Omg, optimization process drive me crazy, haven't slept very long time. Shader compiler behave very strange when code is too much complex for it and i wasted several days only to experiment with different positions of parts of arithmetic, grouping it with other operations, measure performance and check assembly. Completely trials and errors method, my own optimization not work in many cases, simply broken by invalid compilation. Fortunately i found combination when less registers used (from 31 reduced to 25) and at same time it not generate additional useless operations. Also found one interesting thing when performance is much higher even if code is the same, but have several instructions added, of course it's NVidia specific, so i tried to find such areas and optimized at least supersampling mode of ambient occlusions. I don't know what is that exactly, only NVidia engineers can say for sure. Also did another code which reduce triangulation of ssao, you may see it when sampling range is low and camera is near player body, especially legs or some other low polygonal surfaces. This "detriangulation" also makes ssao more clear, it less darken for mostly flat surfaces which have normal maps. Spoilers in editor are closed, except few i thought important. I didn't separate window for save/load in weather properties, may be later, need to move all windows for that, write code for keeping their positions, etc. Fixed some bugs in ssao/ssil code and of course optimized it more.
Gabber
Try new release which i publish soon. Can't give you test files now, i'm very tired atm.
|
