PyCuda使用Streams执行Thrust

Question

I'm trying to adapt the code found here: https://wiki.tiker.net/PyCuda/Examples/ThrustInterop ...to use cuda streams.

(Please excuse that I'm new to c++, and have a few weeks experience with cuda only.)

My main attempt and sticking point has been along the lines of adjusting the NVCC function like so to receive a cuda stream as an arg, and supply to the Thrust call:

nvcc_function = FunctionBody(
   FunctionDeclaration(Value('void', 'my_sort'),
                       [Value('CUdeviceptr', 'input_ptr'),
                        Value('int', 'length'),
                        Value('cudaStream_t','stream')]),
   Block([Statement('thrust::device_ptr<float> thrust_ptr((float*)input_ptr)'),
          Statement('thrust::sort(thrust::cuda::par.on(stream),thrust_ptr, thrust_ptr+length)')]))

I'm getting the error "'cudaStream_t' has not been declared" (referring to the NVCC function argument).

I've tried adding 'cuda_runtime.h' to both the host and device includes lists but to no avail.

Answer 1

I am not familiar with pyCUDA or thrust but I am familiar with CUDA. One of the possible things that come to mind is that some reason the "cuda_runtime.h" might not be included despite being specified. Are you sure that the pyCUDA framework will indicate an error when it cannot find a specific include?

Also another thing that caught my attention is that you are using CUdeviceptr which is a part of the driver api , whereas cudaStream_t is a part of the runtime api , which operates on a different level.

From NVIDIA documentation, it seems that the driver api equivalent type would be CUstream . Source: http://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__STREAM.html#group__CUDA__STREAM

So the problem might be in mixing the functionality of apis on different levels. As I said though, I am not familiar with the exact framework you're using, those are just some suggestions that might or might not turn out useful.

Good luck with debugging!