NSIMD documentation
Index | Tutorial | FAQ | Contribute | API overview | API reference | Wrapped intrinsics | Modules

Documentation can be found here.

What is NSIMD?

At its core, NSIMD is a vectorization library that abstracts SIMD programming. It was designed to exploit the maximum power of processors at a low development cost. NSIMD comes with modules. As of now two of them adds support for GPUs to NSIMD. The direction that NSIMD is taking is to provide several programming paradigms to address different problems and to allow a wider support of architectures. With two of its modules NSIMD provides three programming paradigms:

Supported architectures

Architecture NSIMD core TET1D module SPMD module
CPU (SIMD emulation) Y Y Y
Intel SSE 2 Y Y Y
Intel SSE 4.2 Y Y Y
Intel AVX Y Y Y
Intel AVX2 Y Y Y
Intel AVX-512 for KNLs Y Y Y
Intel AVX-512 for Skylake processors Y Y Y
Arm NEON 128 bits (ARMv7 and earlier) Y Y Y
Arm NEON 128 bits (ARMv8 and later) Y Y Y
Arm SVE (original sizeless SVE) Y Y Y
Arm fixed sized SVE Y Y Y

How it works?

To achieve maximum performance, NSIMD mainly relies on the inline optimization pass of the compiler. Therefore using any mainstream compiler such as GCC, Clang, MSVC, XL C/C++, ICC and others with NSIMD will give you a zero-cost SIMD abstraction library.

To allow inlining, a lot of code is placed in header files. Small functions such as addition, multiplication, square root, etc, are all present in header files whereas big functions such as I/O are put in source files that are compiled as a .so/.dll library.

NSIMD provides C89, C++98, C++11, C++14 and C++20 APIs. All APIs allow writing generic code. For the C API this is achieved through a thin layer of macros; for the C++ APIs it is achieved using templates and function overloading. The C++ APIs are split into two. The first part is a C-like API with only function calls and direct type definitions for SIMD types while the second one provides operator overloading, higher level type definitions that allows unrolling. C++11, C++14 APIs add for instance templated type definitions and templated constants while the C++20 API uses concepts for better error reporting.

Binary compatibility is guaranteed by the fact that only a C ABI is exposed. The C++ API only wraps the C calls.

Supported compilers

NSIMD is tested with GCC, Clang, MSVC, NVCC, HIPCC and ARMClang. As a C89 and a C++98 API are provided, other compilers should work fine. Old compiler versions should work as long as they support the targeted SIMD extension. For instance, NSIMD can compile SSE 4.2 code with MSVC 2010.

Build the library


As CMake is widely used as a build system, we have added support for building the library only and the corresponding find module.

mkdir build
cd build
cmake .. -Dsimd=SIMD_EXT
make install

where SIMD_EXT is one of the following: CPU, SSE2, SSE42, AVX, AVX2, AVX512_KNL, AVX512_SKYLAKE, NEON128, AARCH64, SVE, SVE128, SVE256, SVE512, SVE1024, SVE2048, CUDA, ROCM.

Note that when compiling for NEON128 on Linux one has to choose the ABI, either armel or armhf. Default is armel. As CMake is unable to autodetect this parameter one has to tell CMake manually.

cmake .. -Dsimd=neon128                               # for armel
cmake .. -Dsimd=neon128 -DNSIMD_ARM32_IS_ARMEL=OFF    # for armhf

We provide in the scripts directory a CMake find module to find NSIMD on your system. One can let the module find NSIMD on its own, if several versions for different SIMD extensions of NSIMD are installed then the module will find and return one. There is no guaranty on which versions will be chosen by the module.


If one wants a specific version of the library for a given SIMD extension then use the COMPONENTS part of find_package. Only one component is supported at a time.

find_package(NSIMD COMPONENTS avx2)         # find only NSIMD for Intel AVX2
find_package(NSIMD COMPONENTS sve)          # find only NSIMD for Arm SVE
find_package(NSIMD COMPONENTS sse2 sse42)   # unsupported


The support for CMake has been limited to building the library only. If you wish to run tests or contribute you need to use nsconfig as CMake has several flaws:

Dependencies (nsconfig only)

Generating C/C++ files is done by the Python3 code contained in the egg. Python should be installed by default on any Linux distro. On Windows it comes with the latest versions of Visual Studio on Windows (https://visualstudio.microsoft.com/vs/community/), you can also download and install it directly from https://www.python.org/.

The Python code can call clang-format to properly format all generated C/C++ source. On Linux you can install it via your package manager. On Windows you can use the official binary at https://llvm.org/builds/.

