x64: Deduplicate RDTSC usage

This commit is contained in:
Morph 2023-04-22 22:55:03 -04:00
parent 9c6fc44a59
commit dd12dd4c67
5 changed files with 82 additions and 19 deletions

View File

@ -172,6 +172,8 @@ if(ARCHITECTURE_x86_64)
x64/cpu_wait.h x64/cpu_wait.h
x64/native_clock.cpp x64/native_clock.cpp
x64/native_clock.h x64/native_clock.h
x64/rdtsc.cpp
x64/rdtsc.h
x64/xbyak_abi.h x64/xbyak_abi.h
x64/xbyak_util.h x64/xbyak_util.h
) )

View File

@ -14,6 +14,7 @@
#include "common/common_types.h" #include "common/common_types.h"
#include "common/logging/log.h" #include "common/logging/log.h"
#include "common/x64/cpu_detect.h" #include "common/x64/cpu_detect.h"
#include "common/x64/rdtsc.h"
#ifdef _WIN32 #ifdef _WIN32
#include <windows.h> #include <windows.h>
@ -187,6 +188,8 @@ static CPUCaps Detect() {
caps.tsc_frequency = static_cast<u64>(caps.crystal_frequency) * caps.tsc_frequency = static_cast<u64>(caps.crystal_frequency) *
caps.tsc_crystal_ratio_numerator / caps.tsc_crystal_ratio_numerator /
caps.tsc_crystal_ratio_denominator; caps.tsc_crystal_ratio_denominator;
} else {
caps.tsc_frequency = X64::EstimateRDTSCFrequency();
} }
} }

View File

@ -9,19 +9,11 @@
#include "common/x64/cpu_detect.h" #include "common/x64/cpu_detect.h"
#include "common/x64/cpu_wait.h" #include "common/x64/cpu_wait.h"
#include "common/x64/rdtsc.h"
namespace Common::X64 { namespace Common::X64 {
#ifdef _MSC_VER #ifdef _MSC_VER
__forceinline static u64 FencedRDTSC() {
_mm_lfence();
_ReadWriteBarrier();
const u64 result = __rdtsc();
_mm_lfence();
_ReadWriteBarrier();
return result;
}
__forceinline static void TPAUSE() { __forceinline static void TPAUSE() {
// 100,000 cycles is a reasonable amount of time to wait to save on CPU resources. // 100,000 cycles is a reasonable amount of time to wait to save on CPU resources.
// For reference: // For reference:
@ -32,16 +24,6 @@ __forceinline static void TPAUSE() {
_tpause(0, FencedRDTSC() + PauseCycles); _tpause(0, FencedRDTSC() + PauseCycles);
} }
#else #else
static u64 FencedRDTSC() {
u64 eax;
u64 edx;
asm volatile("lfence\n\t"
"rdtsc\n\t"
"lfence\n\t"
: "=a"(eax), "=d"(edx));
return (edx << 32) | eax;
}
static void TPAUSE() { static void TPAUSE() {
// 100,000 cycles is a reasonable amount of time to wait to save on CPU resources. // 100,000 cycles is a reasonable amount of time to wait to save on CPU resources.
// For reference: // For reference:

39
src/common/x64/rdtsc.cpp Normal file
View File

@ -0,0 +1,39 @@
// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#include <thread>
#include "common/steady_clock.h"
#include "common/uint128.h"
#include "common/x64/rdtsc.h"
namespace Common::X64 {
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();
std::this_thread::sleep_for(std::chrono::milliseconds{1});
FencedRDTSC();
// Get the current time.
const auto start_time = RealTimeClock::Now();
const u64 tsc_start = FencedRDTSC();
// Wait for 100 milliseconds.
std::this_thread::sleep_for(std::chrono::milliseconds{100});
const auto end_time = RealTimeClock::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 RoundToNearest<100'000>(tsc_freq);
}
} // namespace Common::X64

37
src/common/x64/rdtsc.h Normal file
View File

@ -0,0 +1,37 @@
// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#pragma once
#ifdef _MSC_VER
#include <intrin.h>
#endif
#include "common/common_types.h"
namespace Common::X64 {
#ifdef _MSC_VER
__forceinline static u64 FencedRDTSC() {
_mm_lfence();
_ReadWriteBarrier();
const u64 result = __rdtsc();
_mm_lfence();
_ReadWriteBarrier();
return result;
}
#else
static inline u64 FencedRDTSC() {
u64 eax;
u64 edx;
asm volatile("lfence\n\t"
"rdtsc\n\t"
"lfence\n\t"
: "=a"(eax), "=d"(edx));
return (edx << 32) | eax;
}
#endif
u64 EstimateRDTSCFrequency();
} // namespace Common::X64