On 06/04/2012 10:53 PM, Dave Airlie wrote: >> >> I've run these on various workloads and saw nothing worth mentioning. > > Nothing at all? no speedups, slowdowns, etc > > why should we merge all this code then :-) > > Dave. Preserving hardware state across batches is going to be necessary for: * Transform feedback in the presence of geometry shaders Right now, Mesa counts the number of primitives emitted on the CPU, relies on the kernel resetting the register to 0, and offset values in software. With GS, we obviously can't count primitives on the CPU. * Primitive restart in hardware and seems to be increasingly necessary for new features as we move forward. So we'd like to get it in place; we can cut more state uploads and tune Mesa further once it's there.
