summaryrefslogtreecommitdiffstats
path: root/src/common/x64
diff options
context:
space:
mode:
authorFernando Sahmkow <fsahmkow27@gmail.com>2021-01-02 02:24:49 +0100
committerFernando Sahmkow <fsahmkow27@gmail.com>2021-01-02 04:00:27 +0100
commit53d92318b82cd4a9e08f814fcb8aab624d795c6c (patch)
tree4f5236ffebdcf947297d8ace42151b36810f3144 /src/common/x64
parentX86/NativeClock: Improve performance of clock calculations on hot path. (diff)
downloadyuzu-53d92318b82cd4a9e08f814fcb8aab624d795c6c.tar
yuzu-53d92318b82cd4a9e08f814fcb8aab624d795c6c.tar.gz
yuzu-53d92318b82cd4a9e08f814fcb8aab624d795c6c.tar.bz2
yuzu-53d92318b82cd4a9e08f814fcb8aab624d795c6c.tar.lz
yuzu-53d92318b82cd4a9e08f814fcb8aab624d795c6c.tar.xz
yuzu-53d92318b82cd4a9e08f814fcb8aab624d795c6c.tar.zst
yuzu-53d92318b82cd4a9e08f814fcb8aab624d795c6c.zip
Diffstat (limited to 'src/common/x64')
-rw-r--r--src/common/x64/native_clock.cpp41
-rw-r--r--src/common/x64/native_clock.h22
2 files changed, 41 insertions, 22 deletions
diff --git a/src/common/x64/native_clock.cpp b/src/common/x64/native_clock.cpp
index e246432d0..a65f6b832 100644
--- a/src/common/x64/native_clock.cpp
+++ b/src/common/x64/native_clock.cpp
@@ -17,6 +17,7 @@
#include <x86intrin.h>
#endif
+#include "common/atomic_ops.h"
#include "common/uint128.h"
#include "common/x64/native_clock.h"
@@ -102,8 +103,8 @@ NativeClock::NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequen
: WallClock(emulated_cpu_frequency_, emulated_clock_frequency_, true), rtsc_frequency{
rtsc_frequency_} {
_mm_mfence();
- last_measure = __rdtsc();
- accumulated_ticks = 0U;
+ time_point.inner.last_measure = __rdtsc();
+ time_point.inner.accumulated_ticks = 0U;
ns_rtsc_factor = GetFixedPoint64Factor(1000000000, rtsc_frequency);
us_rtsc_factor = GetFixedPoint64Factor(1000000, rtsc_frequency);
ms_rtsc_factor = GetFixedPoint64Factor(1000, rtsc_frequency);
@@ -112,23 +113,35 @@ NativeClock::NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequen
}
u64 NativeClock::GetRTSC() {
- std::scoped_lock scope{rtsc_serialize};
- _mm_mfence();
- const u64 current_measure = __rdtsc();
- u64 diff = current_measure - last_measure;
- diff = diff & ~static_cast<u64>(static_cast<s64>(diff) >> 63); // max(diff, 0)
- if (current_measure > last_measure) {
- last_measure = current_measure;
- }
- accumulated_ticks += diff;
+ TimePoint new_time_point{};
+ TimePoint current_time_point{};
+ do {
+ current_time_point.pack = time_point.pack;
+ _mm_mfence();
+ const u64 current_measure = __rdtsc();
+ u64 diff = current_measure - current_time_point.inner.last_measure;
+ diff = diff & ~static_cast<u64>(static_cast<s64>(diff) >> 63); // max(diff, 0)
+ new_time_point.inner.last_measure = current_measure > current_time_point.inner.last_measure
+ ? current_measure
+ : current_time_point.inner.last_measure;
+ new_time_point.inner.accumulated_ticks = current_time_point.inner.accumulated_ticks + diff;
+ } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack,
+ current_time_point.pack));
/// The clock cannot be more precise than the guest timer, remove the lower bits
- return accumulated_ticks & inaccuracy_mask;
+ return new_time_point.inner.accumulated_ticks & inaccuracy_mask;
}
void NativeClock::Pause(bool is_paused) {
if (!is_paused) {
- _mm_mfence();
- last_measure = __rdtsc();
+ TimePoint current_time_point{};
+ TimePoint new_time_point{};
+ do {
+ current_time_point.pack = time_point.pack;
+ new_time_point.pack = current_time_point.pack;
+ _mm_mfence();
+ new_time_point.inner.last_measure = __rdtsc();
+ } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack,
+ current_time_point.pack));
}
}
diff --git a/src/common/x64/native_clock.h b/src/common/x64/native_clock.h
index a7b1ee9e0..7cbd400d2 100644
--- a/src/common/x64/native_clock.h
+++ b/src/common/x64/native_clock.h
@@ -6,7 +6,6 @@
#include <optional>
-#include "common/spin_lock.h"
#include "common/wall_clock.h"
namespace Common {
@@ -32,22 +31,29 @@ public:
private:
u64 GetRTSC();
+ union alignas(16) TimePoint {
+ TimePoint() : pack{} {}
+ u128 pack{};
+ struct Inner {
+ u64 last_measure{};
+ u64 accumulated_ticks{};
+ } inner;
+ };
+
/// value used to reduce the native clocks accuracy as some apss rely on
/// undefined behavior where the level of accuracy in the clock shouldn't
/// be higher.
static constexpr u64 inaccuracy_mask = ~(UINT64_C(0x400) - 1);
- SpinLock rtsc_serialize{};
- u64 last_measure{};
- u64 accumulated_ticks{};
- u64 rtsc_frequency;
-
+ TimePoint time_point;
// factors
+ u64 clock_rtsc_factor{};
+ u64 cpu_rtsc_factor{};
u64 ns_rtsc_factor{};
u64 us_rtsc_factor{};
u64 ms_rtsc_factor{};
- u64 clock_rtsc_factor{};
- u64 cpu_rtsc_factor{};
+
+ u64 rtsc_frequency;
};
} // namespace X64