CUDA - GPU memory

[Sylvanar]

Hello guyz,

In CUDA kernel call, I am passing an array of points and which is then divided into 4 different quadrants in the GPU depending on their value. After the kernel call is over some more points are added to the initial array of points and then the array is resent.

I want to ask when the array is resent, whether all the computations done during the first kernel call be present in the device memory or if all the computations would have to be done again?

Thank you.
 

[mathijs727]

Just to be sure that I got the question:

• Have an array of points
• Split the array using a GPU kernel
• You have more data and you want to run the same kernel and append te result to the previous results?

 

What you can do is keep track of the size of the output quadrants.

Add an offset for each quadrant to your kenrel so that you write the results after the previous results (instead of overwriting them).

 

Another solution is to allocate new GPU arrays for borth the new input data and the output quadrants.

You can then copy those resuls to the CPU and merge them with your existing results.