[FFmpeg-devel] Unable to compile with cuda

[Dennis Mungai]

On Sun, 23 May 2021 at 15:08, Dylan Fernando <dylanf123 at gmail.com> wrote:

> I got it to work, with --enable-cuda as well, using:
>
> PKG_CONFIG_PATH="/home/dylan/Files/nv-codec-headers" ./configure
> --enable-opencl --enable-vulkan --enable-libglslang --disable-stripping
> --enable-nonfree --enable-cuda --enable-cuda-nvcc
> --extra-cflags=-I/opt/local/cuda/include --nvccflags="-gencode
> arch=compute_52,code=sm_52 -O2"
>
> Thanks,
> Dylan
>
>

Dylan,

That has to be a very old build of FFmpeg. The option --enable-cuda has
been deprecated for months. Try building from a current release or git
master.
For the LLVM route, you  can also pass nvccflags  in this fashion:
--enable-cuda-llvm --nvccflags="--cuda-gpu-arch=sm_52 -O2".
These flags will differ based on your current GPU and the CUDA version in
use. See
https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#options-for-steering-gpu-code-generation
FFmpeg's ./cconfigure script picks a very old set of flags that are
deprecated in newer CUDA versions, hence the need to override these flags
with CUDA builds. See:
https://github.com/FFmpeg/FFmpeg/blob/master/configure#L4382
The difference between the build with and without the proprietary CUDA SDK
will be the inclusion of the scale_npp filters.
With the proprietary SDK, both the scale_npp and scale_cuda filters will be
available.
With the cuda llvm route, only scale_cuda will be available.
For more on their usage, see https://superuser.com/a/1650962/473795
And for more on compiling cuda with llvm, see the llvm wiki entry here:
https://llvm.org/docs/CompileCudaWithLLVM.html