Nikilite/suyu
1
1
Fork 0
forked from suyu/suyu
Code Pull requests Activity

Merge pull request #5121 from bunnei/optimize-core-timing

Browse source 
core: Optimize core timing utility functions to avoid unnecessary math
This commit is contained in:
bunnei 2021-02-16 13:17:22 -08:00 committed by GitHub
commit 6be0975bf2
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
8 changed files with 140 additions and 240 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
src/common/CMakeLists.txt
View file

@ -168,7 +168,6 @@ add_library(common STATIC
time_zone.cpp
time_zone.h
tree.h
uint128.cpp
uint128.h
uuid.cpp
uuid.h

71
src/common/uint128.cpp
View file

@ -1,71 +0,0 @@
// Copyright 2019 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#ifdef _MSC_VER
#include <intrin.h>
#pragma intrinsic(_umul128)
#pragma intrinsic(_udiv128)
#endif
#include <cstring>
#include "common/uint128.h"
namespace Common {
#ifdef _MSC_VER
u64 MultiplyAndDivide64(u64 a, u64 b, u64 d) {
u128 r{};
r[0] = _umul128(a, b, &r[1]);
u64 remainder;
#if _MSC_VER < 1923
return udiv128(r[1], r[0], d, &remainder);
#else
return _udiv128(r[1], r[0], d, &remainder);
#endif
}
#else
u64 MultiplyAndDivide64(u64 a, u64 b, u64 d) {
const u64 diva = a / d;
const u64 moda = a % d;
const u64 divb = b / d;
const u64 modb = b % d;
return diva * b + moda * divb + moda * modb / d;
}
#endif
u128 Multiply64Into128(u64 a, u64 b) {
u128 result;
#ifdef _MSC_VER
result[0] = _umul128(a, b, &result[1]);
#else
unsigned __int128 tmp = a;
tmp *= b;
std::memcpy(&result, &tmp, sizeof(u128));
#endif
return result;
}
std::pair<u64, u64> Divide128On32(u128 dividend, u32 divisor) {
u64 remainder = dividend[0] % divisor;
u64 accum = dividend[0] / divisor;
if (dividend[1] == 0)
return {accum, remainder};
// We ignore dividend[1] / divisor as that overflows
const u64 first_segment = (dividend[1] % divisor) << 32;
accum += (first_segment / divisor) << 32;
const u64 second_segment = (first_segment % divisor) << 32;
accum += (second_segment / divisor);
remainder += second_segment % divisor;
if (remainder >= divisor) {
accum++;
remainder -= divisor;
}
return {accum, remainder};
}
} // namespace Common

89
src/common/uint128.h
View file

@ -4,19 +4,98 @@
#pragma once
#include <cstring>
#include <utility>
#ifdef _MSC_VER
#include <intrin.h>
#pragma intrinsic(__umulh)
#pragma intrinsic(_umul128)
#pragma intrinsic(_udiv128)
#else
#include <x86intrin.h>
#endif
#include "common/common_types.h"
namespace Common {
// This function multiplies 2 u64 values and divides it by a u64 value.
[[nodiscard]] u64 MultiplyAndDivide64(u64 a, u64 b, u64 d);
[[nodiscard]] static inline u64 MultiplyAndDivide64(u64 a, u64 b, u64 d) {
#ifdef _MSC_VER
u128 r{};
r[0] = _umul128(a, b, &r[1]);
u64 remainder;
#if _MSC_VER < 1923
return udiv128(r[1], r[0], d, &remainder);
#else
return _udiv128(r[1], r[0], d, &remainder);
#endif
#else
const u64 diva = a / d;
const u64 moda = a % d;
const u64 divb = b / d;
const u64 modb = b % d;
return diva * b + moda * divb + moda * modb / d;
#endif
}
// This function multiplies 2 u64 values and produces a u128 value;
[[nodiscard]] u128 Multiply64Into128(u64 a, u64 b);
[[nodiscard]] static inline u128 Multiply64Into128(u64 a, u64 b) {
u128 result;
#ifdef _MSC_VER
result[0] = _umul128(a, b, &result[1]);
#else
unsigned __int128 tmp = a;
tmp *= b;
std::memcpy(&result, &tmp, sizeof(u128));
#endif
return result;
}
// This function divides a u128 by a u32 value and produces two u64 values:
// the result of division and the remainder
[[nodiscard]] std::pair<u64, u64> Divide128On32(u128 dividend, u32 divisor);
[[nodiscard]] static inline u64 GetFixedPoint64Factor(u64 numerator, u64 divisor) {
#ifdef __SIZEOF_INT128__
const auto base = static_cast<unsigned __int128>(numerator) << 64ULL;
return static_cast<u64>(base / divisor);
#elif defined(_M_X64) || defined(_M_ARM64)
std::array<u64, 2> r = {0, numerator};
u64 remainder;
#if _MSC_VER < 1923
return udiv128(r[1], r[0], divisor, &remainder);
#else
return _udiv128(r[1], r[0], divisor, &remainder);
#endif
#else
// This one is bit more inaccurate.
return MultiplyAndDivide64(std::numeric_limits<u64>::max(), numerator, divisor);
#endif
}
[[nodiscard]] static inline u64 MultiplyHigh(u64 a, u64 b) {
#ifdef __SIZEOF_INT128__
return (static_cast<unsigned __int128>(a) * static_cast<unsigned __int128>(b)) >> 64;
#elif defined(_M_X64) || defined(_M_ARM64)
return __umulh(a, b); // MSVC
#else
// Generic fallback
const u64 a_lo = u32(a);
const u64 a_hi = a >> 32;
const u64 b_lo = u32(b);
const u64 b_hi = b >> 32;
const u64 a_x_b_hi = a_hi * b_hi;
const u64 a_x_b_mid = a_hi * b_lo;
const u64 b_x_a_mid = b_hi * a_lo;
const u64 a_x_b_lo = a_lo * b_lo;
const u64 carry_bit = (static_cast<u64>(static_cast<u32>(a_x_b_mid)) +
static_cast<u64>(static_cast<u32>(b_x_a_mid)) + (a_x_b_lo >> 32)) >>
32;
const u64 multhi = a_x_b_hi + (a_x_b_mid >> 32) + (b_x_a_mid >> 32) + carry_bit;
return multhi;
#endif
}
} // namespace Common

