How to pass device function as an input argument to host-side function?

Question

I just want to pass device function as argument of a host function, of cause, the host function then can launch some kernels with this device side function.

I tried the usual C++ way (pass by pointer/reference) and the CUDA debugger told me the kernel cannot launch.

Update:

What I want to do is:

__host__ void hostfunction(int a, int (*DeviceFunction)(int))
{
   /...do something.../
   somekernel<<<blocks, threads>>>(int * in, DeviceFunction);
}

And launch the host with:

hostfunction(x, &SomeDeviceFunctionTemplate<int>);

Your question is not entirely clear, at least to me. If you could post an example of what you tried, where the kernel did not launch, that might help. — Robert Crovella, Aug 28 '13 at 13:57
Since `__host__` functions can't take the addresses of `__device__` functions, you basically need to write a short `__global__` function that takes the address of the `__device__` function of interest, and then stores it to memory. Your `__host__` function can then read that function pointer from memory and then pass it to `somekernel`. — Jared Hoberock, Aug 29 '13 at 01:08

Robert Crovella · Answer 1 · 2013-08-29T21:34:48.280

2

This example might be of interest:

$ cat t237.cu
#include <stdio.h>


__device__ int f1(){ printf("dev f1\n"); return 0;}
__device__ int f2(){ printf("dev f2\n"); return 0;}
__device__ int f3(){ printf("dev f3\n"); return 0;}

__device__ int *fptrf1 = (int *)f1;
__device__ int *fptrf2 = (int *)f2;
__device__ int *fptrf3 = (int *)f3;


__global__ void mykernel(int (*fptr)()){

  fptr();
  printf("executed\n");
}

int main(){

  int *hf1, *hf2, *hf3;
  cudaMemcpyFromSymbol(&hf1, fptrf1, sizeof(int *));
  cudaMemcpyFromSymbol(&hf2, fptrf2, sizeof(int *));
  cudaMemcpyFromSymbol(&hf3, fptrf3, sizeof(int *));
  mykernel<<<1,1>>>((int (*)())hf1);
  cudaDeviceSynchronize();
  mykernel<<<1,1>>>((int (*)())hf2);
  cudaDeviceSynchronize();
  mykernel<<<1,1>>>((int (*)())hf3);
  cudaDeviceSynchronize();
  return 0;
}
$ nvcc -arch=sm_20 -O3 -o t237 t237.cu
$ ./t237
dev f1
executed
dev f2
executed
dev f3
executed
[bob@cluster1 misc]$

I think this is roughly along the lines of what Jared was suggesting. As he mentioned, this will not be possible in host code:

&SomeDeviceFunctionTemplate<int>

Assuming SomeDeviceFunctionTemplate refers to a __device__ function.

edited Aug 29 '13 at 21:34

answered Aug 29 '13 at 21:29

Robert Crovella

143,785
11
213
257

Can I adapt this somehow for templated code instead the code you have in `main()`? Or is there no way around explicitly defining the device-side pointer for each and every function I plan to use? – einpoklum Feb 13 '20 at 16:07
some examples linked [here](https://stackoverflow.com/questions/31057870/passing-host-function-as-a-function-pointer-in-global-or-device-function/31058123#31058123) use templating. – Robert Crovella Feb 13 '20 at 16:10
So, [this example](https://stackoverflow.com/a/26813443/1593077) replaces the need for a global pointer variable with the need for a global setup-kernel which copies the device function's pointer to a location of the launcher's choosing. That doesn't help much... reading on. There's the question about kernel addresses, which are usable. But how would I leverage that to let me get at the address of arbitrary device functions? – einpoklum Feb 13 '20 at 16:16

score 1 · Answer 2 · answered Aug 28 '13 at 14:37

It'd be helpful if you could post an example of what you are trying to do, but one thing to check is that you are compiling and running on Fermi (sm_20) or later since older GPUs did not support non-inlined function calls.

Check the compute capability of your device (needs 2.0 or later) and check your nvcc command line (needs -arch=sm_20 or later, or the -gencode equivalent).

How to pass device function as an input argument to host-side function?

2 Answers2

Linked