bypass only works with DMABufBuffer

[AlanGriffiths]

It would be more useful if it worked with ShmBuffer (e.g. glmark2-wayland uses shm)

            auto dmabuf_image = dynamic_cast<mg::DMABufBuffer*>(bypass_buffer->native_buffer_base());
            if (dmabuf_image &&
                bypass_buffer->size() == surface.size())

Is there a technical limitation here? Or just legacy?

[RAOF]

Kind of a bit of both?

DMABufBuffers are easy to bypass with - they're already handles to GPU memory, we just need to get the driver to associate a framebuffer handle with them in order to submit them to scanout.

ShmBuffers aren't in GPU memory. To scanout of them we'd need to create a DRM dumb buffer, map it, memcpy the content of the ShmBuffer into the dumb buffer, then get the driver to associate a fb handle with it.

This is absolutely possible, and it's probably still faster than GL compositing, but it's not a clear win - when using GL the driver can plausibly use GPU hardware resources for memcpy, will likely be able to handle cache invalidation better, and so on.