native_clock.cpp 4.8 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136
  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/uint128.h"
  8. #include "common/x64/native_clock.h"
  9. #ifdef _MSC_VER
  10. #include <intrin.h>
  11. #endif
  12. namespace Common {
  13. #ifdef _MSC_VER
  14. __forceinline static u64 FencedRDTSC() {
  15. _mm_lfence();
  16. _ReadWriteBarrier();
  17. const u64 result = __rdtsc();
  18. _mm_lfence();
  19. _ReadWriteBarrier();
  20. return result;
  21. }
  22. #else
  23. static u64 FencedRDTSC() {
  24. u64 result;
  25. asm volatile("lfence\n\t"
  26. "rdtsc\n\t"
  27. "shl $32, %%rdx\n\t"
  28. "or %%rdx, %0\n\t"
  29. "lfence"
  30. : "=a"(result)
  31. :
  32. : "rdx", "memory", "cc");
  33. return result;
  34. }
  35. #endif
  36. u64 EstimateRDTSCFrequency() {
  37. // Discard the first result measuring the rdtsc.
  38. FencedRDTSC();
  39. std::this_thread::sleep_for(std::chrono::milliseconds{1});
  40. FencedRDTSC();
  41. // Get the current time.
  42. const auto start_time = std::chrono::steady_clock::now();
  43. const u64 tsc_start = FencedRDTSC();
  44. // Wait for 200 milliseconds.
  45. std::this_thread::sleep_for(std::chrono::milliseconds{200});
  46. const auto end_time = std::chrono::steady_clock::now();
  47. const u64 tsc_end = FencedRDTSC();
  48. // Calculate differences.
  49. const u64 timer_diff = static_cast<u64>(
  50. std::chrono::duration_cast<std::chrono::nanoseconds>(end_time - start_time).count());
  51. const u64 tsc_diff = tsc_end - tsc_start;
  52. const u64 tsc_freq = MultiplyAndDivide64(tsc_diff, 1000000000ULL, timer_diff);
  53. return tsc_freq;
  54. }
  55. namespace X64 {
  56. NativeClock::NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequency_,
  57. u64 rtsc_frequency_)
  58. : WallClock(emulated_cpu_frequency_, emulated_clock_frequency_, true), rtsc_frequency{
  59. rtsc_frequency_} {
  60. time_point.inner.last_measure = FencedRDTSC();
  61. time_point.inner.accumulated_ticks = 0U;
  62. ns_rtsc_factor = GetFixedPoint64Factor(NS_RATIO, rtsc_frequency);
  63. us_rtsc_factor = GetFixedPoint64Factor(US_RATIO, rtsc_frequency);
  64. ms_rtsc_factor = GetFixedPoint64Factor(MS_RATIO, rtsc_frequency);
  65. clock_rtsc_factor = GetFixedPoint64Factor(emulated_clock_frequency, rtsc_frequency);
  66. cpu_rtsc_factor = GetFixedPoint64Factor(emulated_cpu_frequency, rtsc_frequency);
  67. }
  68. u64 NativeClock::GetRTSC() {
  69. TimePoint current_time_point{};
  70. TimePoint new_time_point{};
  71. current_time_point.pack = Common::AtomicLoad128(time_point.pack.data());
  72. do {
  73. const u64 current_measure = FencedRDTSC();
  74. u64 diff = current_measure - current_time_point.inner.last_measure;
  75. diff = diff & ~static_cast<u64>(static_cast<s64>(diff) >> 63); // max(diff, 0)
  76. new_time_point.inner.last_measure = current_measure > current_time_point.inner.last_measure
  77. ? current_measure
  78. : current_time_point.inner.last_measure;
  79. new_time_point.inner.accumulated_ticks = current_time_point.inner.accumulated_ticks + diff;
  80. } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack,
  81. current_time_point.pack, current_time_point.pack));
  82. return new_time_point.inner.accumulated_ticks;
  83. }
  84. void NativeClock::Pause(bool is_paused) {
  85. if (!is_paused) {
  86. TimePoint current_time_point{};
  87. TimePoint new_time_point{};
  88. current_time_point.pack = Common::AtomicLoad128(time_point.pack.data());
  89. do {
  90. new_time_point.pack = current_time_point.pack;
  91. new_time_point.inner.last_measure = FencedRDTSC();
  92. } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack,
  93. current_time_point.pack, current_time_point.pack));
  94. }
  95. }
  96. std::chrono::nanoseconds NativeClock::GetTimeNS() {
  97. const u64 rtsc_value = GetRTSC();
  98. return std::chrono::nanoseconds{MultiplyHigh(rtsc_value, ns_rtsc_factor)};
  99. }
  100. std::chrono::microseconds NativeClock::GetTimeUS() {
  101. const u64 rtsc_value = GetRTSC();
  102. return std::chrono::microseconds{MultiplyHigh(rtsc_value, us_rtsc_factor)};
  103. }
  104. std::chrono::milliseconds NativeClock::GetTimeMS() {
  105. const u64 rtsc_value = GetRTSC();
  106. return std::chrono::milliseconds{MultiplyHigh(rtsc_value, ms_rtsc_factor)};
  107. }
  108. u64 NativeClock::GetClockCycles() {
  109. const u64 rtsc_value = GetRTSC();
  110. return MultiplyHigh(rtsc_value, clock_rtsc_factor);
  111. }
  112. u64 NativeClock::GetCPUCycles() {
  113. const u64 rtsc_value = GetRTSC();
  114. return MultiplyHigh(rtsc_value, cpu_rtsc_factor);
  115. }
  116. } // namespace X64
  117. } // namespace Common