Re: That's not the Answer to Your Question

On Apr 16, 7:43 pm, "aruzinsky" <aruzin...@xxxxxxxxxxxxxxx
cathexis.com> wrote:
"The lack of appreciation of the essential factors involved in block
processing is entirely yours."

Absolutely, and that lack of appreciation is warranted if block processing
is unnecessary to solve the OP's speed problem. And, it is you, not I, who
is remiss for not trying to determining the exact cause of the OP's speed
problem before pontificating on block processing.

You are blissfully ignorant of the technical issues and apparently
unwilling to learn.

The whole problem of the OPs speed with the long stride length
implicit in the vertical direction of access to his image array is
entirely due to cache misses. The performance difference on my P4 for
optimised image code ranges between 20-30x faster than naive code
accessing a 1024 square 8bit image using the wrong direction depending
on the algorithm, blocksize and level of optimisation. Don't take my
word for it do the experiment!!!

Meditate on this:

OP: "as processing pixel in a vertical fashion was _very_ slow (0.31 vs.
0.009 for the horizontal)."

And that is about right for naive code on a large array that forces a
read cache miss delay on every access. There might be something else
wrong but solving his memory access problems would get things a factor
of 10-20 faster in the vertical direction.

Tight sequential access to local blocks of memory in level 1 cache as
far as possible and then in level 2 *is* the solution. The OP at least
knew enough to ask the right question. You still don't seem to
understand the problem.

How is it that I am the only one who suspects an outstandingly dumb error
in the OP's code?
I have never encountered anywhere near such a large disparity between
horizontal and vertical processing speeds on a personal computer, even

Your lack of experience is showing. The outstandingly dumb error is
naive obvious coding.

"And also annoyingly by modifying the header you break the threading on
some newsreaders that
don't handle followups too elegantly."

You should blame the newsreaders instead of using me as a scapegoat for
your annoyance. You apparently missed two corrections to my post.

If you change the headers willy nilly you must expect some posts to
get lost. I read via free access NNTP but since my ISP no longer
carries Usenet I post via Google. Your messed up subject headers fool
Google.

Very few optimising compilers are quite so dumb but there are better ways
to cater for odd N."

I wouldn't know, because I haven't done a survey of compilers, but I
suspect calculation in the loop might be part of the C or C++ language
standard because the philosophy is to give the programmer control. I'll
ask some programmers and get back to you.

Optimising compilers have been taking unecessary stuff out of loops
since the mid 70's. Most PC compilers caught up with mainframe
optimising techniques in the late 80's and after that almost any loop
invariant expression would be spotted by the compiler and moved
outside the loop. This was sometimes dangerous if for example a
function call to a random number generator was erroneously optimised
out of a loop (because it wasn't declared volatile).

C provides too many ways for programmers to make life difficult for
optimising compilers but that is a separate issue.

Regards,
Martin Brown

.

Relevant Pages

  • Re: Release Dll Problem - Optimization Issue
    ... If optimising for speed on smaller size processors, ... This limits the compiler in using fast registers for the ... and speeds up access to the variable. ...
    (microsoft.public.pocketpc.developer)
  • Re: Benchmarking Quake II .NET Exposed
    ... The Quake II native Vertigo ... code ran at similar speeds, so the .NET *can* deliver similar ... quite differently - the nature of optimising is different on the two ... idioms with short-lived temporaries is going to clog up the garbage ...
    (borland.public.delphi.non-technical)
  • Re: Thats not the Answer to Your Question
    ... and that lack of appreciation is warranted if block processing ... horizontal and vertical processing speeds on a personal computer, ... I now suspect the OP made the whole thing up to ...
    (sci.image.processing)