CUDA Globel variables [duplicate]

This question already has an answer here: Creating global variables in CUDA (1 answer) Closed 7 y开发者_运维技巧ears ago.

How can I create global variables in CUDA?

__device__ float *devD;
cudaMalloc((void**)&devD, s);
calculateDT_T2B<<<dimGrid, dimBlock>>>();
cudaMemcpy(dtr, devD, s, cudaMemcpyDeviceToHost);
print(dtr);

It doesnot give the correct answer (gives some random numbers). But when I call

calculateDT_T2B<<<dimGrid, dimBlock>>>(devD); instead of
calculateDT_T2B<<<dimGrid, dimBlock>>>();

It gives the correct answer.. why?

---

You cannot directly use cudaMalloc to allocate onto a __device__ symbol in GPU memory. When you do so, you are allocating only in host memory. See my answer to your own, almost identical question which you posted within a minute of this one. The short version is to use cudaMemcpyToSymbol to write a dynamically allocated device pointer onto a statically declared symbol.