Hi ----- Original Message ----- > Hi, > > > The benchmarks are quite encouraging, since I get from +-25% for > > xonotic up to +-100% for glmark. (fwiw, vhost-user-gpu had similar > > results too). Finally, I tried to make it acceptable for upstream. > > Hmm, I don't feel like adding too many modes to virgl ... > > With vhost-user-gpu giving us simliar benefits I'd prefer to focus on > that, for the additional security benefits (sandboxing the separate > process).
Makes sense, I wanted to have a fair comparison between the two. vhost-user-gpu will probably take longer to happen, since it changes the way qemu must be managed. I also expect a lag in vhost-user whenever virtio ring/dataplane gets optimization in qemu, that will need to be adapted for vhost-user. So the benefit of having a virgl thread in qemu is that it is less complicated than vhost-user-gpu, it brings immediate performance boost, and it serves as a comparison point. I can keep it in a branch, but it would be nice to consider this as an "experimental" option too.