Busy.

I didn’t forget to blog. I know you don’t believe me, but I’ve been accumulating items to blog about for the past month. Powering up. Preparing. And now, finally, it’s time to begin opening the valves.

Insanity Returns

When I got back from hibernation, I was immediately accosted by a developer I’d forgotten. One with whom I spent an amount of time consuming adult beverages at XDC again. One who walks with a perpetual glint of madness in his eyes, ready at the drop of a hat to tackle the nearest driver developer and begin raving about the benefits of supporting OpenCL.

Obviously I’m talking about Karol “HOW IS THE PUB ALREADY CLOSED IT’S ONLY 10:30???” Herbst.

I was minding my own business, fixing bugs and addressing perf issues when he assaulted me with a vicious nerdsnipe late one night in January. “Hey, why can’t I run DaVinci Resolve on Zink?” he casually asked me, knowing full well the ramifications of such a question.

I tried to put him off, but he persisted. “You know, RadeonSI supports all those features,” he said next, and my entire week was ruined. As everyone knows, Zink can only ever be compared to one driver, and the comparisons can’t be too uneven.

So it was that I started looking at the CL CTS for the first time this year to implement cl_khr_gl_sharing. This extension is basically EXT_external_objects for CL. It should “just work”. Right?

Right…

The thing is, this mechanism (on Linux) uses dmabufs. You know, that thing we all love because they make display servers go vroom. dmabufs allow sharing memory regions between processes through file descriptors. Or just within the same process. Anywhere, really. One side exports the memory object to the FD, and the other side imports it.

But that’s how normal people use dmabufs. 2D image import/export for display server usage. Or, occasionally, some crazy multi-process browser engine thing. But still 2D.

You know who uses dmabufs with all-the-Ds? OpenCL.

You know who doesn’t implement all-the-Ds? Any Vulkan drivers. Probably. Case in point, I had to hack it in for RADV before I could get CTS to pass and VVL to stop screaming at me.

From there, it turned out zink mostly supported everything already. A minor bugfix and some conditionals to enable raw buffer import/export, and it just works.

Brace yourselves, because this is the foundation for getting Cthulhu-level insane next time.