From 2856fadaa016c6fe378c5ac377aedec156131803 Mon Sep 17 00:00:00 2001 From: Morph <39850852+Morph1984@users.noreply.github.com> Date: Sun, 23 Apr 2023 00:01:08 -0400 Subject: core_timing: Use CNTPCT as the guest CPU tick Previously, we were mixing the raw CPU frequency and CNTFRQ. The raw CPU frequency (1020 MHz) should've never been used as CNTPCT (whose frequency is CNTFRQ) is the only counter available. --- src/core/core_timing.cpp | 35 ++++++++--------------------------- 1 file changed, 8 insertions(+), 27 deletions(-) (limited to 'src/core/core_timing.cpp') diff --git a/src/core/core_timing.cpp b/src/core/core_timing.cpp index 4f2692b05c..9a1d5a69aa 100644 --- a/src/core/core_timing.cpp +++ b/src/core/core_timing.cpp @@ -16,7 +16,6 @@ #include "common/microprofile.h" #include "core/core_timing.h" -#include "core/core_timing_util.h" #include "core/hardware_properties.h" namespace Core::Timing { @@ -45,9 +44,7 @@ struct CoreTiming::Event { } }; -CoreTiming::CoreTiming() - : cpu_clock{Common::CreateBestMatchingClock(Hardware::BASE_CLOCK_RATE, Hardware::CNTFREQ)}, - event_clock{Common::CreateStandardWallClock(Hardware::BASE_CLOCK_RATE, Hardware::CNTFREQ)} {} +CoreTiming::CoreTiming() : clock{Common::CreateOptimalClock()} {} CoreTiming::~CoreTiming() { Reset(); @@ -180,7 +177,7 @@ void CoreTiming::AddTicks(u64 ticks_to_add) { void CoreTiming::Idle() { if (!event_queue.empty()) { const u64 next_event_time = event_queue.front().time; - const u64 next_ticks = nsToCycles(std::chrono::nanoseconds(next_event_time)) + 10U; + const u64 next_ticks = Common::WallClock::NSToCNTPCT(next_event_time) + 10U; if (next_ticks > ticks) { ticks = next_ticks; } @@ -193,18 +190,11 @@ void CoreTiming::ResetTicks() { downcount = MAX_SLICE_LENGTH; } -u64 CoreTiming::GetCPUTicks() const { - if (is_multicore) [[likely]] { - return cpu_clock->GetCPUCycles(); - } - return ticks; -} - u64 CoreTiming::GetClockTicks() const { if (is_multicore) [[likely]] { - return cpu_clock->GetClockCycles(); + return clock->GetCNTPCT(); } - return CpuCyclesToClockCycles(ticks); + return ticks; } std::optional CoreTiming::Advance() { @@ -297,9 +287,7 @@ void CoreTiming::ThreadLoop() { } paused_set = true; - event_clock->Pause(true); pause_event.Wait(); - event_clock->Pause(false); } } @@ -315,25 +303,18 @@ void CoreTiming::Reset() { has_started = false; } -std::chrono::nanoseconds CoreTiming::GetCPUTimeNs() const { - if (is_multicore) [[likely]] { - return cpu_clock->GetTimeNS(); - } - return CyclesToNs(ticks); -} - std::chrono::nanoseconds CoreTiming::GetGlobalTimeNs() const { if (is_multicore) [[likely]] { - return event_clock->GetTimeNS(); + return clock->GetTimeNS(); } - return CyclesToNs(ticks); + return std::chrono::nanoseconds{Common::WallClock::CNTPCTToNS(ticks)}; } std::chrono::microseconds CoreTiming::GetGlobalTimeUs() const { if (is_multicore) [[likely]] { - return event_clock->GetTimeUS(); + return clock->GetTimeUS(); } - return CyclesToUs(ticks); + return std::chrono::microseconds{Common::WallClock::CNTPCTToUS(ticks)}; } } // namespace Core::Timing -- cgit v1.2.3 From 1b83c7eab4a1d69526c89dd4b4931907cfa121db Mon Sep 17 00:00:00 2001 From: Morph <39850852+Morph1984@users.noreply.github.com> Date: Sun, 28 May 2023 17:45:47 -0400 Subject: (wall, native)_clock: Add GetGPUTick