Make it parallel
Most machines these days have multiple cores, some even support hyper threading. If you want your code to be executed as fast as possible, try to use all CPU cores all the time.
- Split your code into tasks and create DAG(‘s). (oneTBB: flow::graph)
- Try to parallelize the code inside those tasks. (oneTBB: parallel_for)
Those mechanisms should be operated by the same schedular based on the same thread pool, which should match the amount of CPU cores of the system.
Reduce the locks
To prevent race conditions/crashes, you possibly have to modify code to ensure safety on parallel access to shared resources. Try to apply the following points for your changes.
- Reduce the amount and size of locks.
- Use atomics instead of explicit locks.
- Use lock free data structures. (Check out oneTBB)1
1 e.g.: concurrent_vector and concurrent_unordered_map
And even more …
- Use even the GPU for suitable tasks, if available. (See CUDA and ROCm)