I would like a simulink block which have double precision floating point in input and output but internal computations in this block are made in quadruple precision: it is possible to use with MP Toolbox ?

it is not possible to use the Matlab Coder while Advanpix is installed.

(Windows 7, Matlab 2018b.)

Target for coder is turbine20.m below which calls no multiprecision functions or definitions,

Only solution is to uninstall multiprecision  toolbox.

Does anyone know how to get around this?

Pavel,

Given as you say (in https://mct.userecho.com/communities/1/topics/125-mp-linprog-support) the high demand for optimization solvers and the complexity and difficulty of implementing high quality optimization solvers "from scratch", i.e., without access to existing double precision code as starting point for modification, I suggest a different development approach.:

Work with developers//vendors of high-end optimization solvers to modify their code to produce an MPT version. Perhaps the vendor would agree to make such a version available to its customers at no extra cost. What would be in it for the vendor would be potential extra sales. This could still be a non-trivial effort to do well, because even though solvers have a large number of settable parameters related to various tolerances, there may be some algorithmic and implementation decisions which might be implicitly predicated on an assumption that the code is double precision. I have no idea of the do-ability of an MPT version, for instance, if the solver code is in C or C++ and the vendors produce mex file, as I think is the case for the following.

Some suggestions  - these are all top-end, under active development, and all have MATLAB toolbox versions.

Note: LP = Linear Program, QP = (linearly constrained) Quadratic Program. QCQP = Quadrqatically Constrained Quadratic Program.. Mi = mixed integer. SOCP = Second Order Cone Problem, which subsumes LP, convex QP and convex QCQP. LMI = Linear Matrix Inequality. = Linear Semidefinite Program.

CPLEX (see https://www.ibm.com/support/knowledgecenter/en/SSSA5P_12.6.3/ilog.odms.cplex.help/CPLEX/MATLAB/topics/gs_ov_toolbox.html ) - LP, MILP,, convex QP, MIQP, convex QCQP, convex MIQCQP, SOCP, MISOCP, local optimal and globally optimal solution of non-convex QP and MIQP, drop-in replacement for lsqlin (linearly constrained linear least squares)

Mosek - LP, MILP, convex QP, convex MIQP, SOCP, MISOCP, LMI, exponential cone (in  forthcoming Mosek version 9.0)

KNITRO -all problem classes handled by FMINCON + integer constraints + special handling of complementarity constraints (MPEC) and SOC constraints.  Higher performance and more robust than corresponding FMINCON algorithms.

Gurobi - same problem classes as CPLEX, except no non-concex QP and MIQP capability

When we can expect latest version of MCT for linux?

Hello, I just found strange behavior (bug?) of MATLAB R2018a (update 3) with latest version of MCT on the Path (Ubuntu Linux 16.04, 64bit).

If I run the bench() function with MCT on the path I get the following results:

>> bench(5)

ans =

    0.0743    0.1146    0.0462    0.0832    0.4225    0.5393
    0.0750    0.1108    0.0462    0.0845    0.3682    0.4330
    0.0739    0.1176    0.0521    0.0895    0.4052    0.4539
    0.0785    0.1275    0.0473    0.0826    0.3355    0.4691
    0.0744    0.1155    0.0464    0.1066    0.3874    0.4228

The same benchmark run without MCT on the path produce the following results:

>> bench(5)

ans =

    0.0745    0.1184    0.0130    0.0839    0.4043    0.5116
    0.0798    0.1139    0.0133    0.0834    0.3796    0.4141
    0.0809    0.1275    0.0141    0.1051    0.4728    0.4096
    0.0850    0.1182    0.0132    0.0839    0.3576    0.4142
    0.0884    0.1132    0.0134    0.0859    0.4168    0.4460

As you can see, the 3rd column (corresponding to the ODE's part of MATLAB benchmark) shows significantly (~4x) slower runs in a case of MCT on the path.

Could you clarify and/or reproduce this situation, please?

After installation of R2018a_update3 (Ubuntu 16.04 64 bit) was observed significant slow down of some MCT built-in functions (especially fft, ifft,fft2 and ifft2).

>> mp
Invalid MEX-file '/home/kva/Documents/MATLAB/AdvanpixMCT/mpimpl.mexa64': /home/kva/Documents/MATLAB/AdvanpixMCT/mpimpl.mexa64: undefined symbol:
_Z19mxGetPropertySharedPK11mxArray_tagmPKc.

Error in mp (line 1255)
           if nargin == 0,     this.id = mpimpl(0);             % create mp-entity of default precision

The latest version of AdvanpixMCT 4.4.6.12719 is not compatible with Matlab R2018a (linux version). Please could you fix it?

Hello Pavel.

I noticed some performance issues with the besselh-function.

I used mp.Digits(34), mp.GuardDigits(0) and got the following

value=mp(rand(100,100));
tic
value1=besselh(mp('2'),2,mp('1i').*value);
toc
tic
value2=besselj(mp('2'),mp('1i').*value)-mp('1i')*bessely(mp('2'),mp('1i').*value);
toc
tic
value3=besselk(mp('2'),-value)*mp('2')/(mp('1i')*mp('pi'));
toc
Elapsed time is 3.986968 seconds.
Elapsed time is 1.879375 seconds.
Elapsed time is 0.772351 seconds.

which gives of course the same results for value1, value2 and value3. Any possibilty to enhance the performance of besselh to that of besselk?

Best regrads,

Michael

Will the Multiprecision Computing Toolbox work under WindowsXP x32?

Hello Pavel.

I found another bug in the isreal() function.

If we store real and complex data in the same array, the function isreal() gives always 0, even for entries with 0 imaginary part.

test(1,1)=mp('1i');
test(1,2)=mp('1');
isreal(test(1,1))
isreal(test(1,2))

ans =

  logical

   0

ans =

  logical

   0

This is differnt compared to the standard Matlab behaviour:

test(1,1)=1i;
test(1,2)=1;
isreal(test(1,1))
isreal(test(1,2))

ans =

  logical

   0

ans =

  logical

   1

Best regards,

Michael