Back in the dark ages, when I was in school, we were "encouraged" to take a
numerical analysis course if we were interested in computers. (I was an EE
major.) It was not an easy topic, but it made us well aware of the
difference between correct results and computational precision. I was
recently astonished to find that most computer science students have no
concept of this area and even less interest in it.

These current thoughts extend to other areas:

- C is more accurate than Fortran (or Basic, or what whatever)
- Obtaining "stable" numeric results means you get the same answer if
you run the program twice
- C produces the fastest programs
- if C is good then C++ is better
- Using all the obscure C operators produces a better program

(Anyone remember the IBM 7030 system? The user could control the rounding
direction of the floating point LSB. In this case running a program twice
(with different rounding options) really was relevant.)

Bill
W2WO

Dear Bill:
I too am appalled at the abandonment of a solid numerical analysis
course in engineering education. Consider the common problem of solving a
set of linear, independent algebraic equations. Students have to be shown
that Cramer's rule will not work when using the (inevitable) finite
resolution of a computer or calculator. Of course, some of the time
Cramer's rule does work so it is important to teach students why it does not
work in general.
This is relevant to antennas where we routinely need to solve large sets
of equations. When using a computer to perform calculations, one needs to
think differently about methods than in the day when one needed to use large
sheets of paper and a pen.
If one is to use numbers, one needs to know the limitations of methods
of use.
73 Mac N8TT

--
J. Mc Laughlin; Michigan U.S.A.
Home:
"Bill Ogden" wrote in message
...
Back in the dark ages, when I was in school, we were "encouraged" to take
a
numerical analysis course if we were interested in computers. (I was an
EE
major.) It was not an easy topic, but it made us well aware of the
difference between correct results and computational precision. I was
recently astonished to find that most computer science students have no
concept of this area and even less interest in it.


snip

Bill
W2WO

This doesn't really matter anymore in the U.S. but it is important that
other countries do not abandon it.We do not rely on home grown engineers
as we have in the past since a
simple telephone call offshore
meets our economic needs.
Art
"J. Mc Laughlin" wrote in message
...
Dear Bill:
I too am appalled at the abandonment of a solid numerical analysis
course in engineering education. Consider the common problem of solving a
set of linear, independent algebraic equations. Students have to be shown
that Cramer's rule will not work when using the (inevitable) finite
resolution of a computer or calculator. Of course, some of the time
Cramer's rule does work so it is important to teach students why it does
not
work in general.
This is relevant to antennas where we routinely need to solve large
sets
of equations. When using a computer to perform calculations, one needs to
think differently about methods than in the day when one needed to use
large
sheets of paper and a pen.
If one is to use numbers, one needs to know the limitations of methods
of use.
73 Mac N8TT

--
J. Mc Laughlin; Michigan U.S.A.
Home:
"Bill Ogden" wrote in message
...
Back in the dark ages, when I was in school, we were "encouraged" to take
a
numerical analysis course if we were interested in computers. (I was an
EE
major.) It was not an easy topic, but it made us well aware of the
difference between correct results and computational precision. I was
recently astonished to find that most computer science students have no
concept of this area and even less interest in it.


snip

Bill
W2WO

"J. Mc Laughlin" wrote in message
...
I too am appalled at the abandonment of a solid numerical analysis
course in engineering education. Consider the common problem of solving a
set of linear, independent algebraic equations. Students have to be shown
that Cramer's rule will not work when using the (inevitable) finite
resolution of a computer or calculator.

I was never shown that, but I do remember it being drilled into our heads that
Cramer's rule was the bogosort of linear system solving -- just about the
least efficient means you could possibly choose, and that it existed primarily
because it can be useful to have a closed form solution to a system of
equations.

Numeric analysis of linear systems is an incredibly in-depth topic, as far as
I can tell. Books such as SIAM's "Numerical Linear Algebra" spends hundreds
of pages going over it all.

---Joel

"Bill Ogden" wrote in message
...
Back in the dark ages, when I was in school, we were "encouraged" to take a
numerical analysis course if we were interested in computers. (I was an EE
major.) It was not an easy topic, but it made us well aware of the
difference between correct results and computational precision. I was
recently astonished to find that most computer science students have no
concept of this area and even less interest in it.

Realistically 90+% of CS students are going to end up in jobs programming web
pages, databases, and other applications where it just isn't going to matter.
There just isn't time in the curriculum these days to cover everything... For
that matter, these days something like understanding the effects of finite
precision in integer arithmetic and how it relates to fixed point DSP
calcuations is probably applicable to a larger number of students!