Allows us to directly calculate the GPU tick without double conversion to and from the host clock tick. --- src/common/wall_clock.cpp | 8 ++++++-- src/common/wall_clock.h | 20 +++++++++++++++++++- src/common/x64/native_clock.cpp | 7 ++++++- src/common/x64/native_clock.h | 3 +++ src/core/core_timing.cpp | 7 +++++++ src/core/core_timing.h | 3 +++ src/video_core/gpu.cpp | 11 +++-------- 7 files changed, 47 insertions(+), 12 deletions(-) (limited to 'src/core/core_timing.cpp') diff --git a/src/common/wall_clock.cpp b/src/common/wall_clock.cpp index ad8db06b08..dc0dcbd687 100644 --- a/src/common/wall_clock.cpp +++ b/src/common/wall_clock.cpp @@ -32,6 +32,10 @@ public: return GetHostTicksElapsed() * NsToCNTPCTRatio::num / NsToCNTPCTRatio::den; } + u64 GetGPUTick() const override { + return GetHostTicksElapsed() * NsToGPUTickRatio::num / NsToGPUTickRatio::den; + } + u64 GetHostTicksNow() const override { return static_cast(SteadyClock::Now().time_since_epoch().count()); } @@ -52,12 +56,12 @@ std::unique_ptr CreateOptimalClock() { #ifdef ARCHITECTURE_x86_64 const auto& caps = GetCPUCaps(); - if (caps.invariant_tsc && caps.tsc_frequency >= WallClock::CNTFRQ) { + if (caps.invariant_tsc && caps.tsc_frequency >= WallClock::GPUTickFreq) { return std::make_unique(caps.tsc_frequency); } else { // Fallback to StandardWallClock if the hardware TSC // - Is not invariant - // - Is not more precise than CNTFRQ + // - Is not more precise than GPUTickFreq return std::make_unique(); } #else diff --git a/src/common/wall_clock.h b/src/common/wall_clock.h index 56c18ca250..fcfdd637ca 100644 --- a/src/common/wall_clock.h +++ b/src/common/wall_clock.h @@ -13,7 +13,8 @@ namespace Common { class WallClock { public: - static constexpr u64 CNTFRQ = 19'200'000; // CNTPCT_EL0 Frequency = 19.2 MHz + static constexpr u64 CNTFRQ = 19'200'000; // CNTPCT_EL0 Frequency = 19.2 MHz + static constexpr u64 GPUTickFreq = 614'400'000; // GM20B GPU Tick Frequency = 614.4 MHz virtual ~WallClock() = default; @@ -29,6 +30,9 @@ public: /// @returns The guest CNTPCT ticks since the construction of this clock. virtual u64 GetCNTPCT() const = 0; + /// @returns The guest GPU ticks since the construction of this clock. + virtual u64 GetGPUTick() const = 0; + /// @returns The raw host timer ticks since an indeterminate epoch. virtual u64 GetHostTicksNow() const = 0; @@ -46,6 +50,10 @@ public: return us * UsToCNTPCTRatio::num / UsToCNTPCTRatio::den; } + static inline u64 NSToGPUTick(u64 ns) { + return ns * NsToGPUTickRatio::num / NsToGPUTickRatio::den; + } + static inline u64 CNTPCTToNS(u64 cntpct) { return cntpct * NsToCNTPCTRatio::den / NsToCNTPCTRatio::num; } @@ -54,6 +62,14 @@ public: return cntpct * UsToCNTPCTRatio::den / UsToCNTPCTRatio::num; } + static inline u64 GPUTickToNS(u64 gpu_tick) { + return gpu_tick * NsToGPUTickRatio::den / NsToGPUTickRatio::num; + } + + static inline u64 CNTPCTToGPUTick(u64 cntpct) { + return cntpct * CNTPCTToGPUTickRatio::num / CNTPCTToGPUTickRatio::den; + } + protected: using NsRatio = std::nano; using UsRatio = std::micro; @@ -63,6 +79,8 @@ protected: using NsToMsRatio = std::ratio_divide; using NsToCNTPCTRatio = std::ratio; using UsToCNTPCTRatio = std::ratio; + using NsToGPUTickRatio = std::ratio; + using CNTPCTToGPUTickRatio = std::ratio; }; std::unique_ptr