71 replies to this topic

[Kambiz]

Today I saw the D's pi computing sample (\dmd\samples\d\pi.d) and thought that porting it to c++ and c# for making a little benchmark should be easy:
d:

import std.c.stdio;
import std.c.stdlib;
import std.c.time;

const int LONG_TIME=4000;

byte[] p;
byte[] t;
int q;

int main(char[][] args)
{
	int startime, endtime;
	int i;

	if (args.length == 2) {
		sscanf(&args[1][0],"%d",&q);
	} else {
		printf("Usage: pi [precision]\n");
		exit(55);
	}

	if (q < 0)
	{
		printf("Precision was too low, running with precision of 0.\n");
		q = 0;
	}

	if (q > LONG_TIME)
	{
	    printf("Be prepared to wait a while...\n");
	}

	// Compute one more digit than we display to compensate for rounding
	q++;

	p.length = q + 1;
	t.length = q + 1;

	/* compute pi */

	std.c.time.time(&startime);
	arctan(2);
	arctan(3);
	mul4();
	std.c.time.time(&endtime);

	// Return to the number of digits we want to display
	q--;

	/* print pi */

	printf("pi = %d.",cast(int)(p[0]));
	for (i = 1; i <= q; i++)
	printf("%d",cast(int)(p[i]));
	printf("\n");
	printf("%ld seconds to compute pi with a precision of %d digits.\n",endtime-startime,q);

	return 0;
}

void arctan(int s)
{
	int n;

	t[0] = 1;
	div(s); /* t[] = 1/s */
	add();
	n = 1;
	do {
		mul(n);
		div(s * s);
		div(n += 2);
		if (((n-1) / 2) % 2 == 0)
			add();
		else
			sub();
	} while (!tiszero());
}

void add()
{
	int j;

	for (j = q; j >= 0; j--)
	{
		if (t[j] + p[j] > 9) {
			p[j] += t[j] - 10;
			p[j-1] += 1;
		} else
			p[j] += t[j];
	}
}

void sub()
{
	int j;

	for (j = q; j >= 0; j--)
		if (p[j] < t[j]) {
			p[j] -= t[j] - 10;
			p[j-1] -= 1;
		} else
			p[j] -= t[j];
}

void mul(int multiplier)
{
	int b;
	int i;
	int carry = 0, digit = 0;

	for (i = q; i >= 0; i--) {
		b = (t[i] * multiplier + carry);
		digit = b % 10;
		carry = b / 10;
		t[i] = digit;
	}
}

/* t[] /= l */

void div(int divisor)
{
	int i, b;
	int quotient, remainder = 0;

	for (i = 0; i <= q; i++) {
		b = (10 * remainder + t[i]);
		quotient = b / divisor;
		remainder = b % divisor;
		t[i] = quotient;
	}
}

void div4()
{
	int i, c, d = 0;

	for (i = 0; i <= q; i++) {
		c = (10 * d + p[i]) / 4;
		d = (10 * d + p[i]) % 4;
		p[i] = c;
	}
}

void mul4()
{
	int i, c, d;

	d = c = 0;

	for (i = q; i >= 0; i--) {
		d = (p[i] * 4 + c) % 10;
		c = (p[i] * 4 + c) / 10;
		p[i] = d;
	}
}

int tiszero()
{
	int k;

	for (k = 0; k <= q; k++)
		if (t[k] != 0)
			return false;
	return true;
}

(I compiled pi.d with dmd -O -release pi.d)

c++:

#include <stdio.h>
#include <stdlib.h>
#include <time.h>

#define LONG_TIME 4000

__int8* p;
__int8* t;
int q;


void arctan(int s);
void add();
void sub();
void mul(int);
void div(int);
void mul4();
void div4();
int tiszero();

int main(int argv,char** args)
{
	time_t startime, endtime;
	int i;

	if (argv == 2) {
		sscanf(&args[1][0],"%d",&q);
	} else {
		printf("Usage: pi [precision]\n");
		exit(55);
	}

	if (q < 0)
	{
		printf("Precision was too low, running with precision of 0.\n");
		q = 0;
	}

	if (q > LONG_TIME)
	{
	    printf("Be prepared to wait a while...\n");
	}

	// Compute one more digit than we display to compensate for rounding
	q++;

	p = new __int8[ q + 1 ];
	t = new __int8[ q + 1 ];
	for(int tt=0;tt<=q;tt++) 
	{
		p[tt]=0;
		t[tt]=0;
	}
	/* compute pi */

	time(&startime);
	arctan(2);
	arctan(3);
	mul4();
	time(&endtime);

