Optimized Fully Fused Differentiable SSIM

This repository provides an optimized version of fused-ssim, achieving better performance while maintaining the same interface. Below are performance comparisons measured using tests/test.py.

Performance Comparison (measured on RTX 4090)

Implementation	Forward Time (ms)	Backward Time (ms)	Inference Time (ms)
Optimized Fused-SSIM	2.49	2.68	1.43
Original Fused-SSIM	3.66	3.52	2.59

Optimizations

1. Fused Computations

• The original implementation computes multiple statistics (mean, variance, covariance) in separate passes, requiring multiple memory accesses.
• Optimization: The computations are fused into a single pass, reducing redundant memory accesses by using shared memory for intermediate results.

2. Exploiting Symmetry in Gaussian Convolution

• The Gaussian filter is symmetric (e.g., G_00 = G_10), but the original implementation does not take advantage of this.
• Optimization: Pairs symmetric elements to halve the number of multiplications from 11 to 6.

3. Constant Memory for Gaussian Coefficients

• The original implementation uses #define macros for Gaussian coefficients, increasing register pressure.
• Optimization: Stores coefficients in CUDA constant memory (__constant__ float cGauss[11]) for:
  • Faster memory access.
  • Reduced register pressure.
  • Improved scalability across GPUs.

4. Smaller Block Size

• The original implementation uses 32x32 thread blocks.
• Optimization: Uses 16x16 blocks, which improves GPU occupancy by reducing resource usage per block.

5. Efficient Shared Memory Usage

• The original implementation loads img1 and img2 into separate shared memory arrays.
• Optimization:
  • Uses a 3D shared memory array (sTile[SHARED_Y][SHARED_X][2]) to load both images simultaneously, reducing global memory accesses.
  • Stores intermediate sums in a unified shared memory array, improving data locality and minimizing memory fragmentation.

6. Reduced Computational Redundancy

• The original implementation performs separate convolution operations for each statistic.
• Optimization: Uses a single convolution pass for all statistics, reducing redundant computations.

Special thanks to Florian Hahlbohm for helping me verifying that my optimization don't break anything.

Radiance Fields Dev Discord

📖 Citation

If you use this optimized fused-SSIM implementation for your research, please cite both the original paper and this implementation:

@inproceedings{optimized-fused-ssim,
    author = {Janusch Patas},
    title = {Optimized Fused-SSIM},
    year = {2025},
    url = {https://github.com/MrNeRF/optimized-fused-ssim},
}
@inproceedings{taming3dgs,
    author = {Mallick, Saswat Subhajyoti and Goel, Rahul and Kerbl, Bernhard and Steinberger, Markus and Carrasco, Francisco Vicente and De La Torre, Fernando},
    title = {Taming 3DGS: High-Quality Radiance Fields with Limited Resources},
    year = {2024},
    url = {https://doi.org/10.1145/3680528.3687694},
    doi = {10.1145/3680528.3687694},
    booktitle = {SIGGRAPH Asia 2024 Conference Papers},
    series = {SA '24}
}