AuroraRuntime/Source/Threading/Primitives/SMTYield.cpp

/***
    Copyright (C) 2023-2024 J Reece Wilson (a/k/a "Reece"). All rights reserved.

    File: SMTYield.cpp
    Date: 2023-3-12
    Author: Reece
***/
#include <Source/RuntimeInternal.hpp>
#include "SMTYield.hpp"
#include <Source/HWInfo/AuCpuId.hpp>

namespace Aurora::Threading
{
    AUKN_SYM void SetSpinCountTimeout(AuUInt32 uTimeout)
    {
        gRuntimeConfig.threadingConfig.uSpinLoopPowerA = uTimeout;
    }

    AUKN_SYM AuUInt32 GetSpinCountTimeout()
    {
        return gRuntimeConfig.threadingConfig.uSpinLoopPowerA;
    }

    AUKN_SYM void SetThreadLocalAdditionalSpinCountTimeout(AuUInt32 uTimeout)
    {
        gHasThreadLocalTimeout = 1;
        tlsSpinCountLocal = uTimeout;
    }

    AUKN_SYM AuUInt32 GetTotalSpinCountTime()
    {
        AuUInt32 uCount {};

        if (!Primitives::ThrdCfg::gPlatformIsSMPProcessorOptimized)
        {
            return 0;
        }

        uCount = AuUInt32(gRuntimeConfig.threadingConfig.uSpinLoopPowerA);
        if (gHasThreadLocalTimeout)
        {
            uCount += tlsSpinCountLocal;
        }

    //#if defined(AURORA_ARCH_X64) || defined(AURORA_ARCH_X86)
    //    uCount *= 4;
    //#endif
        // well, i guess not because intel recommends going by rdtsc ticks, and not to assume eventual uop sleep of an unspecified period
        return uCount;
    }

    AUKN_SYM const ThreadingConfig *GetThreadingConfig()
    {
        return &gRuntimeConfig.threadingConfig;
    }

    AUKN_SYM void SetThreadingConfig(const ThreadingConfig *pUpdateConfig)
    {
        if (!pUpdateConfig)
        {
            return;
        }
        
        ThreadingConfig cpy(*pUpdateConfig);
        cpy.bPreferFutexRWLock           = Primitives::ThrdCfg::gPreferFutexRWLock;
        cpy.bPreferFutexEvent            = Primitives::ThrdCfg::gPreferFutexEvent;
        cpy.bPreferEmulatedWakeOnAddress = Primitives::ThrdCfg::gPreferEmulatedWakeOnAddress;
        gRuntimeConfig.threadingConfig   = decltype(gRuntimeConfig.threadingConfig)(cpy);

        Primitives::InitCfg();
        Primitives::InitAdaptiveThreshold();
    }

    bool IsNativeWaitOnSupported();
}

namespace Aurora::Threading::Primitives
{
    void InitAdaptiveThreshold()
    {
        auto uCores = AuHwInfo::GetCPUInfo().uThreads;

        ThrdCfg::gCountOfPCores = AuHwInfo::GetCPUInfo().maskPCores.CpuBitCount();

        bool bPermitWOAInternal = IsNativeWaitOnSupported();

        gUseFutexRWLock = ThrdCfg::gPreferFutexRWLock &&
            bPermitWOAInternal;

        gPreferFutexEvent = ThrdCfg::gPreferFutexEvent &&
            bPermitWOAInternal;

        gSpinAdaptiveThreadCount = uCores;

        if (uCores == 1)
        {
            gSpinAdaptiveThreshold = 0;
            ThrdCfg::gPlatformIsSMPProcessorOptimized = false;
            return;
        }

