World: r4wp

Still prefer to make the choice as a programmer. How it is done behind 
the curtains is up to the compiler. Compare this with the address 
used in memory. Changing your array from 0-based to 1-based, it ends 
up in the same memory address (say that is the only change you made 
in the source). 

Besides using the 0-based structure was perhaps a struggle for the 
original programmer!!

We, no I, limited myself to keep as close to the original as possible. 
Also because of the obfusction factor of the source. I believe much 
of the construction could be made more like in the REBOL way of doing 
things. A C program just is not a Red/System program. My goal was 
to get a working program reproducing the same result in the shortest 
time. Because I had little idea what the original program did because 
it is complicating things on purpose this was the way to do it. Now 
based on this useable program it can be made more Red/REBOL-like 
step by step.

I do feel for you, for indeed a lot of stuff out there is C or 0-based 
and it brings its problems along, but the next step in bringing Red/System 
to 0-based is having Red being 0-based as well?

[...] the next step in bringing Red/System to 0-based is having Red 
being 0-based as well?


No, it's not directly related. Implementing low-level algorithms 
that require working with index 0 is not the common usage of Red. 
Also, the +/-1 offset required in such case has no noticeable performance 
impact in a Red program while in Red/System, in a tight nested loop, 
the impact could be significant.


The most important part in considering a 0-based indexing system 
for Red/System is mainly helping the user and avoiding common programming 
errors (even I get caught by it from time to time, but it's probably 
due to my strong C background).

The problem is that there is no final solution. In practice, I have 
mixed Red and Red/System code in many places, and I'm already often 
mixed up between the two. So if you change Red/System to be 0 based, 
there will be a new class of errors where programmers think in Red 
in their Red/System code, or in Red/System in their Red code. And 
as Arnold hints at, this will create an eternal pressure from technical 
people to make Red 0 based, and then we have given up on the human 
centered goals of REBOL and are back at square one - which will then 
have to be renamed square 0

Arnold, if you want to provide your random function as a library 
to other programmers, then yes, you should include a function to 
clean up its internal state

Doc, I don't know if you've done other tests, but in the benchmarks 
so far Red/System was only 4 times slower than optimised C in floating 
point code. For integer code, it was around twice slower than optimised 
C, so the Mersenne Twister confirms this

I thought it was twice slower than unoptimised C?

this will create an eternal pressure from technical people to make 
Red 0 based

 In fact, I've decided since a while to add PICKZ and POKEZ to Red 
 so 0-base algorithms would be more natural to implement. I need to 
 add an entry in Trello about that or I will keep forgetting about 
 it...

It's a bit confusing, but Red/System integer is as fast as unoptimised 
C (x 1 in function calling, 1.3 slower in hard integer as measured 
by XieQ), twice as slow as optimised C, and Red/System floating point 
is four times as slow as optimised C

Which also means that Red/System floating point is twice as slow 
as unoptimised C

Any fine examples of how to program refinements in Red(/System) I 
can use to copycat?

Red/System doesn't have refinements. Refinements in Red are almost 
exactly like in REBOL

Hmm random/seed needs an input, random does not, but has an output 
that random/seed has not.

As a function, you'd have to call that SEED or seed-random in Red/System

If you want to offer multiple functions as a library, you can wrap 
their shared state in a CONTEXT

Then you can have random/seed, but it won't be a refinement :-)

And it will definitely stop looking like C :-)

It certainly would be nice to wrap all possibilities up in one package.

I think that's something different than what I meant

I think about a flexible solution to use the one random routine you 
want or need.

You can implement several algorithms with the same interface. CONTEXT 
is useful there. Then you can #include the one implementation you 
want

There are many such constructs in my bindings and Red itself

Hi Doc, I noticed that you published Android support yesterday. Will 
you publish also some example how to use it?

Note that it's preliminary, in the development branch

Is it a problem? :)

I consider all branches to be a development as all Red/S is still 
in alpha (or is it already beta?) version.

I think Red/System counts as beta, as far as v1 is concerned

I don't know if it's a problem. What I do know is that the Android 
port doesn't fully work yet. If you consider it developmental, then 
you can also expect Doc to add examples later. Anyway, I was just 
providing a preliminary answer for you because Doc is offline

Hi Oldes, I have pushed online only the low-level layers of the Android 
port. I will push the rest after some cleanups most probably tomorrow. 
It's working fine on pre-4.1 Android systems, but I'm still struggling 
with low-level bugs with 4.1+.

I will provide the full source code of the preview demo I've posted 
and will probably add another demo. It's still low-level manipulations 
of the Android Java framework, the final layers of abstractions are 
not yet there (dialects, schemes, ...).

I will need to go back to object! implementation in Red before implementing 
those layers.

Quick testing of the MT Red/S and C versions on my machine give a 
~4x difference in speed.

Odd. Is that on a Mac?

Linux

Which GCC version?

Red/System (current master, g0c9c1bf): 2.5s
GCC 4.8.1 -O0: 1.9s
Clang 3.3 -O0: 1.6s
Clang 3.3 -O2: 0.6s
GCC 4.8.1 -O2: 0.5s

So there's a much bigger difference between optimised and unoptimised 
versions than in previous tests

I was on an older GCC version then, 4.1.2 or 4.2.3

Xie is on Windows, and his compiler is apparently called cl. XieQ, 
is that CLang? Which version?

That's MSVC.

Perhaps GCC has gotten better at optimisation

Compared to 6-year old GCC 4.1.2, it certainly has improved a lot.

Just dug one of your Fibonacci results from back in Feb-2012, that 
used GCC 4.43.

dug up*

Oh, don't remember that

Ah, it must have been GoboLinux's GCC

You mentioned the benchmarks running on Syllable Server.

Yes, that's why I thought 4.2.3. But I would have tried them on Syllable 
Workstation on top of Syllable Server

That's an unreleased GoboLinux-on-Syllable-Server FrankenLinux

So it's a pretty big difference in optimisation, and MSVC doesn't 
show it