native_clock.cpp 5.9 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164
  1. // SPDX-FileCopyrightText: Copyright 2020 yuzu Emulator Project
  2. // SPDX-License-Identifier: GPL-2.0-or-later
  3. #include <array>
  4. #include <chrono>
  5. #include <thread>
  6. #include "common/atomic_ops.h"
  7. #include "common/steady_clock.h"
  8. #include "common/uint128.h"
  9. #include "common/x64/native_clock.h"
  10. #ifdef _MSC_VER
  11. #include <intrin.h>
  12. #endif
  13. namespace Common {
  14. #ifdef _MSC_VER
  15. __forceinline static u64 FencedRDTSC() {
  16. _mm_lfence();
  17. _ReadWriteBarrier();
  18. const u64 result = __rdtsc();
  19. _mm_lfence();
  20. _ReadWriteBarrier();
  21. return result;
  22. }
  23. #else
  24. static u64 FencedRDTSC() {
  25. u64 eax;
  26. u64 edx;
  27. asm volatile("lfence\n\t"
  28. "rdtsc\n\t"
  29. "lfence\n\t"
  30. : "=a"(eax), "=d"(edx));
  31. return (edx << 32) | eax;
  32. }
  33. #endif
  34. template <u64 Nearest>
  35. static u64 RoundToNearest(u64 value) {
  36. const auto mod = value % Nearest;
  37. return mod >= (Nearest / 2) ? (value - mod + Nearest) : (value - mod);
  38. }
  39. u64 EstimateRDTSCFrequency() {
  40. // Discard the first result measuring the rdtsc.
  41. FencedRDTSC();
  42. std::this_thread::sleep_for(std::chrono::milliseconds{1});
  43. FencedRDTSC();
  44. // Get the current time.
  45. const auto start_time = Common::RealTimeClock::Now();
  46. const u64 tsc_start = FencedRDTSC();
  47. // Wait for 250 milliseconds.
  48. std::this_thread::sleep_for(std::chrono::milliseconds{250});
  49. const auto end_time = Common::RealTimeClock::Now();
  50. const u64 tsc_end = FencedRDTSC();
  51. // Calculate differences.
  52. const u64 timer_diff = static_cast<u64>(
  53. std::chrono::duration_cast<std::chrono::nanoseconds>(end_time - start_time).count());
  54. const u64 tsc_diff = tsc_end - tsc_start;
  55. const u64 tsc_freq = MultiplyAndDivide64(tsc_diff, 1000000000ULL, timer_diff);
  56. return RoundToNearest<1000>(tsc_freq);
  57. }
  58. namespace X64 {
  59. NativeClock::NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequency_,
  60. u64 rtsc_frequency_)
  61. : WallClock(emulated_cpu_frequency_, emulated_clock_frequency_, true), rtsc_frequency{
  62. rtsc_frequency_} {
  63. // Thread to re-adjust the RDTSC frequency after 10 seconds has elapsed.
  64. time_sync_thread = std::jthread{[this](std::stop_token token) {
  65. // Get the current time.
  66. const auto start_time = Common::RealTimeClock::Now();
  67. const u64 tsc_start = FencedRDTSC();
  68. // Wait for 10 seconds.
  69. if (!Common::StoppableTimedWait(token, std::chrono::seconds{10})) {
  70. return;
  71. }
  72. const auto end_time = Common::RealTimeClock::Now();
  73. const u64 tsc_end = FencedRDTSC();
  74. // Calculate differences.
  75. const u64 timer_diff = static_cast<u64>(
  76. std::chrono::duration_cast<std::chrono::nanoseconds>(end_time - start_time).count());
  77. const u64 tsc_diff = tsc_end - tsc_start;
  78. const u64 tsc_freq = MultiplyAndDivide64(tsc_diff, 1000000000ULL, timer_diff);
  79. rtsc_frequency = tsc_freq;
  80. CalculateAndSetFactors();
  81. }};
  82. time_point.inner.last_measure = FencedRDTSC();
  83. time_point.inner.accumulated_ticks = 0U;
  84. CalculateAndSetFactors();
  85. }
  86. u64 NativeClock::GetRTSC() {
  87. TimePoint new_time_point{};
  88. TimePoint current_time_point{};
  89. current_time_point.pack = Common::AtomicLoad128(time_point.pack.data());
  90. do {
  91. const u64 current_measure = FencedRDTSC();
  92. u64 diff = current_measure - current_time_point.inner.last_measure;
  93. diff = diff & ~static_cast<u64>(static_cast<s64>(diff) >> 63); // max(diff, 0)
  94. new_time_point.inner.last_measure = current_measure > current_time_point.inner.last_measure
  95. ? current_measure
  96. : current_time_point.inner.last_measure;
  97. new_time_point.inner.accumulated_ticks = current_time_point.inner.accumulated_ticks + diff;
  98. } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack,
  99. current_time_point.pack, current_time_point.pack));
  100. return new_time_point.inner.accumulated_ticks;
  101. }
  102. void NativeClock::Pause(bool is_paused) {
  103. if (!is_paused) {
  104. TimePoint current_time_point{};
  105. TimePoint new_time_point{};
  106. current_time_point.pack = Common::AtomicLoad128(time_point.pack.data());
  107. do {
  108. new_time_point.pack = current_time_point.pack;
  109. new_time_point.inner.last_measure = FencedRDTSC();
  110. } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack,
  111. current_time_point.pack, current_time_point.pack));
  112. }
  113. }
  114. std::chrono::nanoseconds NativeClock::GetTimeNS() {
  115. const u64 rtsc_value = GetRTSC();
  116. return std::chrono::nanoseconds{MultiplyHigh(rtsc_value, ns_rtsc_factor)};
  117. }
  118. std::chrono::microseconds NativeClock::GetTimeUS() {
  119. const u64 rtsc_value = GetRTSC();
  120. return std::chrono::microseconds{MultiplyHigh(rtsc_value, us_rtsc_factor)};
  121. }
  122. std::chrono::milliseconds NativeClock::GetTimeMS() {
  123. const u64 rtsc_value = GetRTSC();
  124. return std::chrono::milliseconds{MultiplyHigh(rtsc_value, ms_rtsc_factor)};
  125. }
  126. u64 NativeClock::GetClockCycles() {
  127. const u64 rtsc_value = GetRTSC();
  128. return MultiplyHigh(rtsc_value, clock_rtsc_factor);
  129. }
  130. u64 NativeClock::GetCPUCycles() {
  131. const u64 rtsc_value = GetRTSC();
  132. return MultiplyHigh(rtsc_value, cpu_rtsc_factor);
  133. }
  134. void NativeClock::CalculateAndSetFactors() {
  135. ns_rtsc_factor = GetFixedPoint64Factor(NS_RATIO, rtsc_frequency);
  136. us_rtsc_factor = GetFixedPoint64Factor(US_RATIO, rtsc_frequency);
  137. ms_rtsc_factor = GetFixedPoint64Factor(MS_RATIO, rtsc_frequency);
  138. clock_rtsc_factor = GetFixedPoint64Factor(emulated_clock_frequency, rtsc_frequency);
  139. cpu_rtsc_factor = GetFixedPoint64Factor(emulated_cpu_frequency, rtsc_frequency);
  140. }
  141. } // namespace X64
  142. } // namespace Common