How to copy a device variable to host in CUDA [duplicate] - cuda

Question

I have problem with copying data from GPU to CPU. At the beginning I create variable in GPU space:
__device__ float gpu_array;
In this GPU function, I want to copy the data from od.fS[gi] (value = 0,43...) to gpu_array:
__global__ void Collide(.....){
....
//Streaming
od.fS[gi] = fi_S;
od.fN[gi] = fi_N;
od.fE[gi] = fi_E;
od.fW[gi] = fi_W;
gpu_array = od.fC[gi];
}
End here is function to copy data from GPU to CPU:
void showData(){
cudaDeviceSynchronize();
float* temp_array = NULL;
cudaMemcpyFromSymbol((void**)&temp_array, "gpu_array", sizeof(temp_array), 0, cudaMemcpyDeviceToHost);
float* cpu_array = (float*)malloc(sizeof(float));
cudaMemcpy(cpu_array, temp_array, sizeof(float), cudaMemcpyDeviceToHost);
printf("h_array: %f\n", *cpu_array);
}
In finally copies but wrong values (0.00000, -0.00000 etc.). What am I doing wrong?

Answer 1

There are multiple problems with your use of cudaMemcpyFromSymbol:
float* temp_array = NULL;
cudaMemcpyFromSymbol((void**)&temp_array, "gpu_array", sizeof(temp_array), 0, cudaMemcpyDeviceToHost);
you haven't allocated storage for data pointed to by temp_array
we don't use quotes around the device symbol name anymore
the destination pointer should be just a pointer, not the address of a pointer
the remaining code around cpu_array is completely unnecessary
Correct usage would be something like this:
void showData(){
cudaDeviceSynchronize();
float* temp_array = (float *)malloc(sizeof(float));
cudaMemcpyFromSymbol(temp_array, gpu_array, sizeof(float));
printf("h_array: %f\n", *temp_array);
}
Note that in this example, gpu_array is a scalar quantity, and so we could do something similar with temp_array:
void showData(){
cudaDeviceSynchronize();
float temp_array;
cudaMemcpyFromSymbol(&temp_array, gpu_array, sizeof(float));
printf("h_array: %f\n", temp_array);
}