cudaHostalloc necessary?

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hi there,

If I have the following example code

float *u_d_array_host;

float *u_d_array_device;

u_d_array_host = (float *)cudaHostAlloc((u_d_array_length)*sizeof(float)); //???????

cudaMalloc((void **) &u_d_array_device,(u_d_array_length)*sizeof(float));

for(int j = 0;j<u_d_array_length;j++)                                      //??????????

    {

		u_d_array_host[j] = u_d_array[j];	

    }

cudaMemcpy(u_d_array_device,u_d_array_host,(u_d_array_length)*sizeof(float),cudaMemcpyHostToDevice);

Do I need the step over u_d_array_host or can I just do sth like

float *u_d_array_device;

cudaMalloc((void **) &u_d_array_device,(u_d_array_length)*sizeof(float));

cudaMemcpy(u_d_array_device,u_d_array,(u_d_array_length)*sizeof(float),cudaMemcpyHostToDevice);

I know that my u_d_array has e.g 10 values that are in line, does the step over u_d_array_host offer me any benefits speed wise ???

thanks in advance

#2

Hi Markuss,

The u_d_array_host is not necessary, the second piece of code does work.

But page-locked memory will have good performance in some cases.

Best regards!