Re: For loop benchmarks between Octave versions

[Julien Bect]

Le 22/06/2014 15:32, Julien Bect a écrit :
> Le 17/06/2014 17:11, Julien Bect a écrit :
> > > I don't much like having to manually match up new/delete, so I'd 
> > > rather see that go inside a constructor/destructor pair.  Done 
> > > properly and with inlining, I don't see why it would be inefficient 
> > > and perhaps we could eliminate the BEGIN/END macros?
> >
> > I try that right now. 
>
> John,
>
> As you know, using a constructor/destructor pair is actually what the 
> BEGIN/END macros currently do. I have tried improving this 
> constructor/destructor approach with inlining, but couldn't get a 
> significant improvement in this direction. Perhaps didn't I try hard 
> enough... 

Ok, I have tried harder and finally obtained a (significant ?) speed 
improvement using a constructor/destructor pair and templated classes 
(see attached patch).

The BEGIN/END macros are still there, but of course they can be removed 
(I find them rather convenient, though).

Here are the results that I have obtained on 3.8.2-RC1, compiled with 
--enable-jit (but with jit OFF) :

                        without patch       with patch
   doubleForLoop/a=1    2609.8 [  5.7]    2250.0 [  9.7] evol: -13.8%
   doubleForLoop/a=b    3034.4 [  5.1]    2570.3 [ 12.0] evol: -15.3%
   doubleForLoop/a+b    2084.3 [  6.4]    1379.3 [  1.0] evol: -33.8%
     vectorOfSquares    6864.3 [ 23.5]    5963.7 [ 20.2] evol: -13.1%
              mandel     226.9 [  1.6]     180.5 [  1.6] evol: -20.4%
                 fib     667.8 [  3.4]     539.6 [  2.1] evol: -19.2%
           parse_int    2376.0 [  7.7]    2143.0 [  5.1] evol:  -9.8%
           quicksort    1230.7 [ 10.7]    1066.8 [ 16.4] evol: -13.3%
              pi_sum   18151.0 [ 39.9]   11030.2 [ 21.5] evol: -39.2%
       rand_mat_stat     559.0 [  1.6]     519.9 [  6.9] evol:  -7.0%
        rand_mat_mul     367.4 [  2.0]     366.8 [  2.6] evol:  -0.2%
             printfd    2972.6 [  9.0]    2833.3 [  8.0] evol:  -4.7%

All results are in ms: median over 10 replications [mean absolute 
deviation about the median].

i686-linux-gnu / gcc version 4.7.3 (Ubuntu/Linaro 4.7.3-1ubuntu1)

We're still a long way from the performances of Octave 3.4.3 (or even 
3.6.0, for some tests):

                         Octave 3.4.3     Octave 3.6.0 3.8.2-RC1 + patch
   doubleForLoop/a=1    1487.4 [  4.7]    1632.5 [ 11.9]    2250.0 [  9.7]
   doubleForLoop/a=b    2206.9 [  8.9]    2239.6 [ 21.1]    2570.3 [ 12.0]
   doubleForLoop/a+b    1136.0 [  3.6]    1575.0 [  8.3]    1379.3 [  1.0]
     vectorOfSquares    3435.3 [ 10.8]    3783.2 [  4.2]    5963.7 [ 20.2]
              mandel     130.7 [  0.6]     139.1 [  0.6]     180.5 [  1.6]
                 fib     346.7 [  1.3]     409.9 [  1.2]     539.6 [  2.1]
           parse_int    1461.8 [  3.1]    1675.3 [  8.2]    2143.0 [  5.1]
           quicksort     675.3 [  9.4]     723.4 [  8.8]    1066.8 [ 16.4]
              pi_sum    9197.3 [ 55.3]   14386.8 [ 20.1]   11030.2 [ 21.5]
       rand_mat_stat     322.8 [  3.1]     306.7 [  5.3]     519.9 [  6.9]
        rand_mat_mul     367.7 [ 10.3]     367.7 [ 26.8]     366.8 [  2.6]
             printfd    2426.1 [  5.2]    2097.2 [ 10.4]    2833.3 [  8.0]


I also have interesting results about the behaviour of JIT on this 
benchmark suite, but that's for another thread I think.

jbProfilerTemplate.patch
Description: Text Data

perf.m
Description: Text Data