Patrick Baggett      2011-02-16 22:58:33 PST

This enhancement is for both x86/x64 Windows.

The SDL implementation of mutexes uses the Win32 API interprocess
synchronization primitive called a "Mutex". This implementation is subpar
because it has a much higher overhead than an intraprocess mutex. The exact
technical details are below, but my tests have shown that for reasonably high
contention (10 threads on 4 physical cores), it has 13x higher overhead than
the Win32 CriticalSection API.

If this enhancement is accepted, I will write a patch to implement SDL mutexes
using the critical section API, which should dramatically reduce overhead and
improve scalability.


-- Tech details --
Normally, Win32 Mutexes are used across process boundaries to synchronize
separate processes. In order to lock or unlock them, a user->kernel space
transition is necessary, even in the uncontented case on a single CPU machine.
Win32 CriticalSection objects can only be used within the same process virtual
address space and thus to lock one, does not require a user->kernel space
transition for the uncontended case, and additionally may spin a short while
before going into kernel wait. This small spin allows a thread to obtain the
lock if the mutex is released shortly after the thread starts spinning, in
effect bypassing the overhead of user->kernel space transition which has higher
overhead than the spinning itself.