	// Return to the number of digits we want to display
	q--;

	/* print pi */

	printf("pi = %d.",(int)(p[0]));
	for (i = 1; i <= q; i++)
	printf("%d",(int)(p[i]));
	printf("\n");
	printf("%ld seconds to compute pi with a precision of %d digits.\n",(long)endtime-(long)startime,q);

	delete [] p;
	delete [] t;
	return 0;
}

void arctan(int s)
{
	int n;

	t[0] = 1;
	div(s); /* t[] = 1/s */
	add();
	n = 1;
	do {
		mul(n);
		div(s * s);
		div(n += 2);
		if (((n-1) / 2) % 2 == 0)
			add();
		else
			sub();
	} while (!tiszero());
}

void add()
{
	int j;

	for (j = q; j >= 0; j--)
	{
		if (t[j] + p[j] > 9) {
			p[j] += t[j] - 10;
			p[j-1] += 1;
		} else
			p[j] += t[j];
	}
}

void sub()
{
	int j;

	for (j = q; j >= 0; j--)
		if (p[j] < t[j]) {
			p[j] -= t[j] - 10;
			p[j-1] -= 1;
		} else
			p[j] -= t[j];
}

void mul(int multiplier)
{
	int b;
	int i;
	int carry = 0, digit = 0;

	for (i = q; i >= 0; i--) {
		b = (t[i] * multiplier + carry);
		digit = b % 10;
		carry = b / 10;
		t[i] = digit;
	}
}

/* t[] /= l */

void div(int divisor)
{
	int i, b;
	int quotient, remainder = 0;

	for (i = 0; i <= q; i++) {
		b = (10 * remainder + t[i]);
		quotient = b / divisor;
		remainder = b % divisor;
		t[i] = quotient;
	}
}

void div4()
{
	int i, c, d = 0;

	for (i = 0; i <= q; i++) {
		c = (10 * d + p[i]) / 4;
		d = (10 * d + p[i]) % 4;
		p[i] = c;
	}
}

void mul4()
{
	int i, c, d;

	d = c = 0;

	for (i = q; i >= 0; i--) {
		d = (p[i] * 4 + c) % 10;
		c = (p[i] * 4 + c) / 10;
		p[i] = d;
	}
}

int tiszero()
{
	int k;

	for (k = 0; k <= q; k++)
		if (t[k] != 0)
			return false;
	return true;
}

c#:

using System;
using System.Diagnostics;

class Pi
{
    static void Main(string[] args)
    {
        Pi pi = new Pi();
        pi.run(args);
    }
    private const int LONG_TIME = 4000;
    sbyte[] p;
    sbyte[] t;
    int q;

    void run(string[] args)
    {
        Stopwatch timer = new Stopwatch();
        int i;

        if (args.Length == 1)
        {
            q = int.Parse(args[0]);
        }
        else
        {
            Console.WriteLine("Usage: pi [precision]");
            return;
        }
        
        if (q < 0)
        {
            Console.WriteLine("Precision was too low, running with precision of 0.");
            q = 0;
        }

        if (q > LONG_TIME)
        {
            Console.WriteLine("Be prepared to wait a while...");
        }

        // Compute one more digit than we display to compensate for rounding
        q++;

        p = new sbyte[q + 1];
        t = new sbyte[q + 1];
        /* compute pi */

        timer.Start();
        arctan(2);
        arctan(3);
        mul4();
        timer.Stop();

        // Return to the number of digits we want to display
        q--;

        /* print pi */

        Console.Write("pi = {0}.", p[0]);
        for (i = 1; i <= q; i++)
            Console.Write(p[i]);
        Console.WriteLine();
        Console.WriteLine("{0} seconds to compute pi with a precision of {1} digits.", timer.ElapsedMilliseconds / 1000.0, q);

        return;
    }

    void arctan(int s)
    {
        int n;

        t[0] = 1;
        div(s); /* t[] = 1/s */
        add();
        n = 1;
        do
        {
            mul(n);
            div(s * s);
            div(n += 2);
            if (((n - 1) / 2) % 2 == 0)
                add();
            else
                sub();
        } while (!tiszero());
    }

    void add()
    {
        int j;

        for (j = q; j >= 0; j--)
        {
            if (t[j] + p[j] > 9)
            {
                p[j] += (sbyte)(t[j] - 10);
                p[j - 1] += 1;
            }
            else
                p[j] += t[j];
        }
    }

    void sub()
    {
        int j;

        for (j = q; j >= 0; j--)
            if (p[j] < t[j])
            {
                p[j] -= (sbyte)(t[j] - 10);
                p[j - 1] -= 1;
            }
            else
                p[j] -= t[j];
    }

    void mul(int multiplier)
    {
        int b;
        int i;
        int carry = 0, digit = 0;

        for (i = q; i >= 0; i--)
        {
            b = (t[i] * multiplier + carry);
            digit = b % 10;
            carry = b / 10;
            t[i] = (sbyte)digit;
        }
    }

