[Scilab-Dev] New xcos blocks executing scilab primitive functions

Eric GERNOT gernot.eric at wanadoo.fr
Tue Apr 10 10:31:06 CEST 2018 

Clément, 

 





Would it be different if we had to create a X_cos block executing a single Scilab expression (block output = RHS (block input, t, state derivative)) ?

 

I noted that the actual Scifunc Block is very powerful and educational thus to me it should not be changed.

 

Also I started suspecting that, although Scifunc blocks alone are slow compared to c or already provided blocks, they may not always be the bottleneck in the context of a specific calculation. 

I have to compare a few things and look at how many times each block is called by each solver.


 

Regards,  

Eric

 

 

 

> Message du 09/04/18 15:27
> De : "Clément David" 
> A : "dev at lists.scilab.org" 
> Copie à : 
> Objet : Re: [Scilab-Dev] New xcos blocks executing scilab primitive functions
> 
> Hello Eric,
> 
> As described in some guides [1][2], the interpreter will call a gateway passing values using the C++
> / C object-oriented / C stack-oriented APIs depending how you declared it. On the gateway you are
> free to do whatever you want with the data. 
> 
> [1]: https://wiki.scilab.org/howto/Create%20a%20toolbox#Gateway
> [2]: https://wiki.scilab.org/Developers
> 
> However for Xcos, the story is a bit different as the blocks will be called by a simulator (ODE /
> IDA solver) not an interpreter ; the API is different and called scicos_block4 ; using a C-block you
> have to describe the interface and write some code using the data passed as arguments.
> 
> Note: Scilab2C is a toolbox which might (or might not) follow Scilab coding conventions.
> 
> --
> Clément
> 
> 
> Le lundi 09 avril 2018 à 12:33 +0200, Eric GERNOT a écrit :
> > Hello Clement,
> > 
> > Thank you for this very clear answer.
> > How does Scilab Gateways work? When we tap enter after entering the command line on the prompt,
> > let's say to call a primitive, does the gateway also need the API ?
> > Or is there a more direct way to do so ?
> > 
> > Regarding my questions on the libraries.
> > Well, let's say I am a simple user of scilab, with no code development tools installed.
> > Editing a c-block, seeing the #include math.h line was a bit scary, because I thought "this will
> > require that I have put the math.h file at the good place on my computer".
> > So I insert nothing in the code, click "ok" and the compilation is succesful. 
> > Now after running Scilab 2 c toolbox on my custom code, a number of subdirectories with multiple
> > files were created in my custom directory. I suppose if I want this scilab 2 c created code
> > running in the c-block, I have to put these files where the block can access them at compilation
> > time.
> > Same if I created the code on my own. That is why I asked where the c-block compiler looks to find
> > the files, or how do we configure this, as users.
> > 
> > Best regards,
> > Eric
> > 
> > 
> > 
> > 
> > > > Message du 09/04/18 09:37
> > > > De : "Clément David" 
> > > > A : "dev at lists.scilab.org" 
> > > > Copie à : 
> > > > Objet : Re: [Scilab-Dev] New xcos blocks executing scilab primitive functions
> > > > 
> > > > Hi Eric and welcome :)
> > > > 
> > > > Nice to have some power users (and maybe futur developers) using big Xcos models, Xcos in
> > > 6.0.x have
> > > > some issues only visible on big diagrams and we currently do not spend too much time
> > > benchmarking
> > > > it.
> > > > 
> > > > About the scifunc vs Superbloc vs C++ implementation speedup, the clear winner there would be
> > > C/C++
> > > > if (and only if) there is no event management (either discrete or zero-crossing) to perform.
> > > Zero
> > > > crossing detection is hard to implement even for small sub-systems and discrete events are
> > > only easy
> > > > for synchronous subsystem. Superblocs are straightforward to implement and manage very complex
> > > > events / continuous states intrication in a clean manner however they can be confusing to
> > > maintains
> > > > (for exemple if you want to preserve the same behavior across different blocks). Scifunc are
> > > in-
> > > > between to me in term of implementation however very poor in term of performance (and not
> > > optimized
> > > > at all).
> > > > 
> > > > You could use the scicos_block4 API (the default one) in C or C++ ; if you take a look at our
> > > source
> > > > code there is only a few in C++ but they can be easily converted to. About the libraries, I
> > > did not
> > > > get it, do you want to link to different libraries ? The xcos_toolbox_skeleton and
> > > ilib_for_build()
> > > > flags are working well.
> > > > 
> > > > To call Scilab script within the sciblk2 [1] and sciblk4 [2] blocks we are indeed using the
> > > Scilab
> > > > C++ API, the performance overhead is quiet clear as we are currently allocating Scilab
> > > types::Double
> > > > or a types::MList through createblklist() for each output on each simulation step. This does
