Direct rendering is already present using PBO (GL_EXT_pixel_buffer_object) and
persistent buffers (GL_EXT_buffer_storage). But these extensions are available
only since OpenGL 4.4.

For older OpenGL versions (OpenGL 2.0 and OpenGL ES3 (or as an extension since
OpenGL ES2)), we can still improve quite a bit the performances by using a
non-mapped PBO and updating it via glBufferSubData. This won't be true direct
rendering since you must copy the data from the decoded picture to the GPU.

By using 2 PBOs (double buffering), we make sure that we can upload to the GPU
while the previous picture is beeing rendered.

Performances gains measured by playing 4K30fps video during 1 minute: 7 seconds
spent in vout->prepare() instead of 10 seconds. This result may vary with the
platform/device/drivers.

Thanks to sesse for the advices.