Re: optimizing method dispatch

[Roger H. <rog...@mi...>]

On Sunday, April 25, 2004, at 12:13  am, James W. Walker wrote:

> How much of an improvement are you seeing?

It is difficult for me to separate that optimisation from the others I 
have done. My e-mails back in
September talk of gains from 11% less time up to double the speed in 
GeomTest, and that was not
the end of the work I did. One of the changes is now checked in and I 
am awaiting approval for
the others. The most important methods seems to be creation and 
deletion of objects, even when
the object is still referenced, scanning down the class structure to 
the root object data (dispose
is a method of root) takes a lot of time in the cvs version. Allocating 
all the instance data for each
level of an objects parent classes was another thing I improved from 
the cvs - I now just do one
allocation and all the data lives in the one block of RAM. This may 
also give less problems with
cache misses as all the data is in consecutive locations. This of 
course also save time when the
object gets deleted and helps with memory fragmentation. The downside 
is that the structs have
to be moved from the E3xxxx.c to the E3xxxx.h files, though of course 
they are still not in the SDK's
Q3xxxx.h files.


>
> I tried a simple approach of storing a few of the most heavily-used 
> methods in the E3ClassInfo record.  This change improved the multibox 
> frame rate by 6% or so, and the changes are confined to E3ClassTree.c.

With the method tables, my implementation has to start at root and 
works towards the leaf classes. I have no done
everything yet so I would be interested to know which methods you found 
were important. I could then do any I
have not already done.

Roger.