Merge 083e28b222 into e0c72cd963

Common/Timer: Add a SteadyAwakeClock class to make play time tracking ignore time while suspended.

Source/Core/Common/Swap.h

@@ -19,6 +19,7 @@
 #include <fmt/format.h>
 
 #include "Common/CommonTypes.h"
+#include "Common/Intrinsics.h"
 
 namespace Common
 {
@@ -168,6 +169,64 @@ inline T FromBigEndian(T data)
   return data;
 }
 
+#ifdef __AVX__
+// Byte-swap patterns for PSHUFB.
+template <size_t ByteSize>
+inline __m128i GetSwapShuffle128()
+{
+  if constexpr (ByteSize == 2)
+    return _mm_set_epi64x(0x0e0f0c0d0a0b0809, 0x0607040502030001);
+  else if constexpr (ByteSize == 4)
+    return _mm_set_epi64x(0x0c0d0e0f08090a0b, 0x0405060700010203);
+  else if constexpr (ByteSize == 8)
+    return _mm_set_epi64x(0x08090a0b0c0d0e0f, 0x0001020304050607);
+  else
+    static_assert(false);
+}
+#endif
+
+#ifdef __AVX2__
+// Byte-swap patterns for VPSHUFB.
+template <size_t ByteSize>
+inline __m256i GetSwapShuffle256()
+{
+  __m128i pattern = GetSwapShuffle128<ByteSize>();
+  return _mm256_set_m128i(pattern, pattern);
+}
+#endif
+
+// Templated functions for byteswapped copies.
+template <typename T>
+inline void CopySwapped(T* dst, const T* src, size_t byte_size)
+{
+  constexpr size_t S = sizeof(T);
+  const size_t count = byte_size / S;
+  size_t i = 0;
+
+#ifdef __AVX2__
+  for (; i + 32 / S <= count; i += 32 / S)
+  {
+    const auto vdst = reinterpret_cast<__m256i*>(dst + i);
+    const auto vsrc = reinterpret_cast<const __m256i*>(src + i);
+    const auto swap = GetSwapShuffle256<S>();
+    _mm256_storeu_si256(vdst, _mm256_shuffle_epi8(_mm256_loadu_si256(vsrc), swap));
+  }
+#endif
+
+#ifdef __AVX__
+  for (; i + 16 / S <= count; i += 16 / S)
+  {
+    const auto vdst = reinterpret_cast<__m128i*>(dst + i);
+    const auto vsrc = reinterpret_cast<const __m128i*>(src + i);
+    const auto swap = GetSwapShuffle128<S>();
+    _mm_storeu_si128(vdst, _mm_shuffle_epi8(_mm_loadu_si128(vsrc), swap));
+  }
+#endif
+
+  for (; i < count; ++i)
+    dst[i] = Common::FromBigEndian(src[i]);
+}
+
 template <typename value_type>
 struct BigEndianValue
 {

Source/Core/Common/Timer.cpp

@@ -5,6 +5,7 @@
 
 #include <chrono>
 #include <thread>
+#include "Common/CommonFuncs.h"
 
 #ifdef _WIN32
 #include <Windows.h>
@@ -194,4 +195,37 @@ void PrecisionTimer::SleepUntil(Clock::time_point target)
   }
 }
 
+// Results are appropriately slewed on Linux, but not on Windows, macOS, or FreeBSD.
+// Clocks with that functionality seem to not be available there.
+auto SteadyAwakeClock::now() -> time_point
+{
+#if defined(_WIN32)
+  // The count is system time "in units of 100 nanoseconds".
+  using InterruptDuration = std::chrono::duration<ULONGLONG, std::ratio<100, std::nano::den>::type>;
+
+  ULONGLONG interrupt_time{};
+  if (!QueryUnbiasedInterruptTime(&interrupt_time))
+    ERROR_LOG_FMT(COMMON, "QueryUnbiasedInterruptTime");
+
+  return time_point{InterruptDuration{interrupt_time}};
+#else
+  // Note that Linux's CLOCK_MONOTONIC "does not count time that the system is suspended".
+  // This is in contrast to the behavior on macOS and FreeBSD.
+  static constexpr auto clock_id =
+#if defined(__linux__)
+      CLOCK_MONOTONIC;
+#elif defined(__APPLE__)
+      CLOCK_UPTIME_RAW;
+#else
+      CLOCK_UPTIME;
+#endif
+
+  timespec ts{};
+  if (clock_gettime(clock_id, &ts) != 0)
+    ERROR_LOG_FMT(COMMON, "clock_gettime: {}", LastStrerrorString());
+
+  return time_point{std::chrono::seconds{ts.tv_sec} + std::chrono::nanoseconds{ts.tv_nsec}};
+#endif
+}
+
 }  // Namespace Common

Source/Core/Common/Timer.h

@@ -50,4 +50,19 @@ private:
 #endif
 };
 
+// Similar to std::chrono::steady_clock except this clock
+// specifically does *not* count time while the system is suspended.
+class SteadyAwakeClock
+{
+public:
+  using rep = s64;
+  using period = std::nano;
+  using duration = std::chrono::duration<rep, period>;
+  using time_point = std::chrono::time_point<SteadyAwakeClock>;
+
+  static constexpr bool is_steady = true;
+
+  static time_point now();
+};
+
 }  // Namespace Common

Source/Core/Core/HW/CPU.cpp

@@ -72,8 +72,8 @@ void CPUManager::StartTimePlayedTimer()
 {
   Common::SetCurrentThreadName("Play Time Tracker");
 
-  // Steady clock for greater accuracy of timing
-  std::chrono::steady_clock timer;
+  // Use a clock that will appropriately ignore suspended system time.
+  Common::SteadyAwakeClock timer;
   auto prev_time = timer.now();
 
   while (true)

Source/Core/Core/HW/Memmap.h

@@ -130,29 +130,18 @@ public:
   void Write_U32_Swap(u32 var, u32 address);
   void Write_U64_Swap(u64 var, u32 address);
 
-  // Templated functions for byteswapped copies.
   template <typename T>
-  void CopyFromEmuSwapped(T* data, u32 address, size_t size) const
+  void CopyFromEmuSwapped(T* dst, u32 address, size_t size)
   {
-    const T* src = reinterpret_cast<T*>(GetPointerForRange(address, size));
-
-    if (src == nullptr)
-      return;
-
-    for (size_t i = 0; i < size / sizeof(T); i++)
-      data[i] = Common::FromBigEndian(src[i]);
+    if (const T* src = reinterpret_cast<T*>(GetPointerForRange(address, size)))
+      Common::CopySwapped(dst, src, size);
   }
 
   template <typename T>
-  void CopyToEmuSwapped(u32 address, const T* data, size_t size)
+  void CopyToEmuSwapped(u32 address, const T* src, size_t size)
   {
-    T* dest = reinterpret_cast<T*>(GetPointerForRange(address, size));
-
-    if (dest == nullptr)
-      return;
-
-    for (size_t i = 0; i < size / sizeof(T); i++)
-      dest[i] = Common::FromBigEndian(data[i]);
+    if (T* dst = reinterpret_cast<T*>(GetPointerForRange(address, size)))
+      Common::CopySwapped(dst, src, size);
   }
 
 private: