Make Your LINQ Up to 10x Faster!
HTML-код
- Опубликовано: 28 фев 2024
- Use code MODULAR20 and get 20% off the brand new "Deep Dive into Modular Monoliths" course on Dometrain: dometrain.com/course/deep-div...
Become a Patreon and get special perks: / nickchapsas
Hello, everybody, I'm Nick, and in this video, I will show you how you can improve LINQ's performance in .NET by using hardware acceleration with a Nuget package called SimdLinq.
Give SimdLinq a star on GitHub: github.com/Cysharp/SimdLinq
Workshops: bit.ly/nickworkshops
Don't forget to comment, like and subscribe :)
Social Media:
Follow me on GitHub: github.com/Elfocrash
Follow me on Twitter: / nickchapsas
Connect on LinkedIn: / nick-chapsas
Keep coding merch: keepcoding.shop
#csharp #dotnet
"... a bunch of advanced stuff which I don't understand" - love that honesty.
Always thought that one of the benefits of Linq was that the operations could be improved over time, without breaking the API.
Yes, exactly this! Adding a 3rd party dependency is not what I would consider "making my LINQ faster." Beyond that, this seems to rely on SIMD, which might not work as expected if you are targeting some embedded systems or WASM.
I like you change shirt and hairstyle for the ads
Always
Helps skipping ahead to the end of the ad ;P
@@pavlindrom this guy living in 2034.
Thank you Nick. Yes the min/max thing is really cool.
I learned about a very useful library I didn't know about, thank you.
Yep, cool stuff.
7:28 What we see here looks more magical than it is. It's just some good old C style pointer arithmetic, wrapped in a bunch of dotnet helper methods, which look quite verbose compared to actual C code. Each ref variable is literally a pointer. The function works with these in the for loops.
The code basically copies the Span data to the Vector, does the Sum calculation with that, and then copies the result back.
But what I actually silently hope for with though is that all the methods like Select, Where and Aggregate will get efficient Span based versions as well in the future, so that we won't have loops at all, making C# code look a bit more like Rust code, which uses zero loops most of the time.
So in other words, I believe in zero cost functional programming.
Very informative. Thanks for sharing
That's pretty impressive that runtimes are getting that much smarter to handle basic operations that much quicker!
what are the breaking changes compared to the Linq implementation by MS?
The need for all these specialized methods actually point to a weakness in the C#/.NET compiler. For example, if you write the simple for-loop example you used in C or C++, a compiler like gcc would automatically detect that it could be vectorized, and generate SIMD instructions during compilation, eliminating the need for writing specific specialized methods in the first place. I wish Microsoft spent more time optimizing it's compiler rather than essentially forcing people to come up with complex solutions to speed up your code, whereas a smart compiler could do all this for you. Optimization should be a big part of a compiler recognizing code patterns, rather than having the programmer coming up with ways to write more performant code.
Wow! It took over 2500ns with LINQ on .NET 6. And since .NET 8 it is around 45x faster (57ns on my PC)? I would not believe if I would not run it myself locally.
Nice content! Love .Net!
Just wanna bring up another point of view, sometimes the optimisation could come from side way.
For eg, algorithm of adding number from 1 to n.
Could be simplified mathematically to formula
n (n + 1) / 2
An interesting benchmark to see would be how fast it is when doing these operations over data from a model object.
Eg. You have a list of users and your want an average of their age, so you'd first have to select the age from the model and then do the average. How much faster would LinQ be then a loop then?
It would be the exact same
LongSum you say. Swedish has a word thats pronounced the same way and its Långsam. Difference is that långsam means slow ^^
Pre 64 bit machines long addition was really slow, so it’s an apt name. They effectively took 4 times as long as a regular sum. Now with 64 bit, it’s effectively just a regular add. That said unless you were doing hundreds or thousands of them it probably didn’t matter much.
In German "langsam" means slow too 😁
This is nice.
Honestly this isn't going to be the bottleneck for most people though. Most people aren't calling Sum tens of thousands of times.
Can you ask MS if they will be incorporating these changes in core?
Refreshing to hear frome someone like Nick saying he doesent understand some C# code/techniques.
The last traces of imposter syndrom went suddenly away 😀
Thank you
Is there a way to apply this library to all linq operations in a old .framework (4.7) solution?
Well, if you get pointer to the continuous array of (unmanaged) memory you are summing, it is really fast to calculate sum using simple pointer arithmetic. And that's what this library seems to do. But that's what .NET don't want you to do 😄 because otherwise you'd be writing C.
