[cfe-dev] Operator new in CUDA kernels

[Ralph Kube via cfe-dev]

Hi,
I’m trying to port my CUDA project to clang. Thanks for the fantastic work,
clang is a charm to work with.

One problem however, maybe you guys can help me out. I already posted this
to cfe-users but no answer so far.

I need to allocate memory for a class in a CUDA kernel on the heap. 
So I created a myclass** in the host program, pass it to the kernel and attempt to
allocate with new. Works fine in nvcc, in clang I get the error

[1] % clang++ -std=c++14 -o test_new_device test_new_device.cu -L/Developer/NVIDIA/CUDA-8.0/lib -lcudart
ptxas fatal   : Unresolved extern function '_Znwm'
clang-4.0: error: ptxas command failed with exit code 255 (use -v to see invocation)

I’m on osx 10.12.5, using
% clang++ --version
clang version 4.0.0 (tags/RELEASE_400/final 297808)
Target: x86_64-apple-darwin16.6.0
Thread model: posix

So, is there a way to tell clang that ::operator new() in device function should call
the device version of malloc. As described here:
http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#memory-allocation-and-lifetime


Below is a minimal program that reproduces the error.

Thanks,
Ralph


#include <iostream>
#include <cuda_runtime_api.h>

class myclass
{
public:
   __host__ __device__ myclass(const double _data) : data(_data) {}
   __host__ __device__ ~myclass() 
   {   
       printf("Deleting myclass\n");
   }   
   __host__ __device__ double get_data() const {return(data);}

private:
   const double data;
};



__global__
void init_myclass(myclass** mycs_ptr)
{
   (*mycs_ptr) = new myclass(14.0);
}


__global__
void access_myclass(myclass** mycs_ptr)
{
   printf("I am using data with value = %f\n", (*mycs_ptr) -> get_data());
}

__global__
void delete_myclass(myclass** mycs_ptr)
{
   delete (*mycs_ptr);
}


int main(void)
{
   myclass** myclass_ptr{nullptr};


   init_myclass<<<1, 1>>>(myclass_ptr);

   access_myclass<<<1, 1>>>(myclass_ptr);

   delete_myclass<<<1, 1>>>(myclass_ptr);

   return(0);
}