Вроде как неплохие изменения в последнем билде от кд на 3000 строк.
"This PR essentially does 2 things:
Refactors FIFO processing into something more readable, including separation of the main loop from the flattener.
Significantly reworks how draw calls are processed. This part is still very unfinished, but the idea is to completely ignore the FIFO begin/end commands and use our own mechanisms to decide where to mark the scope of draw calls. The reorganization and optimization here results in quite a good fps boost in applications that are limited by RSX performance.
Other minor improvements include:
Fixes some minor bug in the surface_store that could return an empty set when asked to get an overlapping list with at least one confirmed hit.
Make the FIFO flattener engage dynamically depending on load to avoid slowdown when the extra processing overhead is not required.
At some point late in development, I scrapped the older system with the prebuffered queues and optimization passes. The performance hit of FIFO preprocessing was quite high and the prebuffering caused race conditions in poorly synchronized applications. This should be resolved in the final version, which has greatly simplified FIFO processing.
So, what does this all mean? Well, you could potentially get a ~10% boost in applications with an RSX bottleneck. This may not mean much for some applications, but the gains are significant at high draw call counts. Most other applications may only show a few fps gain."