17
src/common/wall_clock.cpp
View file

@ -2,6 +2,8 @@
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include <cstdint>
#include "common/uint128.h"
#include "common/wall_clock.h"
@ -18,7 +20,9 @@ using base_time_point = std::chrono::time_point<base_timer>;
class StandardWallClock final : public WallClock {
public:
explicit StandardWallClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequency_)
: WallClock(emulated_cpu_frequency_, emulated_clock_frequency_, false) {
: WallClock(emulated_cpu_frequency_, emulated_clock_frequency_, false),
emulated_clock_factor{GetFixedPoint64Factor(emulated_clock_frequency, 1000000000)},
emulated_cpu_factor{GetFixedPoint64Factor(emulated_cpu_frequency, 1000000000)} {
start_time = base_timer::now();
}
@ -41,16 +45,11 @@ public:
}
u64 GetClockCycles() override {
std::chrono::nanoseconds time_now = GetTimeNS();
const u128 temporary =
Common::Multiply64Into128(time_now.count(), emulated_clock_frequency);
return Common::Divide128On32(temporary, 1000000000).first;
return MultiplyHigh(GetTimeNS().count(), emulated_clock_factor);
}
u64 GetCPUCycles() override {
std::chrono::nanoseconds time_now = GetTimeNS();
const u128 temporary = Common::Multiply64Into128(time_now.count(), emulated_cpu_frequency);
return Common::Divide128On32(temporary, 1000000000).first;
return MultiplyHigh(GetTimeNS().count(), emulated_cpu_factor);
}
void Pause([[maybe_unused]] bool is_paused) override {
@ -59,6 +58,8 @@ public:
private:
base_time_point start_time;
const u64 emulated_clock_factor;
const u64 emulated_cpu_factor;
};
#ifdef ARCHITECTURE_x86_64

