forked from ShuriZma/suyu
1
0
Fork 0

Merge pull request #9889 from Morph1984/time-is-ticking

core_timing: Reduce CPU usage on Windows
This commit is contained in:
liamwhite 2023-03-07 10:54:13 -05:00 committed by GitHub
commit a7792e5ff8
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
16 changed files with 324 additions and 65 deletions

View File

@ -477,8 +477,8 @@ if (APPLE)
find_library(COCOA_LIBRARY Cocoa)
set(PLATFORM_LIBRARIES ${COCOA_LIBRARY} ${IOKIT_LIBRARY} ${COREVIDEO_LIBRARY})
elseif (WIN32)
# WSAPoll and SHGetKnownFolderPath (AppData/Roaming) didn't exist before WinNT 6.x (Vista)
add_definitions(-D_WIN32_WINNT=0x0600 -DWINVER=0x0600)
# Target Windows 10
add_definitions(-D_WIN32_WINNT=0x0A00 -DWINVER=0x0A00)
set(PLATFORM_LIBRARIES winmm ws2_32 iphlpapi)
if (MINGW)
# PSAPI is the Process Status API

6
dist/yuzu.manifest vendored
View File

@ -36,12 +36,6 @@ SPDX-License-Identifier: GPL-2.0-or-later
<application>
<!-- Windows 10 -->
<supportedOS Id="{8e0f7a12-bfb3-4fe8-b9a5-48fd50a15a9a}"/>
<!-- Windows 8.1 -->
<supportedOS Id="{1f676c76-80e1-4239-95bb-83d0f6d0da78}"/>
<!-- Windows 8 -->
<supportedOS Id="{4a2f28e3-53b9-4441-ba9c-d69d4a4a6e38}"/>
<!-- Windows 7 -->
<supportedOS Id="{35138b9a-5d96-4fbd-8e2d-a2440225f93a}"/>
</application>
</compatibility>
<trustInfo

View File

@ -113,6 +113,8 @@ add_library(common STATIC
socket_types.h
spin_lock.cpp
spin_lock.h
steady_clock.cpp
steady_clock.h
stream.cpp
stream.h
string_util.cpp
@ -142,6 +144,14 @@ add_library(common STATIC
zstd_compression.h
)
if (WIN32)
target_sources(common PRIVATE
windows/timer_resolution.cpp
windows/timer_resolution.h
)
target_link_libraries(common PRIVATE ntdll)
endif()
if(ARCHITECTURE_x86_64)
target_sources(common
PRIVATE

View File

@ -0,0 +1,56 @@
// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#if defined(_WIN32)
#include <windows.h>
#else
#include <time.h>
#endif
#include "common/steady_clock.h"
namespace Common {
#ifdef _WIN32
static s64 WindowsQueryPerformanceFrequency() {
LARGE_INTEGER frequency;
QueryPerformanceFrequency(&frequency);
return frequency.QuadPart;
}
static s64 WindowsQueryPerformanceCounter() {
LARGE_INTEGER counter;
QueryPerformanceCounter(&counter);
return counter.QuadPart;
}
#endif
SteadyClock::time_point SteadyClock::Now() noexcept {
#if defined(_WIN32)
static const auto freq = WindowsQueryPerformanceFrequency();
const auto counter = WindowsQueryPerformanceCounter();
// 10 MHz is a very common QPC frequency on modern PCs.
// Optimizing for this specific frequency can double the performance of
// this function by avoiding the expensive frequency conversion path.
static constexpr s64 TenMHz = 10'000'000;
if (freq == TenMHz) [[likely]] {
static_assert(period::den % TenMHz == 0);
static constexpr s64 Multiplier = period::den / TenMHz;
return time_point{duration{counter * Multiplier}};
}
const auto whole = (counter / freq) * period::den;
const auto part = (counter % freq) * period::den / freq;
return time_point{duration{whole + part}};
#elif defined(__APPLE__)
return time_point{duration{clock_gettime_nsec_np(CLOCK_MONOTONIC_RAW)}};
#else
timespec ts;
clock_gettime(CLOCK_MONOTONIC, &ts);
return time_point{std::chrono::seconds{ts.tv_sec} + std::chrono::nanoseconds{ts.tv_nsec}};
#endif
}
}; // namespace Common

23
src/common/steady_clock.h Normal file
View File

@ -0,0 +1,23 @@
// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#pragma once
#include <chrono>
#include "common/common_types.h"
namespace Common {
struct SteadyClock {
using rep = s64;
using period = std::nano;
using duration = std::chrono::nanoseconds;
using time_point = std::chrono::time_point<SteadyClock>;
static constexpr bool is_steady = true;
[[nodiscard]] static time_point Now() noexcept;
};
} // namespace Common

