- Notifications
You must be signed in to change notification settings - Fork334
Performance-portable, length-agnostic SIMD with runtime dispatch
License
Apache-2.0, BSD-3-Clause licenses found
Licenses found
google/highway
Folders and files
Name | Name | Last commit message | Last commit date | |
---|---|---|---|---|
Repository files navigation
Highway is a C++ library that provides portable SIMD/vector intrinsics.
Previously licensed under Apache 2, now dual-licensed as Apache 2 / BSD-3.
We are passionate about high-performance software. We see major untappedpotential in CPUs (servers, mobile, desktops). Highway is for engineers who wantto reliably and economically push the boundaries of what is possible insoftware.
CPUs provide SIMD/vector instructions that apply the same operation to multipledata items. This can reduce energy usage e.g.fivefold because fewerinstructions are executed. We also often see5-10x speedups.
Highway makes SIMD/vector programming practical and workable according to theseguiding principles:
Does what you expect: Highway is a C++ library with carefully-chosenfunctions that map well to CPU instructions without extensive compilertransformations. The resulting code is more predictable and robust to codechanges/compiler updates than autovectorization.
Works on widely-used platforms: Highway supports five architectures; thesame application code can target various instruction sets, including those with'scalable' vectors (size unknown at compile time). Highway only requires C++11and supports four families of compilers. If you would like to use Highway onother platforms, please raise an issue.
Flexible to deploy: Applications using Highway can run on heterogeneousclouds or client devices, choosing the best available instruction set atruntime. Alternatively, developers may choose to target a single instruction setwithout any runtime overhead. In both cases, the application code is the sameexcept for swappingHWY_STATIC_DISPATCH
withHWY_DYNAMIC_DISPATCH
plus oneline of code. See also @kfjahnke'sintroduction to dispatching.
Suitable for a variety of domains: Highway provides an extensive set ofoperations, used for image processing (floating-point), compression, videoanalysis, linear algebra, cryptography, sorting and random generation. Werecognise that new use-cases may require additional ops and are happy to addthem where it makes sense (e.g. no performance cliffs on some architectures). Ifyou would like to discuss, please file an issue.
Rewards data-parallel design: Highway provides tools such as Gather,MaskedLoad, and FixedTag to enable speedups for legacy data structures. However,the biggest gains are unlocked by designing algorithms and data structures forscalable vectors. Helpful techniques include batching, structure-of-arraylayouts, and aligned/padded allocations.
We recommend these resources for getting started:
- SIMD for C++ Developers
- Algorithms for Modern Hardware
- Optimizing software in C++
- Improving performance with SIMD intrinsics in three use cases
Online demos using Compiler Explorer:
- multiple targets with dynamic dispatch(more complicated, but flexible and uses best available SIMD)
- single target using -m flags(simpler, but requires/only uses the instruction set enabled by compilerflags)
We observe that Highway is referenced in the following open source projects,found via sourcegraph.com. Most are GitHub repositories. If you would like toadd your project or link to it directly, feel free to raise an issue or contactus via the below email.
- Audio:Zimtohrli perceptual metric
- Browsers: Chromium (+Vivaldi), Firefox (+floorp / foxhound / librewolf /Waterfox)
- Computational biology:RNA analysis
- Computer graphics:Sparse voxel renderer
- Cryptography: google/distributed_point_functions, google/shell-encryption
- Data structures: bkille/BitLib
- Image codecs: eustas/2im,Grok JPEG 2000,JPEG XL,JPEGenc,Jpegli, OpenHTJ2K
- Image processing: cloudinary/ssimulacra2, m-ab-s/media-autobuild_suite,libvips
- Image viewers: AlienCowEatCake/ImageViewer, diffractor/diffractor,mirillis/jpegxl-wic,Lux panorama/image viewer
- Information retrieval:iresearch database index,michaeljclark/zvec,nebula interactive analytics / OLAP,ScaNN Scalable Nearest Neighbors,vectorlite vector search
- Machine learning:gemma.cpp,Tensorflow, Numpy, zpye/SimpleInfer
- Robotics:MIT Model-Based Design and Verification
Other
- Evaluation of C++ SIMD Libraries:"Highway excelled with a strong performance across multiple SIMD extensions[..]. Thus, Highway may currently be the most suitable SIMD library for manysoftware projects."
- zimt: C++11 template library to process n-dimensional arrays with multi-threaded SIMD code
- vectorized Quicksort (paper)
If you'd like to get Highway, in addition to cloning from this GitHub repositoryor using it as a Git submodule, you can also find it in the following packagemanagers or repositories:
- alpinelinux
- conan-io
- conda-forge
- DragonFlyBSD,
- fd00/yacp
- freebsd
- getsolus/packages
- ghostbsd
- microsoft/vcpkg
- MidnightBSD
- MSYS2
- NetBSD
- openSUSE
- opnsense
- Xilinx/Vitis_Libraries
- xmake-io/xmake-repo
See also the list athttps://repology.org/project/highway-simd-library/versions.
