I'm wondering things like, how are parallel GPU workloads scheduled? Is it better to use 1 process and batch GPU calls? Looking for material that will help me better understand this topic.
I can provide more implementation details if necessary, thanks!
(currently using OpenGL, exporting each frame with `glReadPixels `)
In terms of actual scheduling, it's kind of complicated, but for the most part the APIs require that draws happen in-order. Don't really expect to get many savings out of multi-threading your render loop in OpenGL, so that means we move to more standard game optimizations. Memory bandwidth is likely to be your biggest bottleneck, so make sure to do any buffer transfers at the start of the frame or scene, and definitely don't touch them between draws. Do a clear at the start of a frame (GPUs can do "fast clears" which lets the ROP unit simply overwrite it instead of having to pull in the contents of the previous frame). Use modern GL best practices: that means uniform buffer objects, since the classic uniform setters will do transfers mid-frame. Use a good profiler for your platform to determine whether any bottlenecks or slowness is on the CPU in your code, on the CPU in the driver, or on the GPU. If it's on the CPU, consider moving to a modern API like D3D12, Metal or Vulkan, but expect that to take some work.
TensorFire
On a typical Windows PC, pretty simple. OS just throws tasks at hardware, if they take more than a couple of seconds to complete, the OS concludes the GPU has hang and resets the driver.
Virtualization-targeted GPUs have real schedulers, but such systems are rare.
> exporting each frame with `glReadPixels`
I recommend researching how to stop doing that and switch to the hardware video encoder (intel quick sync, nvidia nvenc, amd has one, too). Unfortunately for you, integrating them with OpenGL can be very hard. Quite easy for Direct3D BTW, because media foundation.