View File

@ -1,6 +1,7 @@
// SPDX-FileCopyrightText: Copyright 2020 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#include "common/steady_clock.h"
#include "common/uint128.h"
#include "common/wall_clock.h"
@ -11,45 +12,32 @@
namespace Common {
using base_timer = std::chrono::steady_clock;
using base_time_point = std::chrono::time_point<base_timer>;
class StandardWallClock final : public WallClock {
public:
explicit StandardWallClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequency_)
: WallClock(emulated_cpu_frequency_, emulated_clock_frequency_, false) {
start_time = base_timer::now();
}
: WallClock{emulated_cpu_frequency_, emulated_clock_frequency_, false},
start_time{SteadyClock::Now()} {}
std::chrono::nanoseconds GetTimeNS() override {
base_time_point current = base_timer::now();
auto elapsed = current - start_time;
return std::chrono::duration_cast<std::chrono::nanoseconds>(elapsed);
return SteadyClock::Now() - start_time;
}
std::chrono::microseconds GetTimeUS() override {
base_time_point current = base_timer::now();
auto elapsed = current - start_time;
return std::chrono::duration_cast<std::chrono::microseconds>(elapsed);
return std::chrono::duration_cast<std::chrono::microseconds>(GetTimeNS());
}
std::chrono::milliseconds GetTimeMS() override {
base_time_point current = base_timer::now();
auto elapsed = current - start_time;
return std::chrono::duration_cast<std::chrono::milliseconds>(elapsed);
return std::chrono::duration_cast<std::chrono::milliseconds>(GetTimeNS());
}
u64 GetClockCycles() override {
std::chrono::nanoseconds time_now = GetTimeNS();
const u128 temporary =
Common::Multiply64Into128(time_now.count(), emulated_clock_frequency);
return Common::Divide128On32(temporary, 1000000000).first;
const u128 temp = Common::Multiply64Into128(GetTimeNS().count(), emulated_clock_frequency);
return Common::Divide128On32(temp, NS_RATIO).first;
}
u64 GetCPUCycles() override {
std::chrono::nanoseconds time_now = GetTimeNS();
const u128 temporary = Common::Multiply64Into128(time_now.count(), emulated_cpu_frequency);
return Common::Divide128On32(temporary, 1000000000).first;
const u128 temp = Common::Multiply64Into128(GetTimeNS().count(), emulated_cpu_frequency);
return Common::Divide128On32(temp, NS_RATIO).first;
}
void Pause([[maybe_unused]] bool is_paused) override {
@ -57,7 +45,7 @@ public:
}
private:
base_time_point start_time;
SteadyClock::time_point start_time;
};
#ifdef ARCHITECTURE_x86_64
@ -93,4 +81,9 @@ std::unique_ptr<WallClock> CreateBestMatchingClock(u64 emulated_cpu_frequency,
#endif
std::unique_ptr<WallClock> CreateStandardWallClock(u64 emulated_cpu_frequency,
u64 emulated_clock_frequency) {
return std::make_unique<StandardWallClock>(emulated_cpu_frequency, emulated_clock_frequency);
}
} // namespace Common

View File

@ -55,4 +55,7 @@ private:
[[nodiscard]] std::unique_ptr<WallClock> CreateBestMatchingClock(u64 emulated_cpu_frequency,
u64 emulated_clock_frequency);
[[nodiscard]] std::unique_ptr<WallClock> CreateStandardWallClock(u64 emulated_cpu_frequency,
u64 emulated_clock_frequency);
} // namespace Common