    /* t[] /= l */

    void div(int divisor)
    {
        int i, b;
        int quotient, remainder = 0;

        for (i = 0; i <= q; i++)
        {
            b = (10 * remainder + t[i]);
            quotient = b / divisor;
            remainder = b % divisor;
            t[i] = (sbyte)quotient;
        }
    }

    void div4()
    {
        int i, c, d = 0;

        for (i = 0; i <= q; i++)
        {
            c = (10 * d + p[i]) / 4;
            d = (10 * d + p[i]) % 4;
            p[i] = (sbyte)c;
        }
    }

    void mul4()
    {
        int i, c, d;

        d = c = 0;

        for (i = q; i >= 0; i--)
        {
            d = (p[i] * 4 + c) % 10;
            c = (p[i] * 4 + c) / 10;
            p[i] = (sbyte)d;
        }
    }

    bool tiszero()
    {
        int k;

        for (k = 0; k <= q; k++)
            if (t[k] != 0)
                return false;
        return true;
    }

 
}

I used the release builds for the benchmark, here are the results:
d : 26 seconds to compute pi with a precision of 10000 digits.
c# : 34,745(34.745) seconds to compute pi with a precision of 10000 digits.
c++ : 15 seconds to compute pi with a precision of 10000 digits.


I do not want to start a language war... we had enough such wars in the past days and I have just made this benchmark because I saw the “What do you think of the D language?“ thread and was testing d.

I'm just surprised a little: I thought that D shouldn't be much slower than c++ and I thought that c# would be much faster. Maybe there is some optimization option I have not used(?)
What do you think about the results?

(maybe some one can test java)
(I really like the invariant and unittest features of D, are there equivalents for c#?)
-I can't reply until tomorrow-

[Edited by - Kambiz on October 7, 2006 2:02:45 AM]

[TrueTom]

The comma is the new dot. (At least in germany.)

[Ezbez]

In most/all of Europe, actually. So, 3.14159 would be 3,14159. You really should know that, since chances are it will bite you one time (as you saw, it made a vast difference here).

What compiler where you using for C++ and what optimizations did you enable?

[Raghar]

One of traditions in the computer technology was use of "." for 0.030. The "," was used for something else.

"," was used traditionally only in math classes. However teachers would survive "." instead if they weren't completely retarded.

BTW if you'd like to test Java, should it be a direct port, or should it use BigDecimal instead? You might like to add an 64 bit ASM program as well.

[Promit]

Alright, managed code haters -- don't feel too vindicated yet. I sat down with the C# code for about 15 minutes, and made some modifications to the mul and div functions.

using System;

using System.Diagnostics;



class Pi

{

	static void Main( string[] args )

	{

		Pi pi = new Pi();

		pi.run( args );

	}



	private const int LONG_TIME = 4000;

	sbyte[] p;

	sbyte[] t;

	int q;



	void run( string[] args )

	{

		Stopwatch timer = new Stopwatch();

		int i;



		if( args.Length == 1 )

		{

			q = int.Parse( args[0] );

		}

		else

		{

			Console.WriteLine( "Usage: pi [precision]" );

			return;

		}



		if( q < 0 )

		{

			Console.WriteLine( "Precision was too low, running with precision of 0." );

			q = 0;

		}



		if( q > LONG_TIME )

		{

			Console.WriteLine( "Be prepared to wait a while..." );

		}



		// Compute one more digit than we display to compensate for rounding

		q++;



		p = new sbyte[q + 1];

		t = new sbyte[q + 1];

		/* compute pi */



		timer.Start();

		arctan( 2 );

		arctan( 3 );

		mul4();

		timer.Stop();



		// Return to the number of digits we want to display

		q--;



		/* print pi */



		Console.Write( "pi = {0}.", p[0] );

		for( i = 1; i <= q; i++ )

			Console.Write( p[i] );

		Console.WriteLine();

		Console.WriteLine( "{0} seconds to compute pi with a precision of {1} digits.", timer.ElapsedMilliseconds / 1000.0, q );



		return;

	}



	void arctan( int s )

