summaryrefslogtreecommitdiff
path: root/src/common/x64
diff options
context:
space:
mode:
Diffstat (limited to 'src/common/x64')
-rw-r--r--src/common/x64/cpu_detect.cpp3
-rw-r--r--src/common/x64/cpu_wait.cpp20
-rw-r--r--src/common/x64/native_clock.cpp166
-rw-r--r--src/common/x64/native_clock.h59
-rw-r--r--src/common/x64/rdtsc.cpp39
-rw-r--r--src/common/x64/rdtsc.h37
6 files changed, 126 insertions, 198 deletions
diff --git a/src/common/x64/cpu_detect.cpp b/src/common/x64/cpu_detect.cpp
index 72ed6e96c..c998b1197 100644
--- a/src/common/x64/cpu_detect.cpp
+++ b/src/common/x64/cpu_detect.cpp
@@ -14,6 +14,7 @@
14#include "common/common_types.h" 14#include "common/common_types.h"
15#include "common/logging/log.h" 15#include "common/logging/log.h"
16#include "common/x64/cpu_detect.h" 16#include "common/x64/cpu_detect.h"
17#include "common/x64/rdtsc.h"
17 18
18#ifdef _WIN32 19#ifdef _WIN32
19#include <windows.h> 20#include <windows.h>
@@ -187,6 +188,8 @@ static CPUCaps Detect() {
187 caps.tsc_frequency = static_cast<u64>(caps.crystal_frequency) * 188 caps.tsc_frequency = static_cast<u64>(caps.crystal_frequency) *
188 caps.tsc_crystal_ratio_numerator / 189 caps.tsc_crystal_ratio_numerator /
189 caps.tsc_crystal_ratio_denominator; 190 caps.tsc_crystal_ratio_denominator;
191 } else {
192 caps.tsc_frequency = X64::EstimateRDTSCFrequency();
190 } 193 }
191 } 194 }
192 195
diff --git a/src/common/x64/cpu_wait.cpp b/src/common/x64/cpu_wait.cpp
index cfeef6a3d..c53dd4945 100644
--- a/src/common/x64/cpu_wait.cpp
+++ b/src/common/x64/cpu_wait.cpp
@@ -9,19 +9,11 @@
9 9
10#include "common/x64/cpu_detect.h" 10#include "common/x64/cpu_detect.h"
11#include "common/x64/cpu_wait.h" 11#include "common/x64/cpu_wait.h"
12#include "common/x64/rdtsc.h"
12 13
13namespace Common::X64 { 14namespace Common::X64 {
14 15
15#ifdef _MSC_VER 16#ifdef _MSC_VER
16__forceinline static u64 FencedRDTSC() {
17 _mm_lfence();
18 _ReadWriteBarrier();
19 const u64 result = __rdtsc();
20 _mm_lfence();
21 _ReadWriteBarrier();
22 return result;
23}
24
25__forceinline static void TPAUSE() { 17__forceinline static void TPAUSE() {
26 // 100,000 cycles is a reasonable amount of time to wait to save on CPU resources. 18 // 100,000 cycles is a reasonable amount of time to wait to save on CPU resources.
27 // For reference: 19 // For reference:
@@ -32,16 +24,6 @@ __forceinline static void TPAUSE() {
32 _tpause(0, FencedRDTSC() + PauseCycles); 24 _tpause(0, FencedRDTSC() + PauseCycles);
33} 25}
34#else 26#else
35static u64 FencedRDTSC() {
36 u64 eax;
37 u64 edx;
38 asm volatile("lfence\n\t"
39 "rdtsc\n\t"
40 "lfence\n\t"
41 : "=a"(eax), "=d"(edx));
42 return (edx << 32) | eax;
43}
44
45static void TPAUSE() { 27static void TPAUSE() {
46 // 100,000 cycles is a reasonable amount of time to wait to save on CPU resources. 28 // 100,000 cycles is a reasonable amount of time to wait to save on CPU resources.
