CUDA C Programming Guide 在线教程学习笔记 Part 11

▶ 数学函数

● 舍入函数，考虑被舍入参数有双精度浮点和单精度浮点，舍入方式有区别，舍入结果有整形、长整形和长长整形，所以共有以下舍入函数。

 // math_functions.h
 extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ __CUDA_MATH_CRTIMP double __cdecl round(double x);
 extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ __CUDA_MATH_CRTIMP float  __cdecl roundf(float x);
 extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ __CUDA_MATH_CRTIMP long int __cdecl lround(double x);
 extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ __CUDA_MATH_CRTIMP long int __cdecl lroundf(float x);
 extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ __CUDA_MATH_CRTIMP long long int __cdecl llround(double x);
 extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ __CUDA_MATH_CRTIMP long long int __cdecl llroundf(float x);

 extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ __CUDA_MATH_CRTIMP double __cdecl rint(double x);
 extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ __CUDA_MATH_CRTIMP float  __cdecl rintf(float x);
 extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ __CUDA_MATH_CRTIMP long int __cdecl lrint(double x);
 extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ __CUDA_MATH_CRTIMP long int __cdecl lrintf(float x);
 extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ __CUDA_MATH_CRTIMP long long int __cdecl llrint(double x);
 extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ __CUDA_MATH_CRTIMP long long int __cdecl llrintf(float x);

 extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ __CUDA_MATH_CRTIMP double __cdecl nearbyint(double x);
 extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ __CUDA_MATH_CRTIMP float  __cdecl nearbyintf(float x);
 extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ __CUDA_MATH_CRTIMP double __cdecl ceil(double x) __THROW;
 extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ __CUDA_MATH_CRTIMP double __cdecl trunc(double x);
 extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ __CUDA_MATH_CRTIMP float  __cdecl truncf(float x);

● 推荐单精度浮点使用函数 rint() 系列（单指令计算），而不用函数 round() 系列（8 指令计算）。其他几个用于单精度浮点的舍入函数也是单指令计算完成的。

● 基本算术运算的舍入精度

　　

● 指令级函数（__ 前缀数学函数），使用编译器选项 -use_fast_math 可以强制使用以下指令级函数。

　　

● 有的指令级函数名有可选后缀，表明舍入方向