[Base] Optional raw Clock source.

New cvar clock_source_raw allows to use the cpu cycle counter itself as an alternative time source, if system timing resolution is to low and causes problems.

5 files changed, 145 insertions, 11 deletions

diff --git a/src/xenia/base/clock.cc b/src/xenia/base/clock.cc
index c0f5eff9a..af18967f2 100644
--- a/src/xenia/base/clock.cc
+++ b/src/xenia/base/clock.cc
@@ -20,13 +20,17 @@ DEFINE_bool(clock_no_scaling, false,
             "Disable scaling code. Time management and locking is bypassed. "
             "Guest system time is directly pulled from host.",
             "CPU");
+DEFINE_bool(clock_source_raw, false,
+            "Use the RDTSC instruction as the time source. "
+            "Host CPU must support invariant TSC. ",
+            "CPU");
 
 namespace xe {
 
 // Time scalar applied to all time operations.
 double guest_time_scalar_ = 1.0;
 // Tick frequency of guest.
-uint64_t guest_tick_frequency_ = Clock::host_tick_frequency();
+uint64_t guest_tick_frequency_ = Clock::host_tick_frequency_platform();
 // Base FILETIME of the guest system from app start.
 uint64_t guest_system_time_base_ = Clock::QueryHostSystemTime();
 // Combined time and frequency ratio between host and guest.
@@ -44,7 +48,7 @@ std::mutex tick_mutex_;
 void RecomputeGuestTickScalar() {
   // Create a rational number with numerator (first) and denominator (second)
   auto frac =
-      std::make_pair(guest_tick_frequency_, Clock::host_tick_frequency());
+      std::make_pair(guest_tick_frequency_, Clock::QueryHostTickFrequency());
   // Doing it this way ensures we don't mess up our frequency scaling and
   // precisely controls the precision the guest_time_scalar_ can have.
   if (guest_time_scalar_ > 1.0) {
@@ -104,6 +108,21 @@ inline uint64_t QueryGuestSystemTimeOffset() {
   return guest_tick_count * numerator / denominator;
 }
 
+uint64_t Clock::QueryHostTickFrequency() {
+  if (cvars::clock_source_raw) {
+    return host_tick_frequency_raw();
+  } else {
+    return host_tick_frequency_platform();
+  }
+}
+uint64_t Clock::QueryHostTickCount() {
+  if (cvars::clock_source_raw) {
+    return host_tick_count_raw();
+  } else {
+    return host_tick_count_platform();
+  }
+}
+
 double Clock::guest_time_scalar() { return guest_time_scalar_; }
 
 void Clock::set_guest_time_scalar(double scalar) {
diff --git a/src/xenia/base/clock.h b/src/xenia/base/clock.h
index aec1cfcc6..00f1ff449 100644
--- a/src/xenia/base/clock.h
+++ b/src/xenia/base/clock.h
@@ -15,13 +15,22 @@
 #include "xenia/base/cvar.h"
 
 DECLARE_bool(clock_no_scaling);
+DECLARE_bool(clock_source_raw);
 
