Can't copy device memory to host memory

Accelerated Computing CUDA CUDA Programming and Performance
1

Hey guys,

I made my first tries with CUDA. Now I have a little problem with a really simple application. It should just fill an array with some numbers.

[codebox]

#include <cuda.h>

#include <cutil.h>

#include <stdio.h>

#define N 100

#define BLOCK_SIZE 8

#define BLOCK_COUNT N / BLOCK_SIZE

typedef struct {

int size;

float* elements;

} test_t;

global void testCudaMethod( test_t * t ) {

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

if( x < t->size ) {

	t->elements[ x ] = (float)x;

}

}

int main( int argc, char ** argv ) {

test_t host;

test_t device;

size_t size = sizeof( float ) * N;

host.size = device.size = N;

host.elements = (float*)malloc( size );

CUDA_SAFE_CALL( cudaSetDevice( 0 ) );

CUDA_SAFE_CALL( cudaMalloc( (void**)&device.elements, size ) );

CUDA_SAFE_CALL( cudaMemset( device.elements, 0, size ) );

testCudaMethod<<< BLOCK_COUNT, BLOCK_SIZE>>>( &device );

CUDA_SAFE_CALL( cudaMemcpy( host.elements, device.elements, size, cudaMemcpyDeviceToHost ) );

CUDA_SAFE_CALL( cudaFree( device.elements ) );

for( int i = 0; i < host.size; i++ ) {

	printf( "-> %.1f\n", host.elements[ i ] );

}

free( host.elements );

return 0;

}[/codebox]

After compiling the posted code with “nvcc test.cu -o test -I~/NVIDIA_CUDA_SDK/common/inc” and executing the application, I always get the following error message: “Cuda error in file ‘test.cu’ in line 34 : unspecified launch failure”. That’s exactly the line where I try to copy the device memory to the host memory. I think that I made a stupid failure in this sample code but I can’t figure it out. Can you give me some advices?

Regards,

Tim

2

Hi!

&device is actually a pointer whose value is to be interpreted as an adress in host memory since device is a variable residing in host mem. But the kernel

__global__ void testCudaMethod( test_t * t ) {

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

	if( x < t->size ) {

		t->elements[ x ] = (float)x;

	}

}

will interpret the parameter t as an address in device mem and will therefore cause a segmentation fault.

Regards

Navier

3

Hey Navier,
thank you for your advice. Now it works as it should work :)

Regars,
Tim