Lighthugger 🚀

A rendering of the San Miguel 2.0 scene from the McGuire Computer Graphics Archive. CC BY 3.0.

• A modern Vulkan 1.3 renderer
• Fully bindless with extensive use of Buffer Device Address (BDA), buffers are never bound.
• Meshes are split up into (via meshoptimizer)
• Instances are culled and a single-pass prefix sum over the number of meshlets in each instance is computed using a 64-bit atomic.
• A per-meshlet indirect dispatch is run to further cull meshlets, essentially emulating mesh shaders in compute.
• Triangles are rasterized into a visibility buffer, and lighting for the whole screen is resolved in a single compute pass.
• Only block-compressed .DDS and .KTX2 textures are supported for extemely fast load times.
• Min and Max depth values are computed each frame to tightly bind the cascaded shadowmap frustums.
• Written in C++20 and Vulkan-Hpp.
• GLSL shaders (I'd use HLSL if it had 8-bit int support and if atomics worked on unstructured buffers)

MIT license available upon request.

Clockwise from top-left: Rendered scene, visibility buffer triangle IDs, visibility buffer meshlet IDs, shadow cascades

Rendered view showing all 4 shadowmap cascades. Not that objects that fit in smaller cascades are culled from the larger ones.

This project started out with my writing a C++ Vulkan starter project. That code is available on the starter branch.