 namespace xe {
 
 class Clock {
  public:
-  // Host ticks-per-second.
-  static uint64_t host_tick_frequency();
+  // Host ticks-per-second. Generally QueryHostTickFrequency should be used.
+  // Either from platform suplied time source or from hardware directly.
+  static uint64_t host_tick_frequency_platform();
+  static uint64_t host_tick_frequency_raw();
+  // Host tick count. Generally QueryHostTickCount() should be used.
+  static uint64_t host_tick_count_platform();
+  static uint64_t host_tick_count_raw();
+
+  // Queries the host tick frequency.
+  static uint64_t QueryHostTickFrequency();
   // Queries the current host tick count.
   static uint64_t QueryHostTickCount();
   // Host time, in FILETIME format.
@@ -43,6 +52,7 @@ class Clock {
   // Sets the guest time base, used for computing the system time.
   // By default this is the current system time.
   static void set_guest_system_time_base(uint64_t time_base);
+
   // Queries the current guest tick count, accounting for frequency adjustment
   // and scaling.
   static uint64_t QueryGuestTickCount();
diff --git a/src/xenia/base/clock_posix.cc b/src/xenia/base/clock_posix.cc
index e2f1cf627..f18082b0e 100644
--- a/src/xenia/base/clock_posix.cc
+++ b/src/xenia/base/clock_posix.cc
@@ -2,7 +2,7 @@
  ******************************************************************************
  * Xenia : Xbox 360 Emulator Research Project                                 *
  ******************************************************************************
- * Copyright 2017 Ben Vanik. All rights reserved.                             *
+ * Copyright 2019 Ben Vanik. All rights reserved.                             *
  * Released under the BSD license - see LICENSE in the root for more details. *
  ******************************************************************************
  */
@@ -14,14 +14,14 @@
 
 namespace xe {
 
-uint64_t Clock::host_tick_frequency() {
+uint64_t Clock::host_tick_frequency_platform() {
   timespec res;
   clock_getres(CLOCK_MONOTONIC_RAW, &res);
 
   return uint64_t(res.tv_sec) + uint64_t(res.tv_nsec) * 1000000000ull;
 }
 
-uint64_t Clock::QueryHostTickCount() {
+uint64_t Clock::host_tick_count_platform() {
   timespec res;
   clock_gettime(CLOCK_MONOTONIC_RAW, &res);
 
@@ -40,7 +40,7 @@ uint64_t Clock::QueryHostSystemTime() {
 }
 
 uint64_t Clock::QueryHostUptimeMillis() {
-  return QueryHostTickCount() / (host_tick_frequency() / 1000);
+  return host_tick_count_platform() * 1000 / host_tick_frequency_platform();
 }
 
 }  // namespace xe
\ No newline at end of file
diff --git a/src/xenia/base/clock_win.cc b/src/xenia/base/clock_win.cc
index 7d33d0c93..4a0c8aeb5 100644
--- a/src/xenia/base/clock_win.cc
+++ b/src/xenia/base/clock_win.cc
@@ -13,13 +13,13 @@
 
 namespace xe {
 
-uint64_t Clock::host_tick_frequency() {
+uint64_t Clock::host_tick_frequency_platform() {
   LARGE_INTEGER frequency;
   QueryPerformanceFrequency(&frequency);
   return frequency.QuadPart;
 }
 
-uint64_t Clock::QueryHostTickCount() {
+uint64_t Clock::host_tick_count_platform() {
   LARGE_INTEGER counter;
   uint64_t time = 0;
   if (QueryPerformanceCounter(&counter)) {
@@ -35,7 +35,7 @@ uint64_t Clock::QueryHostSystemTime() {
 }
 
 uint64_t Clock::QueryHostUptimeMillis() {
-  return QueryHostTickCount() / (host_tick_frequency() / 1000);
+  return host_tick_count_platform() * 1000 / host_tick_frequency_platform();
 }
 
