Being an engine programmer usually means being a bit of a jack of all trades. There’s always something weird going on and you have to be pretty familiar with a bunch of low level details that come in handy in unexpected ways. Recently I went down a somewhat unexpected rabbit hole where those skills came in extremely hand. In an effort to blog more and also because it seems like I was the first to run into this issue, I figured I should sit down and just write about it so future people can benefit from it.
Almost a year ago, with macOS 10.15, Apple finally added GPU timer queries to Metal. The documentation for it is incredibly lacklustre to this day, saying the following about [MTLCounterResultTimestamp timestamp]:
A timestamp.
That’s pretty useless, there’s no unit attached to it!
I actually put down GPU timers on Metal because they were pretty damn broken in general, besides their absolute lack of documentation. For example, they would just crash on Intel. And then they would crash on AMD hardware. It wasn’t until the release of macOS 10.15.5 a few days before Apple announced macOS 11, that GPU timers were actually fixed. The documentation, however, was not updated.
If you are working with Vulkan, chances are that at some point you’ll run into a VK_ERROR_DEVICE_LOST error. It’s the worst kind of error, chances are that your GPU choked on some data sometime about a frame or two ago and the position where you received the error is nowhere near where the GPU actually decided to throw up its hands and give up. This is of course because GPUs and CPUs are inherently decoupled from each other, and when you submit your work from the CPU to the GPU, the GPU will start crunching your numbers while in the meantime the CPU goes on with its busy life doing other things. Now, maybe while crunching your numbers the GPU encountered a page fault or did some other computation where it just couldn’t proceed anymore afterwards. Or maybe it took too long to compute results and the kernel watchdog decided that enough was enough and restarted the device. Or something completely different! Either way, at some point this error will have propagated all the way back to your application in userspace and you’ll have no idea why it happened. And are only left to guess about what went wrong and where. Well, Nvidia has your back with the extremely verbosely named VK_NV_device_diagnostic_checkpoints extension, which is a lot like Nvidias Aftermath for DirectX, except it works on Vulkan. And because for some reason nobody on the internet seems to sing its praise, I will do so now in the form of this blog post!
Seems like everybody is excited about the fake interior mapping in Spiderman these days. Here is a Kotaku video showing them off:
Now, the technique itself is nothing new, it’s been proposed by Joost van Dongen who published a paper about it here. The idea is to divide the space into evenly sized “rooms” and then cast a ray in the fragment shader to figure out whether the ceiling/floor or a wall was hit, and then map a texture on top. It’s super cool because it works without requiring extra geometry, the fragment shader is relatively cheap and easy to implement and like the Kotaku reviewer mentioned; The fact that the window is no longer just flat or a mirror adds depth (hah) to the world.
Let’s say that you occasionally look at your application with NSight, VTune, or any other profiling tool of choice. Naturally you want to add debug markers into your application, but you might not necessarily ship with them or have them run at all when no profiling tool is attached. You could put them behind a command line flag, but I prefer automatic discovery: One build, when run with NSight, having all the debug markers I need to dissect a frame, and when run without NSight not doing any of that overhead. With X-Plane, there is another layer to that, and that is that specifically for NSight some features have to be disabled since they are making unsupported GL calls.