	{

		int n;



		t[0] = 1;

		div( s ); /* t[] = 1/s */

		add();

		n = 1;

		do

		{

			mul( n );

			div( s * s );

			div( n += 2 );

			if( ( ( n - 1 ) / 2 ) % 2 == 0 )

				add();

			else

				sub();

		} while( !tiszero() );

	}



	void add()

	{

		int j;



		for( j = q; j >= 0; j-- )

		{

			if( t[j] + p[j] > 9 )

			{

				p[j] += (sbyte) ( t[j] - 10 );

				p[j - 1] += 1;

			}

			else

				p[j] += t[j];

		}

	}



	void sub()

	{

		int j;



		for( j = q; j >= 0; j-- )

			if( p[j] < t[j] )

			{

				p[j] -= (sbyte) ( t[j] - 10 );

				p[j - 1] -= 1;

			}

			else

				p[j] -= t[j];

	}



	void mul( int multiplier )

	{

		int b;

		int i;

		int carry = 0, digit;



		for( i = q; i >= 0; i-- )

		{

			b = ( t[i] * multiplier + carry );

			carry = (int) ( b * 0.1f );

			digit = b - 10 * carry;

			t[i] = (sbyte) digit;

		}

	}



	/* t[] /= l */



	void div( int divisor )

	{

		int i, b;

		int quotient, remainder = 0;

		float fdiv = 1.0f / divisor;



		for( i = 0; i <= q; i++ )

		{

			b = ( 10 * remainder + t[i] );

			quotient = (int) ( b * fdiv );

			remainder = b - divisor * quotient;

			t[i] = (sbyte) quotient;

		}

	}



	void div4()

	{

		int i, c, d = 0;



		for( i = 0; i <= q; i++ )

		{

			c = ( 10 * d + p[i] ) / 4;

			d = ( 10 * d + p[i] ) % 4;

			p[i] = (sbyte) c;

		}

	}



	void mul4()

	{

		int i, c, d;



		d = c = 0;



		for( i = q; i >= 0; i-- )

		{

			d = ( p[i] * 4 + c ) % 10;

			c = ( p[i] * 4 + c ) / 10;

			p[i] = (sbyte) d;

		}

	}



	bool tiszero()

	{

		int k;



		for( k = 0; k <= q; k++ )

			if( t[k] != 0 )

				return false;

		return true;

	}

}

Initial figures on my system before modification were in the vicinity of about 19s (probably closer to 20) for the C++ and 38s for the C#. This version of the C# runs in just under 22s. I brought the same optimizations back to the C++ version, but they didn't appear to have any particular effect.

Perhaps someone can step up and defend D's honor. 26s is much worse than I expected. (Although the fact that the D code was written by Digital Mars does not bode well.)

[Edited by - Promit on October 6, 2006 9:28:51 PM]

[deathkrush]

The fastest way to calculate PI in C is of course:

#define _ -F<00||--F-OO--;
int F=00,OO=00;main(){F_OO();printf("%1.3f\n",4.*-F/OO/OO);}F_OO()
{
            _-_-_-_
       _-_-_-_-_-_-_-_-_
    _-_-_-_-_-_-_-_-_-_-_-_
  _-_-_-_-_-_-_-_-_-_-_-_-_-_
 _-_-_-_-_-_-_-_-_-_-_-_-_-_-_
 _-_-_-_-_-_-_-_-_-_-_-_-_-_-_
_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_
_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_
_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_
_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_
 _-_-_-_-_-_-_-_-_-_-_-_-_-_-_
 _-_-_-_-_-_-_-_-_-_-_-_-_-_-_
  _-_-_-_-_-_-_-_-_-_-_-_-_-_
    _-_-_-_-_-_-_-_-_-_-_-_
        _-_-_-_-_-_-_-_
            _-_-_-_
}

[Kambiz]

Quote:

Original post by TrueTom The comma is the new dot. (At least in germany.)

yes... (I have a German Windows.)

Quote:

Original post by Promit Alright, managed code haters -- don't feel too vindicated yet. I sat down with the C# code for about 15 minutes, and made some modifications to the mul and div functions. ...

19.443 seconds to compute pi with a precision of 10000 digits.
But because of the modifications we should not compare these result with the others.

Quote:

Original post by Ezbez What compiler where you using for C++ and what optimizations did you enable?

