This Algorithm is 1,606,240% FASTER
ฝัง
- เผยแพร่เมื่อ 5 ม.ค. 2023
- 7 Steps it took to make an algorithm 1,606,242% faster!!!!
Become a backend engineer. Its my favorite site
boot.dev/?promo=PRIMEAGENYT
This is also the best way to support me is to support yourself becoming a better backend engineer.
David Perez
github.com/david-a-perez
B3NNY
/ _b_3_n_n_y_
Bit video
• Bit and Byte Explained...
Code used in video
github.com/ThePrimeagen/aoc/b...
Twitch
Everything is built live on twitch
Twitch : bit.ly/3xhFO3E
Discord: discord.gg/ThePrimeagen
Spotify DevHour: open.spotify.com/show/4MLaDrQ...
Editor
All my videos are edited by Flip. Give him a follow! / flipmediaprod He is also open to do more editing, so slide deeeeeeeeep into his dms.
Join this channel to get access to perks:
/ @theprimeagen
Links
Linode: linode.com/prime
/ discord
Twitch: / theprimeagen
Insta: / theprimeagen
Twitter: / theprimeagen
VimRC & i3: github.com/ThePrimeagen/.dotf...
Keyboard 15% off bit.ly/Prime360 USE CODE PRIME360
#coding #neovim #typescript #programming #vim #softwareengineering #codinglife #webdesign #webdevelopment #webdev #javascript #rustlang #rust #twitch #twitchstreamer #programmerhumor #codinghumor #software #softwareengineer #softwaredeveloper #softwaredevelopment - วิทยาศาสตร์และเทคโนโลยี
Exactly. It's not about Rust vs C vs Cpp vs Zig.
It's about how much you understand system programming.
yayaya
Rust vs Cpp vs C vs Zig is still an important question, but the differences are stuff like memory safety and maintainability
it's not understand of system programming. it's general understanding how a hardware works irl
that's true for algorithms but not for data. how your data is laid out in memory is just as important as the algorithm, and some languages limit the control you have over memory allocation, which makes them unsuitable for any program that deals with large amount of data, like games, audio systems etc.
Zig has clean simple syntax like Go with 90% of Rusts safety and are continually decreasing the margin between Rust and Zig's safety gap. It may even get to parity party soon on safety but with a much cleaner API and the entire C lib interop. I have a strong feeling Zig will take the cake within the next few years.
as a beginner developer, seeing you go from an array to bit manipulation was like watching black magic. incredible to see what we can do with these technologies and break them down to their atomic pieces.
Oh you should definetly look into bit and byte representation of several native datatypes. Even if you later code in something really high level aligning data correctly makes everything soo much faster. And some compilers will choose bitmanipulation over actual multiplication or division if you can keep something in a power of two for example.
@@MrHaggyy
Could you not replace the mod 32(decimal) with And 011111(binary) , I can't make out from the disassembly( been a long time ) how the modulus operator is compiled. But being mod 32 the remainder of that will be the result of And 11111. Mod if compiled as a instruction using division/remainder takes longer clock cycles to execute than an And.
At least that is how it seems to me.
But manipulation is a lot simpler than it seems. The hard part is determining where to use it. Definitely test for performance as you use it.
It can be hard and tricky to implement, but if you can learn how to quickly go up and down in levels of abstraction, it can make all the difference.
For anyone who is interested into these kinds of algorithms. Go look into genome matching algorithms and also in-memory processing. Both categories heavily depend on these kinds of bit manipulation algorithms to optimize known problems like searching for a exact or fuzzy match.
This is really awesome because you don't just show the final kick ass solution, but also present a mental model on how to get from the naive solution to the fast one.
Yayaya
@@ThePrimeagen That's why you're the best waifu
This is actually to increase watch depth of the video. You know the next one will be better so you keep watching.
Really like videos going over practical algerithm optimization. I feel in school there is a lot of focus on theory and very little on how to make things run fast in the real world. Love to see more of these types of videos!
fast? you mean blazingly fast?
so true...cs teachers are setting us up to fail oh well thank god for chtGPT
Theory is important
I feel like we talk a lot about solving problems, and if it’s an acceptable result, that’s good enough. Most of the time, I am completely unaware that things even can be optimised this much. Blows mind.