47 // For reference: 29 // For reference:
diff --git a/src/common/x64/native_clock.cpp b/src/common/x64/native_clock.cpp
index 277b00662..7d2a26bd9 100644
--- a/src/common/x64/native_clock.cpp
+++ b/src/common/x64/native_clock.cpp
@@ -1,164 +1,50 @@
1// SPDX-FileCopyrightText: Copyright 2020 yuzu Emulator Project 1// SPDX-FileCopyrightText: Copyright 2020 yuzu Emulator Project
2// SPDX-License-Identifier: GPL-2.0-or-later 2// SPDX-License-Identifier: GPL-2.0-or-later
3 3
4#include <array>
5#include <chrono>
6#include <thread>
7
8#include "common/atomic_ops.h"
9#include "common/steady_clock.h"
10#include "common/uint128.h" 4#include "common/uint128.h"
11#include "common/x64/native_clock.h" 5#include "common/x64/native_clock.h"
6#include "common/x64/rdtsc.h"
12 7
13#ifdef _MSC_VER 8namespace Common::X64 {
14#include <intrin.h>
15#endif
16
17namespace Common {
18 9
19#ifdef _MSC_VER 10NativeClock::NativeClock(u64 rdtsc_frequency_)
20__forceinline static u64 FencedRDTSC() { 11 : start_ticks{FencedRDTSC()}, rdtsc_frequency{rdtsc_frequency_},
21 _mm_lfence(); 12 ns_rdtsc_factor{GetFixedPoint64Factor(NsRatio::den, rdtsc_frequency)},
22 _ReadWriteBarrier(); 13 us_rdtsc_factor{GetFixedPoint64Factor(UsRatio::den, rdtsc_frequency)},
23 const u64 result = __rdtsc(); 14 ms_rdtsc_factor{GetFixedPoint64Factor(MsRatio::den, rdtsc_frequency)},
24 _mm_lfence(); 15 cntpct_rdtsc_factor{GetFixedPoint64Factor(CNTFRQ, rdtsc_frequency)},
25 _ReadWriteBarrier(); 16 gputick_rdtsc_factor{GetFixedPoint64Factor(GPUTickFreq, rdtsc_frequency)} {}
26 return result;
27}
28#else
29static u64 FencedRDTSC() {
30 u64 eax;
31 u64 edx;
32 asm volatile("lfence\n\t"
33 "rdtsc\n\t"
34 "lfence\n\t"
35 : "=a"(eax), "=d"(edx));
36 return (edx << 32) | eax;
37}
38#endif
39 17
40template <u64 Nearest> 18std::chrono::nanoseconds NativeClock::GetTimeNS() const {
41static u64 RoundToNearest(u64 value) { 19 return std::chrono::nanoseconds{MultiplyHigh(GetHostTicksElapsed(), ns_rdtsc_factor)};
42 const auto mod = value % Nearest;
43 return mod >= (Nearest / 2) ? (value - mod + Nearest) : (value - mod);
44} 20}
45 21
46u64 EstimateRDTSCFrequency() { 22std::chrono::microseconds NativeClock::GetTimeUS() const {
47 // Discard the first result measuring the rdtsc. 23 return std::chrono::microseconds{MultiplyHigh(GetHostTicksElapsed(), us_rdtsc_factor)};
48 FencedRDTSC();
49 std::this_thread::sleep_for(std::chrono::milliseconds{1});
50 FencedRDTSC();
51
52 // Get the current time.
53 const auto start_time = Common::RealTimeClock::Now();
54 const u64 tsc_start = FencedRDTSC();
55 // Wait for 250 milliseconds.
56 std::this_thread::sleep_for(std::chrono::milliseconds{250});
57 const auto end_time = Common::RealTimeClock::Now();
58 const u64 tsc_end = FencedRDTSC();
59 // Calculate differences.
60 const u64 timer_diff = static_cast<u64>(
61 std::chrono::duration_cast<std::chrono::nanoseconds>(end_time - start_time).count());
62 const u64 tsc_diff = tsc_end - tsc_start;
63 const u64 tsc_freq = MultiplyAndDivide64(tsc_diff, 1000000000ULL, timer_diff);
64 return RoundToNearest<1000>(tsc_freq);
65} 24}
66 25
67namespace X64 { 26std::chrono::milliseconds NativeClock::GetTimeMS() const {
68NativeClock::NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequency_, 27 return std::chrono::milliseconds{MultiplyHigh(GetHostTicksElapsed(), ms_rdtsc_factor)};
69 u64 rtsc_frequency_)
70 : WallClock(emulated_cpu_frequency_, emulated_clock_frequency_, true), rtsc_frequency{
71 rtsc_frequency_} {
72 // Thread to re-adjust the RDTSC frequency after 10 seconds has elapsed.
73 time_sync_thread = std::jthread{[this](std::stop_token token) {
74 // Get the current time.
75 const auto start_time = Common::RealTimeClock::Now();
76 const u64 tsc_start = FencedRDTSC();
77 // Wait for 10 seconds.
78 if (!Common::StoppableTimedWait(token, std::chrono::seconds{10})) {
79 return;
80 }
81 const auto end_time = Common::RealTimeClock::Now();
82 const u64 tsc_end = FencedRDTSC();
83 // Calculate differences.