58
src/common/x64/native_clock.cpp
View file

@ -8,68 +8,10 @@
#include <mutex>
#include <thread>
#ifdef _MSC_VER
#include <intrin.h>
#pragma intrinsic(__umulh)
#pragma intrinsic(_udiv128)
#else
#include <x86intrin.h>
#endif
#include "common/atomic_ops.h"
#include "common/uint128.h"
#include "common/x64/native_clock.h"
namespace {
[[nodiscard]] u64 GetFixedPoint64Factor(u64 numerator, u64 divisor) {
#ifdef __SIZEOF_INT128__
const auto base = static_cast<unsigned __int128>(numerator) << 64ULL;
return static_cast<u64>(base / divisor);
#elif defined(_M_X64) || defined(_M_ARM64)
std::array<u64, 2> r = {0, numerator};
u64 remainder;
#if _MSC_VER < 1923
return udiv128(r[1], r[0], divisor, &remainder);
#else
return _udiv128(r[1], r[0], divisor, &remainder);
#endif
#else
// This one is bit more inaccurate.
return MultiplyAndDivide64(std::numeric_limits<u64>::max(), numerator, divisor);
#endif
}
[[nodiscard]] u64 MultiplyHigh(u64 a, u64 b) {
#ifdef __SIZEOF_INT128__
return (static_cast<unsigned __int128>(a) * static_cast<unsigned __int128>(b)) >> 64;
#elif defined(_M_X64) || defined(_M_ARM64)
return __umulh(a, b); // MSVC
#else
// Generic fallback
const u64 a_lo = u32(a);
const u64 a_hi = a >> 32;
const u64 b_lo = u32(b);
const u64 b_hi = b >> 32;
const u64 a_x_b_hi = a_hi * b_hi;
const u64 a_x_b_mid = a_hi * b_lo;
const u64 b_x_a_mid = b_hi * a_lo;
const u64 a_x_b_lo = a_lo * b_lo;
const u64 carry_bit = (static_cast<u64>(static_cast<u32>(a_x_b_mid)) +
static_cast<u64>(static_cast<u32>(b_x_a_mid)) + (a_x_b_lo >> 32)) >>
32;
const u64 multhi = a_x_b_hi + (a_x_b_mid >> 32) + (b_x_a_mid >> 32) + carry_bit;
return multhi;
#endif
}
} // namespace
namespace Common {
u64 EstimateRDTSCFrequency() {

1
src/core/CMakeLists.txt
View file

@ -19,7 +19,6 @@ add_library(core STATIC
core.h
core_timing.cpp
core_timing.h
core_timing_util.cpp
core_timing_util.h
cpu_manager.cpp
cpu_manager.h

84
src/core/core_timing_util.cpp
View file

@ -1,84 +0,0 @@
// Copyright 2008 Dolphin Emulator Project / 2017 Citra Emulator Project
// Licensed under GPLv2+
// Refer to the license.txt file included.
#include "core/core_timing_util.h"
#include <cinttypes>
#include <limits>
#include "common/logging/log.h"
#include "common/uint128.h"
#include "core/hardware_properties.h"
namespace Core::Timing {
constexpr u64 MAX_VALUE_TO_MULTIPLY = std::numeric_limits<s64>::max() / Hardware::BASE_CLOCK_RATE;
s64 msToCycles(std::chrono::milliseconds ms) {
if (static_cast<u64>(ms.count() / 1000) > MAX_VALUE_TO_MULTIPLY) {
LOG_ERROR(Core_Timing, "Integer overflow, use max value");
return std::numeric_limits<s64>::max();
}
if (static_cast<u64>(ms.count()) > MAX_VALUE_TO_MULTIPLY) {
LOG_DEBUG(Core_Timing, "Time very big, do rounding");
return Hardware::BASE_CLOCK_RATE * (ms.count() / 1000);
}
return (Hardware::BASE_CLOCK_RATE * ms.count()) / 1000;
}
s64 usToCycles(std::chrono::microseconds us) {
if (static_cast<u64>(us.count() / 1000000) > MAX_VALUE_TO_MULTIPLY) {
LOG_ERROR(Core_Timing, "Integer overflow, use max value");
return std::numeric_limits<s64>::max();
}
if (static_cast<u64>(us.count()) > MAX_VALUE_TO_MULTIPLY) {
LOG_DEBUG(Core_Timing, "Time very big, do rounding");
return Hardware::BASE_CLOCK_RATE * (us.count() / 1000000);
}
return (Hardware::BASE_CLOCK_RATE * us.count()) / 1000000;
}
s64 nsToCycles(std::chrono::nanoseconds ns) {
const u128 temporal = Common::Multiply64Into128(ns.count(), Hardware::BASE_CLOCK_RATE);
return Common::Divide128On32(temporal, static_cast<u32>(1000000000)).first;
}
u64 msToClockCycles(std::chrono::milliseconds ns) {
const u128 temp = Common::Multiply64Into128(ns.count(), Hardware::CNTFREQ);
return Common::Divide128On32(temp, 1000).first;
}
u64 usToClockCycles(std::chrono::microseconds ns) {
const u128 temp = Common::Multiply64Into128(ns.count(), Hardware::CNTFREQ);
return Common::Divide128On32(temp, 1000000).first;
}
u64 nsToClockCycles(std::chrono::nanoseconds ns) {
const u128 temp = Common::Multiply64Into128(ns.count(), Hardware::CNTFREQ);
return Common::Divide128On32(temp, 1000000000).first;
}
u64 CpuCyclesToClockCycles(u64 ticks) {
const u128 temporal = Common::Multiply64Into128(ticks, Hardware::CNTFREQ);
return Common::Divide128On32(temporal, static_cast<u32>(Hardware::BASE_CLOCK_RATE)).first;
}
std::chrono::milliseconds CyclesToMs(s64 cycles) {
const u128 temporal = Common::Multiply64Into128(cycles, 1000);
u64 ms = Common::Divide128On32(temporal, static_cast<u32>(Hardware::BASE_CLOCK_RATE)).first;
return std::chrono::milliseconds(ms);
}
std::chrono::nanoseconds CyclesToNs(s64 cycles) {
const u128 temporal = Common::Multiply64Into128(cycles, 1000000000);
u64 ns = Common::Divide128On32(temporal, static_cast<u32>(Hardware::BASE_CLOCK_RATE)).first;
return std::chrono::nanoseconds(ns);
}
std::chrono::microseconds CyclesToUs(s64 cycles) {
const u128 temporal = Common::Multiply64Into128(cycles, 1000000);
u64 us = Common::Divide128On32(temporal, static_cast<u32>(Hardware::BASE_CLOCK_RATE)).first;
return std::chrono::microseconds(us);
}
} // namespace Core::Timing