How can u do optimization without knowing some basic theories?
Tell me you were surprised by using a vector to speed up instead of using a hashset but telling me i am beautiful (and liking the video)
Slight correction.
When I say left shift is multiplied by 10, I meant 10 in binary. Therefore it is two in decimal
it did surprise me. I was always confused how a hashset was O(1) just like a vector. I didn't realize it was O(1*c) which makes wayyyy more sense. Also the bits I didn't know so thank you :)
SipHash 1-3 (rust collections' default hashing algorithm) is fairly slow for integer keys, so this is kinda to be expected, I am quite surprised a vec is faster though, genius move.
It does make sense. My mind immediately jumped to the u32 "set", as I have seen it used before :D The threading was cool, but c'mon, only 64 threads? Should've used a 128 core ;)
I have started reading the Learning Rust with Entirely Too Many Linked Lists, and it has a link to a clip of mr Bjarne S. Talking about vectors vs lists, and why one should use vectors most of the time
You're beautiful
In the (relatively common) case where there's a 14-character window with more than one duplicate, going backwards within the window allows you to find the duplicate furthest along in the input. This means you can skip ahead further in those cases. It is a good example of a 'greedy' optimization looking for the best case (a late duplicate) first.
But what is the difference between finding the duplicate from back to front first or from front to back? You can skip ahead either way, right? Or am i missing something (which is likely because its 4AM)?
@@nimmidev9122 Consider the example "oabcaefghijlml...".
Looking at the first window, 'oAbcAefghijLmL'. If we check from the front, once we encounter the second 'A', we will start the next window after the first occurrence of that character, skipping 1 window. But, if we check from the back, once we find the second (right-to-left) 'L', we will start the next window before the second occurrence of that character, skipping 10 windows.
Since we are predictably accessing only 14 bytes at a time, prefetching and, by extension, reading backwards should not be an issue. On average, we will end up skipping more windows going backwards.
@@nimmidev9122 try the easiest case of "aaaaa aaaaa aaaa...", if you go from the front, you can't skip ahead, if you go from the back, you can skip ahead, so they are clearly different. But I don't know if there are any case where going from the front is beneficial
@@TheJocadasa Yee, makes much more sense now. Idk what i was thinking. Should probably stop watching such content in the middle of the night 😅 Thanks for the explanation.
But what if you have abcdefghijklman? The first duplicate is the second 'a' in the 14th position, but the first unique sequence is the b-n run. If you skip to the second 'a', you miss that entire sequence. Or if you want more than one duplicate, change the 'c' to a 'b' and add an 'o' at the end (abbdefghijklmano). Now the correct sequence is b-o. Either way, it feels like it would miss correct solutions.
More of this is needed! I never knew I needed to know this stuff. Best part of TH-cam is getting mentored by a senior dev without having to be in the same office!
can you recommend some other senior devs on youtube who teach good concepts...
Awesome way of thinking!!
And a note of caution: Although yes the performance increases significantly, the more optimised it becomes, the more specialised and hard to maintain it will also be.
For the purposes of the Advent of Code challenge, it's perfectly suitable.
However for daily development, the tradeoffs need to be considered really carefully.
Finally, before going down the rabbit hole of extreme optimisation, don't forget to *profile your code first* ! You don't want to spend an inordinate amount of time hyper-optimizing a portion of your code that only affects 1% of total execution time. Find pieces of code that affects 10% or more of execution time and focus your efforts there first.
You hit the nail on the head
this mindset is why today we need 500MB to install a shitty app on a phone. back when resources were scarse, it was people job to come up with clever optimizations and techniques. nowadays people are just lazy and hide behind the "premature optimization" bullshit
for a task of the sort as presented in the video, there is NO REASON AT ALL to stick with the "simple" solution. the clever algorithm reminds of en.wikipedia.org/wiki/Knuth%E2%80%93Morris%E2%80%93Pratt_algorithm and is something any decent programmer is supposed to understand easily (possibly with introductory comments to explain some details). the best practice would be to code the optimized algorithm and unittest it against the simple reference solution for correctness.