Highway supports 24 targets, listed in alphabetical order of platform:
- Any:
EMU128
,SCALAR
; - Armv7+:
NEON_WITHOUT_AES
,NEON
,NEON_BF16
,SVE
,SVE2
,SVE_256
,SVE2_128
; - IBM Z:
Z14
,Z15
; - POWER:
PPC8
(v2.07),PPC9
(v3.0),PPC10
(v3.1B, not yet supported dueto compiler bugs, see #1207; also requires QEMU 7.2); - RISC-V:
RVV
(1.0); - WebAssembly:
WASM
,WASM_EMU256
(a 2x unrolled version of wasm128,enabled ifHWY_WANT_WASM2
is defined. This will remain supported until itis potentially superseded by a future version of WASM.); - x86:
SSE2
SSSE3
(~Intel Core)SSE4
(~Nehalem, also includes AES + CLMUL).AVX2
(~Haswell, also includes BMI2 + F16 + FMA)AVX3
(~Skylake, AVX-512F/BW/CD/DQ/VL)AVX3_DL
(~Icelake, includesBitAlg
+CLMUL
+GFNI
+VAES
+VBMI
+VBMI2
+VNNI
+VPOPCNT
),AVX3_ZEN4
(AVX3_DL plus BF16, optimized for AMD Zen4; requires opt-inby definingHWY_WANT_AVX3_ZEN4
if compiling for static dispatch, butenabled by default for runtime dispatch),AVX3_SPR
(~Sapphire Rapids, includes AVX-512FP16)
Our policy is that unless otherwise specified, targets will remain supported aslong as they can be (cross-)compiled with currently supported Clang or GCC, andtested using QEMU. If the target can be compiled with LLVM trunk and testedusing our version of QEMU without extra flags, then it is eligible for inclusionin our continuous testing infrastructure. Otherwise, the target will be manuallytested before releases with selected versions/configurations of Clang and GCC.
SVE was initially tested using farm_sve (see acknowledgments).
Highway releases aim to follow the semver.org system (MAJOR.MINOR.PATCH),incrementing MINOR after backward-compatible additions and PATCH afterbackward-compatible fixes. We recommend using releases (rather than the Git tip)because they are tested more extensively, see below.
The current version 1.0 signals an increased focus on backwards compatibility.Applications using documented functionality will remain compatible with futureupdates that have the same major version number.
Continuous integration tests build with a recent version of Clang (running onnative x86, or QEMU for RISC-V and Arm) and MSVC 2019 (v19.28, running on nativex86).
Before releases, we also test on x86 with Clang and GCC, and Armv7/8 via GCCcross-compile. See thetesting process fordetails.
Thecontrib
directory contains SIMD-related utilities: an image class withaligned rows, a math library (16 functions already implemented, mostlytrigonometry), and functions for computing dot products and sorting.
If you only require x86 support, you may also use Agner Fog'sVCL vector class library. It includes manyfunctions including a complete math library.
If you have existing code using x86/NEON intrinsics, you may be interested inSIMDe, which emulates thoseintrinsics using other platforms' intrinsics or autovectorization.
This project uses CMake to generate and build. In a Debian-based system you caninstall it via:
sudo apt install cmake
Highway's unit tests usegoogletest.By default, Highway's CMake downloads this dependency at configuration time.You can avoid this by setting theHWY_SYSTEM_GTEST
CMake variable to ON andinstalling gtest separately:
sudo apt install libgtest-dev
Alternatively, you can defineHWY_TEST_STANDALONE=1
and remove all occurrencesofgtest_main
in each BUILD file, then tests avoid the dependency on GUnit.
Running cross-compiled tests requires support from the OS, which on Debian isprovided by theqemu-user-binfmt
package.
To build Highway as a shared or static library (depending on BUILD_SHARED_LIBS),the standard CMake workflow can be used:
mkdir -p build&&cd buildcmake ..make -j&& maketest
Or you can runrun_tests.sh
(run_tests.bat
on Windows).
Bazel is also supported for building, but it is not as widely used/tested.
When building for Armv7, a limitation of current compilers requires you to add-DHWY_CMAKE_ARM7:BOOL=ON
to the CMake command line; see #834 and #1032. Weunderstand that work is underway to remove this limitation.
