Fast 3D LUT on CUDA

3D LUT Transform is massively used for color grading and gamut mapping applications. To solve the task of 3D LUT grading, we have developed fast kernels that run on existing CUDA hardware from NVIDIA. We have implemented various formats for 3D LUTs and achieved very high performance for color grading.

3D LUT Transform Features

• Input data up to 16-bit with arbitrary width and height
• 2.5D and 3D LUT formats: cube
• Color representation: RGB, HSV
• Internal color cube resolution up to 65×65×65
• Compatibility with Fastvideo Image & Video Processing SDK

Hardware and software

• CPU Intel Core i7-5930K (Haswell-E, 6 cores, 3.5–3.7 GHz)
• GPU NVIDIA GeForce GTX 1080 (Pascal, 20 SMM, 2560 cores, 1.6–1.7 GHz)
• OS Windows 10 (x64)
• CUDA Toolkit 9.2

Performance for 2.5D and 3D LUT Transforms on CUDA

Test images: 16-bit RGB, 2432×1366 (2.5K) and 4032×2192 (4K), fast trilinear interpolation
Test info: all data (input and output) in GPU memory, timing measurements include GPU computations only, timing for 2.5K/4K images

• 2.5D LUT (HSV, 90×30 points) – 0.26 ms / 0.64 ms
• 2.5D LUT (HSV, 90×117 points) – 0.26 ms / 0.65 ms
• 3D LUT (HSV, 36×8×8 points) – 0.29 ms / 0.65 ms
• 3D LUT (HSV, 36×29×16 points) – 0.31 ms / 0.68 ms
• 3D LUT (HSV, 36×57×61 points) – 0.44 ms / 0.77 ms
• 3D LUT (RGB, 17×17×17) – 0.22 ms / 0.55 ms
• 3D LUT (RGB, 33×33×33) – 0.22 ms / 0.56 ms
• 3D LUT (RGB, 65×65×65) – 0.30 ms / 0.60 ms

We have designed that software as a part of our CUDA image & video processing SDK. Now our customers have opportunity to use fast 3D LUT transforms on CUDA in their realtime color grading applications with user-defined 3D LUTs.