This sort of mewling adds nothing. "Be sure to check whether the floor is dirty before mopping the floor or else it will be a waste of time" is not the brilliant insight you want it to be. What's really going on is that people like you are often coming from a corporate managerial sort of mindset where you want all your cogs to churn out the most uniformly basic code humanly possible with no intelligent or individualized thought given to things like engineering or efficiency, but saying that out loud in so many words doesn't sound very sexy so instead you go around offering your contempt for people who enjoy the prospect of writing better code as if it is some sort of counter-intuitively enlightened insight you have descended from the heavens to grace us with, but it is no such thing.
This reminds me of the FISR algorithm in Quake 3 that enabled it to run blazingly fast. That was a beautiful abomination peice of code.
I wonder whether there are tools that calculate which of your functions take the most overall time.
Awesome video, thank you! I love it when you dive into the lower-level aspects of programming. Also a nice reminder that approximations can obfuscate significant differences.
More videos like this please! Love the deep dive into blazingly fast computer science topics
Love this content. Really cool and even though I haven't looked at some of the concepts for a while, i can tell you've got fundamental knowledge to bank on. Can't wait to see more!
Tbh this might be my favorite video of you so far. Very well explained, hats off to you! Also congratz to David for the solution, very clever stuff!
Long time fan and this is my favorite video I’ve seen from you, more of this please!!
This is the kind of breakdown I’ve been looking for!! Thank you and your family for this content 😅
Amazing, I really love this type of content and inspire me to always thinking about performance
my man! i forgot to say thanks for the comment.
I like making these videos because its fun!
@@ThePrimeagen Anytime!! I'm waiting for more videos like that!!!
Loved the format of this video, it was explained so nicely and very easy to follow! Ty
This deep dive into optimization was fantastic! Keep up the great content Prime.
Nice explanation! I pretty much went for full speed with no regards for readability / maintainability with this AoC (using rust ofc lol). I managed to get my day 6 solution down to about 4 microseconds on one of the discord callenge bots and I had a rather similar approach to you. The performance on the bot was a bit flakey, but some guy had it down to 2us on the bot which was what I got locally on my ryzen 7 2700x
It's not mentioned, so one way to split the problem for multiple threads, is to divide the string into 64 pieces and ask each thread to solve one piece. You would also need to extend the boundaries of each piece by 14 (maybe 13?) characters on one side so that you can catch 14 size windows that are in multiple pieces. Report the min index from all threads as the final answer.
That would parallelize the list, but has 1 minor issue. We are looking for the first instance of a unique string, and one of the threads could find a unique string later in the sequence before the threads looking through the beginning of the sequence found the first unique string. Maybe, set up a queue for the threads. There are different ways to do so, and in this instance, you could set up two variables.
queue_end = total_string_length - window length
queue_value = 0
A thread would take a window with its starting position at queue_value and increment queue_value by one, so that the next thread would take the next window, and so on, until one of the threads found a unique sequence or queue_value reaches queue_end
@@daedalus_00just do one window per thread. Mark the order of the threads and return the lowest occurrence that no windows to complete before it.
@@LtdJorge And that all will be slower, because threads in programming language are abstractions, that involve allocation lots of memory and creating bunch of stuff, but in the fastest solution in the video there's direct use of simd instructions of the processor, which is on the whole another level
@@daedalus_00 Every thread will return the index where it found the first unique string. This index is the index in the original string and not the smaller piece. So we can take the minimum of the indexes returned by all the threads at the end as the final answer.
@@loam Multi threading works along with SIMD. The function that solves each part of the whole string is exactly the same as the one that solves the entire string with SIMD alone.
If the input is large enough, the multi threading optimization can be taken a step further by using GPUs and distributed systems. Both optimizations have their pros and cons.
I just had my imposter syndrome gone, and then I watched this video. Damn it!
Great video, a lot of small details and bits of information. Being a nerd is a lot of fun. Thanks ;)
This is the best content on TH-cam. I'm about to do your algo course on front-end masters, cant wait.
Thank you . Great content. I’m not very proficient with material related to any STEM type of work but I was very interested in this video
The stack allocated array and the vector array both have cache locality benefits due to them _both_ being contiguous. The difference is the vector is making a syscall (potentially) to get memory off of the heap. You could reuse the vector after an iteration and you could speed up that variant some. The stack allocated array, gets further cache locality, since you don't even need the first fetch of a cache line - its already in the cache.
