The file CUDALERP.h exposes two extremely high performance GPU resize operations, CUDALERP (bilinear interpolation), and CUDANERP (nearest neighbor interpolation), for 8-bit unsigned integer (i.e. grayscale) data.
For 32-bit float data, see the CUDAFLERP project instead.
CUDALERP offers superior accuracy to CUDA's built-in texture interpolator at comparable performance. The accuracy if compiled with -use-fast-math off is nearly equivalent to my CPU interpolator, KLERP, while still being as fast as the built-in interpolation.
Particularly for large images, CUDALERP dramatically outperforms even the highly tuned CPU AVX2 versions.
All functionality is contained in the header 'CUDALERP.h' and the source file 'CUDALERP.cu' and has no external dependencies at all.
Note that these are intended for computer vision use(hence the speed) and are designed for grayscale images.
The file 'main.cpp' is an example and speed test driver.