I was not able to get any significant speedup in the kperf inner loops after 
all. It turns out that variations in profile performance from run to run are 
greater than the differences in performance between loop versions. It would 
take multiple runs plus statistical analysis to reveal any real differences, 
which would be on the order of about a percentage point just for the code 
running in the kperf function not counting opcodes called from it; I do not 
currently have time to perform this level of analysis for what promises to 
be a very small gain in performance.

The BLAS does not have functions that directly perform the operations 
required by arate/krate arithmetic; BLAS is designed to supply routines for 
scientific vector and matrix arithmetic, not simple element by element 
vector arithmetic; the code that currently exists is optimal, although there 
are elementwise vector operations in the Intel performance libraries that 
would almost certainly be somewhat faster.

In short, I was not able with a modest investment of time to identify any 
places that could afford significant speedups in existing Csound code.

Looks like the engine developers have done a very good job!

Regards,
Mike





-------------------------------------------------------------------------
This SF.net email is sponsored by the 2008 JavaOne(SM) Conference 
Don't miss this year's exciting event. There's still time to save $100. 
Use priority code J8TL2D2. 
http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone
_______________________________________________
Csound-devel mailing list
Csound-devel@lists.sourceforge.net

Thanks for reporting this Mike. Even though you didn't find some major  
schism,  it is good to check these things sometimes, and provides an  
answer to some of those niggling doubts.

David
On 21/04/2008, at 12:46 AM, Michael Gogins wrote:
> I was not able to get any significant speedup in the kperf inner  
> loops after
> all. It turns out that variations in profile performance from run to  
> run are
> greater than the differences in performance between loop versions.  
> It would
> take multiple runs plus statistical analysis to reveal any real  
> differences,
> which would be on the order of about a percentage point just for the  
> code
> running in the kperf function not counting opcodes called from it; I  
> do not
> currently have time to perform this level of analysis for what  
> promises to
> be a very small gain in performance.
>
> The BLAS does not have functions that directly perform the operations
> required by arate/krate arithmetic; BLAS is designed to supply  
> routines for
> scientific vector and matrix arithmetic, not simple element by element
> vector arithmetic; the code that currently exists is optimal,  
> although there
> are elementwise vector operations in the Intel performance libraries  
> that
> would almost certainly be somewhat faster.
>
> In short, I was not able with a modest investment of time to  
> identify any
> places that could afford significant speedups in existing Csound code.
>
> Looks like the engine developers have done a very good job!
>
> Regards,
> Mike
>
>
>
>
>
> -------------------------------------------------------------------------
> This SF.net email is sponsored by the 2008 JavaOne(SM) Conference
> Don't miss this year's exciting event. There's still time to save  
> $100.
> Use priority code J8TL2D2.
> http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone
> _______________________________________________
> Csound-devel mailing list
> Csound-devel@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/csound-devel


-------------------------------------------------------------------------
This SF.net email is sponsored by the 2008 JavaOne(SM) Conference 
Don't miss this year's exciting event. There's still time to save $100. 
Use priority code J8TL2D2. 
http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone
_______________________________________________
Csound-devel mailing list
Csound-devel@lists.sourceforge.net

I managed a very small kperf improvement (2,525,405,620 instructions
down from 2,529,244,588 before -- tested on trapped to file).

There are improvements in spoutsf (1,591,343,985->1,538,815,089),
reverb (1,581,859,503->1,463,215,005) and oscka
(1,472,011,800->1,405,458,075)

Overall I got 1.1% improvement in instructions.  So Mike is correct --
a lot of work for little gain!

==John ffitch

-------------------------------------------------------------------------
This SF.net email is sponsored by the 2008 JavaOne(SM) Conference 
Don't miss this year's exciting event. There's still time to save $100. 
Use priority code J8TL2D2. 
http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone
_______________________________________________
Csound-devel mailing list
Csound-devel@lists.sourceforge.net