xsharp.eu • "Round()" error final Test - Page 2

xsharp.eu

The XSharp Forum
https://beta.xsharp.it/phpBB3/

"Round()" error final Test

https://beta.xsharp.it/phpBB3/viewtopic.php?t=1117

Page 2 of 3

"Round()" error, probably a bug in USUAL/FLOAT to Real8 conversion rather?

Posted: Sun Apr 21, 2019 3:52 pm
by Karl-Heinz
Jamal wrote: P.S. For my info, may be someone can tell what situation(s) one may one use a negative number in the decimal places parameter in the Round() function. I've never seen it or needed it.
Hi Jamal,

Based on this link, I've created a statistic that uses a negative value to round the number of the population

https://ec.europa.eu/eurostat/statistic ... ad_example

Code: Select all


FUNCTION Statistic() AS VOID
LOCAL aPopulation AS ARRAY 
LOCAL dwCurrentDecSep, dwLen, i AS DWORD  

   	aPopulation := {{ "BE" , 9660154 } , {"DE" , 61194591 } , { "IT" , 53685300 } , { "LU" , 338500m } , { "NL" , 12957621 } }
   	
   	dwLen := alen ( aPopulation ) 
    
	?
	
  	FOR i := 1 UPTO dwLen 
		? aPopulation [ i , 1 ] + _chr(9) + transform (Round ( aPopulation [ i , 2 ] , -4 ) / 1000 , "999 999" )

   	NEXT
   	
   	?
   	
	dwCurrentDecSep := setdecimalsep ( asc ( "." ) )  
	
	? 
	                         
	FOR i := 1 UPTO dwLen
		? aPopulation [ i , 1 ] + _chr(9) + transform ( ( Round ( aPopulation [ i , 2 ] , -4 ) / 10000 ) / 100.00 , "9999.99" )
		
	NEXT

	setdecimalsep ( dwCurrentDecSep )  
	
	?
	
	RETURN

regards
Karl-Heinz

"Round()" error, probably a bug in USUAL/FLOAT to Real8 conversion rather?

Posted: Sun Apr 21, 2019 4:05 pm
by Karl-Heinz
Upps,

just noticed that in the array the number "338500m" still contains the "m" - it was just a test , that i forgot to remove ;-)

"Round()" error, probably a bug in USUAL/FLOAT to Real8 conversion rather?

Posted: Sun Apr 21, 2019 4:09 pm
by Karl-Heinz
Hi Chris/Robert,

will the Decimal type handling happen in a Round() overload ?

FUNCTION Round ( nNumber AS DECIMAL , siDecimals AS INT ) AS DECIMAL

regards
Karl-Heinz

"Round()" error, time to resolve this forever

Posted: Mon Apr 22, 2019 12:19 am
by Chris
lumberjack wrote:Chris,
Chris wrote:But, remember, VO does not have a decimal type, but still returns the expected results in this area, with REAL and FLOAT data types and of course there is also so much existing code that we need to support.
I am a bit confused, why: "VO does not have a decimal type"? Surely we now in X# with VO language support, and the aim is to not use VO anymore, but X#?

I can't imagine anybody wanting a banker's rounding when RoundAwayFromZero is specified when a number cannot be exactly represented...
Johan, please read my posts above, the problem does not have to do with banker's rounding, it's just the inherent insufficient precision of the Double/REAL/FLOAT data types when representing decimal numbers.

We need to support properly all the VO data types, because there are millions of lines of existing code that uses them, we cannot simply ask people to go in a million places and change their code to now use the Decimal data type, make infinite amount of tests to make sure they still get the same results etc.

But, I agree, for new code, everywhere we need to use decimal numbers and there are no critical performance considerations, then we should always be using System.Decimal.

"Round()" error, probably a bug in USUAL/FLOAT to Real8 conversion rather?

Posted: Mon Apr 22, 2019 12:23 am
by Chris
Hi Karl-Heinz,
Karl-Heinz wrote:
will the Decimal type handling happen in a Round() overload ?