But both have the benefits of spatial and temporal locality.
This was extremely useful!
Please more of this when you can. Lots of Advent of Code days to play with xD
I'd really love to see more videos like this about how crazy optimization can be. It was really intriguing to see how much overhead there is in what seems like the best approach, and how quickly a solution can run if you remove all of that.
Honestly this was such a great video congrats and thanks!
I watched this all unfold on twitch but love having the video here to review. As someone that didn't get a formal education in CS, this low level stuff is so helpful to be able to return to.
I'd love more of these code optimization videos, like okay maybe we're not using these techniques and optimizations everyday but so much fun to nerd out about them
Also worth noting that reverse iterators in rust don’t first go forwards, they rely on double-ended iterators, which for things like slices and vectors can start from the end
That's only sometimes true. You just have to make sure you have a data structure with a known size.
good thing it only did it for 11.9 milli seconds because you'd have a pretty hard time keeping your CPU fed at 617GB/s
@9:40 You don't actually have to collect up all the elements in the iterator for this! In Rust, if an iterator implements DoubleEndedIterator then it can iterate in reverse. std::slice::iter implements this trait, so calling any of the "r" iterator methods will just start counting from the end instead of from the start.
The assembly suggests that optimization is being applied by the compiler, since it is already iterating in reverse (over completely unrolled loops). I might be misunderstanding you: are you saying there are more things that don't need iterated (from your first sentence), and the second sentence about reverse iterator was a separate note? I don't use Rust, so "collect up all the elements in the iterator" might have a Rust technical/idiomatic meaning that I am missing.
I love this type of content.
Really cool video, Prime. Learned a lot. Ty!
David's version is a clever use of the Boyer-Moore heuristic of good suffix for string pattern matching
3:51 you get cache locality with vectors too, it's more about accessing the heap which takes longer than accessing on the stack since everything is known compile time as opposed to having to look up the page (might not or might not be in the TLB), deferencing, etc.
I think a step you missed in your explanation that would make the approach more clarifying is to mention explicitly that you want to use each bit to represent whether a certain character has been seen, and that you use the character's ordinal representation modulo 32 to determine the index of the bit this information is stored in. I think it's worth remarking that you can only do this because all the characters in your input are adjacent in their ordinals and there is less than 32 of them (or else you'd get position clashes).
i really enjoyed this video. I would love to see more of these
I just started learning about hash tables, and immediately saw incredible performance boost. I would have never imagined that an array or a vector could do the trick
If you can use arrays, they are always the fastest
This video: A+ I really enjoyed seeing some interesting optimisation techniques and in Rust which I am currently learning. Looking forward to some more [amazingly fast] optimisation algorithms and techniques 😁 and in Rust
I know I’ve seen you do this many times, but I think it would be SUPER cool to see you do the setup and generation of stats/graphs for perf. The more I perf test, the more I see people are often optimizing for the wrong things
I like the comments about the order of hashing. I think in university alot of the time they fail to mention the constants in the complexity conversation and you often end up with people always thinking hashmaps and sets being the fastest completely unaware of the cost of hashing and the common list iterator for collisions.
When I went to university, hash sets didn't even have linked lists inside of containers. We actually used linear and exponential back off to determine it's insertion location.
It's a neat improvement to how much space you need for a set, and it also seems pretty fun to implement.
@@ThePrimeagen there's a cpp con lecture about googles flat hashmap implementation I watched years ago that was pretty interesting in terms of the decisions they made with writing a faster implementation. One of their optimizations was to use no pointers at all so once the hash gave an address you would just iterate across the key value pairs till you found your equality. On average you wouldn't need to iterate across many elements.
Implementation details matter. It’s important to have a rough idea how data structures work to understand the performance implications.
6:33 That is NOT the Logical AND Operator (which is &&) but actually the Bitwise AND Operator (&).
Great video! I learnt a few things here!
I'm going to need to re-watch this a few times to get my head around those improvements. 😀
I have 0 understanding of wtf are you doing but i'm enjoying this so much for some reason, great vid!
The idea of iterating from the back of the window reminds me of the Boyer Moore algorithm for substring search.
