NVIDIA nvcc编译标志，用于constexpr深度和IEEE 754指数计算

Question

Consider the following code that computes the exponent of a double floating point number as a constant expression (in the format specified by the IEEE 754 standard).

    template <typename T>  constexpr T abs_CE(const T x){return x>=0?x:-x;}
    constexpr unsigned long long int __double_exponent_CE_(const double x){return x==0?0:(x>=2.?(__double_exponent_CE_(x/2.)+1):(x<1?__double_exponent_CE_(x*2.)-1:0));}
    constexpr unsigned long long int __double_exponent_CE(const double x){return (x==0)?0:(__double_exponent_CE_(abs_CE(x))+1023);}

That code fails to compile as constant expression in gcc under normal compilation flags circumstances for certain inputs like std::numeric_limits< double >::max. The reason it fails to compile is because it exceeds the max recursion depth for a constant expression (512 is the default value). For example std::numeric_limits< double >::max requires 1024 calls, exceeding the limit.

If the flag -fconstexpr-depth=2048 is added, then the code compiles perfectly, and evaluates to a constant expression that can be passed as a template parameter.

That code fails to compile under nvcc with the flag -Xcompiler -fconstexpr-depth=2048 (specifically it crashes when nvcc issues the cicc command), so is there any way to change the depth limit in nvcc? I have not found any flag to change it in NVCC options .

Just in case there is no such equivalent flag in nvcc, does anybody know any other way to compute the exponent of a double in compile time with less than 512 recursions calls?

Answer 1

You should consider using " --expt-relaxed-constexpr " flag:

Experimental flag:

Allow host code to invoke __device__constexpr functions, and device code to invoke __host__constexpr functions.

Note that the behavior of this flag may change in future compiler releases.

Also in CUDA C Programming Guide :

By default, a constexpr function cannot be called from a function with incompatible execution space. The experimental nvcc flag --expt-relaxed-constexpr removes this restriction. When this flag is specified, host code can invoke a __device__ constexpr function and device code can invoke a __host__ constexpr function.