Use directly restrict in C99 mode (__restrict does not have exactly the same semantics).<div><br></div><div>For a valgrind profil, you can check my blog (<a href="http://matt.eifelle.com/2009/04/07/profiling-with-valgrind/">http://matt.eifelle.com/2009/04/07/profiling-with-valgrind/</a>)</div>
<div>Basically, if you have a python script, you can valgrind --optionsinmyblog python myscript.py</div><div><br></div><div>For PAPI, you have to install several packages (perf module for kernel for instance) and a GUI to analyze the results (in Eclispe, it should be possible).</div>
<div><br></div><div>Matthieu</div><div><br><div class="gmail_quote">2011/2/15 Sebastian Haase <span dir="ltr"><<a href="mailto:seb.haase@gmail.com">seb.haase@gmail.com</a>></span><br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
Thanks Matthieu,<br>
using __restrict__ with g++ did not change anything. How do I use<br>
valgrind with C extensions?<br>
I don't know what "PAPI profil" is ...?<br>
<font color="#888888">-Sebastian<br>
</font><div><div></div><div class="h5"><br>
<br>
On Tue, Feb 15, 2011 at 4:54 PM, Matthieu Brucher<br>
<<a href="mailto:matthieu.brucher@gmail.com">matthieu.brucher@gmail.com</a>> wrote:<br>
> Hi,<br>
> My first move would be to add a restrict keyword to dist (i.e. dist is the<br>
> only pointer to the specific memory location), and then declare dist_ inside<br>
> the first loop also with a restrict.<br>
> Then, I would run valgrind or a PAPI profil on your code to see what causes<br>
> the issue (false sharing, ...)<br>
> Matthieu<br>
><br>
> 2011/2/15 Sebastian Haase <<a href="mailto:seb.haase@gmail.com">seb.haase@gmail.com</a>><br>
>><br>
>> Hi,<br>
>> I assume that someone here could maybe help me, and I'm hoping it's<br>
>> not too much off topic.<br>
>> I have 2 arrays of 2d point coordinates and would like to calculate<br>
>> all pairwise distances as fast as possible.<br>
>> Going from Python/Numpy to a (Swigged) C extension already gave me a<br>
>> 55x speedup.<br>
>> (.9ms vs. 50ms for arrays of length 329 and 340).<br>
>> I'm using gcc on Linux.<br>
>> Now I'm wondering if I could go even faster !?<br>
>> My hope that the compiler might automagically do some SSE2<br>
>> optimization got disappointed.<br>
>> Since I have a 4 core CPU I thought OpenMP might be an option;<br>
>> I never used that, and after some playing around I managed to get<br>
>> (only) 50% slowdown(!) :-(<br>
>><br>
>> My code in short is this:<br>
>> (My SWIG typemaps use obj_to_array_no_conversion() from numpy.i)<br>
>> -------<Ccode> ----------<br>
>> void dists2d(<br>
>> double *a_ps, int nx1, int na,<br>
>> double *b_ps, int nx2, int nb,<br>
>> double *dist, int nx3, int ny3) throw (char*)<br>
>> {<br>
>> if(nx1 != 2) throw (char*) "a must be of shape (n,2)";<br>
>> if(nx2 != 2) throw (char*) "b must be of shape (n,2)";<br>
>> if(nx3 != nb || ny3 != na) throw (char*) "c must be of shape (na,nb)";<br>
>><br>
>> double *dist_;<br>
>> int i, j;<br>
>><br>
>> #pragma omp parallel private(dist_, j, i)<br>
>> {<br>
>> #pragma omp for nowait<br>
>> for(i=0;i<na;i++)<br>
>> {<br>
>> //num_threads=omp_get_num_threads(); --> 4<br>
>> dist_ = dist+i*nb; // dists_ is only<br>
>> introduced for OpenMP<br>
>> for(j=0;j<nb;j++)<br>
>> {<br>
>> *dist_++ = sqrt( sq(a_ps[i*nx1] - b_ps[j*nx2]) +<br>
>> sq(a_ps[i*nx1+1]<br>
>> - b_ps[j*nx2+1]) );<br>
>> }<br>
>> }<br>
>> }<br>
>> }<br>
>> -------</Ccode> ----------<br>
>> There is probably a simple mistake in this code - as I said I never<br>
>> used OpenMP before.<br>
>> It should be not too difficult to use OpenMP correctly here<br>
>> or - maybe better -<br>
>> is there a simple SSE(2,3,4) version that might be even better than<br>
>> OpenMP... !?<br>
>><br>
>> I supposed, that I did not get the #pragma omp lines right - any idea ?<br>
>> Or is it in general not possible to speed this kind of code up using<br>
>> OpenMP !?<br>
>><br>
>> Thanks,<br>
>> Sebastian Haase<br>
>> _______________________________________________<br>
>> NumPy-Discussion mailing list<br>
>> <a href="mailto:NumPy-Discussion@scipy.org">NumPy-Discussion@scipy.org</a><br>
>> <a href="http://mail.scipy.org/mailman/listinfo/numpy-discussion" target="_blank">http://mail.scipy.org/mailman/listinfo/numpy-discussion</a><br>
><br>
><br>
><br>
> --<br>
> Information System Engineer, Ph.D.<br>
> Blog: <a href="http://matt.eifelle.com" target="_blank">http://matt.eifelle.com</a><br>
> LinkedIn: <a href="http://www.linkedin.com/in/matthieubrucher" target="_blank">http://www.linkedin.com/in/matthieubrucher</a><br>
><br>
> _______________________________________________<br>
> NumPy-Discussion mailing list<br>
> <a href="mailto:NumPy-Discussion@scipy.org">NumPy-Discussion@scipy.org</a><br>
> <a href="http://mail.scipy.org/mailman/listinfo/numpy-discussion" target="_blank">http://mail.scipy.org/mailman/listinfo/numpy-discussion</a><br>
><br>
><br>
_______________________________________________<br>
NumPy-Discussion mailing list<br>
<a href="mailto:NumPy-Discussion@scipy.org">NumPy-Discussion@scipy.org</a><br>
<a href="http://mail.scipy.org/mailman/listinfo/numpy-discussion" target="_blank">http://mail.scipy.org/mailman/listinfo/numpy-discussion</a><br>
</div></div></blockquote></div><br><br clear="all"><br>-- <br>Information System Engineer, Ph.D.<br>Blog: <a href="http://matt.eifelle.com" target="_blank">http://matt.eifelle.com</a><br>LinkedIn: <a href="http://www.linkedin.com/in/matthieubrucher" target="_blank">http://www.linkedin.com/in/matthieubrucher</a><br>
</div>