jp9000/obs-studio
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

libobs/util: Improve os_sleepto_ns on Windows

Browse Source 
Avoid nanosecond abstraction to reduce math operations.

Replace Sleep(0) with YieldProcessor(). We want the thread to remain
scheduled, the current CPU core to do less work, and the hyperthread
sibling to perform better.

Hacky profiling showed maybe 10-25 µs skid reduction per function call.
I think power/performance gains would be hard to measure, so I haven't
tried, but it would be shocking if they got worse.
This commit is contained in:
jpark37 2021-09-19 22:35:29 -07:00 committed by Jim
parent cf6e106939
commit e69f051736
1 changed files with 19 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

35
libobs/util/platform-windows.c
View File

@ -330,27 +330,30 @@ void os_cpu_usage_info_destroy(os_cpu_usage_info_t *info)
bool os_sleepto_ns(uint64_t time_target)
{
uint64_t t = os_gettime_ns();
uint32_t milliseconds;
const double freq = (double)get_clockfreq();
const LONGLONG count_target =
(LONGLONG)((double)time_target * freq / 1000000000.0);
if (t >= time_target)
return false;
LARGE_INTEGER count;
QueryPerformanceCounter(&count);
milliseconds = (uint32_t)((time_target - t) / 1000000);
if (milliseconds > 1)
Sleep(milliseconds - 1);
const bool stall = count.QuadPart < count_target;
if (stall) {
const DWORD milliseconds = (DWORD)(
((count_target - count.QuadPart) * 1000.0) / freq);
if (milliseconds > 1)
Sleep(milliseconds - 1);
for (;;) {
t = os_gettime_ns();
if (t >= time_target)
return true;
for (;;) {
QueryPerformanceCounter(&count);
if (count.QuadPart >= count_target)
break;
#if 0
Sleep(1);
#else
Sleep(0);
#endif
YieldProcessor();
}
}
return stall;
}
void os_sleep_ms(uint32_t duration)