More examples

[fxia22]

Hi pytorch team,

I am looking to port https://github.com/qassemoquab/stnbhwd to pytorch with effi. Do you know is it possible? Is the mechanism of writing extension for torch and pytorch similar or in other words, can I reuse some of the code from that repo. THanks.

[apaszke]

Yes, it should be quite easy to reuse it. You'd only need to copy over the C files and change the functions to accept tensors as arguments instead of parsing them out of the Lua state. Then just use the package example and that should be it.

[fxia22]

Awesome, thank you! I will let you know how things go.

[fxia22]

@apaszke I am trying to get the data from CudaTensor, I changed the example library to the following but that gives me seg fault:

int my_lib_add_forward_cuda(THCudaTensor *input1, THCudaTensor *input2,
               THCudaTensor *output)
{
  if (!THCudaTensor_isSameSizeAs(state, input1, input2))
    return 0;
  float * input_data = THCudaTensor_data(state, input1);
  printf("data %f\n", input_data[0]);
  THCudaTensor_resizeAs(state, output, input1);
  THCudaTensor_cadd(state, output, input1, 1.0, input2);
  return 1;
}

I need similar operations for spatial transformer network to work with CUDA (cpu version already works). Can you share with me how to do this extraction? Thanks in advance.

[fxia22]

I guess my question is about how to reuse cuda code. When I attempted to do so, it tells me threadId.x is not defined.

[soumith]

you cannot printf a cuda pointer, it will segfault. Maybe you can lightly read the CUDA programming guide: docs.nvidia.com/cuda/cuda-c-programming-guide

[apaszke]

Can't you just copy the code from the original repo? You shouldn't need to change any code that computes the function, only change the argument parsing.

[fxia22]

Thanks for your reply.

@apaszke Yes, I finished the CPU version porting and it was quite intuitive. And I read the CUDA programming guide. But how can I build a .cu extension with extension-ffi? I am able to use some torch CUDA functions like THCudaTensor_cadd, but how can I write my own CUDA functions?

For example, when I try to write on own add function, it gives me this error:

/home/fei/Development/extension-ffi/script/src/my_lib_cuda.c: In function ‘VecAdd’:
/home/fei/Development/extension-ffi/script/src/my_lib_cuda.c:9:64: error: ‘threadIdx’ undeclared (first use in this function)
 __global__ void VecAdd(float* A, float* B, float* C) { int i = threadIdx.x; C[i] = A[i] + B[i];  }
                                                                ^

[mattmacy]

@fxia22 torch.utils.ffi doesn't appear to have any knowledge of nvcc or .cu. I think you need to build your cuda sources separately (see the example Makefiles that come with the CUDA SDK) and then add the built object(s) to 'extra_objects' through kwargs when creating an extension.

See:
https://docs.python.org/3/distutils/apiref.html#distutils.core.Extension

[fxia22]

@mattmacy Thanks, I will give it a shot!