Compile error on "gridsample_grad2"

[Burningdust21]

Hi, thanks for your great work! I try to run the code on a linux server and came across compile error when compiling the CUDA source. The complete error log can be found here, some critical errors are:

/root/.cache/torch_extensions/gridsample_grad2/f6f3395827c77b6f68bed38af0c94ab3-tesla-v100-sxm2-32gb/gridsample_cuda.cu(560): error: too many arguments in function call

/root/.cache/torch_extensions/gridsample_grad2/f6f3395827c77b6f68bed38af0c94ab3-tesla-v100-sxm2-32gb/gridsample_cuda.cu(189): error: no instance of function template "at::native::safe_add_2d" matches the argument list
            argument types are: (double *, int, int, int, int, int, int, double, int, const int)
          detected during instantiation of "void at::native::<unnamed>::grid_sampler_2d_grad2_kernel(index_t, at::cuda::detail::TensorInfo<scalar_t, index_t>, at::cuda::detail::TensorInfo<scalar_t, index_t>, at::cuda::detail::TensorInfo<scalar_t, index_t>, at::cuda::detail::TensorInfo<scalar_t, index_t>, at::cuda::detail::TensorInfo<scalar_t, index_t>, at::cuda::detail::TensorInfo<scalar_t, index_t>, at::cuda::detail::TensorInfo<scalar_t, index_t>, at::cuda::detail::TensorInfo<scalar_t, index_t>, at::native::detail::GridSamplerPadding, __nv_bool, index_t) [with scalar_t=double, index_t=int]" 
(560): here

/root/.cache/torch_extensions/gridsample_grad2/f6f3395827c77b6f68bed38af0c94ab3-tesla-v100-sxm2-32gb/gridsample_cuda.cu(474): error: no instance of function template "at::native::safe_add_3d" matches the argument list
            argument types are: (double *, int64_t, int64_t, int64_t, int64_t, int64_t, int64_t, int64_t, int64_t, int64_t, double, int64_t, const int64_t)
          detected during instantiation of "void at::native::<unnamed>::grid_sampler_3d_grad2_kernel(index_t, at::cuda::detail::TensorInfo<scalar_t, index_t>, at::cuda::detail::TensorInfo<scalar_t, index_t>, at::cuda::detail::TensorInfo<scalar_t, index_t>, at::cuda::detail::TensorInfo<scalar_t, index_t>, at::cuda::detail::TensorInfo<scalar_t, index_t>, at::cuda::detail::TensorInfo<scalar_t, index_t>, at::cuda::detail::TensorInfo<scalar_t, index_t>, at::cuda::detail::TensorInfo<scalar_t, index_t>, at::native::detail::GridSamplerPadding, __nv_bool, index_t) [with scalar_t=double, index_t=int64_t]" 
(627): here

It seems the complier cannot found the correct function definition of safe_add_2d, safe_add_3d and AT_DISPATCH_FLOATING_TYPES_AND_HALF.

My environment is torch 1.8.0 with CUDA 11.1, Ubuntu 20.04.

Do you have any idea what maybe the problem? Thanks!

[cChrister]

I have the same problem as you, do you fix it?

[StardustLu]

It seems to be the CUDA version.
I came across the same problem when compiling it with CUDA 11.1.
Then I downgrade to 11.0 or upgrade to 11.3. It works.