Codeforces: С++17 64bit Support

5 years ago, # |

+507

I would donate for this.

→ Reply

Vanilla_

5 years ago, # ^ |

Agree, if Mike would start a patreon for this, the goal would be achieved in no time

→ Reply

Sonic_AR

5 years ago, # |

+10

i would pray for this

→ Reply

5 years ago, # |

how to initialize int128?

→ Reply

5 years ago, # ^ |

-6

To use it you should initialize variable as __int128 and use %llu with scanf and printtf

Tested on problem 4A — Watermelon

There is my submission : https://mirror.codeforces.com/contest/4/submission/73821176

→ Reply

5 years ago, # ^ |

+10

Sorry it doesnt work My bad

→ Reply

U can use two long longs for doing it faster

5 years ago, # ^ |

← Rev. 5 →

-13

I found out that in order to use it every number it work with should be __int128 too. But you cannot read int128 or print it by using cout or cin, scanf and printf dont work too.

Example:

__int128 num, hundred = 100, ten = 10, zero = 0, one = 1;
string s;

int main() {
  num = one<<hundred;
  while (num > zero) {
    s += "0123456789"[num % ten];
    num /= ten;
  }
  reverse(all(s));
  cout << s;
}

→ Reply

dmkozyrev

5 years ago, # ^ |

+15

→ Reply

purplesyringa

5 years ago, # ^ |

+11

Adding anything to std namespace is undefined behaviour. While it probably won't cause any problem, it's better to be safe and define to_string outside std (argument-dependent lookup makes the usage even easier).

→ Reply

snowflame

5 years ago, # ^ |

-8

We can use quickread and quickwrite to initialize __int128.

→ Reply

Laggay

5 years ago, # |

← Rev. 2 →

+285

What's something to say when starting a blog or getting a handjob?

Mike: I hope you wash your hands and feel great.

→ Reply

Prometheus

5 years ago, # |

-15

Finally! I can't wait to try this out.

Thanks Mike

→ Reply

dalex

5 years ago, # |

+83

FYI, Rust supports int128. I once tried to learn it, and... now I'm washing my hands.

→ Reply

taran_1407

5 years ago, # |

-124

"I am slightly worried that the presence of such features may widen the gap between C ++ and other languages. Wait and see."

It'd anyway compensate if time multipliers would be considered for slower languages. (Feature Request :P)

→ Reply

f22raptor

5 years ago, # ^ |

+18

The Church of C strikes out with the infamous downvote :P

→ Reply

5 years ago, # ^ |

← Rev. 2 →

+19

I'm always stunned by how much people dislike the idea of time multipliers. It decreases the gap between languages. I often want to give bigger TL in div2-ABC for Python, and sometimes slightly bigger TL for Java in hard problems.

Time multipliers can help to maximize the ratio slowestIntendedSolution / fastestNaiveSolution (assuming that running time is divided by time multiplier).

→ Reply

_dobby_

5 years ago, # ^ |

+71

This reminds me of a local contest we hosted on HackerRank, the expected solution for the hardest problem (related to strings) was O(nlog(n)), but we had several Bruteforce (O(n^2)) solutions accepted due to the large multiplier for python language (x10) on HackerRank.

And it turns out that python is not so slow with strings.

→ Reply

5 years ago, # ^ |

+12

This reminds me of thousands of times when people can't solve Codeforces problems in Python.

Yes, organizers need to be careful with multipliers. It's stupid to give Python a x10 multipier. Same for Java, some simple operations can be equally fast as in C++.

→ Reply

_dobby_

5 years ago, # ^ |

I wasn't taking any side. I am not some learned person to wisely put forward my views on this matter.

It was just some incident that I thought might be worth noting in this debate.

Sure there must be workarounds or ways to handle this situation, but we didn't consider this as a possible problem before setting up.

→ Reply

dunpeal

5 years ago, # ^ |

← Rev. 7 →

You bring up a valid point, that a constant multiplier can be too blunt: the performance gap between Python and C++ isn't constant, so a constant multiplier in some cases would be too much, and in other cases too little — i.e. you wouldn't be able to write an accepted solution in Python with a multiplier that's too low.

Personally I love using Python, and I wish solving this problem was simpler.