I love this style of rust. It looks pretty confusing at first sight but it's much better than for loops with random looking operations. The algorithms kick ass. And the browsing through at low level was pretty neat as well. It's pretty wild to consider how a human would just look at the 15th value and say "I've seen that already" instead of comparing all of them individually. Despite being pretty slow at some things, human brain just keeps amazing with all its optimizations.
The Perez algorithm was really cool, I only recently saw it explained somewhere but I didn't understand it originally nor did I remember it. What surprises me is that it isn't part of standard algorithms.
Hi, love your content! Can you share your thought process behind the use of excel and how you populate the data there?
thank you for this blazing fast series.
Thanks for the awesome content! Minor quibble: left shift always increases by the base, in binary that’s 2, not 10, it just looks like you multiply by 10 because the only digits in the numbers are 1 and 0
I actually look forward to watching your videos... something rare for me on youtube nowadays
Wow! That one I'll need to reawatch a bunch of times to get what's happening on those improvements.
Hey man I’ve been really enjoying your content. I’ve been developing since 1994 and just came across your channel and have spent quite a few hours listening to you. Especially enjoying the friendly controversy with Theo. Ha! Also, awesome stache. Keep on keepin on.
“Binary ten thousand” is a phrase that makes be very uncomfortable. Lol
man you just say "blazingly fast" and i automatically click the like button.
omg i just realized i did that too
More technical videos on algorithms. This was fantastic
Nice tricks. I would appreciate a video like that but optimizing multiple network calls like an api gateway
Great stuff, a lot of fun stuff happens at the bit and byte level
Your understanding of concepts always amazes me. I just got off learning JavaScript and now learning c++ and data structures for job placement but my true goal is to be like you. Have a deep knowledge of code and write blazingly fast solutions.
blazingly fast is fun, but it's really all not that important in the workforce
@@codesymphony it depends on who you work for. in my job having to process 50+ million records does require inner knowledge to deliver better experiences
@@epicgameryt4052 Correction to symphony's comment: *not that important in the majority of the workforce.
@@earthling_parth not sure I agree. The fact that front end developers have a reputation for being less educated on CS fundamentals doesn’t excuse the utility of knowing them.
@@TehKarmalizer Games is a great example, programmers used to rely on fast and optimized code to the game even run, nowadays it seems that it as shifted and more and more optimized garbage shows up every day.
this is super interesting, upload more content like that!
Hey man, really like your videos. BTW, which laptop do you use?
S76 Lemur Pro
This is absolutely fascinating. I’ve a few questions:
1) How does one goes about developing these fast solutions? In order words, how does one learn to develop the skill of algorithmic thinking?
2) There’s always a tradeoff between performance and readability. So how realistic is it to use the fastest solution in a real world code base?
1. well you have to determine if the speed up is meaningful
something simple as "just because its O(1) doesn't mean its fast" is a great thing to remember. Experiment to find what makes a difference
For a real code base, you write your code in a readable way, then measure it. If it's not fast enough, find where it's slowest in the hot path and make that bit faster. Repeat until it's fast enough.
@@ThePrimeagen
I think that a good take that big O notation is not about whether an algorithm is fast or slow. It is about if it scales.
A O(n) solution that is slow but viable at small scales will continue to be slow and viable at large scales.
1) Never stop programming new things/projects. This will challenge your intuition and improve it.
2) It depends on the use case. When performance is critical, readability is sometimes sacrificed, though this can be mitigated by, for example, using comments explaining what's going on. However, it's often more encouraged to prioritze readability when performance isn't a concern, so as to make the code easier to scale
@@georgehelyar This, or see if you can abstract the problem to a more general problem where optimizations are already known. An easy example example: sorting characters can be abstracted to sorting numbers, and sorting algorithms for numbers are well known.
savage. went from web dev to game dev in 7 steps. blazingly fast indeed, ty for sharing sir!
Oh wow thank you, this opened my eyes to much.
The array solution is the one I would have come up with, but I think there may be a little room for improvement there, depending on what the compiler is already doing with that code.
You are presenting everything with such an ease, and it is highly understandable to me, self-taught (no university degree) DevOps engineer. If you were my teacher in high school I would be probably working for NASA or Roskosmos now... Make a course related to Rust development, Rust for beginners or similar...