CreateOptimalClock(); diff --git a/src/common/x64/native_clock.cpp b/src/common/x64/native_clock.cpp index 5d1eb05907..7d2a26bd9a 100644 --- a/src/common/x64/native_clock.cpp +++ b/src/common/x64/native_clock.cpp @@ -12,7 +12,8 @@ NativeClock::NativeClock(u64 rdtsc_frequency_) ns_rdtsc_factor{GetFixedPoint64Factor(NsRatio::den, rdtsc_frequency)}, us_rdtsc_factor{GetFixedPoint64Factor(UsRatio::den, rdtsc_frequency)}, ms_rdtsc_factor{GetFixedPoint64Factor(MsRatio::den, rdtsc_frequency)}, - cntpct_rdtsc_factor{GetFixedPoint64Factor(CNTFRQ, rdtsc_frequency)} {} + cntpct_rdtsc_factor{GetFixedPoint64Factor(CNTFRQ, rdtsc_frequency)}, + gputick_rdtsc_factor{GetFixedPoint64Factor(GPUTickFreq, rdtsc_frequency)} {} std::chrono::nanoseconds NativeClock::GetTimeNS() const { return std::chrono::nanoseconds{MultiplyHigh(GetHostTicksElapsed(), ns_rdtsc_factor)}; @@ -30,6 +31,10 @@ u64 NativeClock::GetCNTPCT() const { return MultiplyHigh(GetHostTicksElapsed(), cntpct_rdtsc_factor); } +u64 NativeClock::GetGPUTick() const { + return MultiplyHigh(GetHostTicksElapsed(), gputick_rdtsc_factor); +} + u64 NativeClock::GetHostTicksNow() const { return FencedRDTSC(); } diff --git a/src/common/x64/native_clock.h b/src/common/x64/native_clock.h index d6f8626c1e..334415eff3 100644 --- a/src/common/x64/native_clock.h +++ b/src/common/x64/native_clock.h @@ -19,6 +19,8 @@ public: u64 GetCNTPCT() const override; + u64 GetGPUTick() const override; + u64 GetHostTicksNow() const override; u64 GetHostTicksElapsed() const override; @@ -33,6 +35,7 @@ private: u64 us_rdtsc_factor; u64 ms_rdtsc_factor; u64 cntpct_rdtsc_factor; + u64 gputick_rdtsc_factor; }; } // namespace Common::X64 diff --git a/src/core/core_timing.cpp b/src/core/core_timing.cpp index 9a1d5a69aa..e57bca70af 100644 --- a/src/core/core_timing.cpp +++ b/src/core/core_timing.cpp @@ -197,6 +197,13 @@ u64 CoreTiming::GetClockTicks() const { return ticks; } +u64 CoreTiming::GetGPUTicks() const { + if (is_multicore) [[likely]] { + return clock->GetGPUTick(); + } + return Common::WallClock::CNTPCTToGPUTick(ticks); +} + std::optional CoreTiming::Advance() { std::scoped_lock lock{advance_lock, basic_lock}; global_timer = GetGlobalTimeNs().count(); diff --git a/src/core/core_timing.h b/src/core/core_timing.h index fdacdd94a6..1873852c40 100644 --- a/src/core/core_timing.h +++ b/src/core/core_timing.h @@ -119,6 +119,9 @@ public: /// Returns the current CNTPCT tick value. u64 GetClockTicks() const; + /// Returns the current GPU tick value. + u64 GetGPUTicks() const; + /// Returns current time in microseconds. std::chrono::microseconds GetGlobalTimeUs() const; diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp index 70762c51a4..db385076da 100644 --- a/src/video_core/gpu.cpp +++ b/src/video_core/gpu.cpp @@ -193,18 +193,13 @@ struct GPU::Impl { } [[nodiscard]] u64 GetTicks() const { - // This values were reversed engineered by fincs from NVN - // The GPU clock is 614.4 MHz - using NsToGPUTickRatio = std::ratio<614'400'000, std::nano::den>; - static_assert(NsToGPUTickRatio::num == 384 && NsToGPUTickRatio::den == 625); - - u64 nanoseconds = system.CoreTiming().GetGlobalTimeNs().count(); + u64 gpu_tick = system.CoreTiming().GetGPUTicks(); if (Settings::values.use_fast_gpu_time.GetValue()) { - nanoseconds /= 256; + gpu_tick /= 