Code review request for 6908131 Pure Java implementations of StrictMath.floor(double) &StrictMath.ceil(double)

[Joe Darcy]

A few comments.

First, part and parcel of doing numerical libraries work is developing 
tests; see, "Test where the failures are likely to be" 
(http://blogs.sun.com/darcy/entry/test_where_the_failures_are).  An 
obvious set of tests to run are the regression tests I wrote as part of 
the fix for this bug 
(http://hg.openjdk.java.net/jdk7/jdk7/jdk/rev/ad1e30930c6c).

Second, micro-benchmarking Java programs is a subtle art; see Cliff 
Click's various writeups for guidance, including

http://www.azulsystems.com/events/javaone_2002/microbenchmarks.pdf
http://blogs.azulsystems.com/cliff/2008/03/another-round-o.html
http://www.azulsystems.com/events/javaone_2009/session/2009_J1_Benchmark.pdf

Various benchmarks are examined by the performance team.  You can assume 
  benchmarks Sun publishes numbers on are included in the set of 
benchmarks of interest :-)

-Joe

Jeff Hain wrote:
>  
> Dmitry,
>  
>  From your bench, I see FloorCeil methods get in trouble
> with 64 bits architecture, but I never ran anything on 64 bits yet.
> 
> I ran your bench on my laptop:
> - DELL inspiron 9400, with (two) Genuine Intel T2300 @ 1.666GHz.
> - java.runtime.version=1.6.0_11-b03
> - java.class.version=50.0
> - os.name=Windows XP
> - os.arch=x86
> - os..version=5.1
> 
> I added two patterns:
> - pattern3 ("common" case (for doubles) of non-integer values) {
>             0.1, -0.1,
>         1e4+0.1, -1e4-0.1,
>         1e8+0.1, -1e8-0.1,
>        1e12+0.1, -1e12-0.1,
>    (1L<<52)/1.1, -(1L<<52)/1.1}
> - pattern4 (case of integer values) {
>             0.0, -0.0,
>             1e4, -1e4,
>             1e8, -1e8,
>        (1L<<52), -(1L<<52),
>        (1L<<60), -(1L<<60)}
> 
> I also added test of the pure Java StrictMath class (which is the one 
> you ran
> as "StrictMath" in your bench?) posted on this subject: "StrictMath_JD".
> 
> Here is what I got:
>  
> -----------------------------------------
> - CLIENT
> -----------------------------------------
>  
> pattern 1:
> empty                took       4,42 nsec 
> Math.ceil            took      63,77 nsec 
> Math.floor           took      67,55 nsec 
> StrictMath.ceil      took      64,20 nsec 
> StrictMath.floor     took      68,76 nsec 
> StrictMath_JD.ceil   took      52,58 nsec 
> StrictMath_JD.floor  took      53,57 nsec 
> FloorCeil.ceil       took      24,31 nsec 
> FloorCeil.floor      took      18,37 nsec 
>  
> pattern 2:
> empty                took       4,50 nsec
> Math.ceil            took     110,95 nsec
> Math.floor           took     110,42 nsec
> StrictMath.ceil      took     111,38 nsec
> StrictMath.floor     took     110,97 nsec
> StrictMath_JD.ceil   took      48,95 nsec
> StrictMath_JD.floor  took      49,38 nsec
> FloorCeil.ceil       took      23,29 nsec
> FloorCeil.floor      took      20,81 nsec
>  
> pattern 3:
> empty                took       4,39 nsec
> Math.ceil            took     111,72 nsec
> Math.floor           took     114,48 nsec
> StrictMath.ceil      took     113,68 nsec
> StrictMath.floor     took     114,32 nsec
> StrictMath_JD.ceil   took      60,23 nsec
> StrictMath_JD.floor  took      60,13 nsec
> FloorCeil.ceil       took      27,11 nsec
> FloorCeil.floor      took      23,77 nsec
>  
> pattern 4:
> empty                took       4,38 nsec
> Math.ceil            took     113,49 nsec
> Math.floor           took     113,34 nsec
> StrictMath.ceil      took     110,35 nsec
> StrictMath.floor     took     112,95 nsec
> StrictMath_JD.ceil   took      43,69 nsec
> StrictMath_JD.floor  took      44,28 nsec
> FloorCeil.ceil       took      22,09 nsec
> FloorCeil.floor      took      18,71 nsec
>  
> -----------------------------------------
> - SERVER
> -----------------------------------------
>  
> pattern 1:
> empty                took       2,40 nsec 
> Math.ceil            took      64,61 nsec 
> Math.floor           took      65,14 nsec 
> StrictMath.ceil      took      63,93 nsec 
> StrictMath.floor     took      65,99 nsec 
> StrictMath_JD.ceil   took      37,52 nsec 
> StrictMath_JD.floor  took      39,77 nsec 
> FloorCeil.ceil       took      21,36 nsec 
> FloorCeil.floor      took      13,96 nsec 
>  
> pattern 2:
> empty                took       2,23 nsec
> Math.ceil            took     112,42 nsec
> Math.floor           took     110,88 nsec
> StrictMath.ceil      took     112,05 nsec
> StrictMath.floor     took     113,54 nsec
> StrictMath_JD.ceil   took      33,98 nsec
> StrictMath_JD.floor  took      33,97 nsec
> FloorCeil.ceil       took      22,09 nsec
> FloorCeil.floor      took      16,75 nsec
>  
> pattern 3:
> empty                took       2,19 nsec
> Math.ceil            took     113,67 nsec
> Math.floor           took     113,53 nsec
> StrictMath.ceil      took     114,46 nsec
> StrictMath.floor     took     115,02 nsec
> StrictMath_JD.ceil   took      44,19 nsec
> StrictMath_JD.floor  took      43,57 nsec
> FloorCeil.ceil       took      25,57 nsec
> FloorCeil.floor      took      20,95 nsec
>  
> pattern 4:
> empty                took       2,45 nsec
> Math.ceil            took     108,56 nsec
> Math.floor           took     112,96 nsec
> StrictMath.ceil      took     109,43 nsec
> StrictMath.floor     took     113,14 nsec
> StrictMath_JD.ceil   took      30,18 nsec
> StrictMath_JD.floor  took      31,07 nsec
> FloorCeil.ceil       took      18,11 nsec
> FloorCeil.floor      took      15,91 nsec
>  
> -----------------------------------------
> 
> Jeff
> PS : The ceil and floor I posted here are an extract of a FastMath class 
> (I already
>      see some people in this mailing list invoking the gods against me 
> while having
>      a heart attack at the hearing of this class name) I did, and put on 
> source
>      forge (I already told about it in this list few months ago): maybe 
> some of
>      you might want to take a look at it (it's called "FastMath", but 
> not ALL of
>      its treatments are wrong...).
>  
>