Screenshot Thread

There have been a lot of updates from last video. The AO is now fast, like HELLA fast. The biggest part of the performance impact of any AO solution is the cache inefficiency. Takes ages to explain but basically, the higher the sample radius, the lower the fps. I just did a major update to ensure maximum efficiency. To put it in numbers:

GTASA without AO, GTX 970: 90-110 fps
GTASA without AO, GT 920: 50-70 fps
old implementation: 64 samples, GTX 970: 50-60 fps
old implementation: 150 samples, GT 920: 5-8 fps
new implementation: 64 samples: GTX 970: 85-100 fps
new implementation: 150 samples: GT 920: 45-55 fps

Incredible, right? And this in fullscreen with fairly high sample count and radius. Before, I could kill my GTX 970 with the AO shader, now I cannot force-drop the fps below 70, even with 255 samples and screen-size big radius. This is no hoax, as you can see in the video where I used a radius twice as big as Nvidia HBAO+ has and 72 samples (3x HBAO+ very high) with less fps impact. Not to forget that HBAO+ is only AO while this calculates Indirect Lighting along the way which at least doubles the GPU load. Using a special dither texture (took me half an hour to write down a 16x16 bayer matrix), the resulting pattern needs less blur steps in the bilateral postblur than the noisy result of usual pseudo-random generators.
Even if that's not enough, the user can render the AO/IL in half screen resolution to reduce the number of processed pixels by three quarters. Then the implementation runs with unnoticeable fps impact at any setting.