84 const u64 timer_diff = static_cast<u64>(
85 std::chrono::duration_cast<std::chrono::nanoseconds>(end_time - start_time).count());
86 const u64 tsc_diff = tsc_end - tsc_start;
87 const u64 tsc_freq = MultiplyAndDivide64(tsc_diff, 1000000000ULL, timer_diff);
88 rtsc_frequency = tsc_freq;
89 CalculateAndSetFactors();
90 }};
91
92 time_point.inner.last_measure = FencedRDTSC();
93 time_point.inner.accumulated_ticks = 0U;
94 CalculateAndSetFactors();
95} 28}
96 29
97u64 NativeClock::GetRTSC() { 30u64 NativeClock::GetCNTPCT() const {
98 TimePoint new_time_point{}; 31 return MultiplyHigh(GetHostTicksElapsed(), cntpct_rdtsc_factor);
99 TimePoint current_time_point{};
100
101 current_time_point.pack = Common::AtomicLoad128(time_point.pack.data());
102 do {
103 const u64 current_measure = FencedRDTSC();
104 u64 diff = current_measure - current_time_point.inner.last_measure;
105 diff = diff & ~static_cast<u64>(static_cast<s64>(diff) >> 63); // max(diff, 0)
106 new_time_point.inner.last_measure = current_measure > current_time_point.inner.last_measure
107 ? current_measure
108 : current_time_point.inner.last_measure;
109 new_time_point.inner.accumulated_ticks = current_time_point.inner.accumulated_ticks + diff;
110 } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack,
111 current_time_point.pack, current_time_point.pack));
112 return new_time_point.inner.accumulated_ticks;
113} 32}
114 33
115void NativeClock::Pause(bool is_paused) { 34u64 NativeClock::GetGPUTick() const {
116 if (!is_paused) { 35 return MultiplyHigh(GetHostTicksElapsed(), gputick_rdtsc_factor);
117 TimePoint current_time_point{};
118 TimePoint new_time_point{};
119
120 current_time_point.pack = Common::AtomicLoad128(time_point.pack.data());
121 do {
122 new_time_point.pack = current_time_point.pack;
123 new_time_point.inner.last_measure = FencedRDTSC();
124 } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack,
125 current_time_point.pack, current_time_point.pack));
126 }
127} 36}
128 37
129std::chrono::nanoseconds NativeClock::GetTimeNS() { 38u64 NativeClock::GetHostTicksNow() const {
130 const u64 rtsc_value = GetRTSC(); 39 return FencedRDTSC();
131 return std::chrono::nanoseconds{MultiplyHigh(rtsc_value, ns_rtsc_factor)};
132} 40}
133 41
134std::chrono::microseconds NativeClock::GetTimeUS() { 42u64 NativeClock::GetHostTicksElapsed() const {
135 const u64 rtsc_value = GetRTSC(); 43 return FencedRDTSC() - start_ticks;
136 return std::chrono::microseconds{MultiplyHigh(rtsc_value, us_rtsc_factor)};
137} 44}
138 45
139std::chrono::milliseconds NativeClock::GetTimeMS() { 46bool NativeClock::IsNative() const {
140 const u64 rtsc_value = GetRTSC(); 47 return true;
141 return std::chrono::milliseconds{MultiplyHigh(rtsc_value, ms_rtsc_factor)};
142} 48}
143 49
144u64 NativeClock::GetClockCycles() { 50} // namespace Common::X64
145 const u64 rtsc_value = GetRTSC();
146 return MultiplyHigh(rtsc_value, clock_rtsc_factor);
147}
148
149u64 NativeClock::GetCPUCycles() {
150 const u64 rtsc_value = GetRTSC();
151 return MultiplyHigh(rtsc_value, cpu_rtsc_factor);
152}
153
154void NativeClock::CalculateAndSetFactors() {
155 ns_rtsc_factor = GetFixedPoint64Factor(NS_RATIO, rtsc_frequency);
156 us_rtsc_factor = GetFixedPoint64Factor(US_RATIO, rtsc_frequency);
157 ms_rtsc_factor = GetFixedPoint64Factor(MS_RATIO, rtsc_frequency);
158 clock_rtsc_factor = GetFixedPoint64Factor(emulated_clock_frequency, rtsc_frequency);
159 cpu_rtsc_factor = GetFixedPoint64Factor(emulated_cpu_frequency, rtsc_frequency);
160}
161
162} // namespace X64
163
164} // namespace Common
diff --git a/src/common/x64/native_clock.h b/src/common/x64/native_clock.h
index 03ca291d8..334415eff 100644
--- a/src/common/x64/native_clock.h
+++ b/src/common/x64/native_clock.h
@@ -3,58 +3,39 @@
3 3
4#pragma once 4#pragma once
5 5
6#include "common/polyfill_thread.h"