256; } - return nanoseconds * NsToGPUTickRatio::num / NsToGPUTickRatio::den; + return gpu_tick; } [[nodiscard]] bool IsAsync() const { -- cgit v1.2.3 From 52a219cfe97f0667a9f2e815ec6ff592334f4293 Mon Sep 17 00:00:00 2001 From: Morph <39850852+Morph1984@users.noreply.github.com> Date: Wed, 7 Jun 2023 21:38:28 -0400 Subject: core_timing: Fix SingleCore cycle timer --- src/common/wall_clock.h | 36 ++++++++++++++++++++---------------- src/core/core_timing.cpp | 26 +++++++++----------------- src/core/core_timing.h | 2 +- src/core/hle/kernel/svc/svc_tick.cpp | 10 +--------- 4 files changed, 31 insertions(+), 43 deletions(-) (limited to 'src/core/core_timing.cpp') diff --git a/src/common/wall_clock.h b/src/common/wall_clock.h index fcfdd637ca..f45d3d8c52 100644 --- a/src/common/wall_clock.h +++ b/src/common/wall_clock.h @@ -13,8 +13,9 @@ namespace Common { class WallClock { public: - static constexpr u64 CNTFRQ = 19'200'000; // CNTPCT_EL0 Frequency = 19.2 MHz - static constexpr u64 GPUTickFreq = 614'400'000; // GM20B GPU Tick Frequency = 614.4 MHz + static constexpr u64 CNTFRQ = 19'200'000; // CNTPCT_EL0 Frequency = 19.2 MHz + static constexpr u64 GPUTickFreq = 614'400'000; // GM20B GPU Tick Frequency = 614.4 MHz + static constexpr u64 CPUTickFreq = 1'020'000'000; // T210/4 A57 CPU Tick Frequency = 1020.0 MHz virtual ~WallClock() = default; @@ -46,28 +47,26 @@ public: return ns * NsToCNTPCTRatio::num / NsToCNTPCTRatio::den; } - static inline u64 USToCNTPCT(u64 us) { - return us * UsToCNTPCTRatio::num / UsToCNTPCTRatio::den; - } - static inline u64 NSToGPUTick(u64 ns) { return ns * NsToGPUTickRatio::num / NsToGPUTickRatio::den; } - static inline u64 CNTPCTToNS(u64 cntpct) { - return cntpct * NsToCNTPCTRatio::den / NsToCNTPCTRatio::num; + // Cycle Timing + + static inline u64 CPUTickToNS(u64 cpu_tick) { + return cpu_tick * CPUTickToNsRatio::num / CPUTickToNsRatio::den; } - static inline u64 CNTPCTToUS(u64 cntpct) { - return cntpct * UsToCNTPCTRatio::den / UsToCNTPCTRatio::num; + static inline u64 CPUTickToUS(u64 cpu_tick) { + return cpu_tick * CPUTickToUsRatio::num / CPUTickToUsRatio::den; } - static inline u64 GPUTickToNS(u64 gpu_tick) { - return gpu_tick * NsToGPUTickRatio::den / NsToGPUTickRatio::num; + static inline u64 CPUTickToCNTPCT(u64 cpu_tick) { + return cpu_tick * CPUTickToCNTPCTRatio::num / CPUTickToCNTPCTRatio::den; } - static inline u64 CNTPCTToGPUTick(u64 cntpct) { - return cntpct * CNTPCTToGPUTickRatio::num / CNTPCTToGPUTickRatio::den; + static inline u64 CPUTickToGPUTick(u64 cpu_tick) { + return cpu_tick * CPUTickToGPUTickRatio::num / CPUTickToGPUTickRatio::den; } protected: @@ -78,9 +77,14 @@ protected: using NsToUsRatio = std::ratio_divide; using NsToMsRatio = std::ratio_divide; using NsToCNTPCTRatio = std::ratio; - using UsToCNTPCTRatio = std::ratio; using NsToGPUTickRatio = std::ratio; - using CNTPCTToGPUTickRatio = std::ratio; + + // Cycle Timing + + using CPUTickToNsRatio = std::ratio; + using CPUTickToUsRatio = std::ratio; + using CPUTickToCNTPCTRatio = std::ratio; + using CPUTickToGPUTickRatio = std::ratio; }; std::unique_ptr CreateOptimalClock(); diff --git a/src/core/core_timing.cpp b/src/core/core_timing.cpp index e57bca70af..4f0a3f8eab 100644 --- a/src/core/core_timing.cpp +++ b/src/core/core_timing.cpp @@ -20,7 +20,7 @@ namespace Core::Timing { -constexpr s64 MAX_SLICE_LENGTH = 4000; +constexpr s64 MAX_SLICE_LENGTH = 10000; std::shared_ptr CreateEvent(std::string name, TimedCallback&& callback) { return std::make_shared(std::move(callback), std::move(name)); @@ -65,7 +65,7 @@ void CoreTiming::Initialize(std::function&& on_thread_init_) { on_thread_init = std::move(on_thread_init_); event_fifo_id = 0; shutting_down = false; - ticks = 0; + cpu_ticks = 0; const auto empty_timed_callback = [](std::uintptr_t, u64, std::chrono::nanoseconds) -> std::optional { return std::nullopt; }; ev_lost = CreateEvent("_lost_event", empty_timed_callback); @@ -170,20 +170,12 @@ void CoreTiming::UnscheduleEvent(const std::shared_ptr& event_type, } void CoreTiming::AddTicks(u64 ticks_to_add) { - ticks += ticks_to_add; - downcount -= static_cast(ticks); + cpu_ticks += ticks_to_add; + downcount -= static_cast(cpu_ticks); } void CoreTiming::Idle() { - if (!event_queue.empty()) { - const u64 next_event_time = event_queue.front().time; - const u64 next_ticks = Common::WallClock::NSToCNTPCT(next_event_time) + 10U; - if (next_ticks > ticks) { - ticks = next_ticks; - } - return; - } - ticks += 1000U; + cpu_ticks += 1000U; } void CoreTiming::ResetTicks() { @@ -194,14 +186,14 @@ u64 CoreTiming::GetClockTicks() const { if (is_multicore) [[likely]] { return clock->GetCNTPCT(); } - return ticks; + return Common::WallClock::CPUTickToCNTPCT(cpu_ticks); } u64 CoreTiming::GetGPUTicks() const { if (is_multicore) [[likely]] { return clock->GetGPUTick(); } - return Common::WallClock::CNTPCTToGPUTick(ticks); + return Common::WallClock::CPUTickToGPUTick(cpu_ticks); } std::optional CoreTiming::Advance() { @@ -314,14 +306,14 @@ std::chrono::nanoseconds CoreTiming::GetGlobalTimeNs() const { if (is_multicore) [[likely]] { return clock->GetTimeNS(); } - return std::chrono::nanoseconds{Common::WallClock::CNTPCTToNS(ticks)}; + return std::chrono::nanoseconds{Common::WallClock::CPUTickToNS(cpu_ticks)}; } std::chrono::microseconds CoreTiming::GetGlobalTimeUs() const { if (is_multicore) [[likely]] { return clock->GetTimeUS(); } - return std::chrono::microseconds{Common::WallClock::CNTPCTToUS(ticks)}; + return std::chrono::microseconds{Common::WallClock::CPUTickToUS(cpu_ticks)}; } } // namespace Core::Timing diff --git a/src/core/core_timing.h b/src/core/core_timing.h index 1873852c40..10db1de552 100644 --- a/src/core/core_timing.h +++ b/src/core/core_timing.h @@ -167,7 +167,7 @@ private: s64 pause_end_time{}; /// Cycle timing - u64 ticks{}; + u64 cpu_ticks{}; s64 downcount{}; }; diff --git a/src/core/hle/kernel/svc/svc_tick.cpp b/src/core/hle/kernel/svc/svc_tick.cpp index 5613364821..7dd7c6e518 100644 --- a/src/core/hle/kernel/svc/svc_tick.cpp +++ b/src/core/hle/kernel/svc/svc_tick.cpp @@ -12,16 +12,8 @@ namespace Kernel::Svc { int64_t GetSystemTick(Core::System& system) { LOG_TRACE(Kernel_SVC, "called"); - auto& core_timing = system.CoreTiming(); - // Returns the value of cntpct_el0 (https://switchbrew.org/wiki/SVC#svcGetSystemTick) - const u64 result{core_timing.GetClockTicks()}; - - if (!system.Kernel().IsMulticore()) { - core_timing.AddTicks(400U); - } - - return static_cast(result); + return static_cast(system.CoreTiming().GetClockTicks()); } int64_t GetSystemTick64(Core::System& system) { -- cgit v1.2.3