Building on 32-bit x86 is not officially supported, and AVX2/3 are disabled bydefault there. Note that johnplatts has successfully built and run the Highwaytests on 32-bit x86, including AVX2/3, on GCC 7/8 and Clang 8/11/12. On Ubuntu22.04, Clang 11 and 12, but not later versions, require extra compiler flags-m32 -isystem /usr/i686-linux-gnu/include
. Clang 10 and earlier require theabove plus-isystem /usr/i686-linux-gnu/include/c++/12/i686-linux-gnu
. See#1279.
highway is now available invcpkg
vcpkg install highway
The highway port in vcpkg is kept up to date by Microsoft team members and community contributors. If the version is out of date, pleasecreate an issue or pull request on the vcpkg repository.
You can use thebenchmark
inside examples/ as a starting point.
Aquick-reference page briefly lists all operationsand their parameters, and theinstruction_matrixindicates the number of instructions per operation.
TheFAQ answers questions about portability, API design andwhere to find more information.
We recommend using full SIMD vectors whenever possible for maximum performanceportability. To obtain them, pass aScalableTag<float>
(or equivalentlyHWY_FULL(float)
) tag to functions such asZero/Set/Load
. There are twoalternatives for use-cases requiring an upper bound on the lanes:
For up to
N
lanes, specifyCappedTag<T, N>
or the equivalentHWY_CAPPED(T, N)
. The actual number of lanes will beN
rounded down tothe nearest power of two, such as 4 ifN
is 5, or 8 ifN
is 8. This isuseful for data structures such as a narrow matrix. A loop is still requiredbecause vectors may actually have fewer thanN
lanes.For exactly a power of two
N
lanes, specifyFixedTag<T, N>
. The largestsupportedN
depends on the target, but is guaranteed to be at least16/sizeof(T)
.
Due to ADL restrictions, user code calling Highway ops must either:
- Reside inside
namespace hwy { namespace HWY_NAMESPACE {
; or - prefix each op with an alias such as
namespace hn = hwy::HWY_NAMESPACE; hn::Add()
; or - add using-declarations for each op used:
using hwy::HWY_NAMESPACE::Add;
.
Additionally, each function that calls Highway ops (such asLoad
) must eitherbe prefixed withHWY_ATTR
, OR reside betweenHWY_BEFORE_NAMESPACE()
andHWY_AFTER_NAMESPACE()
. Lambda functions currently requireHWY_ATTR
beforetheir opening brace.
Do not use namespace-scope norstatic
initializers for SIMD vectors becausethis can cause SIGILL when using runtime dispatch and the compiler chooses aninitializer compiled for a target not supported by the current CPU. Instead,constants initialized viaSet
should generally be local (const) variables.
The entry points into code using Highway differ slightly depending on whetherthey use static or dynamic dispatch. In both cases, we recommend that thetop-level function receives one or more pointers to arrays, rather thantarget-specific vector types.
For static dispatch,
HWY_TARGET
will be the best available target amongHWY_BASELINE_TARGETS
, i.e. those allowed for use by the compiler (seequick-reference). Functions insideHWY_NAMESPACE
can be called usingHWY_STATIC_DISPATCH(func)(args)
withinthe same module they are defined in. You can call the function from othermodules by wrapping it in a regular function and declaring the regularfunction in a header.For dynamic dispatch, a table of function pointers is generated via the
HWY_EXPORT
macro that is used byHWY_DYNAMIC_DISPATCH(func)(args)
tocall the best function pointer for the current CPU's supported targets. Amodule is automatically compiled for each target inHWY_TARGETS
(seequick-reference) ifHWY_TARGET_INCLUDE
isdefined andforeach_target.h
is included. Note that the first invocationofHWY_DYNAMIC_DISPATCH
, or each call to the pointer returned by the firstinvocation ofHWY_DYNAMIC_POINTER
, involves some CPU detection overhead.You can prevent this by calling the following before any invocation ofHWY_DYNAMIC_*
:hwy::GetChosenTarget().Update(hwy::SupportedTargets());
.
See also a separateintroduction to dynamic dispatchby @kfjahnke.
When using dynamic dispatch,foreach_target.h
is included from translationunits (.cc files), not headers. Headers containing vector code shared betweenseveral translation units require a special include guard, for example thefollowing taken fromexamples/skeleton-inl.h
:
#if defined(HIGHWAY_HWY_EXAMPLES_SKELETON_INL_H_) == defined(HWY_TARGET_TOGGLE)#ifdef HIGHWAY_HWY_EXAMPLES_SKELETON_INL_H_#undef HIGHWAY_HWY_EXAMPLES_SKELETON_INL_H_#else#define HIGHWAY_HWY_EXAMPLES_SKELETON_INL_H_#endif#include "hwy/highway.h"// Your vector code#endif
By convention, we name such headers-inl.h
because their contents (oftenfunction templates) are usually inlined.
Applications should be compiled with optimizations enabled. Without inliningSIMD code may slow down by factors of 10 to 100. For clang and GCC,-O2
isgenerally sufficient.
