Kolbe.LitMath
0.5.1
See the version list below for details.
dotnet add package Kolbe.LitMath --version 0.5.1
NuGet\Install-Package Kolbe.LitMath -Version 0.5.1
<PackageReference Include="Kolbe.LitMath" Version="0.5.1" />
paket add Kolbe.LitMath --version 0.5.1
#r "nuget: Kolbe.LitMath, 0.5.1"
// Install Kolbe.LitMath as a Cake Addin
#addin nuget:?package=Kolbe.LitMath&version=0.5.1
// Install Kolbe.LitMath as a Cake Tool
#tool nuget:?package=Kolbe.LitMath&version=0.5.1
LitMath
A collection of AVX-256 accelerated mathematical functions for .NET
I rewrote Exp
, Log
, Sin
and a few other useful functions using pure AVX intrinsics, so instead of doing one calculation per core, you can now do 4 doubles or 8 floats per core. I added the Sqrt
, ERF function and a Normal Distribution CDF as well. On doubles, the following accuracies apply:
Exp
andSqrt
run at double precision limitsERF
at1e-13
Sin
andCos
at1e-15
Tan
in $[0,\pi/4]$ at2e-16
ATan
at1e-10
(working on it)
There are examples in the benchmark and tests. But here is one to get you started anyway.
Calculate n
$e^x$'s in chunks of 4 and store the result in y.
int n = 40;
Span<double> x = new Span<double>(Enumerable.Range(0 , n).Select(z => (double)z/n).ToArray());
Span<double> y = new Span<double>(new double[n]);
Lit.Exp(ref x, ref y);
Parallel Processing
LitMath leverages SIMD for instruction level parallelism, but not compute cores. For array sizes large enough, it would be a really good idea to do multicore processing. There's an example called LitExpDoubleParallel
in the ExpBenchmark.cs file to see one way to go about this.
Non-Math Things
Making a library like this involves reinventing the wheel so to speak on very basic concepts. The Util
class includes methods like Max
and Min
and IfElse
, which are key to many programming problems in AVX programming, because it needs to be branch-free.
FAQ
Why does this exist?
The reasons I hear most for why this library is pointless is that the Intel MKL can do everything here better than I or C# can, and that if you want such extreme optimization, you shouldn't be using C# to begin with. I wholeheartedly disagree with both. C# is a great language with increasingly great compilers, and the performance I've been getting in this library is close to what you'd find in the MKL. Choosing C# to do some serious back end math with is a totally fine choice, and if you do math, then you might care about performance, or about cloud computing fees. And for these reasons, optimization to its fullest extent can matter.
The MKL is great. But marshaling objects out of C# into C in order to use the MKL is not great. I have benchmarks that compare the exp
function running on an array in LitMath
and the MKL, and for <2000, LitMath wins on my Zen 3 processor. And it wins by a lot (10x) when you're talking about 256 bit sized arrays. This is the most interesting thing to me. Because with LitMath
you can chain the Vector256
interfaced functions together to make whatever ComplexFunction(Vector256 x)
you would like, then instead of thrashing your cache by running each n
sized array through the MKL over and over to get to your complex function, simply run the full function over each x
to get y
. That is, the MKL would require you to have intermediate results for each basic function run, and these intermediate results would be run over the entire array each time. But by making your entire function a single Vector256
to Vector256
, you only run though the array once. The ERF
is a good example of this.
Product | Versions Compatible and additional computed target framework versions. |
---|---|
.NET | net6.0 is compatible. net6.0-android was computed. net6.0-ios was computed. net6.0-maccatalyst was computed. net6.0-macos was computed. net6.0-tvos was computed. net6.0-windows was computed. net7.0 was computed. net7.0-android was computed. net7.0-ios was computed. net7.0-maccatalyst was computed. net7.0-macos was computed. net7.0-tvos was computed. net7.0-windows was computed. net8.0 was computed. net8.0-android was computed. net8.0-browser was computed. net8.0-ios was computed. net8.0-maccatalyst was computed. net8.0-macos was computed. net8.0-tvos was computed. net8.0-windows was computed. |
-
net6.0
- No dependencies.
NuGet packages
This package is not used by any NuGet packages.
GitHub repositories
This package is not used by any popular GitHub repositories.
Version | Downloads | Last updated |
---|---|---|
0.7.0 | 241 | 11/16/2023 |
0.6.4 | 130 | 10/13/2023 |
0.6.3 | 332 | 1/5/2023 |
0.6.2 | 294 | 1/3/2023 |
0.6.1 | 287 | 1/3/2023 |
0.6.0 | 268 | 12/8/2022 |
0.5.8 | 300 | 12/4/2022 |
0.5.7 | 312 | 11/27/2022 |
0.5.6 | 290 | 11/26/2022 |
0.5.5 | 306 | 11/26/2022 |
0.5.4 | 319 | 11/26/2022 |
0.5.3 | 306 | 11/26/2022 |
0.5.2 | 321 | 11/22/2022 |
0.5.1 | 299 | 11/22/2022 |
0.5.0 | 389 | 9/24/2022 |
0.4.1 | 400 | 9/21/2022 |
0.4.0 | 406 | 9/21/2022 |
0.3.8 | 410 | 9/15/2022 |
0.3.7 | 429 | 9/14/2022 |
0.3.6 | 433 | 9/14/2022 |
0.3.5 | 441 | 9/14/2022 |
0.3.4 | 400 | 7/22/2022 |
0.3.3 | 405 | 7/21/2022 |
0.3.2 | 402 | 7/20/2022 |
0.3.1 | 430 | 7/20/2022 |
0.3.0 | 428 | 7/18/2022 |
0.2.1 | 428 | 7/9/2022 |
0.2.0 | 440 | 6/5/2022 |
0.1.2 | 414 | 6/3/2022 |
0.1.1 | 379 | 6/2/2022 |
0.1.0 | 416 | 6/2/2022 |