgs._C.rasterize_gaussians
   • Sometimes mis-attributed to visibility_filter : (radii > 0).nonzero().
3. With CUDA_LAUNCH_BLOCKING=1 and pipe.debug=True (--debug_from N):
   • Error localizes to cuda_rasterizer/rasterizer_impl.cu around CHECK_CUDA(FORWARD::render(...)).
4. Saved forward args (snapshot_fw.dump via debug path) showed at crash:
   • P ≈ 81461 Gaussians
   • opacity min ≈ 0.001086 (< 1/255 ≈ 0.0039215686)
   • Image ~1600×899

Root cause (proposed)

In cuda_rasterizer/auxiliary.h, duplicateToTilesTouched:

float t = 2.0f * log(con_o.w * 255.0f);
<img width="2396" height="925" alt="image" src="https://github.com/user-attachments/assets/ed45a306-801f-4f10-ac3d-85ad0cd4bdd2" />