The performance of all kinds of C operations

[mrby]

Hi,

Does anyone know of any link which describes the (relative)
performance of all kinds of C operations? e.g: how fast is "add"
comparing with "multiplication " on a typical machine.

Thanks!

--
B. Y.

[Eric Sosman]

mrby wrote:

Hi,

Does anyone know of any link which describes the (relative)
performance of all kinds of C operations? e.g: how fast is "add"
comparing with "multiplication " on a typical machine.

Pages with this sort of information are scattered about
the Web. Most that I've seen have been highly system-specific,
whether they say they are or not. One page I saw made a fairly
serious effort to assign "costs" to various C constructs, but
it seemed rooted in the days when optimizers were less radical
and when the speed disparity between CPU and memory was not so
enormous.

Nowadays it is very nearly meaningless to ask whether an
addition is slower or faster than a multiplication, even if the
data types are specified. If the operands of the addition are
in memory while those of the multiplication are in registers,
the multiplication will likely finish far sooner. A division
will likely finish far sooner; even taking the square root of
a register-resident value will likely be quicker than performing
an addition that incurs two memory references.

You can probably get an answer by ignoring the effects of
memory, of the multiple levels of cache, and of pipelining --
but the answer you get by ignoring reality is not likely to be
very helpful. It's like noticing that jet airplanes are faster
than bicycles, and therefore choosing an airplane for a one-mile
journey.

In all but a tiny and steadily diminishing fraction of cases,
micro-optimization is a waste of time and effort. Choose a good
algorithm without worrying about whether it uses multiplications
or additions, pointer arithmetic or array indexing. Then code it
as clearly and robustly as you can. Go no further unless and
until you have measured the resulting performance and found it
inadequate.

--
Eric Sosman
es*****@acm-dot-org.invalid

[websnarf]

mrby wrote:

Does anyone know of any link which describes the (relative)
performance of all kinds of C operations? e.g: how fast is "add"
comparing with "multiplication " on a typical machine.

Well, things are not that simple, but I have an old page with this kind
of information that still stands up:

http://www.pobox.com/~qed/optimize.html

where I stated that: "Arithmetic operation performance is ordered
roughly by: transcendental functions, square root, modulo, divide,
multiply, add/subtract/mutiply by power of 2/divide by power of
2/modulo by power of 2." I guess I should have added addition,
subtraction and logic operations at the end. Anyhow, this list is
still largely true on pretty much all architectures. The reason why
all architectures seem to perform so similarly on these operations is
that the best or near best techniques for building logic that perform
all those operations in hardware are generally well known. This is a
reality worth noting.

However, it should be noted that over time, memory bandwidth has not
kept pace with modern CPU speeds. Because of this, operations even as
slow as trascendental functions are now no longer slower than first
time uncached memory accesses. So if you think of memory access as an
"operation" , this would be the one major one which has changed its
relative performance over time (by getting slower).

--
Paul Hsieh
http://www.pobox.com/~qed/
http://bstring.sf.net/

[Jack Klein]

On 29 Apr 2006 05:55:30 -0700, "mrby" <bi******@gmail .com> wrote in
comp.lang.c:

Hi,

Does anyone know of any link which describes the (relative)
performance of all kinds of C operations? e.g: how fast is "add"
comparing with "multiplication " on a typical machine.

Thanks!

The real reason that your question is meaningless is that there is no
such thing as a "typical machine", as far as C is concerned. This is
not just theoretical, there is a vast difference between 8-bit
microcontroller s such as an 8051 or AVR on the one hand, and 64-bit
desk top processors like Pentium or PowerPC on the other.

I do a lot of work these days with a DSP where multiplication and
addition take exactly the same amount of time, one clock cycle. And
it can also do MAC, multiply two numbers and add them to an
accumulator, in one clock cycle. Or it can square a number and add it
to an accumulator, all in one clock cycle.

--
Jack Klein
Home: http://JK-Technology.Com
FAQs for
comp.lang.c http://c-faq.com/
comp.lang.c++ http://www.parashift.com/c++-faq-lite/
alt.comp.lang.l earn.c-c++
http://www.contrib.andrew.cmu.edu/~a...FAQ-acllc.html

[Ronald Bruck]

In article <mr************ *************** *****@4ax.com>, Jack Klein
<ja*******@spam cop.net> wrote:
....