FUNCTION Round ( nNumber AS DECIMAL , siDecimals AS INT ) AS DECIMAL
Well, this overload would internally use Decimal.Round() instead of Math.Round(), so I assume it wouln't have that problem. But probably isn't it better to always use Decimal.Round() with Decimal vars directly, instead of introducing a runtime functions which simply just calls that?

"Round()" error, time to resolve this forever

Posted: Mon Apr 22, 2019 6:29 am
by lumberjack
Chris wrote:but we need to check if the speed penalty is too much and if it indeed works well in all cases
Well, as they say the proof of the pudding, is in the X# testing...:

Code: Select all

FUNCTION Start() AS VOID
	LOCAL u := 5.0 AS USUAL
	LOCAL y AS USUAL
	LOCAL z AS REAL8
	LOCAL dDec, dMDiff, dVDiff AS Decimal
	SetFloatDelta(10)
	SetDecimal(9)
	? PadL("USUAL", 15), PadL("REAL8", 15), PadL("ROUND", 5), ;
		PadL("Dec Round", 15), PadL("MathRound", 15), PadL("VORound", 15), "MathFact", "VOFact"
	FOR LOCAL x := 0 TO 9
		FOR LOCAL i := 1 UPTO 9
			y := x + ( u * 10^(-i))
			z := x + ( 5.0 * 10^(-i))
			? PadL(y, 15), PadL(z, 15), PadL(i - 1, 5), ;
				PadL(dDec := Decimal.Round((Decimal)(REAL8)y, i - 1, MidpointRounding.AwayFromZero), 15), ;
				PadL(dMDiff := Math.Round((REAL8)y, i - 1, MidpointRounding.AwayFromZero), 15), ;
				PadL(dVDiff := VORound(y, i-1), 15), PadL(dMDiff / dDec, 8), PadL(dVDiff / dDec, 6)
		NEXT
	NEXT
	BenchMark()
RETURN

FUNCTION VORound(r8 AS REAL8, iDec AS INT) AS REAL8
	VAR dRound := Decimal.Round((Decimal)r8, iDec, MidpointRounding.AwayFromZero)
RETURN (REAL8)dRound

FUNCTION BenchMark() AS VOID
	LOCAL u := 5.0 AS REAL8
	LOCAL dStart AS DateTime
	LOCAL mTick, vTick AS TimeSpan
	dStart := DateTime.Now
	
	FOR LOCAL j := 1 UPTO 10000000
	FOR LOCAL i := 1 UPTO 9
		Math.Round(u, i - 1, MidpointRounding.AwayFromZero)
	NEXT
	NEXT
	? "Math.Round", (mTick := DateTime.Now - dStart):ToString()
	dStart := DateTime.Now
	FOR LOCAL j := 1 UPTO 10000000
	FOR LOCAL i := 1 UPTO 9
		VORound(u, i - 1)
	NEXT
	NEXT
	? "   VORound", (vTick := DateTime.Now - dStart):ToString()
	? "   Penalty", (Decimal)vTick:Ticks / (Decimal)mTick:Ticks
RETURN

"Round()" error final Test

Posted: Mon Apr 22, 2019 9:53 am
by lumberjack
Final testing:

Code: Select all

FUNCTION Start() AS VOID
	LOCAL u := 5.0 AS USUAL
	LOCAL y AS USUAL
	LOCAL z AS REAL8
	LOCAL dDec, dMDiff, dVDiff AS Decimal
	SetFloatDelta(10)
	SetDecimal(9)
	? PadL("USUAL", 15), PadL("REAL8", 15), PadL("ROUND", 5), ;
		PadL("Dec Round", 15), PadL("MathRound", 15), PadL("VORound", 15), "MathFact", "VOFact"
	FOR LOCAL x := 0 AS INT  TO 9
		FOR LOCAL i := 1 AS INT UPTO 9
			y := x + ( u * 10^(-i))
			z := x + ( 5.0 * 10^(-i))
			? PadL(y, 15), PadL(z, 15), PadL(i - 1, 5), ;
				PadL(dDec := Decimal.Round((Decimal)(REAL8)y, i - 1, MidpointRounding.AwayFromZero), 15), ;
				PadL(dMDiff := (Decimal)(REAL8)Round(y, i - 1), 15), ;
				PadL(dVDiff := VORound(y, i-1), 15), PadL(dMDiff - dDec, 8), PadL(dVDiff - dDec, 6)
		NEXT
	NEXT
	BenchMark()
RETURN

