Simple memcpy not working

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello all,

I am trying to do a simple memcpy with Cuda 1.1 on a Windows XP machine with a GeForce 8600 GTS board. I am compiling under MS Visual Studio 2005 and I have used the template application for a test.

Here is the code for the host:

/* [2007.12.29 Phil Pratt-Szeliga] : CUDA */

#include <cutil.h>

#include "gpu_kernel.cu"

double tilted_angle[513] = {0.0};

main (int argc, char *argv[])

{

	CUT_DEVICE_INIT();

          

	dim3  grid( 1, 1, 1);

	dim3  threads( 1, 1, 1);

	double *d_tilted_angle;

	CUDA_SAFE_CALL( cudaMalloc( (void**) &d_tilted_angle, sizeof(tilted_angle)));

	CUDA_SAFE_CALL( cudaMemcpy( d_tilted_angle, tilted_angle, sizeof(tilted_angle),cudaMemcpyHostToDevice) );

	slice_kernel<<<grid, threads>>>(d_tilted_angle);

	CUT_CHECK_ERROR("Kernel execution failed");

	CUDA_SAFE_CALL(cudaThreadSynchronize());

	CUT_CHECK_ERROR("Kernel execution failed");

	//copy the memory back

	CUDA_SAFE_CALL( cudaMemcpy( tilted_angle, d_tilted_angle, sizeof(tilted_angle),cudaMemcpyDeviceToHost) );

	//free the allocated memory

	CUDA_SAFE_CALL(cudaFree(d_tilted_angle));

}

And here is the code for the device:

#ifndef _TEMPLATE_KERNEL_H_

#define _TEMPLATE_KERNEL_H_

#include <stdio.h>

__global__ void slice_kernel(double * tilted_angle)

{

	for(int i = 0; i < 513; i++){

  tilted_angle[i] = 15.5;

	}

	return;

}

#endif // #ifndef _TEMPLATE_KERNEL_H_

Here is my problem: (In Debug mode) Before I call the cudaMemcpy (device to host) tilted_angle on the host is all zeros. After I call cudaMemcpy tilted_angle takes on values like: 5.426734841397e-315#DEN

However, in EmuDebug mode tilted_angle takes on values like 15.5 in main.

Does anyone know what is going on?

Thanks

Phil

2

You are using doubles, the current GPU generation is single-precision only.
The current CUDA compiler will squash doubles to floats. Convert the array before sending to the GPU.

3

Okay thanks

4

Is double supported on host functions? Where is the documentation related to this?

5

host code is compiled by your host compiler, so yes. I believe it is in the programming manual, or otherwise the documentation of nvcc