One possible solution is to ensure that each problem has an acceptable optimal solution in Python, but if that requires a multiplier, also ensure any suboptimal solution won't pass. Obviously that's a lot of work for the setter, and very hard to get right: a Python expert may be able to write a faster naive solution than the setter's fastest naive solution, and then get AC with the multiplier. Even if the setter spends all that effort and gets it exactly right for Python, competitors will complain that their favorite dynamic language doesn't get the same treatment, so the setter will have to spend the same effort on all 20 supported languages, which is simply unrealistic.

So as much as I love Python, I must agree with riadwaw below that the most practical solution is to standardize on C++. Then the problem setters only have to check the above (optimal solution passes, suboptimal solution TLEs) for 1 language instead of 20.

We should also mention that it's possible for a language to be so fast that a naive solution would pass. In practice if you make sure that doesn't happen for C++, then you are pretty safe for all other languages since they tend to be slower than C++ across the board. That feature of C++ is a big reason to standardize on it.

→ Reply

riadwaw

5 years ago, # ^ |

+58

I think any constant multiplier will be bad, because in some problems it's possible to write solution that will spend most time in the compiled arts written in C/C++, which will take time close to time in C++.

This leaves with the option to set the TLs manually, but that requires authors to write solutions in 20 languages and they better know them well, not to leave these loopholes.

I think that leaving hard problems c++-only is lesser of evil. (And for easier problems you can often just reduce the constraints anyway)

→ Reply

5 years ago, # ^ |

+20

It just happens a lot that organizers have solutions in Python for easy problems and in Java for hard problems and they see that slightly different TL for a different language would be better. Time multipliers shouldn't be default for sure.

in some problems it's possible to write solution that will spend most time in the compiled arts written in C/C++, which will take time close to time in C++.

I agree that this is dangerous and a good argument to avoid time multipliers.

→ Reply

riadwaw

5 years ago, # ^ |

+76

After all, "do something so that our computer can answer these questions(tests) in 2s each" sounds reasonable, while "do something so that our computer can answer these questions(tests) in 2s each, but if you know only python, 10s is fine" is more strange

→ Reply

GottaGoFaster

5 years ago, # ^ |

Yeah, that's why it's wrong to give Java x2 multiplier!

→ Reply

plagues

5 years ago, # ^ |

+36

Yes, but languages has some dignities and some shortcomings, and language speed it's one of shortcomings, you can choose other language if you do not like this.

→ Reply

5 years ago, # ^ |

A lot of programmers know Python and they are discouraged from participating in Codeforces because TL is adjusted to C++ and Java. Are you sure that "learn a new language" is the best solution here?

→ Reply

plagues

5 years ago, # ^ |

← Rev. 3 →

+13

Hmmmm, what if I know only C++, but python has some functions what i need and C++ has not it? Codeforces must import this functions in C++? No, because it is shortcomings of C++ and I can choose other language if I do not like this.

→ Reply

5 years ago, # ^ |

+20

C++ is anyway better suited than Python (for CP). You're saying that we can't make two languages equivalent so we shouldn't decrease the gap even if it's easy. I strongly disagree. Making Python viable in div2 would greatly broaden the competitive programming community in the long run.

→ Reply

dunpeal

5 years ago, # ^ |

← Rev. 2 →

You're saying that we can't make two languages equivalent so we shouldn't decrease the gap even if it's easy.

The problem is that it's not easy.

You need to get the multiplier exactly right, so the optimal solution passes but all suboptimal solutions TLE. That's very hard to do even just in Python, and you'll have to be a Python expert to know how to write the fastest naive code. Now multiply by the 20 languages Codeforces supports and it's obviously impractical.

Even just for Python it may be impractical. Because of extreme differences between certain optimal and suboptimal techniques. For some problems, I may be able to use techniques that rely heavily on CPython's C routines and thus approach C performance, and with the multiplier I will AC naive solutions, but you will not be able to remove the multiplier, because if someone doesn't know these specific techniques they won't be able to get an optimal solution to AC without it.

→ Reply

5 years ago, # ^ |

No, you don't need to get the multiplier exactly right. Almost always it's true that naive C++ solution is faster than naive Python solution, and intended C++ is faster than intended Python. Then multiplier of 1.001 for Python always improves the ratio slowestIntended/fastestNaive. And almost always the multiplier of 1.25 will be good too, sometimes even 2 or 5.