View File

@ -0,0 +1,109 @@
// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#include <windows.h>
#include "common/windows/timer_resolution.h"
extern "C" {
// http://undocumented.ntinternals.net/index.html?page=UserMode%2FUndocumented%20Functions%2FTime%2FNtQueryTimerResolution.html
NTSYSAPI LONG NTAPI NtQueryTimerResolution(PULONG MinimumResolution, PULONG MaximumResolution,
PULONG CurrentResolution);
// http://undocumented.ntinternals.net/index.html?page=UserMode%2FUndocumented%20Functions%2FTime%2FNtSetTimerResolution.html
NTSYSAPI LONG NTAPI NtSetTimerResolution(ULONG DesiredResolution, BOOLEAN SetResolution,
PULONG CurrentResolution);
// http://undocumented.ntinternals.net/index.html?page=UserMode%2FUndocumented%20Functions%2FNT%20Objects%2FThread%2FNtDelayExecution.html
NTSYSAPI LONG NTAPI NtDelayExecution(BOOLEAN Alertable, PLARGE_INTEGER DelayInterval);
}
// Defines for compatibility with older Windows 10 SDKs.
#ifndef PROCESS_POWER_THROTTLING_EXECUTION_SPEED
#define PROCESS_POWER_THROTTLING_EXECUTION_SPEED 0x1
#endif
#ifndef PROCESS_POWER_THROTTLING_IGNORE_TIMER_RESOLUTION
#define PROCESS_POWER_THROTTLING_IGNORE_TIMER_RESOLUTION 0x4
#endif
namespace Common::Windows {
namespace {
using namespace std::chrono;
constexpr nanoseconds ToNS(ULONG hundred_ns) {
return nanoseconds{hundred_ns * 100};
}
constexpr ULONG ToHundredNS(nanoseconds ns) {
return static_cast<ULONG>(ns.count()) / 100;
}
struct TimerResolution {
std::chrono::nanoseconds minimum;
std::chrono::nanoseconds maximum;
std::chrono::nanoseconds current;
};
TimerResolution GetTimerResolution() {
ULONG MinimumTimerResolution;
ULONG MaximumTimerResolution;
ULONG CurrentTimerResolution;
NtQueryTimerResolution(&MinimumTimerResolution, &MaximumTimerResolution,
&CurrentTimerResolution);
return {
.minimum{ToNS(MinimumTimerResolution)},
.maximum{ToNS(MaximumTimerResolution)},
.current{ToNS(CurrentTimerResolution)},
};
}
void SetHighQoS() {
// https://learn.microsoft.com/en-us/windows/win32/procthread/quality-of-service
PROCESS_POWER_THROTTLING_STATE PowerThrottling{
.Version{PROCESS_POWER_THROTTLING_CURRENT_VERSION},
.ControlMask{PROCESS_POWER_THROTTLING_EXECUTION_SPEED |
PROCESS_POWER_THROTTLING_IGNORE_TIMER_RESOLUTION},
.StateMask{},
};
SetProcessInformation(GetCurrentProcess(), ProcessPowerThrottling, &PowerThrottling,
sizeof(PROCESS_POWER_THROTTLING_STATE));
}
} // Anonymous namespace
nanoseconds GetMinimumTimerResolution() {
return GetTimerResolution().minimum;
}
nanoseconds GetMaximumTimerResolution() {
return GetTimerResolution().maximum;
}
nanoseconds GetCurrentTimerResolution() {
return GetTimerResolution().current;
}
nanoseconds SetCurrentTimerResolution(nanoseconds timer_resolution) {
// Set the timer resolution, and return the current timer resolution.
const auto DesiredTimerResolution = ToHundredNS(timer_resolution);
ULONG CurrentTimerResolution;
NtSetTimerResolution(DesiredTimerResolution, TRUE, &CurrentTimerResolution);
return ToNS(CurrentTimerResolution);
}
nanoseconds SetCurrentTimerResolutionToMaximum() {
SetHighQoS();
return SetCurrentTimerResolution(GetMaximumTimerResolution());
}
void SleepForOneTick() {
LARGE_INTEGER DelayInterval{
.QuadPart{-1},
};
NtDelayExecution(FALSE, &DelayInterval);
}
} // namespace Common::Windows

View File

@ -0,0 +1,38 @@
// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#pragma once
#include <chrono>
namespace Common::Windows {
/// Returns the minimum (least precise) supported timer resolution in nanoseconds.
std::chrono::nanoseconds GetMinimumTimerResolution();
/// Returns the maximum (most precise) supported timer resolution in nanoseconds.
std::chrono::nanoseconds GetMaximumTimerResolution();
/// Returns the current timer resolution in nanoseconds.
std::chrono::nanoseconds GetCurrentTimerResolution();
/**
* Sets the current timer resolution.
*
* @param timer_resolution Timer resolution in nanoseconds.
*
* @returns The current timer resolution.
*/
std::chrono::nanoseconds SetCurrentTimerResolution(std::chrono::nanoseconds timer_resolution);
/**
* Sets the current timer resolution to the maximum supported timer resolution.
*
* @returns The current timer resolution.
*/
std::chrono::nanoseconds SetCurrentTimerResolutionToMaximum();
/// Sleep for one tick of the current timer resolution.
void SleepForOneTick();
} // namespace Common::Windows

