| Component | Name |
|---|---|
| CPU | Intel Core [email protected] (Turbo Boost: 5.2Ghz, 16C/24T) |
| GPU | NVIDIA GeForce RTX 3080 10GB |
| RAM | 32GB DDR4 (3600 Mhz) |
| Component | Name |
|---|---|
| Name | Apple MacBook Pro 16" 2023 |
| CPU | Apple M2 Pro |
| RAM | 32GB |
The tests have been performed on a set of 9 pictures with a resolution from 1 to 151 MPix and a 3D LUT of size 33.
The trilinear interpolation method provides excellent image quality at the cost of higher mathematical complexity. This is a perfect scenario for a GPU, as it fully utilizes its potential to accelerate heavy floating point compute operations combined with the parallel nature of the image matrix. The memory copying/allocation costs are fully compensated by the performance gains for images of size higher than 1 MPix.
The nearest-value interpolation provides worse image quality, but it's much faster than the previous method. Its computational complexity is not high enough to compensate for the GPU memory I/O latency, so the performance of CUDA kernels is worse than multithreaded implementation.
Nearest-neighbor interpolation is significantly faster on the CPU than the trilinear method.
The results are nearly identical, meaning that almost entire time cost is consumed by the memory I/O operations (copying image matrix between the local RAM and GPU VRAM).