|
|
Blitz Support : |
From: Julian Cummings (cummings_at_[hidden])
Date: 2005-05-17 14:56:38
Hi Frank,
The overhead you are seeing with the blitz Array is due to the fact that
Arrays can have multiple dimensions. The reduction code in the header
file <blitz/array/reduce.cc> does some extra work to loop over the Array
indices for each dimension.
If you replace Array<double,1> with Vector<double> and use the blitz 1D
Vector container, you should see this overhead disappear. I get the
same timing for the blitz Vector and the STL vector class. (BTW, I had
to do some major surgery on your code example to get it to compile, and
I changed the expression used for filling the vector elements to avoid
producing NaNs and a result of "inf" for the mean.)
It is hard to eliminate this sort of overhead without introducing a new
container type such as Vector because you can't have reduction functions
that are partially specialized for N_rank=1.
Regards, Julian C.
On Tue, 2005-05-17 at 10:45 -0700, Frank Astier wrote:
> Hi -
>
> I'm trying to use Blitz for a very simple mean computation on a 1D
> vector of floats. I'm comparing to my own algorithm, which just does
> a linear scan with N additions and a division. Compiling with -O2,
> apparently, Blitz is slower... Am I missing some compilation flag to
> make the thing go faster?
>
> Here are the runtime results and the whole program, followed by my
> implementation of mean (admittedly not safe):
>
> Blitz: N = 5000000 mean = inf in 1.71s
> Utils: N = 5000000 mean = inf in 1.6s
>
> int main(int, char** argv) {
>
> const size_t N = 5000000;
>
> {
> Array<double, 1> a(N);
>
> for (size_t i = 0; i < N; ++i)
> a(i) = double(rand() % 32768) / double(rand() % 65535);
>
> timer t;
> double m1 = mean(a);
> cout << "Blitz: N = " << N << " mean = " << m1
> << " in " << t.elapsed() << endl;
> }
>
> {
> vector<double> a(N);
>
> for (size_t i = 0; i < N; ++i)
> a[i] = double(rand() % 65535) / double(rand() % 32768);
>
> timer t;
> double m1 = mean(a);
> cout << "Utils: N = " << N << " mean = " << m1
> << " in " << t.elapsed() << endl;
> }
>
> return 0;
> }
> //------------------------------------------------------------------------------
> template <typename Iter1>
> inline IteratorValueType mean(Iter1 first, Iter1 last)
> {
> // non empty ranges only
> return sum(first, last) / (last - first);
> }
>
>
> //------------------------------------------------------------------------------
> template <typename Container>
> inline ContainerValueType mean(const Container& v)
> {
> return mean(v.begin(), v.end());
> }
>
> Frank
>
>
> _______________________________________________
> Blitz-support mailing list
> Blitz-support_at_[hidden]
> http://www.oonumerics.org/mailman/listinfo.cgi/blitz-support
>
-- Dr. Julian C. Cummings E-mail: cummings_at_[hidden] California Institute of Technology Phone: 626-395-2543 1200 E. California Blvd., Mail Code 158-79 Fax: 626-584-5917 Pasadena, CA 91125 Office: 125 Powell-Booth