The file CUDAFLERP.h exposes two extremely high performance GPU resize operations, CUDAFLERP (bilinear interpolation), and CUDAFNERP (nearest neighbor interpolation), for 32-bit float grayscale data.
For 8-bit unsigned integer data, see the CUDALERP project instead.
CUDAFLERP 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, CUDAFLERP dramatically outperforms even the highly tuned CPU AVX2 versions.
All functionality is contained in the header 'CUDAFLERP.h' and the source file 'CUDAFLERP.cu' and has no external dependencies at all.
The file 'main.cpp' is an example and speed test driver.