SSAO

Greatly reduced the VRAM usage Architectural changes:

Unit stencil texture

• All features and units model bounding boxes are rendered onto a downsized texture, that may be requested and reused by any subsequent pass.

Gbuffer Fusion pass

• Removed merging of misc tex buffers, saving a fullscreen buffers worth of vram and some gpu load
• Removed merging of map and model normals, saving another fullscreen buffer, since these were also queried only once per fragment in ssao shader
• In the gbuffer fusion shader, map model occlusion is indicated by a positive z in viewpos tex
• The entire gbuffer fusion shader can be turned off, directly sampling the map and model depth buffers, this is most useful when rendering ssao to a downsampled target

SSAO

The ssao shader samples multiple neighbouring texels when downsampling, to reduce shimmer in the downsampled target. Random hemispherical samples are now generated across a fibonacci sphere for even surface distribution Outputs are now:

• RG channel is viewpos xy normals, which is output as zero vector for ground fragments
• B channel is relative depth of fragment
• A is occlusion

Blur

• Blur shader is now designed to work in a single pair of horizontal and vertical passes, and the dir uniform controls wether it - - should perform its final mixdown
• Blur shader has correct offsets and weights
• Compares the neighbouring normals and depths when blurring
• Identifies samples that are outliers with no similar neighbours
• Blur shader reuses the SSAO texture in its second pass, saving more vram
• During final mixdown, occlusion is packed into alpha, and brightening (inverse occlusion) is passed into rgb

Compositing

• A different gl blending operation is used, where the ssao alpha darkens the original image, and the ssao color brightens the original.

Bloom

• Less VRAM usage
• Fixed shimmering on low qualities
• Uses Pixel Quad Message passing to massively reduce the number of passes and texture samples required
• Wider kernels, smoother looks
• Better performance

CAS

• Uses Pixel Quad Message passing to massively reduce the number of texture samples

Unit Stencil API

• For early bailing a lot of near-unit shading stuff. Useable on demand.

Thaks to @rlcevg for every bit of help. Thanks to @icexuick for tuning SSAO.