FUNCTION VORound(r8 AS REAL8, iDec AS INT) AS REAL8
	VAR dRound := Decimal.Round((Decimal)r8, iDec, MidpointRounding.AwayFromZero)
RETURN (REAL8)dRound

FUNCTION BenchMark() AS VOID
	LOCAL u := 5.0 AS USUAL
	LOCAL dStart AS DateTime
	LOCAL vTick, mTick, xTick AS TimeSpan
	dStart := DateTime.Now
	FOR LOCAL j := 1 AS INT UPTO 10000000
	FOR LOCAL i := 1 AS INT UPTO 9
		Round(u, i - 1)
	NEXT
	NEXT
	? "  VO.Round", (vTick := DateTime.Now - dStart):ToString()
	dStart := DateTime.Now
	FOR LOCAL j := 1 AS INT UPTO 10000000
	FOR LOCAL i := 1 AS INT UPTO 9
		Math.Round((REAL8)u, i - 1)
	NEXT
	NEXT
	? "Math.Round", (mTick := DateTime.Now - dStart):ToString()
	dStart := DateTime.Now
	FOR LOCAL j := 1 AS INT UPTO 10000000
	FOR LOCAL i := 1 AS INT UPTO 9
		VORound(u, i - 1)
	NEXT
	NEXT
	? " Dec.Round", (xTick := DateTime.Now - dStart):ToString()
	? "   Penalty", (Decimal)vTick:Ticks / (Decimal)mTick:Ticks, (Decimal)xTick:Ticks / mTick:Ticks
RETURN
/************** Results **************
  VO.Round 00:00:05.9704984
Math.Round 00:00:03.1099580
 Dec.Round 00:00:10.2371290
   Penalty 1.919800332 3.291725805
************ End of Results **********/

"Round()" error final Test

Posted: Mon Apr 22, 2019 11:15 am
by robert
Johan,
I am not sure what exactly your idea behind this message is ?
- Are you suggesting that we replace the current Round() function with your VORound() ?
- Or something else ?

And your code tests the speed, but does it also test the correctness ?
Decimals are limited to 4 positions. Does your code work for Round() with more than 4 positions ?

Without additional comments I don't know what to do with your message.

Robert

"Round()" error final Test

Posted: Mon Apr 22, 2019 11:19 am
by Chris
Thanks Johan! So there's a considerable performance penalty with Decimal.Round(), but it's not a massive one, and since it is very important to have the same results as in VO, I think we should go with it. Just need to make sure that it does indeed work as expected in all cases, but in my tests so for it seems to do so.

"Round()" error final Test

Posted: Mon Apr 22, 2019 11:25 am
by Chris
robert wrote:Johan,
I am not sure what exactly your idea behind this message is ?
- Are you suggesting that we replace the current Round() function with your VORound() ?
- Or something else ?

And your code tests the speed, but does it also test the correctness ?
Decimals are limited to 4 positions. Does your code work for Round() with more than 4 positions ?

Without additional comments I don't know what to do with your message.

Robert
Robert, I think Johan was responding to the "speed considerations" I posted earlier. Apparently the speed difference between Math.Round() and Decimal.Round() is not too huge (including the conversions to/from Decimal), so I think we should go with it (adjust the Round() function in the runtime to use Math.Round()). I have done some tests about correctness and it seems to work fin, but will do some more of course.
All times are UTC
Page 2 of 3

Powered by phpBB® Forum Software © phpBB Limited