Golden bilinear image scaling


Has anyone written host code that they are willing to share that will resize a 8 bit greyscale image and produce identical results compared to a GPU implementation using the tex2D() bilinear interpolation access? I realize that GPU results might depend slightly on kernel implementation, but any efficient implementation would be fine, as long as it produces identical results on CPU and GPU.

At the moment, I need to upload images and do the scaling on the CPU, then download them again in order to verify that the rest of my algorithms produces identical results on CPU/GPU (they do if I use CPU scaling). So, if I could use CPU scaling code that produces identical results to the GPU, I could avoid this step.

Does appendix E in the programming guide give enough details to implement this?


Isn’t it quite useless to recommend a .NET based image processing class to a CUDA/C++ programmer? Besides, the original question was to obtain an efficient host side implementation of the algorithm that runs on the GPU.