I do a lot of work these days with a DSP where multiplication and
addition take exactly the same amount of time, one clock cycle. And
it can also do MAC, multiply two numbers and add them to an
accumulator, in one clock cycle. Or it can square a number and add it
to an accumulator, all in one clock cycle.

VERY interesting. While this is irrelevant to C, do you know how the
DSP accomplishes this? IIRC the complexity of multiplication is higher
than that of addition (perhaps the DSP parallelizes the operation
better?).

--Ron Bruck

[Tim Prince]

Ronald Bruck wrote:

In article <mr************ *************** *****@4ax.com>, Jack Klein
<ja*******@spam cop.net> wrote:
...

I do a lot of work these days with a DSP where multiplication and
addition take exactly the same amount of time, one clock cycle. And
it can also do MAC, multiply two numbers and add them to an
accumulator, in one clock cycle. Or it can square a number and add it
to an accumulator, all in one clock cycle.

VERY interesting. While this is irrelevant to C, do you know how the
DSP accomplishes this? IIRC the complexity of multiplication is higher
than that of addition (perhaps the DSP parallelizes the operation
better?).

--Ron Bruck

So we are to assume that these one cycle quotations apply to peak
throughput, not total latency?

[Roberto Waltman]

On Thu, 15 Jun 2006 18:29:20 GMT, Tim Prince
<ti***********@ sbcglobal.net> wrote:

Ronald Bruck wrote:

In article <mr************ *************** *****@4ax.com>, Jack Klein
<ja*******@spam cop.net> wrote:
...

I do a lot of work these days with a DSP where multiplication and
addition take exactly the same amount of time, one clock cycle. And
it can also do MAC, multiply two numbers and add them to an
accumulator, in one clock cycle. Or it can square a number and add it
to an accumulator, all in one clock cycle.

VERY interesting. While this is irrelevant to C, do you know how the
DSP accomplishes this? IIRC the complexity of multiplication is higher
than that of addition (perhaps the DSP parallelizes the operation
better?).

--Ron Bruck

So we are to assume that these one cycle quotations apply to peak
throughput, not total latency?

<OT>
I am using a TI F2812, a similar processor (I believe) to the one Jack
is referring to. The "one op per clock cycle" can be sustained as long
as the code & operands reside in the CPU internal RAM.

The throughput is reduced drastically when accessing data in internal
FLASH memory, or external memory (of any kind,) since that requires
inserting a few wait states on each memory access cycle.

But still, as long as the location of the data is similar, the time
required to perform an addition, multiplication or MAC operation
remains identical.
</OT>

[spibou]

From: websn...@gmail. com (Paul Hsieh)
Date: Sat, Apr 29 2006 3:57 pm

mrby wrote:

Does anyone know of any link which describes the (relative)
performance of all kinds of C operations? e.g: how fast is "add"
comparing with "multiplication " on a typical machine.

Well, things are not that simple, but I have an old page with this kind
of information that still stands up:

http://www.pobox.com/~qed/optimize.html

I didn't read your whole page but had a look at the table in the
section "Strictly for beginners". Can you explain why would
"x = y << 3" be faster than "x = y * 8" ? Or why would
"if( ((a-b)|(c-d)|(e-f))==0 )" be faster than "if( a==b &&c==d &&e==f
)" ?

Spiros Bousbouras

[Chris Torek]

An off-topic answer to an off-topic question; I shall keep it short:

In article <15************ **********@math .usc.edu>
Ronald Bruck <br***@math.usc .edu> wrote:

... do you know how the DSP accomplishes [single-cycle multiply / MAC]?
IIRC the complexity of multiplication is higher than that of addition
(perhaps the DSP parallelizes the operation better?).

Look up "Booth multiplier".
--
In-Real-Life: Chris Torek, Wind River Systems
Salt Lake City, UT, USA (40°39.22'N, 111°50.29'W) [phone]
email: forget about it http://web.torek.net/torek/index.html
Reading email is like searching for food in the garbage, thanks to spammers.

[Kevin D. Quitt]

On Thu, 15 Jun 2006 18:29:20 GMT, Tim Prince <ti***********@ sbcglobal.net> wrote:

VERY interesting. While this is irrelevant to C, do you know how the
DSP accomplishes this?

A 16x16 multiply can be accomplished in one lookup into a 64K-word table.
--
#include <standard.discl aimer>
_
Kevin D Quitt USA 91387-4454 96.37% of all statistics are made up