djoser
9ea85db5ce
- Hugo static site at root: config, content (home/about/resume/blog + 5 projects), data/resume.yaml, custom layouts (gpu background, resume renderer, projects list), dark theme with canvas behind content - gpu-bg/: extracted WebGPU/WebGL2 spiral engine as a Vite sub-package; slim bootstrap drops router/overlay/text renderer; fixed uniforms (scene0/dim0); self-contained engine.js - archive/: retired Vite SPA preserved (router, overlay, data, engine/text orphaned) - 3-stage Dockerfile (node -> hugo extended/dart-sass -> nginx), static nginx config - DEPLOYMENT.md handover + README rewrite; pre-hugo rollback tag
768 B
+++ title = "GPU-Driven Point Cloud Renderer" description = "TUM — 140M+ points at ~100 FPS" date = "2024-01-01" period = "2024" tags = ["C++", "Vulkan", "GLSL"] categories = ["programming"] +++
Implemented a real-time GPU-driven renderer for large point clouds (140M+ points at ~100 FPS) using a custom 5-stage compute shader pipeline in Vulkan, based on Schütz et al. (2021, 2022).
Created a software rasterization pipeline using atomic uint64 depth-color packing and compute shaders, enabling fine-grained depth testing without a hardware depth buffer.
Implemented batch-based frustum culling, Morton-code spatial sorting, and screen-coverage LOD selection via 10-bit quantized position precision. Used Vulkan subgroup extensions for warp-level optimizations.