device arrays and host

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

I have recently started doing some project on CUDA.
Your help will be highly appreciated.

I am looking for simpler ways to initialize arrays and define arrays in the device. Most of the examples(matrix, template etc) use runtime APIs like CUDAMALLOC and CUDAMEMCPY. What I want is static initialization to be done by the device.

Example: I define two device variables g_idata, g_odata;

template.cu

device float g_idata[32];
float h_idata[32];

I am using the CudaMEMCPY function to read from the device array to the the host.

// copy host memory to device
//Line 97:
CUDA_SAFE_CALL( cudaMemcpy( g_idata, h_idata, mem_size,
cudaMemcpyHostToDevice) );

But I am getting a runtime error when I execute this code.

Cuda error in file ‘template.cu’ in line 97 : invalid device pointer.

I am assuming that the compiler knows that g_idata is a device pointer.
Can the host directly use this pointer in its function calls to do the memcpy thing ?
Am I missing something else ?

Regds, JarJar

2

To copy data from statically allocated memory, you have to use cudaMemcpyFromSymbol().

This function is most commonly used on constant memory, which has to be statically allocated, but the Reference Manual says that it also works for global, i.e. device, memory as well.

3

Thanks. Let me try the approach and see how it works ?