7#include "common/wall_clock.h" 6#include "common/wall_clock.h"
8 7
9namespace Common { 8namespace Common::X64 {
10 9
11namespace X64 {
12class NativeClock final : public WallClock { 10class NativeClock final : public WallClock {
13public: 11public:
14 explicit NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequency_, 12 explicit NativeClock(u64 rdtsc_frequency_);
15 u64 rtsc_frequency_);
16 13
17 std::chrono::nanoseconds GetTimeNS() override; 14 std::chrono::nanoseconds GetTimeNS() const override;
18 15
19 std::chrono::microseconds GetTimeUS() override; 16 std::chrono::microseconds GetTimeUS() const override;
20 17
21 std::chrono::milliseconds GetTimeMS() override; 18 std::chrono::milliseconds GetTimeMS() const override;
22 19
23 u64 GetClockCycles() override; 20 u64 GetCNTPCT() const override;
24 21
25 u64 GetCPUCycles() override; 22 u64 GetGPUTick() const override;
26 23
27 void Pause(bool is_paused) override; 24 u64 GetHostTicksNow() const override;
28 25
29private: 26 u64 GetHostTicksElapsed() const override;
30 u64 GetRTSC();
31
32 void CalculateAndSetFactors();
33
34 union alignas(16) TimePoint {
35 TimePoint() : pack{} {}
36 u128 pack{};
37 struct Inner {
38 u64 last_measure{};
39 u64 accumulated_ticks{};
40 } inner;
41 };
42
43 TimePoint time_point;
44 27
45 // factors 28 bool IsNative() const override;
46 u64 clock_rtsc_factor{};
47 u64 cpu_rtsc_factor{};
48 u64 ns_rtsc_factor{};
49 u64 us_rtsc_factor{};
50 u64 ms_rtsc_factor{};
51 29
52 u64 rtsc_frequency; 30private:
53 31 u64 start_ticks;
54 std::jthread time_sync_thread; 32 u64 rdtsc_frequency;
33
34 u64 ns_rdtsc_factor;
35 u64 us_rdtsc_factor;
36 u64 ms_rdtsc_factor;
37 u64 cntpct_rdtsc_factor;
38 u64 gputick_rdtsc_factor;
55}; 39};
56} // namespace X64
57
58u64 EstimateRDTSCFrequency();
59 40
60} // namespace Common 41} // namespace Common::X64
diff --git a/src/common/x64/rdtsc.cpp b/src/common/x64/rdtsc.cpp
new file mode 100644
index 000000000..9273274a3
--- /dev/null
+++ b/src/common/x64/rdtsc.cpp
@@ -0,0 +1,39 @@
1// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
2// SPDX-License-Identifier: GPL-2.0-or-later
3
4#include <thread>
5
6#include "common/steady_clock.h"
7#include "common/uint128.h"
8#include "common/x64/rdtsc.h"
9
10namespace Common::X64 {
11
12template <u64 Nearest>
13static u64 RoundToNearest(u64 value) {
14 const auto mod = value % Nearest;
15 return mod >= (Nearest / 2) ? (value - mod + Nearest) : (value - mod);
16}
17
18u64 EstimateRDTSCFrequency() {
19 // Discard the first result measuring the rdtsc.
20 FencedRDTSC();
21 std::this_thread::sleep_for(std::chrono::milliseconds{1});
22 FencedRDTSC();
23
24 // Get the current time.
25 const auto start_time = RealTimeClock::Now();
26 const u64 tsc_start = FencedRDTSC();
27 // Wait for 100 milliseconds.
28 std::this_thread::sleep_for(std::chrono::milliseconds{100});
29 const auto end_time = RealTimeClock::Now();
30 const u64 tsc_end = FencedRDTSC();
31 // Calculate differences.
32 const u64 timer_diff = static_cast<u64>(
33 std::chrono::duration_cast<std::chrono::nanoseconds>(end_time - start_time).count());
34 const u64 tsc_diff = tsc_end - tsc_start;
35 const u64 tsc_freq = MultiplyAndDivide64(tsc_diff, 1000000000ULL, timer_diff);
36 return RoundToNearest<100'000>(tsc_freq);
37}
38
39} // namespace Common::X64
diff --git a/src/common/x64/rdtsc.h b/src/common/x64/rdtsc.h
new file mode 100644
index 000000000..0ec4f52f9
--- /dev/null
+++ b/src/common/x64/rdtsc.h
@@ -0,0 +1,37 @@
1// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
2// SPDX-License-Identifier: GPL-2.0-or-later
3
4#pragma once
5
6#ifdef _MSC_VER
7#include <intrin.h>
8#endif
9
10#include "common/common_types.h"
11
12namespace Common::X64 {
13
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
24static inline 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
35u64 EstimateRDTSCFrequency();
36
37} // namespace Common::X64