/////////////////////////////////////////////////////////////////////////////////////////////////// // OpenGL Mathematics Copyright (c) 2005 - 2011 G-Truc Creation (www.g-truc.net) /////////////////////////////////////////////////////////////////////////////////////////////////// // Created : 2007-05-10 // Updated : 2009-06-04 // Licence : This source is under MIT License // File : glm/gtx/number_precision.hpp /////////////////////////////////////////////////////////////////////////////////////////////////// // Dependency: // - GLM core // - GLM_GTC_type_precision // - GLM_GTC_quaternion /////////////////////////////////////////////////////////////////////////////////////////////////// #ifndef glm_gtx_number_precision #define glm_gtx_number_precision // Dependency: #include "../glm.hpp" #include "../gtc/type_precision.hpp" #if(defined(GLM_MESSAGES) && !defined(glm_ext)) # pragma message("GLM: GLM_GTX_number_precision extension included") #endif namespace glm{ namespace gtx{ namespace number_precision ///< GLM_GTX_number_precision extension: Defined size types. { using namespace gtc::type_precision; ///////////////////////////// // Unsigned int vector types /// \addtogroup gtx_number_precision /// @{ typedef u8 u8vec1; //!< \brief 8bit unsigned integer scalar. (from GLM_GTX_number_precision extension) typedef u16 u16vec1; //!< \brief 16bit unsigned integer scalar. (from GLM_GTX_number_precision extension) typedef u32 u32vec1; //!< \brief 32bit unsigned integer scalar. (from GLM_GTX_number_precision extension) typedef u64 u64vec1; //!< \brief 64bit unsigned integer scalar. (from GLM_GTX_number_precision extension) ////////////////////// // Float vector types typedef f16 f16vec1; //!< \brief Half-precision floating-point scalar. (from GLM_GTX_number_precision extension) typedef f32 f32vec1; //!< \brief Single-precision floating-point scalar. (from GLM_GTX_number_precision extension) typedef f64 f64vec1; //!< \brief Single-precision floating-point scalar. (from GLM_GTX_number_precision extension) ////////////////////// // Float matrix types typedef f16 f16mat1; //!< \brief Half-precision floating-point scalar. (from GLM_GTX_number_precision extension) typedef f16 f16mat1x1; //!< \brief Half-precision floating-point scalar. (from GLM_GTX_number_precision extension) typedef f32 f32mat1; //!< \brief Single-precision floating-point scalar. (from GLM_GTX_number_precision extension) typedef f32 f32mat1x1; //!< \brief Single-precision floating-point scalar. (from GLM_GTX_number_precision extension) typedef f64 f64mat1; //!< \brief Double-precision floating-point scalar. (from GLM_GTX_number_precision extension) typedef f64 f64mat1x1; //!< \brief Double-precision floating-point scalar. (from GLM_GTX_number_precision extension) /// @} }//namespace number_precision }//namespace gtx }//namespace glm #include "number_precision.inl" namespace glm{using namespace gtx::number_precision;} #endif//glm_gtx_number_precision