So I’m using TensorRT as post-process in a graphical program. The input to TensorRT is a d3d11 rendered texture.
For cuda to access the texture, I’m using cudaGraphicsD3D11RegisterResource/cudaGraphicsMapResources/cudaGraphicsSubResourceGetMappedArray. Then I find that I have to make a cudaMemcpy2DfromArray, and then access the cuda ptr from the TRT side.
My question: can I access the cudaArray directly without copying?
A cudaArray cannot be indexed into or otherwise accessed like you would access “ordinary” memory via pointers. Beyond that, for options, I have moved this to TensorRT section/area. If TRT is expecting an ordinary memory pointer (seems logical to me), then there is no way you will be able to use a cudaArray as a substitute.