Arne Linder wrote:
>
> Hello everybody,
>
> my question is a bit off topic, but since I know from the conference in Milan, that
>there are a lot of programming specialists among us, I hope, you can help me.
>
> In my real-time application an ISR has to do a matrix multiplication in real time.
>This is done by the following code:
> ===
> for (j=0; j<Np; j++)
> {
> Element=0.0;
> k=(j+1)*Np;
> for (i=0; i<Np; i++)
> Element+= *(P+k+i) * (*(W+i) - *(f+i));
> *(DU+j)=Element;
> }
> ===
>
> *DU, *W and *f are declared as double.
> Compiler options: -O2 -D__KERNEL__ -DMODULE -c -I/usr/src/rtai- 1.3/include
>-funroll-loops
>
> With Np=100 the calculation time is 51�s. This seems to be quite high for an AMD
>Duron processor with 900MHz. So I tried to find out, which part of the routine is
>consuming most of the time. I realized, that deleting the command "*(DU+j)=Element"
>reduces the processing time to 9�s.
> My first thought was, that due to cache misses the time to access *DU gives that
>huge difference, but changing "*(DU+j)=Element" into "Element=*(DU+j)" also resulted
>in 9�s processing time. (Of course the program didn't give the desired results in
>that case.)
>
> In my opinion the gcc compiler (version 2.95.2 19991024) refuses to unroll the loops
>correctly when inserting the line "*(DU+j)=Element;" into the outer for-loop. Does
>someone have an idea to solve this problem?
>
You are getting 200 MFlops (tight multiply/add loop). Under gcc I think
it is a consistent performance for your machine. It goes down to 9
likeley just because gcc sees it as a useless operation, if the result
is not saved, and then does not carry out the related calculations.
double *row;
for (j = Np - 1; j >= 0; j--) {
Element = 0.0;
row = P[j];
for (i = Np - 1; i >= 0; i--) {
Element += row[i]*(W[i] - f[i]);
}
DU[j] = Element;
}
Gcc perfectly understand array indeces and there is no performance
difference with respect to using pointers directly, but you get
something more readable.
Also decrementing indeces can help because the loop test is directly on
zero.
To get more use SIMD support.
Paolo.
-- [rtl] ---
To unsubscribe:
echo "unsubscribe rtl" | mail [EMAIL PROTECTED] OR
echo "unsubscribe rtl <Your_email>" | mail [EMAIL PROTECTED]
--
For more information on Real-Time Linux see:
http://www.rtlinux.org/