> > > an
> > > > extra allocation (for the Scilab datatype) and a copy (for the data copy to scilab) whereas in
> > > C/C++
> > > > the data are passed by reference to the shared buffers. We clearly have to do better there ;
> > > maybe
> > > > by allocating only once (at Initialization time) or passing types:: views to the shared
> > > buffers.
> > > > There is a test case within xcos_toolbox_skeleton which switch the implementation of a NOOP
> > > block
> > > > (that compute nothing) between Scilab and C depending on a parameter in Scilab.
> > > > 
> > > > [1]: http://cgit.scilab.org/scilab/tree/scilab/modules/scicos/src/cpp/sciblk2.cpp#n57
> > > > [2]: http://cgit.scilab.org/scilab/tree/scilab/modules/scicos/src/cpp/sciblk4.cpp#n217
> > > > 
> > > > As a conclusion, yes you can call C++ functions from blocks and Scilab blocks issues comes
> > > from a
> > > > slow implementation of C->Scilab calls not from the Scilab code interpretation.
> > > > 
> > > > NB: yes this a known problem and some users already reports that but no one is currently
> > > working on
> > > > that.
> > > > 
> > > > Thanks,
> > > > 
> > > > --
> > > > Clément
> > > > 
> > > > Le vendredi 06 avril 2018 à 05:47 -0700, ericgernot a écrit :
> > > > > Hi, 
> > > > > 
> > > > > I am new to scilab, and I am not a developper, but I fear I may become one
> > > > > because of Scilab :-;
> > > > > I am on windows 7 /64, and installed some recent versions.
> > > > > 
> > > > > I created a physic model (chemical process with compositions, streams, etc.)
> > > > > from a script and used xcos to solve it, only using Sci_func blocks calling
> > > > > my functions (so that it is easily portable to other scilab versions). It
> > > > > works as desired, so now I want to speed it up.
> > > > > 
> > > > > After all the possible improvements from vectorization, I decided to convert
> > > > > 30% of my scifunc into superblocks, giving a 30% speedup boost to the model.
> > > > > I could have written these in c/c++ also, but I suspect it would have taken
> > > > > longer (for me)
> > > > > 
> > > > > My remaining scifunc blocks are using some scilab gateways such as fsolve,
> > > > > diag, etc, thus the conversion is not as straighforward. 
> > > > > 
> > > > > I tried to convert it in c using the scilab 2 c toolbox, this is great for
> > > > > all operations exept scilab gateways.
> > > > > 
> > > > > I had a look at how to create a new block, how the c-block work, and finally
> > > > > how existing blocks are coded in c++.
> > > > > 
> > > > > If I understand well, the possibilities are
> > > > > 1. copy the c++ primitive function code in a c-block, it will not work
> > > > > because it is not c, and because I don't know which libraries should be
> > > > > included. By the way, where can I put them so that the compiler finds them ?
> > > > > 2. use scilab API from a xcos c-block and send jobs. Will it really skip the
> > > > > interpreter ?
> > > > > 3. create a new block as a user (from within scilab). But you can't use c++
> > > > > can you ? So I have to write or find c equivalent of scilab primitives, this
> > > > > is a pity !
> > > > > 4. create a new block source code, and compile a new version of scilab with
> > > > > the new block. I suppose that the same problem exists: converting to c
> > > > > scilab primitives ?
> > > > > 
> > > > > So this raised a few questions:
> > > > > - In the block source code, why not call scilab built in functions in c/c++
> > > > > (not the gateway), only passing them the block arguments / parameters,
> > > > > instead of pasting the code there. That would ease the creation of new
> > > > > blocks isn't it ?
> > > > > - can users call these built in functions in their custom blocks (e.g.
> > > > > c-block) ? (I mean without invoking the parser etc., without linking already
> > > > > linked libraries, etc.)
> > > > > 
> > > > > I suppose that many users ran through these problems.
> > > > > Thank you very mucn
> > > > > Regards, 
> > > > > Eric
> > > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > > --
> > > > > Sent from: http://mailinglists.scilab.org/Scilab-developers-Mailing-Lists-Archives-f2574944.
> > > html
> > > > > _______________________________________________
> > > > > dev mailing list
> > > > > dev at lists.scilab.org
> > > > > http://lists.scilab.org/mailman/listinfo/dev _______________________________________________
> > > dev mailing list dev at lists.scilab.org http://lists.scilab.org/mailman/listinfo/dev
> > 
> > _______________________________________________
> > dev mailing list
> > dev at lists.scilab.org
> > http://lists.scilab.org/mailman/listinfo/dev _______________________________________________ dev mailing list dev at lists.scilab.org http://lists.scilab.org/mailman/listinfo/dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.scilab.org/pipermail/dev/attachments/20180410/38ccae8f/attachment.htm>

More information about the dev mailing list