raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-29 06:16:55 +01:00
Code Issues Packages Projects Releases Wiki Activity

perf: limit high-resolution sleep busy-waiting

Browse Source
This commit is contained in:
nmzik
2026-08-07 17:44:55 +02:00
parent 7907a50817
commit 216c580d13
3 changed files with 16 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/common/threads.cpp
+5 -1
View File
@@ -29,6 +29,7 @@
#include <windows.h> // IWYU pragma: keep #include <windows.h> // IWYU pragma: keep
// IWYU pragma: no_include <winbase.h> // IWYU pragma: no_include <winbase.h>
constexpr DWORD KYTY_CS_SPIN_COUNT = 4000; constexpr DWORD KYTY_CS_SPIN_COUNT = 4000;
constexpr uint64_t KYTY_SLEEP_SPIN_LIMIT_100NS = 500; // 50 us
#ifndef CREATE_WAITABLE_TIMER_HIGH_RESOLUTION #ifndef CREATE_WAITABLE_TIMER_HIGH_RESOLUTION
#define CREATE_WAITABLE_TIMER_HIGH_RESOLUTION 0x00000002 #define CREATE_WAITABLE_TIMER_HIGH_RESOLUTION 0x00000002
@@ -39,7 +40,10 @@ static void SleepHighResolution100ns(uint64_t units_100ns) {
return; return;
} }
if (units_100ns <= 10000) { // Keep spinning only where a kernel transition is
// likely to cost more than the requested delay; ordinary millisecond sleeps use the
// per-thread high-resolution waitable timer below.
if (units_100ns <= KYTY_SLEEP_SPIN_LIMIT_100NS) {
LARGE_INTEGER frequency {}; LARGE_INTEGER frequency {};
LARGE_INTEGER start {}; LARGE_INTEGER start {};
if (QueryPerformanceFrequency(&frequency) != 0 && QueryPerformanceCounter(&start) != 0 && if (QueryPerformanceFrequency(&frequency) != 0 && QueryPerformanceCounter(&start) != 0 &&
src/graphics/presentation/videoOut.cpp
+4 -8
View File
@@ -788,14 +788,10 @@ void VideoOutDriver::Impl::PresentThread(std::stop_token token) {
while (!token.stop_requested()) { while (!token.stop_requested()) {
const auto sleep_begin = Common::Timer::QueryPerformanceCounter(); const auto sleep_begin = Common::Timer::QueryPerformanceCounter();
if (total_wait > 0) { if (total_wait > 0) {
const auto sleep_end = sleep_begin + static_cast<uint64_t>(total_wait); const auto remaining_us =
auto now = sleep_begin; (static_cast<uint64_t>(total_wait) * 1000000u + frequency - 1) / frequency;
while (!token.stop_requested() && now < sleep_end) { Common::Thread::SleepMicro(static_cast<uint32_t>(
const auto remaining_us = (sleep_end - now) * 1000000u / frequency; std::clamp<uint64_t>(remaining_us, 1, std::numeric_limits<uint32_t>::max())));
Common::Thread::SleepMicro(
static_cast<uint32_t>(std::clamp<uint64_t>(remaining_us, 1, 1000)));
now = Common::Timer::QueryPerformanceCounter();
}
} }
if (token.stop_requested()) { if (token.stop_requested()) {
break; break;
src/graphics/presentation/window/window.cpp
+4 -12
View File
@@ -752,26 +752,18 @@ void WindowContext::Run() {
#if defined(__APPLE__) #if defined(__APPLE__)
DrainMainThreadTasks(); DrainMainThreadTasks();
#endif #endif
if (SDL_PollEvent(&loop.event) != 0) {
ProcessEvent(timer.GetTimeS());
continue;
}
if (loop.paused.load(std::memory_order_acquire)) { if (loop.paused.load(std::memory_order_acquire)) {
if (!timer.IsPaused()) { if (!timer.IsPaused()) {
timer.Pause(); timer.Pause();
} }
} else if (timer.IsPaused()) {
timer.Resume();
}
if (SDL_WaitEvent(&loop.event) == 0) { if (SDL_WaitEvent(&loop.event) == 0) {
EXIT("%s\n", SDL_GetError()); EXIT("%s\n", SDL_GetError());
} }
ProcessEvent(timer.GetTimeS()); ProcessEvent(timer.GetTimeS());
continue;
}
if (timer.IsPaused()) {
timer.Resume();
}
Common::Thread::SleepMicro(1000);
} }
} }