    #if defined(AURORA_ARCH_X64) || defined(AURORA_ARCH_X86)
        {
        #if 0
            auto cpuId = AuHwInfo::cpuid(7);
            ThrdCfg::gIsIntelAlderLakeOrGreater = (cpuId.ecx >> 5) & 1;
        #else
            auto &cpuId = AuHwInfo::GetCPUInfo().cpuId;
            ThrdCfg::gIsIntelAlderLakeOrGreater = AuBitTest(cpuId.f_7_ECX, 5);
            ThrdCfg::gIsZen3OrGreater = AuBitTest(cpuId.f_81_ECX, 29);
        #endif
        }
    #endif

        if (!ThrdCfg::gForceEnableAdaptiveSpin)
        {
            gSpinAdaptiveThreshold = 0;
            return;
        }

        if (uCores >= 16)
        {
            gSpinAdaptiveThreshold = uCores / ThrdCfg::gAdaptiveSpinCUCnt16;
        }
        else if (uCores >= 8 && ThrdCfg::gAdaptiveSpinCUCnt8)
        {
            gSpinAdaptiveThreshold = uCores / ThrdCfg::gAdaptiveSpinCUCnt8;
        }
        else if (uCores >= 4 && ThrdCfg::gAdaptiveSpinCUCnt4)
        {
            gSpinAdaptiveThreshold = uCores / ThrdCfg::gAdaptiveSpinCUCnt4;
        }
        else if (uCores >= 0 && ThrdCfg::gAdaptiveSpinCUCnt0)
        {
            gSpinAdaptiveThreshold = uCores / ThrdCfg::gAdaptiveSpinCUCnt0;
        }
        else
        {
            gSpinAdaptiveThreshold = 0;
        }
    }

    void InitAdaptiveThresholdFirstTime()
    {
        if (!ThrdCfg::gForceEnableAdaptiveSpin &&
            ThrdCfg::gPreferEnableAdaptiveSpin)
        {
        #if defined(AURORA_IS_LINUX_DERIVED)
            if (ThrdCfg::gPreferLinuxAdaptiveSpin)
            {
                ThrdCfg::gForceEnableAdaptiveSpin = true;
            }
        #else
            if (AuSwInfo::IsWindows10OrGreater())
            {
                ThrdCfg::gForceEnableAdaptiveSpin = ThrdCfg::gPreferNewWin32AdaptiveSpin;
            }
            else
            {
                ThrdCfg::gForceEnableAdaptiveSpin = ThrdCfg::gPreferOldWin32AdaptiveSpin;
            }
        #endif
        }

