VeloXel

VeloXel is a native C++20 image/frame-processing engine built around a rigorous optimization story: start from a clean scalar baseline, then earn every speedup — tiled multithreading, then hand-written SIMD — with benchmarks proving each step.

The engine applies a fixed color-grading chain (ColorGrade → 3D LUT → BoxBlur → Gamma) to float-RGBA frames, and ships VeloXel Studio, a DaVinci-Resolve-styled grading GUI on top.

Frames live in 64-byte-aligned flat float buffers. Every processing stage implements a tile interface with an explicit aliasing contract separating in-place point ops from spatial ops that need scratch memory. A pipeline runs the chain either as a scalar reference or fanned out over 128×128 tiles through a custom condition-variable thread pool — with byte-identical output guaranteed between the two.

Correctness is enforced by tooling: an image-diff utility holds every optimized path to within one 8-bit step of the scalar reference.

Per-architecture kernels are hand-vectorized: NEON on arm64 using structured de-interleaving loads, AVX2 + FMA on x86 with in-register interleaving and gathers — selected by runtime CPU detection, never coexisting in one binary.

The flagship trick: a vectorized polynomial pow for gamma correction (exp2(k·log2 x) with a degree-5 minimax polynomial) that beats the libm call the auto-vectorizer can't touch — 4× on that kernel alone. The 3D LUT stage does gather-based trilinear interpolation with Resolve/Adobe .cube parsing.

VeloXel Studio is an ImGui + GLFW/OpenGL app styled after DaVinci Resolve: Lift/Gamma/Gain color wheels, a node graph with live per-stage thumbnails, RGB parade and histogram scopes, drag-and-drop images and .cube LUTs, and live switches for SIMD level and tile size.

Rendering runs on a worker thread with a latest-wins mailbox, so dragging a color wheel re-renders continuously without ever freezing the UI thread.