View File

@ -6,6 +6,7 @@
#include <thread>
#include "common/atomic_ops.h"
#include "common/steady_clock.h"
#include "common/uint128.h"
#include "common/x64/native_clock.h"
@ -39,6 +40,12 @@ static u64 FencedRDTSC() {
}
#endif
template <u64 Nearest>
static u64 RoundToNearest(u64 value) {
const auto mod = value % Nearest;
return mod >= (Nearest / 2) ? (value - mod + Nearest) : (value - mod);
}
u64 EstimateRDTSCFrequency() {
// Discard the first result measuring the rdtsc.
FencedRDTSC();
@ -46,18 +53,18 @@ u64 EstimateRDTSCFrequency() {
FencedRDTSC();
// Get the current time.
const auto start_time = std::chrono::steady_clock::now();
const auto start_time = Common::SteadyClock::Now();
const u64 tsc_start = FencedRDTSC();
// Wait for 200 milliseconds.
std::this_thread::sleep_for(std::chrono::milliseconds{200});
const auto end_time = std::chrono::steady_clock::now();
// Wait for 250 milliseconds.
std::this_thread::sleep_for(std::chrono::milliseconds{250});
const auto end_time = Common::SteadyClock::Now();
const u64 tsc_end = FencedRDTSC();
// Calculate differences.
const u64 timer_diff = static_cast<u64>(
std::chrono::duration_cast<std::chrono::nanoseconds>(end_time - start_time).count());
const u64 tsc_diff = tsc_end - tsc_start;
const u64 tsc_freq = MultiplyAndDivide64(tsc_diff, 1000000000ULL, timer_diff);
return tsc_freq;
return RoundToNearest<1000>(tsc_freq);
}
namespace X64 {

View File

@ -6,6 +6,10 @@
#include <string>
#include <tuple>
#ifdef _WIN32
#include "common/windows/timer_resolution.h"
#endif
#include "common/microprofile.h"
#include "core/core_timing.h"
#include "core/core_timing_util.h"
@ -38,7 +42,8 @@ struct CoreTiming::Event {
};
CoreTiming::CoreTiming()
: clock{Common::CreateBestMatchingClock(Hardware::BASE_CLOCK_RATE, Hardware::CNTFREQ)} {}
: cpu_clock{Common::CreateBestMatchingClock(Hardware::BASE_CLOCK_RATE, Hardware::CNTFREQ)},
event_clock{Common::CreateStandardWallClock(Hardware::BASE_CLOCK_RATE, Hardware::CNTFREQ)} {}
CoreTiming::~CoreTiming() {
Reset();
@ -185,15 +190,15 @@ void CoreTiming::ResetTicks() {
}
u64 CoreTiming::GetCPUTicks() const {
if (is_multicore) {
return clock->GetCPUCycles();
if (is_multicore) [[likely]] {
return cpu_clock->GetCPUCycles();
}
return ticks;
}
u64 CoreTiming::GetClockTicks() const {
if (is_multicore) {
return clock->GetClockCycles();
if (is_multicore) [[likely]] {
return cpu_clock->GetClockCycles();
}
return CpuCyclesToClockCycles(ticks);
}
@ -252,22 +257,21 @@ void CoreTiming::ThreadLoop() {
const auto next_time = Advance();
if (next_time) {
// There are more events left in the queue, wait until the next event.
const auto wait_time = *next_time - GetGlobalTimeNs().count();
auto wait_time = *next_time - GetGlobalTimeNs().count();
if (wait_time > 0) {
#ifdef _WIN32
// Assume a timer resolution of 1ms.
static constexpr s64 TimerResolutionNS = 1000000;
const auto timer_resolution_ns =
Common::Windows::GetCurrentTimerResolution().count();
// Sleep in discrete intervals of the timer resolution, and spin the rest.
const auto sleep_time = wait_time - (wait_time % TimerResolutionNS);
if (sleep_time > 0) {
event.WaitFor(std::chrono::nanoseconds(sleep_time));
}
while (!paused && !event.IsSet() && wait_time > 0) {
wait_time = *next_time - GetGlobalTimeNs().count();
while (!paused && !event.IsSet() && GetGlobalTimeNs().count() < *next_time) {
// Yield to reduce thread starvation.
if (wait_time >= timer_resolution_ns) {
Common::Windows::SleepForOneTick();
} else {
std::this_thread::yield();
}
}
if (event.IsSet()) {
event.Reset();
@ -285,9 +289,9 @@ void CoreTiming::ThreadLoop() {
}
paused_set = true;
clock->Pause(true);
event_clock->Pause(true);
pause_event.Wait();
clock->Pause(false);
event_clock->Pause(false);
}
}
@ -303,16 +307,23 @@ void CoreTiming::Reset() {
has_started = false;
}
std::chrono::nanoseconds CoreTiming::GetCPUTimeNs() const {
if (is_multicore) [[likely]] {
return cpu_clock->GetTimeNS();
}
return CyclesToNs(ticks);
}
std::chrono::nanoseconds CoreTiming::GetGlobalTimeNs() const {
if (is_multicore) {
return clock->GetTimeNS();
if (is_multicore) [[likely]] {
return event_clock->GetTimeNS();
}
return CyclesToNs(ticks);
}
std::chrono::microseconds CoreTiming::GetGlobalTimeUs() const {
if (is_multicore) {
return clock->GetTimeUS();
if (is_multicore) [[likely]] {
return event_clock->GetTimeUS();
}
return CyclesToUs(ticks);
}