59
src/core/core_timing_util.h
View file

@ -1,24 +1,59 @@
// Copyright 2008 Dolphin Emulator Project / 2017 Citra Emulator Project
// Licensed under GPLv2+
// Copyright 2020 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#pragma once
#include <chrono>
#include "common/common_types.h"
#include "core/hardware_properties.h"
namespace Core::Timing {
s64 msToCycles(std::chrono::milliseconds ms);
s64 usToCycles(std::chrono::microseconds us);
s64 nsToCycles(std::chrono::nanoseconds ns);
u64 msToClockCycles(std::chrono::milliseconds ns);
u64 usToClockCycles(std::chrono::microseconds ns);
u64 nsToClockCycles(std::chrono::nanoseconds ns);
std::chrono::milliseconds CyclesToMs(s64 cycles);
std::chrono::nanoseconds CyclesToNs(s64 cycles);
std::chrono::microseconds CyclesToUs(s64 cycles);
namespace detail {
constexpr u64 CNTFREQ_ADJUSTED = Hardware::CNTFREQ / 1000;
constexpr u64 BASE_CLOCK_RATE_ADJUSTED = Hardware::BASE_CLOCK_RATE / 1000;
} // namespace detail
u64 CpuCyclesToClockCycles(u64 ticks);
[[nodiscard]] constexpr s64 msToCycles(std::chrono::milliseconds ms) {
return ms.count() * detail::BASE_CLOCK_RATE_ADJUSTED;
}
[[nodiscard]] constexpr s64 usToCycles(std::chrono::microseconds us) {
return us.count() * detail::BASE_CLOCK_RATE_ADJUSTED / 1000;
}
[[nodiscard]] constexpr s64 nsToCycles(std::chrono::nanoseconds ns) {
return ns.count() * detail::BASE_CLOCK_RATE_ADJUSTED / 1000000;
}
[[nodiscard]] constexpr u64 msToClockCycles(std::chrono::milliseconds ms) {
return static_cast<u64>(ms.count()) * detail::CNTFREQ_ADJUSTED;
}
[[nodiscard]] constexpr u64 usToClockCycles(std::chrono::microseconds us) {
return us.count() * detail::CNTFREQ_ADJUSTED / 1000;
}
[[nodiscard]] constexpr u64 nsToClockCycles(std::chrono::nanoseconds ns) {
return ns.count() * detail::CNTFREQ_ADJUSTED / 1000000;
}
[[nodiscard]] constexpr u64 CpuCyclesToClockCycles(u64 ticks) {
return ticks * detail::CNTFREQ_ADJUSTED / detail::BASE_CLOCK_RATE_ADJUSTED;
}
[[nodiscard]] constexpr std::chrono::milliseconds CyclesToMs(s64 cycles) {
return std::chrono::milliseconds(cycles / detail::BASE_CLOCK_RATE_ADJUSTED);
}
[[nodiscard]] constexpr std::chrono::nanoseconds CyclesToNs(s64 cycles) {
return std::chrono::nanoseconds(cycles * 1000000 / detail::BASE_CLOCK_RATE_ADJUSTED);
}
[[nodiscard]] constexpr std::chrono::microseconds CyclesToUs(s64 cycles) {
return std::chrono::microseconds(cycles * 1000 / detail::BASE_CLOCK_RATE_ADJUSTED);
}
} // namespace Core::Timing