Fine-Grained Load Balancing with Warp-per-Pixel Rendering

[OrangeEarth15]

Introduces fine-grained load balancing optimization for 3DGUT rendering to address GPU load imbalance issues, inspired by Balanced 3DGS.

Implementation

Transitions from thread-per-pixel to warp-per-pixel parallelism:

• Each warp (32 threads) now processes one pixel cooperatively instead of one thread per pixel
• Subdivides each 16×16 tile into 64 virtual tiles (2×2 pixels each)
• Enables particle-level parallelism across warp threads

Configuration

# configs/render/3dgut.yaml
splat:
  fine_grained_load_balancing: true  # default: false

Changes

• Add renderBalanced kernel with warp-per-pixel processing
• Add evalForwardNoKBufferBalanced for cooperative particle traversal
• Add compile-time flag FINE_GRAINED_LOAD_BALANCING
• Add config parameter fine_grained_load_balancing

Performance

Benchmark on MipNeRF360 dataset (8× downsampled):
Render kernel: 3.19ms → 1.78ms (~1.8× speedup, 44% reduction)

[wilsonCernWq]

Awesome