Re: [pydstool-users] Improving code generation for constant mass matrix

[Juan P. C. <jua...@ug...>]

Hi,

I manage to get my hack in by printing to the C code only the non-zero elements.
The new code for FuncSpec is (between lines 1690 and 1699)
                # C integrators expect column-major matrices
                for col in range(n):
                    for row in range(m):
                        try:
                            if specdict[specvars[row]][col] != "0" and
specdict[specvars[row]][col] != "0.0": ##Print only nonzero elements
                              body_processed += "f_[" + str(col) +
"][" + str(row) \
                              + "] = " + specdict[specvars[row]][col] + ";\n"
                        except KeyError:
                            raise ValueError("Mass matrix should be
%sx%s"%(m,n))

The new compilation times are
rows of matrix, compilation time, ODEsystem constructor time, number
of equations
3     2.3    0.3   37
6     2.9    5.0   181
12   6.4    98.8  793

So now the bottle neck is in the constructor of the generator. I see a
lot of dictionary time which I guess corresponds to the entries of the
zero rows of the mass matrix. A solution would be to not make entry
for zero rows, but I guess this will brake all the code.

Is anybody interested in helping with the implementation of fast
assembly and C code compilation for the case of sparse and constant
mass matrix?


Thank you


The only problem with this is that all the other elements of f_ are
not set to zero. Where is the declaration and initialization of the
**double pointer that is passed to the massMatrix function?

On Wed, Mar 13, 2013 at 10:41 PM, Juan Pablo Carbajal
<jua...@ug...> wrote:
> Hello,
>
> When a DAE is defined with a constant mass matrix the C code generated
> has one line per element in the matrix. This causes very long
> compilation times for big matrices
> number of rows in massmatrix, size of c-code (KB), size of .so (KB),
> time to compile (seconds)
>       3     33    963      2
>       4    102   1100     4
>       5    265   1400     12
>       6    591   1800     58
>       7   1200   2400    238
>       8   2100   3200    831
> When replaced with a nested for loop the compilation time for rows==7
> goes down to 4 seconds.
>
> I tried to improve FuncSpec to print a for loop fr the case of
> constant matrix, but I can't cause the for loop is parsed as a macro
> by QuantSpec.
>
> Is there a way to prevent enumeration of the mass matrix and just get
> code of the kind
>
>  # C integrators expect column-major matrices
>   body_processed += "unsigned int i,j;\n" + \
>                                 "for (i=0;i<" + str(n) + ";i++)\n" + \
>                                  "\tfor (j=0;j<" + str(m) + ";j++)\n" + \
>                                  "\t\t f_[i][j] = 0;\n\n"
>  for col in range(n):
>      for row in range(m):
>           try:
>             if specdict[specvars[row]][col]:
>                body_processed += "f_[" + str(col) + "][" + str(row) \
>                                             + "] = " +
> specdict[specvars[row]][col] + ";\n"
>
> (this was my attemp in FuncSpec.py line 1688)
>
> Thanks
>
> --
> ---------------
> Dr. sc. nat. Juan Pablo Carbajal
> Reservoir Computing Lab
> Department of Electronics and Information Systems (ELIS)
> Ghent University
> Sint Pietersnieuwstraat 41
> 9000 Ghent
> Belgium
> web: http://users.elis.ugent.be/~jcarbaja/



-- 
---------------
Dr. sc. nat. Juan Pablo Carbajal
Reservoir Computing Lab
Department of Electronics and Information Systems (ELIS)
Ghent University
Sint Pietersnieuwstraat 41
9000 Ghent
Belgium
web: http://users.elis.ugent.be/~jcarbaja/