DXGI_ERROR_DEVICE_HUNG产生的C++AMP方法

Question

我试图实现一个函数，为Gauss数值积分法计算权重和脓肿，使用C++AMP并行化进程，并在运行时得到一个DXGI_ERROR_DEVICE_HUNG错误。

这是我的助手方法，用于计算GPU上伽马函数的对数：

template <typename T>
T gammaln_fast( T tArg ) restrict( amp )
{
    const T tCoefficients[] = { T( 57.1562356658629235f ), T( -59.5979603554754912f ),
        T( 14.1360979747417471f ), T( -0.491913816097620199f ), T( 0.339946499848118887E-4f ),
        T( 0.465236289270485756E-4f ), T( -0.983744753048795646E-4f ), T( 0.158088703224912494E-3f ),
        T( -0.210264441724104883E-3f ), T( 0.217439618115212643E-3f ), T( -0.164318106536763890E-3f ),
        T( 0.844182239838527433E-4f ), T( -0.261908384015814087E-4f ), T( 0.386991826595316234E-5f ) };

    T y = tArg, tTemp = tArg + T( 5.2421875f );
    tTemp = (tArg + T( 0.5f )) * concurrency::fast_math::log( tTemp ) - tTemp;

    T tSer = T( 0.999999999999997092f );

    for( std::size_t s = 0; s < (sizeof( tCoefficients ) / sizeof( T )); ++s )
    {
        tSer += tCoefficients[s] / ++y;
    }

    return tTemp + concurrency::fast_math::log( T( 2.5066282746310005f ) * tSer / tArg );
}

这是我计算重量和脓肿的函数：

template <typename T>
ArrayPair<T> CalculateGaussLaguerreWeights_fast( const T tExponent, const std::size_t sNumPoints, T tEps = std::numeric_limits<T>::epsilon() )
{
    static_assert(std::is_floating_point<T>::value, "You can only instantiate this function with a floating point data type");
    static_assert(!std::is_same<T, long double>::value, "You can not instantiate this function with long double type"); // The long double type is not currently supported by C++AMP

    T tCurrentGuess, tFatherGuess, tGrandFatherGuess;
    std::vector<T> vecInitialGuesses( sNumPoints );
    for( std::size_t s = 0; s < sNumPoints; ++s )
    {
        if( s == 0 )
        {
            tCurrentGuess = (T( 1.0f ) + tExponent) * (T( 3.0f ) + T( 0.92f ) * tExponent) / (T( 1.0f ) + T( 2.4f ) * sNumPoints + T( 1.8f ) * tExponent);
        }
        else if( s == 1 )
        {
            tFatherGuess = tCurrentGuess;
            tCurrentGuess += (T( 15.0f ) + T( 6.25f ) * tExponent) / (T( 1.0f ) + T( 0.9f ) * tExponent + T( 2.5f ) * sNumPoints);
        }
        else
        {
            tGrandFatherGuess = tFatherGuess;
            tFatherGuess = tCurrentGuess;
            std::size_t sDec = s - 1U;
            tCurrentGuess += ((T( 1.0f ) + T( 2.55f ) * sDec) / (T( 1.9f ) * sDec) + T( 1.26f ) * sDec * tExponent
                / (T( 1.0f ) + T( 3.5f ) * sDec)) * (tCurrentGuess - tGrandFatherGuess) / (T( 1.0f ) + T( 0.3f ) * tExponent);
        }
        vecInitialGuesses[s] = tCurrentGuess;
    }

    concurrency::array<T> arrWeights( sNumPoints ), arrAbsciasses( sNumPoints, std::begin(vecInitialGuesses) );

    try {
        concurrency::parallel_for_each( arrAbsciasses.extent, [=, &arrAbsciasses, &arrWeights]( concurrency::index<1> index ) restrict( amp ) {
            T tVal = arrAbsciasses[index], tIntermediate;
            T tPolynomial1 = T( 1.0f ), tPolynomial2 = T( 0.0f ), tPolynomial3, tDerivative;
            std::size_t sIterationNum = 0;
            do {
                tPolynomial1 = T( 1.0f ), tPolynomial2 = T( 0.0f );

                for( std::size_t s = 0; s < sNumPoints; ++s )
                {
                    tPolynomial3 = tPolynomial2;
                    tPolynomial2 = tPolynomial1;
                    tPolynomial1 = ((2 * s + 1 + tExponent - tVal) * tPolynomial2 - (s + tExponent) * tPolynomial3) / (s + 1);
                }

                tDerivative = (sNumPoints * tPolynomial1 - (sNumPoints + tExponent) * tPolynomial2) / tVal;
                tIntermediate = tVal;
                tVal = tIntermediate - tPolynomial1 / tDerivative;
                ++sIterationNum;

            } while( concurrency::fast_math::fabs( tVal - tIntermediate ) > tEps || sIterationNum < 10 );

            arrAbsciasses[index] = tVal;
            arrWeights[index] = -concurrency::fast_math::exp( gammaln_fast( tExponent + sNumPoints ) - gammaln_fast( T( sNumPoints ) ) ) / (tDerivative * sNumPoints * tPolynomial2);
        } );
    }
    catch( concurrency::runtime_exception& e )
    {
        std::cerr << "Runtime error, code: " << e.get_error_code() << "; message: " << e.what() << std::endl;
    }

    return std::make_pair( std::move( arrAbsciasses ), std::move( arrWeights ) );
}

下面是调试控制台的完整跟踪：

D3D11:移除装置。D3D11错误: ID3D11Device::RemoveDevice:设备删除是由于以下原因触发的(DXGI_ERROR_DEVICE_HUNG:设备执行命令花费了不合理的时间，或者硬件崩溃/挂起)。因此，触发了TDR (超时值检测和恢复)机制。挂起时，当前设备上下文正在执行命令。应用程序可能希望重新出现并退回到不太积极地使用显示硬件)。执行错误#378: DEVICE_REMOVAL_PROCESS_AT_FAULT D3D11错误: ID3D11DeviceContext::Map:返回DXGI_ERROR_DEVICE_REMOVED，当资源试图使用READ或READWRITE映射时。RESOURCE_MANIPULATION错误#2097214: RESOURCE_MAP_DEVICEREMOVED_RETURN

我很抱歉未能举出一个可复制的小例子；我希望这仍然是一个可以接受的问题，因为我无法自己解决这个问题。

Answer 1

在使用DirectCompute时，主要的挑战是编写不与Direct3D自动“GPU挂起”检测超时冲突的计算。默认情况下，系统假设如果一个着色器花费超过几秒钟，GPU实际上是挂起的。这种启发式方法适用于视觉着色器，但是您可以很容易地创建一个DirectCompute着色器，它需要很长时间才能完成。

解决方案是禁用超时检测。您可以通过使用Direct3D创建D3D11_CREATE_DEVICE_DISABLE_GPU_TIMEOUT 11设备来做到这一点，参见在Windows 8上禁用C++ AMP算法的TDR博客文章。主要要记住的是，D3D11_CREATE_DEVICE_DISABLE_GPU_TIMEOUT需要Windows11.1或更高版本的运行时，该运行时包含在Windows8.x中，可以用KB2670838安装在Windows7ServicePack 1上。有关使用DirectX 11.1和Windows 7、DirectX 11.1和Windows7更新和MSDN的一些注意事项，请参阅KB2670838。