(OK, ok, I'd be the first to admit that college courses have been dumbed down
over the years as well, but this is a direct reflection of the fact that
industry just doesn't _think_ they need that many engineers who DO know the
'hard core' bits...)

These current thoughts extend to other areas:

- C is more accurate than Fortran (or Basic, or what whatever)

Arbitrary statement (on the student's part).

- Obtaining "stable" numeric results means you get the same answer if
you run the program twice

Ditto.

- C produces the fastest programs

There is some truth to this, perhaps if only because so much more work (as far
as I can tell) has been done on C optimiziers than for other languages.
Perhaps a better statement would be, "With novice programmers, C tends to
produce the fastest programs."

- if C is good then C++ is better

C++ does have a lot of nice benefits over regular old C. The last time I
programmed in FORTRAN it was FORTRAN 77, but I can only imagine that FORTRAN
90 has some nice improvements over FORTRAN 77 as well. (And Delphi is
purpoertedly a nice improvement to Pascal, etc...)

- Using all the obscure C operators produces a better program

Uggh.

---Joel

On Mon, 25 Apr 2005 17:47:55 -0700, "Joel Kolstad"
wrote:

- C produces the fastest programs

There is some truth to this, perhaps if only because so much more work (as far
as I can tell) has been done on C optimiziers than for other languages.
Perhaps a better statement would be, "With novice programmers, C tends to
produce the fastest programs."

Hi Joel,

I skipped this groaner the first time through. You could program in
almost any language to the same speed of performance if you simply
focused on the 5% bottleneck and coded it in assembler. Nearly every
"optimizer" consists of saving a lazy programmer's bacon when they
sloppily write poor control structures and assignment statements. It
should be called a de-babelizer.

73's
Richard Clark, KB7QHC

Hi Rich,

"Richard Clark" wrote in message
...
I skipped this groaner the first time through. You could program in
almost any language to the same speed of performance if you simply
focused on the 5% bottleneck and coded it in assembler.

Good point, although in the case of highly sophisticated CPUs (superscalar,
VLIW, etc.), the difficulty in getting all the cache and register access
scheduling optimal is difficult enough that there are typically very few
people who can consistently do better in assembly than a high level language
with an optimizer.

In many cases selecting a better algorithm might buy one a lot more!

On Tue, 26 Apr 2005 08:35:35 -0700, "Joel Kolstad"
wrote:

Good point, although in the case of highly sophisticated CPUs (superscalar,
VLIW, etc.), the difficulty in getting all the cache and register access
scheduling optimal is difficult enough that there are typically very few
people who can consistently do better in assembly than a high level language
with an optimizer.

In many cases selecting a better algorithm might buy one a lot more!

Hi Joel,

Almost every performance gain you describe is hardware limited - not
software limited. If you can conspire to make every reference call a
cache hit, you win, but 99.999% of the applications used by everyone
here (including antenna modeling) fail in that one regard and stumble
over the rest of the "optimizations." When I look at my performance
monitor, it is idling along at 0 to 2% usage as I type (no surprise).

When I pull up a page from the New York Times (before I set my
firewall filters to turn off advertising) it would peg at 100% ad
infinitum (I guess there's an ironic pun in that). I dare say that no
one is using optimized code for running Nike ads - or if they are,
that it makes any appreciable difference at 2GHz (with a memory access
running at, what, 10% of that?). What HAS been optimized is the data
compression schemes that make up for sloppy code (the problem is
undoubtedly a memory leak or a failed garbage collection routine).

In a sense, I used client side optimization to kill the advertising
stream.

73's
Richard Clark, KB7QHC

I love "religious wars" over languages and algorithms!!!

grin

Warmest regards,
John

"Richard Clark" wrote in message
...
On Mon, 25 Apr 2005 17:47:55 -0700, "Joel Kolstad"
wrote:

- C produces the fastest programs

There is some truth to this, perhaps if only because so much more work (as
far
as I can tell) has been done on C optimiziers than for other languages.
Perhaps a better statement would be, "With novice programmers, C tends to
produce the fastest programs."

Hi Joel,

I skipped this groaner the first time through. You could program in
almost any language to the same speed of performance if you simply
focused on the 5% bottleneck and coded it in assembler. Nearly every
"optimizer" consists of saving a lazy programmer's bacon when they
sloppily write poor control structures and assignment statements. It
should be called a de-babelizer.

73's
Richard Clark, KB7QHC

On Tue, 26 Apr 2005 08:45:36 -0700, "John Smith"
wrote:
I love "religious wars" over languages and algorithms!!!

Hi Brett,

Well, I've programmed in them all from binary to AI - so that makes me
an agnostic.

73's
Richard Clark, KB7QHC