TORCH::FROM_BLOB memory types

So to my understanding, pytorch has the capability of doing direct GPU inference. For example, if the inputs are already in the GPU I can tell the inference network the input location in GPU memory and also specify where to place the output in GPU memory. The Torch::From_Blob method would be used for this.

My question now is how does the situation differ if the memory is not linear but instead a volume as a Cuda Array. This type of structure doesn’t really have a pointer in the same sense and under normal circumstances one would have to use surf3Dwrite() and surf3Dread() in Cuda kernels. If the input is in this format, what is the recommended way to run inference on it?


Unfortuatly, from_blob can only take raw memory pointer (and size/stride informations). So you won’t be able to wrap complex data types inside a torch Tensor.
But if one copy is ok, you can dump it into a new contiguous memory buffer and use that.

@albanD Thanks for the clarification