        InitAdaptiveThreshold();
    }

    void InitCfg()
    {
        ThrdCfg::gPlatformIsSMPProcessorOptimized   = gRuntimeConfig.threadingConfig.bPlatformIsSMPProcessorOptimized;
        ThrdCfg::gEnableAggressiveScheduling        = gRuntimeConfig.threadingConfig.bEnableAggressiveScheduling;
        ThrdCfg::gEnableAgrSchedulingRatelimit      = gRuntimeConfig.threadingConfig.bEnableAgrSchedulingRatelimit;
        ThrdCfg::gPreferNtCondvarModernWinSpin      = gRuntimeConfig.threadingConfig.bPreferNtCondvarModernWinSpin;
        ThrdCfg::gPreferNtCondvarOlderWinSpin       = gRuntimeConfig.threadingConfig.bPreferNtCondvarOlderWinSpin;
        ThrdCfg::gPreferNtSemaphoreSpinTryLock      = gRuntimeConfig.threadingConfig.bPreferNtSemaphoreSpinTryLock;
        ThrdCfg::gPreferNtMutexSpinTryLock          = gRuntimeConfig.threadingConfig.bPreferNtMutexSpinTryLock;
        ThrdCfg::gPreferNtCondMutexSpinTryLock      = gRuntimeConfig.threadingConfig.bPreferNtCondMutexSpinTryLock;
        ThrdCfg::gPreferLinuxSemaphoreSpinTryLock   = gRuntimeConfig.threadingConfig.bPreferLinuxSemaphoreSpinTryLock;
        ThrdCfg::gPreferLinuxMutexSpinTryLock       = gRuntimeConfig.threadingConfig.bPreferLinuxMutexSpinTryLock;
        ThrdCfg::gPreferLinuxCondMutexSpinTryLock   = gRuntimeConfig.threadingConfig.bPreferLinuxCondMutexSpinTryLock;
        ThrdCfg::gPreferEmulatedWakeOnAddress       = gRuntimeConfig.threadingConfig.bPreferEmulatedWakeOnAddress;
        ThrdCfg::gPreferWaitOnAddressAlwaysSpin     = gRuntimeConfig.threadingConfig.bPreferWaitOnAddressAlwaysSpin;
        ThrdCfg::gPreferWaitOnAddressAlwaysSpinNative = gRuntimeConfig.threadingConfig.bPreferWaitOnAddressAlwaysSpinNative;
        ThrdCfg::gPreferRWLockReadLockSpin          = gRuntimeConfig.threadingConfig.bPreferRWLockReadLockSpin;
        ThrdCfg::gUWPNanosecondEmulationCheckFirst  = gRuntimeConfig.threadingConfig.bUWPNanosecondEmulationCheckFirst;
        ThrdCfg::gUWPNanosecondEmulationMaxYields   = gRuntimeConfig.threadingConfig.uUWPNanosecondEmulationMaxYields;
        ThrdCfg::gForceEnableAdaptiveSpin           = gRuntimeConfig.threadingConfig.bForceEnableAdaptiveSpin;
        ThrdCfg::gPreferEnableAdaptiveSpin          = gRuntimeConfig.threadingConfig.bPreferEnableAdaptiveSpin;
        ThrdCfg::gPreferLinuxAdaptiveSpin           = gRuntimeConfig.threadingConfig.bPreferLinuxAdaptiveSpin;
        ThrdCfg::gPreferOldWin32AdaptiveSpin        = gRuntimeConfig.threadingConfig.bPreferOldWin32AdaptiveSpin;
        ThrdCfg::gPreferNewWin32AdaptiveSpin        = gRuntimeConfig.threadingConfig.bPreferNewWin32AdaptiveSpin;
        ThrdCfg::gAdaptiveSpinCUCnt0                = gRuntimeConfig.threadingConfig.uAdaptiveSpinCUCnt0;
        ThrdCfg::gAdaptiveSpinCUCnt4                = gRuntimeConfig.threadingConfig.uAdaptiveSpinCUCnt4;
        ThrdCfg::gAdaptiveSpinCUCnt8                = gRuntimeConfig.threadingConfig.uAdaptiveSpinCUCnt8;
        ThrdCfg::gAdaptiveSpinCUCnt16               = gRuntimeConfig.threadingConfig.uAdaptiveSpinCUCnt16;
        ThrdCfg::gPreferFutexRWLock                 = gRuntimeConfig.threadingConfig.bPreferFutexRWLock;
        ThrdCfg::gWinXpThrough7BlazeOptimizerPower  = gRuntimeConfig.threadingConfig.bWinXpThrough7BlazeOptimizerPower;
        ThrdCfg::gPreferLinuxPrimitivesFutexNoSpin  = gRuntimeConfig.threadingConfig.bPreferLinuxPrimitivesFutexNoSpin;
        ThrdCfg::gPreferUnixPrimitivesNoSpin        = gRuntimeConfig.threadingConfig.bPreferUnixPrimitivesNoSpin;
        ThrdCfg::gAlwaysRWLockWriteBiasOnReadLock   = gRuntimeConfig.threadingConfig.bAlwaysRWLockWriteBiasOnReadLock;
        ThrdCfg::gEnableRWLockWriteBiasOnReadLock   = gRuntimeConfig.threadingConfig.bEnableRWLockWriteBiasOnReadLock;
        ThrdCfg::gPreferFutexEvent                  = gRuntimeConfig.threadingConfig.bPreferFutexEvent;
    }
}

extern "C" AuUInt32 SMTGetAPICNumber(void)
{
    return AuHwInfo::cpuid(1).ebx >> 24;
}