Even if the setter spends all that effort and gets it exactly right for Python, competitors will complain that their favorite dynamic language doesn't get the same treatment, so the setter will have to spend the same effort on all 20 supported languages, which is simply unrealistic.

(this is quote from your other comment above)
Helping users of just Python is much better than not helping at all. Other languages are screwed in both cases.

→ Reply

dunpeal

5 years ago, # ^ |

← Rev. 5 →

And almost always the multiplier of 1.25 will be good too, sometimes even 2 or 5.

That right there is the problem:

A multiplier of 1.25 sounds nice and harmless. Indeed, you could probably apply a 1.25 constant multiplier safely right now. Why? Because it does almost nothing.

Python is generally slower than C++ by a factor that's far larger than 1.25. So 1.25 won't do much harm (you won't be able to cheese with it). It also won't do much good. If a problem was unsolvable in Python, a multiplier of 1.25 would almost never help you.

A multiplier that would actually make a difference would be closer to 5, which I believe is the multiplier on CodeChef. But then of course you are risking cheese if I can delegate most of the work to C routines.

I'm personally all for improving Python usability on Codeforces, but we should be aware of the potential issues. Incidentally, there's a 3rd-party project to do that, PyRival.

The best solution an OJ could implement would be to compile all solutions to the same Intermediate Representation (IR) that runs on an identical virtual machine. You can then calculate the exact computational cost of each solution, in terms of primitive operations. You could also just run it and expect the running time to be equivalent.

A second fundamentally correct approach would be to compute the actual time complexity of the solution by running it through multiple input sizes and calculating the running time growth curve.

The first approach is a pretty big project, but its result will be a drop-in replacement for the various platforms currently used by all OJs. The second one is a bit more realistic, but of course requires some more work from the setters and will not immediately work on existing test suites.

Notice that by standardizing on C++, as the CP community has effectively done, we get all the benefits of the first approach without the cost of an ambitious and difficult VM project. That's why most people are pushing to keep it as the status quo.

My guess is that we either stay in this current status quo, or we start applying multipliers to Python, and then we will have cases when people cheese naive solutions by delegating to C routines. Personally, I'm sympathetic to the argument that this is a small price to pay for the great benefit of making CP more widely accessible and inviting to newbies. Perhaps the best realistic approach would be to apply multipliers in div2 but not in the more competitive divisions and contests.

→ Reply

https://mirror.codeforces.com/contest/1305/submission/73826085

5 years ago, # ^ |

What I want as a setter is an ability to choose a different TL for Java and Python sometimes. It doesn't matter what be be a good constant multiplier, my examples with numbers were there to show that we don't need to get the number exactly right.

→ Reply

jai_dewani

5 years ago, # ^ |

I agree to that, I have good experience with Python because I use it for development hence I use it for CP as well as it's easy for me to just start writing code in python without even thinking much (Time is of utmost importance here, right? :p).
It is almost like a third language to me after Hindi and English. It took me a long time to get to this comfort zone, learning another language to this fluency will take up a long time for me.

I believe people like me who don't wanna learn a whole new language because they enjoy doing CP for fun in their native language, multipliers should be taken care of at least for div2 because that's where we are most of the time.

→ Reply

shamdin

5 years ago, # ^ |

Wouldn't it be great if numpy was supported?

→ Reply

jai_dewani

5 years ago, # ^ |

Yes, I learned just some months back CodeChef does support numpy!

→ Reply

Chilli

5 years ago, # |

← Rev. 3 →

+65

With this improvement, we can now cheese https://mirror.codeforces.com/contest/1305/problem/F with a fast factoring implementation (taken from KACTL + Montgomery reduction):

Thanks Mike!

→ Reply

hoke_t

5 years ago, # ^ |

← Rev. 2 →

+31

chilli orz

→ Reply

navneet.h

5 years ago, # ^ |

← Rev. 2 →

you could also try delaunay triangulation using int128, my submission for simple question, also i want to ask what is this error,Judgement failed

→ Reply

Ari

5 years ago, # ^ |

+53

All the work that went into preparing that problem is now dead QAQ.

