Neat, this is impressive! The first thing that comes to mind is that it's not very amenable to parallel encoding or decoding, even less so than PNG. The fact t

Neat, this is impressive! The first thing that comes to mind is that it's not very amenable to parallel encoding or decoding, even less so than PNG. The fact that QOI_OP_INDEX could reach arbitrarily far back basically eliminates parallel decompression. But I guess the simplest workaround for that would be to just break up your images into tiles that are each decoded in parallel. You could imagine a simple wrapper format that specifies how to reconstruct a larger image out of such tiles, which would be amenable to parallelism (albeit still not easily implementable on GPU, though I guess you could wring a little bit of parallelism out of it from by using one GPU SIMD lane per RGBA channel). Of course, it's entirely possible that your decompression goes so fast it doesn't matter that it's sequential and you're bottlenecked on I/O in any reasonable use case…