If you just do pointer arithmetics to advance over the items one by one to sum, that would be slower than LINQ with its vectorization. On my computer summing 2000 integers with a for-loop is 555ns, LINQ is 108ns and getting a pointer to the array, advancing the pointer until the pointer of the array end is reached is 378ns.
@@maschyt you are right but if you run the operation so that you split the memory into blocks, process each block in its own thread, and finally sum everything up, you get the speed up - and this is exactly what the library does.
I've implemented some image processing stuff this way. I split the image horizontally and ran the processing parallel in threads. But I wrote that in C to get down to the machine level without .net coming in my way, because .net does not really want you to access things directly.
I'm a bit surprised that in the loop over all values the compiler wasn't able to vectorize it. Pretty straight forward loop to vectorize for a compiler I would think.
I hope that in future Microsoft will add automatic SIMD optimization to normal for loops aswell. C/C++ compiler can vectorize some simple for loops, so Microsoft have here another opportunity for improvement.
if you were forced into a c# version that doesn't have this yet implemented, could one emulate the API and implementation to achieve this? I assume yes
I think you can use interceptors to resolve LINQ in a more performant way (I did this with some things already), but the lack of closure captures makes it very strange. If only microsoft stopped with the "don't modify source code" bs.
"The concept of writing faster code than microsoft is not unique" :D
Yo Nick, what theme are you using for Rider?
my guess is "Visual studio dark"
Which C# and .NET book will be best to read to have good grasp on fundamentals and understand how these internally works?
C# in Depth is still good.
C# in a Nutshell is an amazing book :)
A few years ago I made a video series about the internals here on youtube. It's in German but the auto-generated subtitles seem to work reasonably well.
how about pro C# 10 with .NET 6
Yes, this is kind of interesting, but we all know that linq is most often used to do select, order/sort operations
I'm curious what behavior they are breaking with this, that is allowing this much extra performance 🤔
Memory safety I think? They're using unsafe stuff from what I saw
The github says they don't take things like 'NaN' into account, and that many of the methods are 'unchecked' (meaning they don't check for overflows).
Also mentions there are slight rounding inconsistencies in floating point results, due to values being evaluated in a very different order.
Great Video.
In my i7 4770 CPU for loop is still faster :)
Is it true linq join has poor performance than SQL view??
Back in my Dataset days -- apparently MS intends to download the entire database onto a windows form...
Hi! Could you make a video explaining differences between visual studio and your IDE, plis? Thanks
It’s JetBrains Rider, you can look up comparisons for it
@@JollyGiant19 i know i can look it up. I was asking for It to someone who has expirience with it
VS is from Microsoft
Rider is from Jet brains
Rider is cross platform and works really well. I like Riders default key bindings better.
SIMDeez nuts!
Hi Nick. Could you ban that bot with inappropriate profile picture that keeps posting generic comments like "Wow, the production quality is off the charts" on your videos. I can see it under every video of yours.
It's everywhere on yt. I keep reporting it but yt doesn't give a damn.
I keep banning it and it keeps coming back
I think its hilarious, I reply to it every time with equally nonsensical crap 😂
LoL, try doing something in the Ms or sec range, and check if it's still 10x faster
Wow, the production quality is off the charts
i really think .net ecosystem lacks resources for learning how to write high performance code, memory management etc
I agree.
Fastest way to sum range: ( + ) * ( / 2) :-)
This works only if the interval between each number is the same
Obviously that will work in very particular case when the range represents arithmetic progression, because actually this is the formula for sum of arithmetic progression, but as
@alfflasymphonyx said, not any range is arithmetic progression
@@johnnykeems2911 You are all wrong. Nick talked about the fastest way to get the sum of Enumerable.Range. Well, that's what I wrote. And that was a joke, of course. You don't have to take it seriously.
@@szikig I'm a programmer, which means I'm kind of pedantic person. Your answer joke or not, is generally incorrect, so I dont really understand how I could be wrong. If it is a joke, that's fine, cant read your mind to know when you are joking, I just noticed that this formula only works when the sequence is an arithmetic progression.
@@johnnykeems2911 and this is what is always true for Enumerable.Range, right ?
First() or [0]?
FirstOrDefault usually...
prefer.?Last(to => to.post);
If you have an IList First() will call [0] for you. First operates on IEnumerable so it has other checks to get you the first element for differing implementations, like IPartition.