→ Reply

Narut

5 years ago, # |

← Rev. 2 →

+65

time to

#define int long long

in my template

→ Reply

pajenegod

5 years ago, # ^ |

+41

Make it

#define int int long long

for some extra style points

→ Reply

Priyam2k

5 years ago, # ^ |

← Rev. 3 →

+26

More style points

#define int long int long

→ Reply

UnbeatableMaster

5 years ago, # ^ |

can you explain it?

→ Reply

jef

5 years ago, # ^ |

Order of type modifiers doesn't matter, so #define int long int long replaces int with long int long, which is the same as long long int and long long.

→ Reply

HipHop

5 years ago, # |

+55

Wow! Now we need more contests on the quarantine please.

→ Reply

5 years ago, # |

← Rev. 2 →

+16

Main question: how do 64-bit instructions perform in this mode? Are they still much slower than on 64-bit systems?

Second attempt: 73831074, judgement failed — system failed to run file. Dunno why, but I get the feeling that's not supposed to happen.

→ Reply

okwedook

5 years ago, # ^ |

-8

I also have a similar problem

73994883 and 73994843 are absolutely the same, but C++17 (64) gives WA, while C++17 (32) passes.

→ Reply

5 years ago, # ^ |

That could be undefined behaviour. Either that or a compiler bug, but it makes sense for UB to happen only in one version.

→ Reply

Igorjan94

5 years ago, # ^ |

O_o

const int N = 502;
...
int16_t dp[N][N][N];
...
for (int len = 1; len <= 10000; ++len)
     checkmax(dp[i][j][len], dp[i][j][len - 1]);

→ Reply

okwedook

5 years ago, # ^ |

You obviously miss the point. In each for-loop under it I check i + len <= n and j + len <= m.

→ Reply

Igorjan94

5 years ago, # ^ |

Ah yes. By default skipped these lines "there is a usual for"
So, this is really strange behaviour

→ Reply

ffao

5 years ago, # ^ |

+14

You use the value of R.p without initializing it, which is undefined behavior.

→ Reply

okwedook

5 years ago, # ^ |

Yeah, I suppose that's true. Now I can't trust anything!!

→ Reply

YahiaAglan74

5 years ago, # |

+21

This is the best thing happen in 2020

→ Reply

I_love_natalia

5 years ago, # |

← Rev. 2 →

+62

Imagine how wide the gap between C++ and other languages will be if we have compilers that are not obsoleted (eg 19.24 for MSVC and 9.0 for clang) with C++2a turned on...

→ Reply

NaOH_Frog

5 years ago, # |

Now I would definitely not meet my ancestor because long long isn't enabled. :)

→ Reply

forget_it

5 years ago, # |

Can someone please explain ,what this 64 bit g++ means.

Under which computer science topic ,does this knowledge come. (presently ,i only know how to code in c++,and topics like OS, COA basics are undergoing in my college)

I want to know what is this ,so that i can join this experimentation and testing.

→ Reply

https://github.com/GaurangTandon/codeforces-darktheme

5 years ago, # ^ |

The topics are CPU/low-level programming and OS.

→ Reply

jef

5 years ago, # ^ |

← Rev. 3 →

The relevant topics are probably processor architectures and other low level stuff. 32 bit and 64 bit refer to the word size of the processor.

→ Reply

oversolver

5 years ago, # |

+19

well done, what goal for dark mode?

→ Reply

sushkov.pv

5 years ago, # ^ |

→ Reply

spookywooky

5 years ago, # |

xalanq, I would be super happy to see that option in cf-tool Thanks for the work!

→ Reply

AkiLotus

5 years ago, # |

+16

Just curious, but will this support available in Polygon as well?
Speaking of which, I'd be glad if Polygon supports PyPy, since its behavior compared to Python is completely different (other than just "significantly-faster") and Python solutions should be tested equally on both when preparing problems. ;)

→ Reply

nishant.ns19

5 years ago, # |

← Rev. 2 →

Can anyone tell me how to add

<boost/multiprecision/cpp_int.hpp> and

library for big integers on mac(c++) ?

P.s. I am currently using sublime for coding with olympic fast coding add on.