View File

@ -122,6 +122,9 @@ public:
/// Returns current time in emulated in Clock cycles
u64 GetClockTicks() const;
/// Returns current time in nanoseconds.
std::chrono::nanoseconds GetCPUTimeNs() const;
/// Returns current time in microseconds.
std::chrono::microseconds GetGlobalTimeUs() const;
@ -139,7 +142,8 @@ private:
void Reset();
std::unique_ptr<Common::WallClock> clock;
std::unique_ptr<Common::WallClock> cpu_clock;
std::unique_ptr<Common::WallClock> event_clock;
s64 global_timer = 0;

View File

@ -13,10 +13,8 @@ namespace Core {
namespace Hardware {
// The below clock rate is based on Switch's clockspeed being widely known as 1.020GHz
// The exact value used is of course unverified.
constexpr u64 BASE_CLOCK_RATE = 1019215872; // Switch cpu frequency is 1020MHz un/docked
constexpr u64 CNTFREQ = 19200000; // Switch's hardware clock speed
constexpr u64 BASE_CLOCK_RATE = 1'020'000'000; // Default CPU Frequency = 1020 MHz
constexpr u64 CNTFREQ = 19'200'000; // CNTPCT_EL0 Frequency = 19.2 MHz
constexpr u32 NUM_CPU_CORES = 4; // Number of CPU Cores
// Virtual to Physical core map.

View File

@ -197,7 +197,7 @@ struct GPU::Impl {
constexpr u64 gpu_ticks_num = 384;
constexpr u64 gpu_ticks_den = 625;
u64 nanoseconds = system.CoreTiming().GetGlobalTimeNs().count();
u64 nanoseconds = system.CoreTiming().GetCPUTimeNs().count();
if (Settings::values.use_fast_gpu_time.GetValue()) {
nanoseconds /= 256;
}

View File

@ -91,6 +91,9 @@ static FileSys::VirtualFile VfsDirectoryCreateFileWrapper(const FileSys::Virtual
#include "common/microprofile.h"
#include "common/scm_rev.h"
#include "common/scope_exit.h"
#ifdef _WIN32
#include "common/windows/timer_resolution.h"
#endif
#ifdef ARCHITECTURE_x86_64
#include "common/x64/cpu_detect.h"
#endif
@ -377,6 +380,12 @@ GMainWindow::GMainWindow(std::unique_ptr<Config> config_, bool has_broken_vulkan
LOG_INFO(Frontend, "Host RAM: {:.2f} GiB",
Common::GetMemInfo().TotalPhysicalMemory / f64{1_GiB});
LOG_INFO(Frontend, "Host Swap: {:.2f} GiB", Common::GetMemInfo().TotalSwapMemory / f64{1_GiB});
#ifdef _WIN32
LOG_INFO(Frontend, "Host Timer Resolution: {:.4f} ms",
std::chrono::duration_cast<std::chrono::duration<f64, std::milli>>(
Common::Windows::SetCurrentTimerResolutionToMaximum())
.count());
#endif
UpdateWindowTitle();
show();

View File

@ -42,6 +42,8 @@
#include <windows.h>
#include <shellapi.h>
#include "common/windows/timer_resolution.h"
#endif
#undef _UNICODE
@ -314,6 +316,8 @@ int main(int argc, char** argv) {
#ifdef _WIN32
LocalFree(argv_w);
Common::Windows::SetCurrentTimerResolutionToMaximum();
#endif
MicroProfileOnThreadCreate("EmuThread");