cudaMemcpyToSymbolAsync

Accelerated Computing CUDA CUDA Programming and Performance
#1

I am trying to use cudaMemCpyToSymbolAsync, but must be passing parameters incorrectly. In the function simpleTest, cudaMemCpyToSymbol appears to execute, but cudaMemCpyToSymbolAsync gives error value 11 – invalid argument. What am I doing wrong?

My “big picture” objective is to do calculations in a kernel in stream[1]. When a certain time period passes, I will use stream[2] to upload a flag to the device. stream[1] will poll this variable periodically. When the flag is set, stream[1] will terminate (after doing some finalizing steps). Is there a better way to do this? What is more efficient for polling – global, constant, or pinned memory?

unsigned int h_testVar[2];
device unsigned int d_testVar[2];

void chkCudaReturn(cudaError_t err, unsigned int myErrLoc)
{
if (!err == cudaSuccess)
{
printf("\a\a\n***ERROR CUDA ERROR %u\n", myErrLoc);
printf(“Error Val %u\n”,err);
printf(cudaGetErrorString(err));
}
}

void simpleTest()
{
h_testVar[0] = 100;
h_testVar[1] = 200;
cudaStream_t stream[2];
for (int i = 0; i < 2; ++i) cudaStreamCreate(&stream[i]);
//Good One
chkCudaReturn(cudaMemcpyToSymbol (d_testVar, h_testVar, sizeof(h_testVar), 0, cudaMemcpyHostToDevice),1);
//Bad One
chkCudaReturn(cudaMemcpyToSymbolAsync (d_testVar, h_testVar, sizeof(h_testVar), 0, cudaMemcpyHostToDevice, stream[1]),2);
chkCudaReturn(cudaGetLastError(),3);

}

#2

I got the same problem using CUDA 3.2 and the 263.06 Driver on Windows XP!

The Function will only work with stream = 0 means not asynchron.

#3

Hi,
I have the same problem :/
Do you know if it’s fixed in cuda 4.0 ??

Thanks,
Greg

#4

Hello again,
here’s the news. You don’t need to upgrade to 4.0RC to fix it.
Just install the newest driver (the one that comes with 4.0RC). I believe I have 270.28 for notebooks right now.
It solved the problem and my streams work fine, now.

Hope it’ll help you guys.
Regards,
Greg.

#5

Hello,

The synchronous call (cudaMemcpyToSymbol) works because host memory doesn’t need to be page-locked. However, in the asynchronous call (cudaMemcpyToSymbolAsync), host memory must be page-locked. Thus, h_testVar must be allocated with cudaMallocHost() or cudaHostAlloc() before using it. For example:

int *h_testVar = cudaMallocHost((void**)&h_testVar, sizeof(int) * 2);

h_testVar[0] = 100;

  h_testVar[1] = 200;

cudaMemcpyToSymbolAsync (d_testVar, h_testVar, sizeof(int) * 2, 0, cudaMemcpyHostToDevice );

Regards,

Carlos.