→ Reply

madhur4127

5 years ago, # ^ |

← Rev. 2 →

Boost.multiprecision is a header-only library so brew install boost should place the required headers in boost subdirectory under /usr/include.

Alternatively, you can download the headers from Boost.Multiprecision's GitHub repository and place it wherever you want but make sure to add the path of the downloaded directory or use -I compiler flag instead.

→ Reply

nishant.ns19

5 years ago, # ^ |

I will try it. Thanks :)

→ Reply

jef

5 years ago, # ^ |

+11

You can't use boost or other third party libraries on Codeforces or any other competitive programming site that I know of though.

→ Reply

nishant.ns19

5 years ago, # ^ |

I once used it on Codechef .

→ Reply

BITree

5 years ago, # |

← Rev. 4 →

-34

...

→ Reply

mango_lassi

5 years ago, # ^ |

+14

Should we ban using integers in python because they can be arbitrarily large? This argument is just ridiculous.

→ Reply

BITree

5 years ago, # ^ |

← Rev. 2 →

-19

...

→ Reply

BITree

5 years ago, # |

← Rev. 2 →

Now I can use #define int __int128 ????

→ Reply

Pecco

5 years ago, # ^ |

Note that you have to implement I/O by your own, because __int128 is supported by neither cin/cout nor scanf/printf.

→ Reply

prakhar_111

5 years ago, # ^ |

And how exactly to do that? Can you give a simple example?

→ Reply

Pecco

5 years ago, # ^ |

If you've ever used fast read-in, you may feel familiar with this:

__int128 read()
{
    __int128 ans = 0;
    int sgn = 1;
    char c = getchar();
    while (!isdigit(c))
    {
        if (c == '-')
            sgn *= -1;
        c = getchar();
    }
    while (isdigit(c))
    {
        ans = ans * 10 + c - '0';
        c = getchar();
    }
    return ans * sgn;
}

As for output:

void print(__int128 x)
{
    if (x > 9)
        print(x / 10);
    putchar(x % 10 + '0');
}

→ Reply

mango_lassi

5 years ago, # |

← Rev. 2 →

+25

YES YES YES! Finally convex hull trick isn't obnoxious to code! Finally we can factor large integers! Best feature ever!

→ Reply

ksun48

5 years ago, # ^ |

+184

Featured in next cf round: a[i] <= 5*10^37 in input.

→ Reply

Golovanov399

5 years ago, # ^ |

And how am I supposed to read this

→ Reply

Neroast

5 years ago, # ^ |

+44

As a string and then convert

→ Reply

Spheniscine

3 years ago, # ^ |

+20

Took slightly over two years, but it's happened 1656H - Equal LCM Subsets

→ Reply

Chilli

5 years ago, # ^ |

← Rev. 2 →

What's annoying about either of these? I don't think I've seen people setting obnoxiously large points that require convex hull.

→ Reply

mango_lassi

5 years ago, # ^ |

Whoops, I meant convex hull trick.

→ Reply

Bishwanath

5 years ago, # |

What about for the Linux users? :)

→ Reply

Norrius

5 years ago, # ^ |

Just install g++, your system already supports it. Windows users have to resort to Cygwin or MinGW to get GCC because GCC primarily targets Linux.

You could use alternative compilers such as MSVC, but then the resulting binary might be significantly different between your local machine and the judge.

→ Reply

avm

5 years ago, # |

On Linux, gdc is shipped with gcc9. Does gdc work in given msys2 distribution? If it does, it is possible to consider adding this alternative DLang compiler on Codeforces in a separate slot?

→ Reply

Abdelrahman_Alaa_

5 years ago, # |

** Solution to include bits/stdc++.h in visual c++ **

Because I really suffered from remembering all libraries you can add #include <bits/stdc++.h> which includes all libraries

This is just a method i used to make me able to include bits/stdc++.h
in visual c++.
for those had minGW installed on PC :
C:\MinGW\mingw32\lib\gcc\mingw32\4.8.1\include\c++\mingw32\bits
copy this folder and then go to this adress
C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\include
paste your folder. go to your visual studio project type bits you will see
the auto-complete for the library and then choose stdc++.h
for those don't have minGW:
you should write your own header file and include all libraries in it then
go to C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\include
make new folder name it "bits" and name the header file stdc++.h
then paste it in "bits" folder.
Hope this helps you!
Happy coding

→ Reply

5 years ago, # ^ |

Alternative: don't use bits/stdc++.h because it takes way too long to compile, just use a few common includes.

