Re: Mathematica 20x slower than Java at arithmetic/special functions, is
- To: mathgroup at smc.vnet.net
- Subject: [mg115910] Re: Mathematica 20x slower than Java at arithmetic/special functions, is
- From: Oliver Ruebenkoenig <ruebenko at wolfram.com>
- Date: Tue, 25 Jan 2011 06:30:28 -0500 (EST)
On Tue, 25 Jan 2011, Sjoerd C. de Vries wrote:
> I found a speed-up factor of about 52 using Oliver's method. This is
> on my quad-core laptop.
> So quite an improvement, but not a factor of 100.
>
Autch, I tested against a non N[] version of the problem. Sorry for that. I
see the same speedup as everyone else.
Oliver
> I used Evaluate[] too. I assume it works because with it the Bessel
> functions are evaluated at compile time (it's arguments do not depend
> on x and y), otherwise it's done at runtime.
>
> Cheers -- Sjoerd
>
> On Jan 24, 11:57 am, Leo Alekseyev <dnqu... at gmail.com> wrote:
>> Vivek, Oliver -- thanks for your input! My knowledge in using
>> Compile[] is somewhat lacking (mostly, due to the fact that I was
>> never able to get it to work well for me). In particular, I tried
>> using Compile[] much in the same way that Vivek has suggested, but I
>> neglected to use Evaluate[], which leads to a compiled function taking
>> substantially longer. Is there a quick explanation for why Evaluate[]
>> (or, in Oliver's example, a construct like
>> With[{code=code},Compile[{...},code]] necessary?..
>>
>> On my (very modest) hardware, I indeed get ~25x speedup that Vivek
>> mentions. Oliver's code for me performs about the same (~25x
>> improvement) without parallelism, and 2x faster on a dual-core
>> machine; this actually seems reasonable since the two methods are
>> fairly similar.
>>
>> I should note that it seems that these optimizations are very
>> dependent on Mathematica 8: in particular, cfunc2 (compilation of a
>> compiled function evaluating over some data) in Vivek's example gives
>> no additional gain under Mathematica 7 (makes me curious what changed
>> in version 8), and RuntimeAttributes -> Listable, Parallelization ->
>> True options that Oliver uses are new to version 8.
>>
>> --Leo
>>
>> On Mon, Jan 24, 2011 at 4:14 AM, Vivek J. Joshi <viv... at wolfram.com> wrot=
> e:
>>
>>> Without going into too much detail, a simple compilation of the functio=
> n gives approx 6x to 25x speed up,
>>
>>> ClearAll[grid1dc];
>>> grid1dc[x_,y_]=(With[{d=0.1,NN=50},
>>> Sum[Re[N[d BesselJ[1,2 Pi d Sqrt[m^2+n^2]]/Sqrt[m^2+n^2+10^-7]] Exp[I 2=
> .0Pi (m x+n y)]],{m,-NN,NN,1},{n,-NN,NN,1}]])//N;
>>
>>> gridres1da=With[{delta=0.5,xlim=2.5,ylim=2.5},
>>> Table[{x,y,grid1dc[x,y]},{x,-xlim,xlim,delta},{y,-ylim,ylim,delta}]];//=
> AbsoluteTiming
>>> {7.371354,Null}
>>
>>> Clear[cfunc];
>>> cfunc = Compile[{{x,_Real},{y,_Real}},Evaluate[grid1dc[x,y]]];
>>
>>> gridres1da2=With[{delta=0.5,xlim=2.5,ylim=2.5},
>>> Table[{x,y,cfunc[x,y]},{x,-xlim,xlim,delta},{y,-ylim,ylim,delta}]];//Ab=
> soluteTiming
>>> {1.237029,Null}
>>
>>> Norm[gridres1da[[All,All,3]]-gridres1da2[[All,All,3]]]//Chop
>>> 0
>>
>>> Following gives about 25x speedup,
>>
>>> Clear[cfunc2];
>>> cfunc2= Compile[{{xlim,_Real},{ylim,_Real},{delta,_Real}},
>>> Block[{x,y},
>>> Table[{x,y,cfunc[x,y]},{x,-xlim,xlim,delta},{y,-ylim,ylim,delta}]]];
>>
>>> gridres1da3=cfunc2[2.5,2.5,0.5];//AbsoluteTiming
>>> {0.292562,Null}
>>
>>> Norm[gridres1da[[All,All,3]]-gridres1da3[[All,All,3]]]//Chop
>>> 0
>>
>>> Vivek J. Joshi
>>> Kernel Developer
>>> Wolfram Research Inc.
>>
>>> On Jan 24, 2011, at 4:03 AM, Leo Alekseyev wrote:
>>
>>>> I was playing around with JLink the other day, and noticed that Java
>>>> seems to outperform Mathematica by ~20x in an area where I'd expect
>>>> Mathematica to be rather well optimized -- arithmetic involving specia=
> l
>>>> functions. In my particular example, I am simply evaluating a sum o=
> f
>>>> Bessel functions. I understand that much depends on the underlying
>>>> implementation, but I just want to run this by Mathgroup to see if
>>>> this is to be expected, or maybe if I'm doing something suboptimal in
>>>> Mathematica. Here's the code that I'm running:
>>
>>>> grid1dc[x_,
>>>> y_] = (With[{d = 0.1, NN = 50},
>>>> Sum[Re[N[
>>>> d BesselJ[1, 2 Pi d Sqrt[m^2 + n^2]]/
>>>> Sqrt[m^2 + n^2 + 10^-7]] Exp[
>>>> I 2.0 Pi (m x + n y)]], {m, -NN, NN, 1}, {n, -NN, NN, 1=
> }]=
>> ]) //
>>>> N
>>
>>>> and
>>
>>>> gridres1da =
>>>> With[{delta = 0.5, xlim = 2.5, ylim = 2.5},
>>>> Table[{x, y, grid1dc[x, y]}, {x, -xlim, xlim, delta}, {y, -ylim,
>>>> ylim, delta}]]
>>
>>>> Java implementation uses Colt and Apache common math libraries for the
>>>> Bessels and complex numbers, uses a double for loop, and consistently
>>>> runs 15-20 times faster.
>>
>>>> --Leo
>>
>>
>
>
>