Week 9: what VA-API breaks that NVDEC did not
The hardware turned up this week, so I could finally read the VA-API interop properly instead of guessing at it. It gets its frames in a completely different way from the decoder I already have working.
Week 8: reshaping the branch, then running out of hardware
A rebase from five commits to nine, and then the discovery that my machine cannot test VA-API at all
Week 7: making a software interop
Software video was the one thing my interop system could not handle, and it was needed to have the same interface for both types (software and hardware).
Week 6: writing the frame sync into code
Last week I designed the handshake that keeps CUDA and the renderer from touching the same image at once. This week I wrote it, then spent most of my time making sure it was actually correct.
Week 5: syncing the NVDEC copy and the render
I had to make CUDA's copy and libplacebo's render take turns over one image, using a single timeline counter shared across both APIs with no CPU copy.
Week 4: the green noise that proved the obvious fix was wrong
Last week's validation error said my image can't be both disjoint and dedicated at once. The obvious fix was to drop the dedicated part, so I did, and the color fell apart into green noise across the sky and the water.
Week 3: first light on the Vulkan path
An NVDEC frame decoded on the GPU, imported into Vulkan with no CPU copy, and drawn by libplacebo. It works, the 4K file plays with correct colors. Then I turned on Vulkan's validation layer and it flagged the code as illegal, the same code NVIDIA had been drawing.
Week 2: NVDEC the easy way first
Before porting NVDEC to Vulkan, I got it playing zero-copy through OpenGL first, so I have a known-good reference to compare against. Then a long time reading my mentor's branch.
Week 1: first patch written
Week 1 of my GSoC with VideoLAN: a first patch that changes nothing on purpose, an assumption that fell apart once I read the source.
My GSoC 2026 Project Overview
Starting Google Summer of Code 2026 with VideoLAN, teaching VLC to draw GPU-decoded frames without copying them through the CPU first. The series opener: who I am and the round-trip I'm here to delete.