This video shows perfectly why I love writing high performant code. You get to be creative: Use your existing knowledge, e.g. the Boyer Moore algorithm for string matching, to come up with solutions for new problems. Know the internals of the underlying system to select the right data structures. All theory aside, you have to actually test and experiment. Measure something odd? Then fire up that compiler and check the assembly. It just never gets boring!
Most exciting
Nice video
Quick feedback, i think the explanation at 7:05 is a bit misleading as in in this case with the 0u32 as filter XOR is like a toggle. For me it sounded like you said XOR itself is like a toggle
i explained everything i could in less than 1 minute :)
i can only do so much in so little time
The going backwards to skip large number of characters was likely inspired by the Boyer-Moore string-search algorithm, which is also what powers "grep". It's actually faster to grep for long strings than it is to grep for short strings because of this optimization. Longer search strings let you skip faster.
I love the pipeline from 3 line solution to x86 assembly
Appreciate you King 👑
kind or tempting to try it in assembly now :D
Hey Prime, where should I start to learn about systems programming, any book or side project you recommend? I already build some small CLI applications in Rust, but it is hard to find a use case for me to learn about bits and performance, without it being too much to learn.
I didn't understand most of it but still loved the video! Thanks!
well howdy! you will some day :)
The greatest video I ever seen in TH-cam, Well Done Captain!
This was cool! More content like this please.
Good content Prime, amazing stuff
I use this bit manipulation to store a series of flags as a DWORD to keep track of e.g. user configurations in a database. Much more efficient than storing a zillion of true/false fields. Never thought about using it to compare hashes. Good one 😁
This is amazing thank you
In Julia, the naïve solution (using Sets and moving window just +1) runs in 7-8ms and the fast solution in 7-8µs on chars, or ~3µs if the input is already bit states. (Didn't bother doing a threaded version.)
Great content. 🙏🏻
love optimization Videos! Good one :)
As a new programmer, and an aspiring front end web dev and indie game dev, I don't understand a single fucking thing you're talking about, but you get me excited to learn.
But uh...that really was blazingly fast. Maybe I'll change career paths.
👁️ 👄 👁️
really enjoyed this. Question: In a practical environment, how would you gauge speed vs maintainability? Almost no engineers would understand the final solution whereas I think most would understand the middle array solution. Just curious on your thoughts for that trade off?
Depends. If the readable speed is ok for your use case then go with readable speed. If u really need to get that speed then u bite the bullet, extract this code into an isolated module so that when requirements change then you can rewrite that solution (there is no way you will be able to refactor that code to fit new requirements after a long time, it is easier to write it from scratch)
You make my brain hurt... in the most pleasing way possible.
Woosh. A joy to see!
My mental try before the video ended at the one from David a perrez, cool optimization problem :)
Ok but that's like one of the best layman breakdowns of bit manipulation I've seen
Honestly I expected to see a mind-bending mathematically motivated algo but this is still very impressive
MORE OF THIS!!! luv u
Instant like for performance optimization content!!🎉
Yes 2023 off to a great start. These videos are the secret sauce to this channel in '23 🔥
Blazingly fast, indeed. Love this channel
love this blazingly fast themed videos 😂
Very clever adaptation of the Boyer-Moore algorithm !
this was awesome. more please
Wow, amazing stuff. I love this type of content.
excellent!
It's funny, I didn't even think of the hashset approach and went straight for mask & lower(c) as u8 - 'a' as u8 (bringing the number into 0 and 25). I did use the windows iterator function in rust though, beautiful utility for this kind of thing.
Did you benchmark early return in the bitset vs bitwise or all character indicies and use the popcnt intrinsic?
Sometimes being branchless is better for speculative execution and vectorisation.
correct, i am unsure if i looked for that, its been a few weeks and cannot rememebr :)
The David solution resembles the Boyer-Moore string-search algorithm. Nice video!
Is there any chance for Elixir vs content in near future?
At around 5:40 you use modulo to turn a letter (a..z) to a number between 0..31. Modulo is overkill here. There are only 26 letters, so a simple subtraction c - 'a' would give you a number between 0..26, with 'a' = 0, 'b' = 1, and so on.