---

Modules

NSIMD comes with several additional modules. A module provides a set of functionnalities that are usually not at the same level as SIMD intrinsics and/or that do not provide all C and C++ APIs. These functionnalities are given with the library because they make heavy use of NSIMD core which abstract SIMD intrinsics. Below is the exhaustive list of modules.

• Memory management
  This module provides C-style memory managmenent functions: malloc, calloc, free, copy to/from devices, etc... Its purpose is to facilitate the use of data buffers in a portable way for systems with CPUs only and for systems with CPUs and GPUs.

• Fixed-point arithmetic
  This module provides vectorized fixed-point arithmetic through a C++98 API. The programmer can choose the integral type and the place of the coma for representing its fixed-point numbers. A number of operators are also provided.

• Random number generators
  This module define functions that generate pseudorandom numbers usingalgorithms described in Parallel Random Numbers: As Easy as 1,2,3, byJohn K. Salmon, Mark A. Moraes, Ron O. Dror and David E.Shaw.

• Tiny expression templates 1D
  This module provide a thin layer of expression templates above NSIMD core. It also allows the programmer to target Intel, NVIDIA and AMD GPUs. Expression template are a C++ technique that allows the programmer to write code "à la MATLAB" where variables usually represents vectors and operators are itemwise.

• SPMD programming
  SPMD programming allows the programmer to focus on kernels and the compiler to vectorize kernel code more effectively. Basically this module provides a "à la CUDA" programming C++ DSL to targets CPU SIMD as well as Intel, NVIDIA and AMD GPUs.