→ Reply

p__q

5 years ago, # |

Finally! Thanks, Mike

→ Reply

Falcon__

5 years ago, # |

← Rev. 3 →

-23

emmmm,Sorry

→ Reply

AkiLotus

5 years ago, # ^ |

Butthurt with other languages' fancy libraries?
Keep in mind that referring to other libraries and calling methods from class will be slightly slower, not to mention big integer arithmetic by strings cannot match actual ALU-in-depth arithmetic support. That explains how BigInteger (Java) and Python int with large absolute value is significantly slower when used.

→ Reply

Falcon__

5 years ago, # ^ |

Sorry Sir,I am too excited with int128's appearance in codeforce ,I think I should apologize for every pyer and javaer~ Sorry bros!

→ Reply

5 years ago, # |

Thanks MikeMirzayanov for the kind correspondence and for this good news.

P.S. Please consider fixing the following typo in the first prompt of the pbox command

PBOX is abscent or not the last release. Downloading...

It should be written in English as

PBOX is absent or not the last release. Downloading...

→ Reply

OrientalHorizon

5 years ago, # |

← Rev. 2 →

Is scanf("%I64d", &a); also available while reading a 64-bit integer in this mode, or we have to change it?

→ Reply

5 years ago, # ^ |

← Rev. 4 →

Backward compatibility is often among the features that software developers and engineers aim to maintain when introducing new versions of computer language compilers.

Check the following performance comparison using the same code:

GNU C++17 9.2.0 (64-bit) 30 msec and 12 KB 74014633
GNU C++17 7.3.0 (32-bit) 31 msec and 8 KB 74017705

→ Reply

tribute_to_Ukraine_2022

5 years ago, # |

+34

#include <bits/stdc++.h>
using namespace std;
const int nax = 2e5;
using ll = long long;
bitset <nax>b;
int main() {
	ll ans = 0;
	int p = 0;
	for (int i = 0; i < nax; ++i) {
		b[p] = 1;
		p += 99827;
		if (p >= nax) p -= nax;
		ans += b.count();
	}
	printf("%lld\n", ans);
}

I tried running following code in custom test. Results:

gnu g++17 (64bit):

20000100000
=====
Used: 2807 ms, 20 KB

gnu g++17 (32bit):

20000100000
=====
Used: 2448 ms, 28 KB

Shouldn't bitsets be faster on 64bit machines?

→ Reply

box

5 years ago, # ^ |

← Rev. 3 →

-30

Anything using bitset is profoundly slow. Any container in STL is profoundly slow, actually, but bitset is one of the slowest containers. I stopped using it after getting TLE on several problems from using the STL.

If it's really necessary for an 8x memory optimization, it's possible to write a bitset by hand, with an array of uint8_t.

For reference, this code

#include <bits/stdc++.h>
using namespace std;
const int nax = 2e5;
using ll = long long;
uint8_t b[(nax+7)>>3];
int main() {
	ll ans = 0;
	int p = 0, bc = 0;
	for (int i = 0; i < nax; ++i) {
	    bc += !(bool)(b[p>>3]&(1<<(p&7)));
		b[p>>3] |= 1<<(p&7);
		p += 99827;
		if (p >= nax) p -= nax;
		ans += bc;
	}
	printf("%lld\n", ans);
}

uses the same amount of memory and runs in 0ms.

20000100000

=====
Used: 0 ms, 20 KB

→ Reply

tribute_to_Ukraine_2022

5 years ago, # ^ |

+20

Of course it does run faster because your code has linear complexity and mine has quadratic divided by word size.

→ Reply

5 years ago, # ^ |

← Rev. 14 →

-10

The following implementation of your program using vector<uint64_t> is much faster.

→ Reply

tribute_to_Ukraine_2022

5 years ago, # ^ |

+11

Yes. I do know it is possible to implement bitset in a way, that count works in constant time (and by the way slow down all other operations by some constant). It doesn't matter here. I intentionally implemented something in $$$O(n^2)$$$ to measure time on 32-bit and 64-bit machine.

→ Reply