Dynamic Global Memory Allocation ? faulty documentation - malloc within Kernel

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

In “Cuda C Programming Guide” - chapter B.15. - Dynamic Global Memory Allocation; states that it is possible to dynamically allocate Device memory.

However when I try to use malloc from within the kernel I get following error:

error : calling a host function from a device/global function is not allowed

Is it possible to use malloc from within the kernel function or chapter B.15. is completely wrong?

EDIT:

I have figured out what is the problem. On devices with Compute Capability < 2.0 compile will fail if you have malloc in global or device function. However on compute capability >= 2.0 it works as stated in documentation.

As time goes by I realize that purchase of C1060 a month ago was complete waste of money :(

many thanks

Mirko

2

On Fermi GPUs with the 3.2 toolkit you can. Just include stdlib.h and compile for the sm20 architecture and you should be good to go.

3

thanks avidday,

I just realized that.

I have C1060 and GTX 470, tried it with Fermi and it works OK.

Documentation should be changed to clearly state for what architecture this feature can be implemented.

thanks

Mirko

4

Thanks for pointing that out, I will bring it to the attention of the documentation folks.

5

thanks njuffa