Using Device Pointers from different functions

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi!

I have some trouble using device pointers from different functions. I have a class and defined a (private) member which I want to use as a device pointer to a table in the class. However it only works if the pointer is only used in one function.

So something like

Class::InitTable()

{

XX = cudaMalloc(size);

setTable<<<...>>>();

}

Class::UseTable

{

DoSomething<<<...>>>(XX);

}

doesn’t work and the table seems just to be filled with random values. While on the other hand

Class::InitAndUseTable()

{

if (!XX) {cudaMalloc(size); setTable<<<...>>>();}

DoSomething<<<...>>>(XX);

}

works perfectly fine.

Is this a known problem with CUDA? What else could be the reason?

Thanks!

2

Are these functions potentially called by different host threads? Device pointers are generally not portable between CUDA contexts, and each host thread is implicitly given a different context in the runtime API.

3

Indeed they are, thanks for the hint.

I guess I’ll have to rewrite things to execude all the CUDA stuff from one thread.

4

Improving support for multithreaded applications is something we’re actively working on.