Use of cuMemcpyDtoD in a __global__proc

Accelerated Computing CUDA CUDA Programming and Performance
1

I have a list of integers to copy in an other list.
A solution is

global proc()…
{

tid = blockDim.x * blockIdx.x + threadIdx.x;

for(jj=0;jj<mt_nn;jj++)// {
local_tableau_mots[jj]=d_tableau_mot[index_premier_k+mt_nn*tid+jj];
}

}

I would like to use cuMemcpyDtoD with something like:

cuMemcpyDtoD(local_tableau_mots,&d_tableau_mots[index_premier_k+mt_nntid],mt_nnsizeof(int)));

I get then the erro by VC2055:

error:
1> identifier “cuMemcpyDtoD” is undefined
1> cuMemcpyDtoD(local_tableau_mots,&d_tableau_mots[index_premier_k+mt_nntid],mt_nnsizeof(int));

What is wrong?
Thanks. :) :)

2

global functions run on the device, and you can only call cuMemcpyDtoD (or cudaMemcpy if you are using the Runtime API instead of the Driver API) from the host.

3

Thanks for your answer.
Does it mean there is only the first solution ?

4

Oh wait, I just reread your original post. Please ignore my previous response. I thought you were calling cuMemcpyDtoD from inside your kernel, but that’s not what you said. My apologies. :)

You #include “cuda.h” in your source file, right? cuMemcpyDtoD is defined there.

5

It is better,cuMemcpyDtoD is available with the #include you proposed,…but I have still the problem of type…(see message below)
My tables are unsigned int, and I don’ t know if and how I can
use the type CUdeviceptr to build the table correctly.

Thanks for your help.
Regards.

“c:\Program Files\NVIDIA Corporation\NVIDIA CUDA SDK\projects\sabr2\sabrKernelMT.cu”, line 83: error:
1> argument of type "unsigned int " is incompatible with parameter of
1> type “CUdeviceptr”
1> cuMemcpyDtoD(local_tableau_mots,&d_tableau_mots[index_premier_k+624tid],624*sizeof(unsigned int));

6

Is the solution to create the table with cuMemAlloc which returns the pointer?
But I am still troubled by the fact there will be no definition of the type of elements in the table, and the way to get one element.
Thanks.

7

Ok, from the sound of it, you actually want to be using the Runtime API. See appendix D in the programmer’s guide. Specifically, you should be using cudaMemcpy() in appendix D.5.10.

8

if I use cudaMemcpy in this global function, I have the error:

c:\Program Files\NVIDIA Corporation\NVIDIA CUDA SDK\projects\sabr2\sabrKernelMT.cu", line 88: error:
1> calling a host function from a device/global function is
1> only allowed in device emulation mode

I tried with the include of
#include “cuda.h”
#include “cuda_runtime_api.h”
#include “cuda_runtime.h”

but nothing better.

So, it seems not possible to use RunTime API in a global fucntion.

Only a solution with Driver API seems to be possible, with cuMemcpyDtoD or example.
Am I right, or what did I forgot?
Thanks in advance.

9

Ok, now we come back to my original point: You can’t call cudaMemcpy (or cuMemcpyDtoD) from a global function. Only from functions running on the CPU.

10

Are any functions that will do CUDA memcpy and memset operations inside global or device functions.

Thanks.