For MSVC, we recommend compiling with/Gv
to allow non-inlined functions topass vector arguments in registers. If intending to use the AVX2 target togetherwith half-width vectors (e.g. forPromoteTo
), it is also important to compilewith/arch:AVX2
. This seems to be the only way to reliably generateVEX-encoded SSE instructions on MSVC. Sometimes MSVC generates VEX-encoded SSEinstructions, if they are mixed with AVX, but not always, seeDevCom-10618264.Otherwise, mixing VEX-encoded AVX2 instructions and non-VEX SSE may cause severeperformance degradation. Unfortunately, with/arch:AVX2
option, the resultingbinary will then require AVX2. Note that no such flag is needed for clang andGCC because they support target-specific attributes, which we use to ensureproper VEX code generation for AVX2 targets.
When vectorizing a loop, an important question is whether and how to deal witha number of iterations ('trip count', denotedcount
) that does not evenlydivide the vector sizeN = Lanes(d)
. For example, it may be necessary to avoidwriting past the end of an array.
In this section, letT
denote the element type andd = ScalableTag<T>
.Assume the loop body is given as a functiontemplate<bool partial, class D> void LoopBody(D d, size_t index, size_t max_n)
.
"Strip-mining" is a technique for vectorizing a loop by transforming it into anouter loop and inner loop, such that the number of iterations in the inner loopmatches the vector width. Then, the inner loop is replaced with vectoroperations.
Highway offers several strategies for loop vectorization:
Ensure all inputs/outputs are padded. Then the (outer) loop is simply
for (size_t i = 0; i < count; i += N) LoopBody<false>(d, i, 0);
Here, the template parameter and second function argument are not needed.
This is the preferred option, unless
N
is in the thousands and vectoroperations are pipelined with long latencies. This was the case forsupercomputers in the 90s, but nowadays ALUs are cheap and we see mostimplementations split vectors into 1, 2 or 4 parts, so there is little costto processing entire vectors even if we do not need all their lanes. Indeedthis avoids the (potentially large) cost of predication or partialloads/stores on older targets, and does not duplicate code.Process whole vectors and include previously processed elementsin the last vector:
for (size_t i = 0; i < count; i += N) LoopBody<false>(d, HWY_MIN(i, count - N), 0);
This is the second preferred option provided that
count >= N
andLoopBody
is idempotent. Some elements might be processed twice, buta single code path and full vectorization is usually worth it. Even ifcount < N
, it usually makes sense to pad inputs/outputs up toN
.Use the
Transform*
functions in hwy/contrib/algo/transform-inl.h. Thistakes care of the loop and remainder handling and you simply define ageneric lambda function (C++14) or functor which receives the current vectorfrom the input/output array, plus optionally vectors from up to two extrainput arrays, and returns the value to write to the input/output array.Here is an example implementing the BLAS function SAXPY (
alpha * x + y
):Transform1(d, x, n, y, [](auto d, const auto v, const auto v1) HWY_ATTR { return MulAdd(Set(d, alpha), v, v1);});
Process whole vectors as above, followed by a scalar loop:
size_t i = 0;for (; i + N <= count; i += N) LoopBody<false>(d, i, 0);for (; i < count; ++i) LoopBody<false>(CappedTag<T, 1>(), i, 0);
The template parameter and second function arguments are again not needed.
This avoids duplicating code, and is reasonable if
count
is large.Ifcount
is small, the second loop may be slower than the next option.Process whole vectors as above, followed by a single call to a modified
LoopBody
with masking:size_t i = 0;for (; i + N <= count; i += N) { LoopBody<false>(d, i, 0);}if (i < count) { LoopBody<true>(d, i, count - i);}
Now the template parameter and third function argument can be used inside
LoopBody
to non-atomically 'blend' the firstnum_remaining
lanes ofv
with the previous contents of memory at subsequent locations:BlendedStore(v, FirstN(d, num_remaining), d, pointer);
. Similarly,MaskedLoad(FirstN(d, num_remaining), d, pointer)
loads the firstnum_remaining
elements and returns zero in other lanes.This is a good default when it is infeasible to ensure vectors are padded,but is only safe
#if !HWY_MEM_OPS_MIGHT_FAULT
!In contrast to the scalar loop, only a single final iteration is needed.The increased code size from two loop bodies is expected to be worthwhilebecause it avoids the cost of masking in all but the final iteration.
- Highway introduction (slides)
- Overview of instructions per operation on different architectures
- Design philosophy and comparison
- Implementation details
We have usedfarm-sve by BerengerBramas; it has proved useful for checking the SVE port on an x86 developmentmachine.
This is not an officially supported Google product.Contact:janwas@google.com
About
Performance-portable, length-agnostic SIMD with runtime dispatch