Simple Basic question, please answer.

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

One simple doubt is…

void hostFun()
{
float* d_A;
CUDA_SAFE_CALL(cudaMalloc((void**) &d_A, mem_size_A));
}

My hostFun() is a host function and d_A is a float pointer declared in host function. and
I’m trying to allocte the device memory for d_A using CUDA_SAFE_CALL(cudaMalloc( d_A ));

  • Is memory allocated for d_A is device memory or host memory?

  • If the above question answer is no, then how can I allocate device memory for a pointer ( which is declared in a host function ) sitting in host function?

2

your code should allocate on a CUDA-enabled device, if you have any, and I think it will allocate for the default device=0. YOu can specify which device with cudaSetDevice(deviceID)

what system do you have?

have you run the simpleDeviceQuery in the SDK to find what devices there are in your system?

3

Yes, I’m using NVIDIA QUADRO NVS 290 & G8600 GPU card.

and I’m able to run all the CUDA sample SDK examples.

BTW, I’m not declaring the pointer, float* d_A, as device pointer.

I mean: device float* d_A

so, how can it allocates device memory?

4

If you did not compile it in emulation mode.

then to use cudaMalloc will allocate memory on GPU global memory

5

cudaMalloc allocates device memory, and puts the address into a host variable (in your case d_A).

If you try to dereference the variable from host code (e.g. d_A[100]), you will get an error. The memory is allocated on the device.

You may pass the pointer when invoking a kernel and your kernel may use the device memory.