MSVC++ 2005
/O2 /Ot /GL /D "WIN32" /D "NDEBUG" /D "_CONSOLE" /D "_UNICODE" /D "UNICODE" /FD /EHsc /MD /Fo"Release\\" /Fd"Release\vc80.pdb" /W3 /nologo /c /Wp64 /Zi /TP /errorReport:prompt
Linker:
/OUT:"D:\kambiz\c++\pi\Release\pi.exe" /INCREMENTAL:NO /NOLOGO /MANIFEST /MANIFESTFILE:"Release\pi.exe.intermediate.manifest" /DEBUG /PDB:"d:\kambiz\c++\pi\Release\pi.pdb" /SUBSYSTEM:CONSOLE /OPT:REF /OPT:ICF /LTCG /MACHINE:X86 /ERRORREPORT:PROMPT kernel32.lib user32.lib gdi32.lib winspool.lib comdlg32.lib advapi32.lib shell32.lib ole32.lib oleaut32.lib uuid.lib odbc32.lib odbccp32.lib

[ToohrVyk]

Quote:

Original post by Kambiz But because of the modifications we should not compare these result with the others.

Under which conditions may we compare one result with another?

[Promit]

Quote:

Original post by ToohrVyk Quote: Original post by Kambiz But because of the modifications we should not compare these result with the others. Under which conditions may we compare one result with another?

Presumably, one is only allowed to use the exact implementation originally written in C, without taking into account the fact that a good C implementation may be a terrible implementation for another language.

In short, when comparing the performance of languages, one must assume that all programmers are competent at no languages other than C.

[ToohrVyk]

Quote:

Original post by Promit Presumably, one is only allowed to use the exact implementation originally written in C, without taking into account the fact that a good C implementation may be a terrible implementation for another language.

You seem to be right. I've only just noticed that the version advertised as C++ isn't C++, but is in fact C.

[Stachel]

Quote:

Original post by Kambiz I'm just surprised a little: I thought that D shouldn't be much slower than c++ and I thought that c# would be much faster. Maybe there is some optimization option I have not used(?) What do you think about the results?

Interesting, but you're comparing the Microsoft C++ optimizer with the Digital Mars C++ optimizer. DMD, the Digital Mars D compiler, is built using the Digital Mars C++ optimizer.

To compare apples-apples, it makes sense to compare DMD versus DMC++, or GDC versus g++, as then the optimizer and code generator would be the same for each language. I tried DMD and DMC++, and got essentially the same times for each.

[Promit]

I actually have a suspicion that some of the modifications I applied in the C# version will improve the D version's performance as well. Anyone want to try my edits on D and see what happens? I really don't feel like downloading DMD.

[barakus]

Quote:

Original post by Kambiz d : 26 seconds to compute pi with a precision of 10000 digits. c# : 34,745(34.745) seconds to compute pi with a precision of 10000 digits. c++ : 15 seconds to compute pi with a precision of 10000 digits. ... I do not want to start a language war...

Yeah, right. The fact that you only gave a decimal point only for the C# version (using a comma to represent it, making it look like a huge number) and your dismissal of Promit's benchmark implies that you hold a certain prejudice toward C#. Now, theres nothing inherently wrong with that but trying to spread propaganda in the form of a benchmark is pretty sad if you ask me.

[jollyjeffers]

I've deleted the AP replies that seem intent on derailing this thread. So far the discussion is clean and interesting - whilst these benchmarks may not be the best way to compare languages they are a good platform for learning and discussing various characteristics.

The whole "," or "." thing is a bit pointless - why bother arguing over something so silly [smile]

Cheers,
Jack

[asp_]

I'm not much of a .net freak but to me it seems like .net is going to be slower no matter what in this specific case. It's basically a series of tight loops with bounds checked array accesses which will generate at least an extra branch for each array access, probably with stack unwinding also.

I see nothing wrong with the test and it demonstrates what is a fairly well known weakness of .net, tight loops and a lot of array accesses.

[doctorsixstring]

Quote:

Original post by Anonymous Poster It's not like Promit's code is more idiomatic C#, it's just a tweak that happens to make .NET run faster. Similar tweaks could probably be found in C++ and D versions as well, and we could spend weeks optimizing each one. Why not actually compare versions that are near identical in code to save the trouble? It even looks to me like his code would make the algorithm wrong due to the fact that floating point math isn't as accurate, but I didn't check this..

Different languages work in different ways. It is extremely biased to write a program in one language (C++) and expect a direct port to run equally fast on other languages (D or C#, in this case). Spending "weeks" (hopefully less) on optimizing each language according to their strengths would provide the best benchmarks. After all, I could care less how well C# or D implements C-specific algorithms. What I care about is the performance I'll achieve using those languages in the way they're designed.

- Mike