Multithreaded CPU Rendering and Fluid Sim

As I was done with the fluid simulation and rendering, I was going to move on to working on gameplay. Its workload was mostly going to be on CPU so I looked into CPU multithreading to take advantage of modern gaming CPU with core count of 6-12. 

I came across multiple GDC talks Parallelizing the Naughty Dog Engine Using Fibers - YouTube Destiny's Multithreaded Rendering Architecture - YouTube. I decided that rather than assigning a thread to each task (rendering thread, physics thread, audio thread, etc...), that I should use a method similar to thread pooling. I should have multiple general worker threads corresponding to the number of physical core count on the CPU that are spawned only once at the start of the program. 

1. Multiple worker threads are spawned (same as physical core count on the CPU). Each worker thread is bound to each physical CPU core.
2. Now repeat the following.
   1.  Jobs are submitted to a queue if its dependency counter is 0.
   2.  An idle worker thread grabs a job at the front of the queue then executes it.  

Going parallel means synchronization between workloads. Dependency graph/counter is used such that a job is only submitted to the queue once its dependency counter is set to 0

I know hyperthreading is enabled on most CPUs which are not ideal for game engine workloads in general. For now, I set thread affinity to every other CPU thread (SetThreadAffinityMask on Windows).

I parallelized the fluid and rendering workload by splitting up the workload into 8 chunks, which seemed like a reasonable workload distribution considering processor count on modern CPUs. I also tried to keep each job to granularity of about 50 to 200 microseconds.

This is a timeline graph of 2 frames I captured with Tracy profiler, with each row corresponding a thread.

Reduced execution time from 20ms (single-threaded) to 7ms (eight threads on Ryzen 7800X3D).

This is 256x256 Grid/Particles running in 60fps. I was barely able to run 128x128 in 30fps with a single thread.