 }  // namespace xe
diff --git a/src/xenia/base/clock_x64.cc b/src/xenia/base/clock_x64.cc
new file mode 100644
index 000000000..72a250b9d
--- /dev/null
+++ b/src/xenia/base/clock_x64.cc
@@ -0,0 +1,105 @@
+/**
+ ******************************************************************************
+ * Xenia : Xbox 360 Emulator Research Project                                 *
+ ******************************************************************************
+ * Copyright 2019 Ben Vanik. All rights reserved.                             *
+ * Released under the BSD license - see LICENSE in the root for more details. *
+ ******************************************************************************
+ */
+
+#include "xenia/base/platform.h"
+
+#if XE_ARCH_AMD64
+
+#include "xenia/base/clock.h"
+#include "xenia/base/logging.h"
+
+// Wrap all these different cpu compiler intrinsics.
+// So no inline assembler here and the compiler will remove the clutter.
+#if XE_COMPILER_MSVC
+#define xe_cpu_cpuid(level, eax, ebx, ecx, edx)              \
+  {                                                          \
+    int __xe_cpuid_registers_[4];                            \
+    __cpuid(__xe_cpuid_registers_, (level));                 \
+    (eax) = static_cast<uint32_t>(__xe_cpuid_registers_[0]); \
+    (ebx) = static_cast<uint32_t>(__xe_cpuid_registers_[1]); \
+    (ecx) = static_cast<uint32_t>(__xe_cpuid_registers_[2]); \
+    (edx) = static_cast<uint32_t>(__xe_cpuid_registers_[3]); \
+  }
+#define xe_cpu_rdtsc() __rdtsc()
+#elif XE_COMPILER_CLANG || XE_COMPILER_GNUC
+#include <cpuid.h>
+#define xe_cpu_cpuid(level, eax, ebx, ecx, edx) \
+  __cpuid((level), (eax), (ebx), (ecx), (edx));
+#define xe_cpu_rdtsc() __rdtsc()
+#else
+#error "No cpu instruction wrappers for current compiler implemented."
+#endif
+
+#define CLOCK_FATAL(msg)                                      \
+  xe::FatalError(                                             \
+      "The raw clock source is not supported on your CPU. \n" \
+      "%s \n"                                                 \
+      "Set the cvar 'clock_source_raw' to 'false'.",          \
+      (msg));
+
+namespace xe {
+// Getting the TSC frequency can be a bit tricky. This method here only works on
+// Intel as it seems. There is no easy way to get the frequency outside of ring0
+// on AMD, so we fail gracefully if not possible.
+uint64_t Clock::host_tick_frequency_raw() {
+  uint32_t eax, ebx, ecx, edx;
+
+  // 00H Get max supported cpuid level.
+  xe_cpu_cpuid(0x0, eax, ebx, ecx, edx);
+  auto max_cpuid = eax;
+  // 80000000H Get max extended cpuid level
+  xe_cpu_cpuid(0x80000000, eax, ebx, ecx, edx);
+  auto max_cpuid_ex = eax;
+
+  // 80000007H Get extended power feature info
+  if (max_cpuid_ex >= 0x80000007) {
+    xe_cpu_cpuid(0x80000007, eax, ebx, ecx, edx);
+    // Invariant TSC bit at position 8
+    auto tsc_invariant = edx & (1 << 8);
+    // If the TSC is not invariant it will change its frequency with power
+    // states and across cores.
+    if (!tsc_invariant) {
+      CLOCK_FATAL("The CPU has no invariant TSC.");
+      return 0;
+    }
+  } else {
+    CLOCK_FATAL("Unclear if the CPU has an invariant TSC.")
+    return 0;
+  }
+
+  if (max_cpuid >= 0x15) {
+    // 15H Get TSC/Crystal ratio and Crystal Hz.
+    xe_cpu_cpuid(0x15, eax, ebx, ecx, edx);
+    uint64_t ratio_num = ebx;
+    uint64_t ratio_den = eax;
+    uint64_t cryst_freq = ecx;
+    // For some CPUs, Crystal frequency is not reported.
+    if (ratio_num && ratio_den && cryst_freq) {
+      // If it is, calculate the TSC frequency
+      auto tsc_freq = cryst_freq * ratio_num / ratio_den;
+    }
+  }
+
+  if (max_cpuid >= 0x16) {
+    // 16H Get CPU base frequency MHz in EAX.
+    xe_cpu_cpuid(0x16, eax, ebx, ecx, edx);
+    uint64_t cpu_base_freq = static_cast<uint64_t>(eax) * 1000000;
+    assert(cpu_base_freq);
+    return cpu_base_freq;
+  }
+
+  CLOCK_FATAL("The clock frequency could not be determined.");
+  return 0;
+}
+
+uint64_t Clock::host_tick_count_raw() { return xe_cpu_rdtsc(); }
+
+}  // namespace xe
+
+#endif