C++ std::isnan及std::isinf

std::isnan及std::isinf

摘自isnan，isnan函數的簽名為：

bool isnan (float x);
bool isnan (double x);
bool isnan (long double x);

它的作用為：

Returns whether x is a NaN (Not-A-Number) value.

The NaN values are used to identify undefined or non-representable 
values for floating-point elements, such as the square root of 
negative numbers or the result of 0/0.

即用來判斷一個浮點數是否為未定義或無法表示的值，如負數的根號值，負數的log值或0/0。

要注意的是1.0/0.0或-1.0/0.0並不是NaN，而是Inf（無限）。如果要判斷一個浮點數是否為無限，則應該使用std::isinf這個函數。

筆者寫了一段代碼：cpp-code-snippets/isnan_isinf.cpp，用來測試這兩個函數，以下為其運行結果：

0.0/0.0    is nan? 1, is inf? 0
sqrt(-1.0) is nan? 1, is inf? 0
log(-1.0)  is nan? 1, is inf? 0
1.0/0.0    is nan? 0, is inf? 1
-1.0/0.0   is nan? 0, is inf? 1

在TensorRT/parsers/caffe/caffeWeightFactory/caffeWeightFactory.cpp的函數CaffeWeightFactory::checkForNans中：

template <typename T>
bool CaffeWeightFactory::checkForNans(const void* values, int count, const std::string& layerName)
{
    const T* v = reinterpret_cast<const T*>(values);
    for (int i = 0; i < count; i++)
    {
        if (std::isnan(float(v[i])))
        {
            //...
            return false;
        }
    }
    return true;
}

便用std::isnan函數來判斷float(v[i])是否為NaN。

參考連結

isnan

isinf

cpp-code-snippets/isnan_isinf.cpp