Dear All, Is it possible to compile gretl at own pc with new c functions on Linux, e.g. Ubuntu?

On Sat, 14 Oct 2017, Sven Schreiber wrote: > Am 14.10.2017 um 15:55 schrieb > oleg_komashko(a)ukr.net: >> Dear All, >> Is it possible to compile >> gretl at own pc with new >> c functions on Linux, e.g. >> Ubuntu? > > Do you really mean just a new C function, or do you mean a new > built-in hansl function which uses compiled C code for execution? > > The first case is probably relatively easy, you just have to make > sure that the new function is actually picked up and gets included > into libgretl. About the second case -- well that's the secret of > the masters I guess... > > No seriously, it's a very useful question. (But I would like to > know myself.) Yes, a good question. Adding a new C function to libgretl is just a matter of... doing just that. You write a function in C, stick into an appropriate .c file in lib/src, and -- if it's supposed to be publicly available (and I don't suppose there would be much point otherwise) -- also add a declaration of the function to the corresponding .h header file. Making such a function accessible via hansl requires working with the C files genparse.h, genlex.c and geneval.c (in lib/src). genparse.h is where you insert a unique internal identifier for the function (e.g. F_NUMHESS), putting it into the functions "enum" in the right place. Scanning the existing enum one should get an idea of what "the right place" means. Basically: how many arguments does the new function take? -- other than that it doesn't much matter where it goes. genlex.c is where you create a mapping from the function's internal identifier (e.g. F_NUMHESS) to the name under which users can call it (e.g. "numhess"). Plenty of examples available there. The exact position of the new entry doesn't matter. geneval.c is where you actually "hook up" the hansl function to its C back-end. A hansl-function must have an entry in the function named "eval". This entry should (usually) do some basic type-checking on the arguments provided by the user, then hand off to an implementer function. The best way to grok this is to follow the path of execution for a hansl function that you're familiar with. Ideally you want to find an existing function that is as similar as possible to your new one, in terms of (a) what it returns, and (b) the type of arguments it expects. I (or Jack) can provide more detail as needed, but that's the basic overview. Allin _______________________________________________ Gretl-devel mailing list Gretl-devel@lists.wfu.eduhttp://lists.wfu.edu/mailman/listinfo/gretl-devel

On Sun, 15 Oct 2017, > oleg_komashko(a)ukr.net wrote: > Thank you, Allin > If I do not mistake the thing > I want to try use libc only > so, hopefully, there will > be no problems with include If your thoughts tend in the direction of complex matrices, maybe I should point out that we have a few "hidden" hansl functions for complex operations. They're hidden because we don't have a native complex-matrix type, and also because the API is still negotiable, but... our current internal representation of an m x n complex matrix is a 2*m x n gretl matrix, where the doubling of the number of rows means that each complex element occupies two rows (real, then imaginary). The functions available are (yes, the leading underscore is part of the name): _ceigh: eigen-decomposition of complex (Hermitian) matrix _cinv: inverse of a complex matrix via LU decomposition _cmmult: multiplication of complex matrices _chprod: Hadamard product of complex matrices _cmatrix: compose a complex matrix from its real and imaginary component matrices _cfft: complex Fourier transform _cxtract: extract the real or complex part of a complex matrix _ctran: conjugate transpose of complex matrix You can inspect the C code in lib/src/gretl_cmatrix.c. We use OpenBLAS functionality for the more substantial operations. Allin _______________________________________________ Gretl-devel mailing list Gretl-devel@lists.wfu.eduhttp://lists.wfu.edu/mailman/listinfo/gretl-devel

On Mon, 16 Oct 2017, > oleg_komashko(a)ukr.net wrote: > Thank you, Allin > Also, does it contain > elementary functions > in complex argument e.g. > exp(), log(), etc as in libc? > I can't tell on the future, > what I want to do now > is to make gradient and Hessian > for real functions of real arguments > with imaginary steps inside: > e.g. for gradient: > Im(f(x+i*delta_x)-f(x))/delata_x > > It is very exact, but very slow > if it is not coded in c > Due to the nature of complex numbers > delta_x = $macheps for any real x > The precision does not depends on x > That's why it is so exact I fail to see how this approach can give more precise results than the mechanism we have now. ------------------------------------------------------- Riccardo (Jack) Lucchetti Dipartimento di Scienze Economiche e Sociali (DiSES) Università Politecnica delle Marche (formerly known as Università di Ancona) r.lucchetti@univpm.ithttp://www2.econ.univpm.it/servizi/hpp/lucchetti ------------------------------------------------------- _______________________________________________ Gretl-devel mailing list Gretl-devel@lists.wfu.eduhttp://lists.wfu.edu/mailman/listinfo/gretl-devel

