In the .cu file, you put your CUDA code, and usually also definitions of one or more C functions that are interface for the rest of your code (one of these is usually the function containing the kernel execute statement). So you may put declarations of these functions into corresponding header file; this file could also contain declaration of some datatypes and structures that are common for your both host and kernel code. By convention, you could choose .cuh extension for such header file; you could find number of examples of alike files within code samples in SDK. But it is all matters of convention - these files are processed as any other header files, and I think that none of tools involved care about the extension of these particular files.
Please avoid using path including non-ASCII characters. See NVRTC compilation fails when CuPy is installed under directory path containing non-ASCII · Issue #7581 · cupy/cupy · GitHub .
You don’t exclude .cu files from the build, normally. Those are considered to be compilable modules, ordinarily. A .cuh is a header file, it should be excluded from the build, it only gets compiled when it is included in a compilable module somewhere.
But the interesting thing is that I extracted my input data after the data transfer to the GPU to make sure that I had no problems, I saved the bin file and used the trtexec command to go through the inference, and it seems to be fine, here are the relevant commands, And we can also see the normal inference output afterwards with the --dumpOutput command.
people often use .cuh in my experience as a convention to indicate a header file that also contains cuda-specific syntax, meaning it must be processed by nvcc , and included in a .cu file. But there is nothing that enforces that behavior.
Tags:
Cuh,
2208b,
ps4,
что,
значит,
