New community-maintained version of "Calculators benchmark: add loop"

[Pier Aiello]

So, i just copied (even if is not complete) the original calculator benchmark of Gene Wright that is no more updated (1) on a wiki4hp page.

Do you know other results that weren't added to the original article? If yes, could you link me the discussion? (or could you add it directly on the wiki? (2) )

(1) The other benchmark that i known, the N-Queens, is regularly updated.

(2) Here is, for me, the power of a wiki: concentrate the efforts of multiple users on a single page and it is really useful for FAQs; articles that needs continuous updates; organization in a more usable way of preexisting contents on other sites (ex: an organized index of links); and so on without the limit of a single editor (that can become inactive).

[nina scholz]

fresh test with hp 15c le (ordered by speed) disclaimer: stopped by hand, maybe 2 s +/-

LBL  E + GTO  E		48823

LBL 09 + GTO 09		48763

LBL  A + GTO  A		48721

LBL 01 + GTO 01		48509

LBL .9 + GTO .9		44448

LBL .1 + GTO .1		44083

[Pier Aiello]

I'll add it ASAP. Thanks :)

PS: My hp50g is way slower (if i don't use HPgcc but standard userRPL) than your arm powered 15c. That's unfair!!! :P

Edited: 10 Sept 2013, 9:32 a.m.

[Pier Aiello]

Added!

[Marcel Samek]

With Stack depth 4

LBL'XYZ' + GTO'XYZ'		90891

LBL A    + GTO A                90675

LBL 00   + GTO 00               90369

With Stack depth 8

LBL 00   + GTO 00               78209

[Paul Dale]

The earlier version 2 firmware will be faster than version 3.

There are lots of other possibilities for improvement BACK and INC for example. Also, integer mode might be a little faster.

- Pauli

[Thomas Klemm]

Could you please try the following program:

000 +

001 x=0

Just fill the stack with 1s and hit the R/S key. For this to work all the programs have to be cleared beforehand. So I'd understand if you don't want to do that.

Another variant is to start with 0 in register I and use this program:

000 ISG I

The results are 2038 respectively 1669 using a DM-15CC.

Best regards

Thomas

PS: A false condition on the last line jumps over to the first line.

Edited: 10 Sept 2013, 1:20 p.m.

[Pier Aiello]

#5 and #6 added!

PS: just register on the wiki and contribute directly!

Edited: 10 Sept 2013, 1:49 p.m.

[Tim Wessman]

Kind of hard since the source is locked by administrator. :-?

Prime: v2013.8.13

export LOOP()

begin

 A:=0;

 for A from 1 to 1E9 do

 end;

end;

Recall A after ON to stop running on 1 min. Average of several runs.

6.646.300

TW

Edited: 10 Sept 2013, 4:14 p.m.

[Thomas Klemm]

Reminds me of The Evolution of a Programmer.

[Pier Aiello]

Thanks! (Added)

About editor status, i get it in minutes but i agree, a wiki should be quite open, indeed i create this wiki talk: here

I hope that the admin will ad some new moderators ASAP.

[Gilles Carpentier]

This version (same to the PRIME version) is faster on the HP39GII than a A+1->A version published

EXPORT LOOP()

BEGIN

 A:=0;

 FOR A FROM 1 TO 1E9 DO END;

END;

-> 1062108 loop in 60 sec

The prime is 6.25 x faster

Edited: 11 Sept 2013, 3:20 p.m.

[Pier Aiello]

Thanks! Indeed for loops are faster in general :)

Edited: 11 Sept 2013, 3:37 p.m.

[Gilles Carpentier]

By the way, the A:=0 line can be deleted :

EXPORT LOOP()

BEGIN

 // A:=0; 

 FOR A FROM 1 TO 1E9 DO END;

END;

But this change nothing

[nina scholz]

+ x=0?		59006 or 59182

ISG 00		41062 or 41048

ISG  I		40746

ISG .9		40922

ps: thanks, i did't know.

[Pier Aiello]

added :)

[Thomas Klemm]

Don't know whether that counts but it was fun to do.

Odhner Count: 251.

I can hardly believe that I was faster than a HP-67.

Cheers

Thomas

[Pier Aiello]

Sure it counts!

And, this time, i think that emulators can have their list. So if you have an emulator, use it!

[Pier Aiello]

I have added new, and updated, results found on MoHPC :)

- Some TI performs better with a for loop

- 39gII results

- a 30b with crystals.

Edited: 11 Sept 2013, 11:36 a.m.

[Marcel Samek]

I was playing around with this on the WP-34S and could see that how one codes the loop makes a big difference.

If you run two tests, one with one addition inside the loop, and the second with two additions inside the same loop, then it is possible to deduce the cost of the loop and figure out just the cost of the addition operation.

I ran some experiments on the WP-34S to see how consistent this was. I ran a loop with 1,2,3,4,5, and 6 additions inside it and the cost of the loop was consistent.

When one takes this approach, it turns out that the setup code and the loop can be as complex as need be in order to set up the benchmark of the individual operation. Thus, for multiplication or division, one can make sure the stack is set up with valid numbers. The multiple tests will hide the cost of the setup/loop.

I ran six tests to benchmark addition, and multiplication on the WP-34S. You can see that the average result is that the calculator can do 2428 additions per second and 2422 multiplications per second.

Here is the spreadsheet with my numbers and calculations:

Spreadsheet with results (LibreOffice/OpenOffice)

I would find it interesting to use this kind of approach to compare the most common operations on all the calculators to get an idea of their relative speed. I have played with different modes on the WP-34S and can see the difference they make.

I will continue to benchmark the WP-34S instructions in different modes and will post the results in the next few days.

[Pier Aiello]

Nice idea! (and it is much more consistent to compare calculators on four operations)

Could you setup a wiki page (or an article) about this to store the results? Thanks :)