Testing the library requires the MPFR library that can be found at https://www.mpfr.org/.

Benchmarking the library requires Google Benchmark version 1.3 that can be found at https://github.com/google/benchmark plus all the other SIMD libraries used for comparison:

Compiling the library requires a C++98 compiler. Any version of GCC, Clang or MSVC will do. Note that the produced library and header files for the end-user are C89, C++98, C++11 compatible. Note that C/C++ files are generated by a bunch of Python scripts and they must be executed first before running building the library.

Build for Linux

bash scripts/build.sh for simd_ext1/.../simd_extN with comp1/.../compN

For each combination a directory build-simd_ext-comp will be created and will contain the library. Supported SIMD extension are:

Supported compiler are:

Note that certain combination of SIMD extension/compilers are not supported such as aarch64 with icc, or avx512_skylake with nvcc.

Build on Windows

Make sure you are typing in a Visual Studio prompt. The command is almost the same as for Linux with the same constraints on the pairs SIMD extension/compilers.

scripts\build.bat for simd_ext1/.../simd_extN with comp1/.../compN

More details on building the library

The library uses a tool called nsconfig (https://github.com/agenium-scale/nstools) which is basically a Makefile translator. If you have just built NSIMD following what's described above you should have a nstools directory which contains bin/nsconfig. If not you can generate it using on Linux

bash scripts/setup.sh

and on Windows


Then you can use nsconfig directly it has a syntax similar to CMake at command line. Here is a quick tutorial with Linux command line. We first go to the NSIMD directory and generate both NSIMD and nsconfig.

$ cd nsimd
$ python3 egg/hatch.py -ltf
$ bash scripts/setup.sh
$ mkdir build
$ cd build

Help can be displayed using --help.

$ ../nstools/bin/nsconfig --help
usage: nsconfig [OPTIONS]... DIRECTORY
Configure project for compilation.

  -v              verbose mode, useful for debugging
  -nodev          Build system will never call nsconfig
  -DVAR=VALUE     Set value of variable VAR to VALUE
  -list-vars      List project specific variable
  -GBUILD_SYSTEM  Produce files for build system BUILD_SYSTEM
                  Supported BUILD_SYSTEM:
                    make       POSIX Makefile
                    gnumake    GNU Makefile
                    nmake      Microsot Visual Studio NMake Makefile
                    ninja      Ninja build file (this is the default)
                    list-vars  List project specific variables
  -oOUTPUT        Output to OUTPUT instead of default
  -suite=SUITE    Use compilers from SUITE as default ones
                  Supported SUITE:
                    gcc       The GNU compiler collection
                    msvc      Microsoft C and C++ compiler
                    llvm      The LLVM compiler infrastructure
                    armclang  Arm suite of compilers based on LLVM
                    icc       Intel C amd C++ compiler
                    rocm      Radeon Open Compute compilers
                    cuda, cuda+gcc, cuda+clang, cuda+msvc
                              Nvidia CUDA C++ compiler
                  Use COMPILER when COMMAND is invoked for compilation
                  If VERSION and/or ARCHI are not given, nsconfig will
                  try to determine those. This is useful for cross
                  compiling and/or setting the CUDA host compiler.
                  COMMAND must be in { cc, c++, gcc, g++, cl, icc, nvcc,
                  hipcc, hcc, clang, clang++, armclang, armclang++,
                  cuda-host-c++ } ;
                  VERSION is compiler dependant. Note that VERSION
                  can be set to only major number(s) in which case
                  nsconfig fill missing numbers with zeros.
                  Supported ARCHI:
                    x86      Intel 32-bits ISA
                    x86_64   Intel/AMD 64-bits ISA
                    armel    ARMv5 and ARMv6 32-bits ISA
                    armhf    ARMv7 32-bits ISA
                    aarch64  ARM 64-bits ISA
                  Supported COMPILER:
                    gcc, g++              GNU Compiler Collection
                    clang, clang++        LLVM Compiler Infrastructure
                    msvc, cl              Microsoft Visual C++
                    armclang, armclang++  ARM Compiler
                    icc                   Intel C/C++ Compiler
                    dpcpp                 Intel DPC++ Compiler
                    nvcc                  Nvidia CUDA compiler
                    hipcc                 ROCm HIP compiler
  -prefix=PREFIX  Set path for installation to PREFIX
  -h, --help      Print the current help

NOTE: Nvidia CUDA compiler (nvcc) needs a host compiler. Usually on
      Linux systems it is GCC while on Windows systems it is MSVC.
      If nvcc is chosen as the default C++ compiler via the -suite
      switch, then its host compiler can be invoked in compilation
      commands with 'cuda-host-c++'. The latter defaults to GCC on Linux
      systems and MSVC on Windows systems. The user can of course choose
      a specific version and path of this host compiler via the
      '-comp=cuda-hostc++,... parameters. If nvcc is not chosen as the
      default C++ compiler but is used for compilation then its default
      C++ host compiler is 'c++'. The latter can also be customized via
      the '-comp=c++,...' command line switch.