See "The Complex-Step Derivative Approximation" in ACM Transactions on Mathematical Software, Vol. 29, No. 3, September 2003, Pages 245–262. and the Reference manual of the R numDeriv package (https://cran.r-project.org/web/packages/numDeriv/numDeriv.pdf) assuming that Gretl is not using the complex step method. And assuming that the functions to be differentiated can handle complex values.

On 16 Oct 2017, at 17:35, Allin Cottrell <cottrell(a)wfu.edu&gt; wrote: On Mon, 16 Oct 2017, Riccardo (Jack) Lucchetti wrote: > [T]he quality of the result is comparable to what we already get via Richardson extrapolation. The complex-step technique performs a second-order approximation, which is what we get by setting fdjac_quality to 2. For example, try the following script (which uses the definition of rderiv as per Oleh's previous script): > > <hansl> > set verbose off > > function scalar cube(matrix *x) > return x[1]^3 > end function > > g = rderiv("x^3", 0) > printf "rderiv: %36.33f\n", g > > x = {0} > > set fdjac_quality 1 > g = fdjac(x, cube(&x)) > printf "fdjac(1): %36.33f\n", g > > set fdjac_quality 2 > g = fdjac(x, cube(&x)) > printf "fdjac(2): %36.33f\n", g > </hansl> > > On my system I get > <output> > rderiv: -0.000000000000000000000000000000049 > fdjac(1): 0.000000000000000888178419700125430 > fdjac(2): 0.000000000000000000000000000000000 > </output> OK, but try Oleh's rderiv() examples: <hansl> d1 = rderiv("sin(x)",$pi/6) - cos($pi/6) x = {$pi/6} d2 = fdjac(x,(sin(x))) - cos($pi/6) printf "sin(x): d1 = %g, d2 = %g\n", d1, d2 d1 = rderiv("exp(2*x)",10) - 2*exp(20) x = {10} d2 = fdjac(x,exp(2*x)) - 2*exp(20) printf "exp(2x): d1 = %g, d2 = %g\n", d1, d2 d1 = rderiv("x+x^2+x^3",1) - (1 + 2 + 3) x = {1} d2 = fdjac(x,(x+x^2+x^3)) - (1 + 2 + 3) printf "cubic: d1 = %g, d2 = %g\n", d1, d2 d1 = rderiv("exp(x)",100) - exp(100) x = {100} f = fdjac(x, exp(x)) d2 = f - exp(100) printf "exp(100): d1 = %g, d2 = %g\n", d1, d2 printf " (proportional error = %g)\n", abs(f - exp(100)) / exp(100) d1 = rderiv("100+x",0.1) - 1 x = {0.1} d2 = fdjac(x,(100+x)) - 1 printf "100+x: d1 = %g, d2 = %g\n", d1, d2 d1 = rderiv("sin(x)",(20000+1/4)*$pi) - cos((20000+1/4)*$pi) x = {(20000+1/4)*$pi} d2 = fdjac(x, sin(x)) - cos((20000+1/4)*$pi) printf "sin(x): d1 = %g, d2 = %g\n", d1, d2 </hansl> With fdjac_quality 2 I get: <output> sin(x): d1 = 0, d2 = -2.88153e-11 exp(2x): d1 = 0, d2 = -0.0899086 cubic: d1 = 0, d2 = -2.70636e-10 exp(100): d1 = 0, d2 = -4.09506e+32 (proportional error = 1.52339e-11) 100+x: d1 = 0, d2 = -3.15016e-09 sin(x): d1 = 0, d2 = -1.72044e-07 </output> There's no doubt that the complex method gives greater accuracy. However, as you say, this method is limited to functions that accept complex arguments and produce complex return values.

Allin _______________________________________________ Gretl-devel mailing list Gretl-devel(a)lists.wfu.edu http://lists.wfu.edu/mailman/listinfo/gretl-devel