I am using an iterative error minimization algorithm that utilizes 3D texture interpolation. A simplified workflow of each iteration is:
- Place 3D data into cudaArray
- Sample with linear interpolation from texture of cudaArray
- Update 3D data
I would like to use a cudaTextureObject to link a known (at run-time) section of global memory to a texture that I would send to a kernel every iteration. And after that kernel is finished I would like to update the global memory that is linked to that texture object without binding/unbinding (creating/deleting) the texture (object).
Is this feasible with how texture objects and global memory of cudaArrray’s work?