Too many decode surface would waste GPU memory. Also it seems to be
introducing additional latency depending on stream. Since nvcodec
sdk version 9.0, CUVID parser API has been providing the minimum
required number of surface. By using it, we can save GPU memory
and reduce possible latency.
We weren't using the correct calling convention when calling CUDA and
CUVID APIs. `CUDAAPI` is `__stdcall` on Windows. This was working fine
on x64 because `__stdcall` is ignored and there's no special calling
convention. However, on x86, we need to use `__stdcall`.
By this commit, each codec has its own element factory so the
nvdec element factory is removed. Also, if there are more than one device,
additional nvdec element factory will be created per
device like nvh264device{device-id}dec, so that the element factory
can expose the exact capability of the device for the codec.
... and put them into new nvcodec plugin.
* nvcodec plugin
Now each nvenc and nvdec element is moved to be a part of nvcodec plugin
for better interoperability.
Additionally, cuda runtime API header dependencies
(i.e., cuda_runtime_api.h and cuda_gl_interop.h) are removed.
Note that cuda runtime APIs have prefix "cuda". Since 1.16 release with
Windows support, only "cuda.h" and "cudaGL.h" dependent symbols have
been used except for some defined types. However, those types could be
replaced with other types which were defined by "cuda.h".
* dynamic library loading
CUDA library will be opened with g_module_open() instead of build-time linking.
On Windows, nvcuda.dll is installed to system path by CUDA Toolkit
installer, and on *nix, user should ensure that libcuda.so.1 can be
loadable (i.e., via LD_LIBRARY_PATH or default dlopen path)
Therefore, NVIDIA_VIDEO_CODEC_SDK_PATH env build time dependency for Windows
is removed.