Each project can defined its own set of variable controlling the generation of the ninja file of Makefile.

$ ../nstools/bin/nsconfig .. -list-vars
Project variables list:
name               | description
simd               | SIMD extension to use
cuda_arch_flags    | CUDA target arch flag(s) for tests
mpfr               | MPFR compilation flags (for tests only)
sleef              | Sleef compilation flags (for benchmarks only)
benchmark          | Google benchmark compilation flags (for benchmarks only)
build_library_only | Turn off tests/bench/ulps
static_libstdcpp   | Compile the libstdc++ statically

Finally one can choose what to do and compile NSIMD and its tests.

$ ../nstools/bin/nsconfig .. -Dsimd=avx2
$ ninja
$ ninja tests

Note that MPFR (https://www.mpfr.org/) is needed to compile the tests. If you do not have the MPFR header installed on your system or if you want to use a custom version of MPFR you can tell nsconfig how where to find it.

$ ../nstools/bin/nsconfig .. -Dsimd=avx2 \
      -Dmpfr="-Iwhere/is/mpfr/include -Lwhere/is/mpfr/lib -lmpfr"
$ ninja
$ ninja tests

Nsconfig comes with nstest a small tool to execute tests.

$ ../nstools/bin/nstest -j20

Cross compilation

It is useful to cross-compile for example when you are on a Intel workstation and want to compile for a Raspberry Pi. Nsconfig generate some code, compile and run it to obtain informations on the C or C++ compilers. When cross compiling, unless you configured your Linux box with binfmt_misc to tranparently execute aarch64 binaries on a x86_64 host you need to give nsconfig all the informations about the compilers so that it does not need to run aarch64 code on x86_64 host.

$ ../nstools/bin/nsconfig .. -Dsimd=aarch64 \
      -comp=cc,gcc,aarch64-linux-gnu-gcc,10.0,aarch64 \

Defines that control NSIMD compilation and usage

Several defines control NSIMD.

Philosophy of NSIMD

Originally the library aimed at providing a portable zero-cost abstraction over SIMD vendor intrinsics disregarding the underlying SIMD vector length. NSIMD will of course continue to wrap SIMD intrinsics from various vendors but more efforts will be put into writing NSIMD modules and improving the existing ones especially the SPMD module.

The SPMD paradigm

It is our belief that SPMD is a good paradigm for writing vectorized code. It helps both the developer and the compiler writer. It forces the developers to better arrange its data ion memory more suited for vectorization. On the compiler side it is more simplier to write a "SPMD compiler" than a standard C/C++/Fortran compiler that tries to autovectorize some weird loop with data scattered all around the place. Our priority for our SPMD module are the following:

Our approach can be roughly compared to ISPC (https://ispc.github.io/) but from a library point of view.

Wrapping intrinsics in NSIMD core

NSIMD was designed following as closely as possible the following guidelines:

You may wrap intrinsics that require compile time knowledge of the underlying vector length but this should be done with caution.

Wrapping intrinsics that do not exist for all types is difficult and may require casting or emulation. For instance, 8 bit integer vector multiplication using SSE2 does not exist. We can either process each pair of integers individually or we can cast the 8 bit vectors to 16 bit vectors, do the multiplication and cast them back to 8 bit vectors. In the second case, chaining operations will generate many unwanted casts.

To avoid hiding important details to the user, overloads of operators involving scalars and SIMD vectors are not provided by default. Those can be included explicitely to emphasize the fact that using expressions like scalar + vector might incur an optimization penalty.

The use of nsimd::pack may not be portable to ARM SVE and therefore must be included manually. ARM SVE registers can only be stored in sizeless strucs (__sizeless_struct). This feature (as of 2019/04/05) is only supported by the ARM compiler. We do not know whether other compilers will use the same keyword or paradigm to support SVE intrinsics.

Contributing to NSIMD

The wrapping of intrinsics, the writing of test and bench files are tedious and repetitive tasks. Most of those are generated using Python scripts that can be found in egg.

Please see nsimd/CONTRIBUTE.md for more details.


Copyright (c) 2020 Agenium Scale

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.