These are chat archives for symengine/symengine

Nov 2015
Ralf Stephan
Nov 05 2015 15:43

@certik , multiplying series (cos*sin) is fastest with Flint, see below:

                                N=100   N=1000      

flint fmpq_poly_mul()            0.27       88
flint fmpq_poly_mullow()         0.29       91
piranha psin*pcos trunc.         0.69      690
piranha psin*pcos no trunc.      5.0      4670
pari sin(x)+cos(x)              11        1400


*that is pari sin times cos, not add
Ondřej Čertík
Nov 05 2015 17:53
@rwst great job! This is awesome. How do you like benchpress?
I started playing with Nonius in #634, but I don't like its dependence on Boost.
@bluescarni any ideas why Piranha is so slow?
I noticed at, that it should be better to use the kronecker monomial, wouldn't it?
Ralf Stephan
Nov 05 2015 18:32
@certik , benchpress needs some love, I had to fiddle quite a bit and had to install gcc-4.9, and then it still would only work the way I found myself, not as advertised. But I like its lightweightness. But you're right I should test the kronecker monomial.
Ondřej Čertík
Nov 05 2015 18:37
@rwst another problem is with the length N. What exactly does fmpq_poly_sin_series(x, x, N) do?

In Piranha, the following code

    for (unsigned int i=0; i<100; i++) {
        const short j = 2*i + 1;
        if (i != 0)
            prod *= 1-j;
        prod *= j;
        psin += rational{1,prod} * x.pow(j);

returns a series in sin(x)sin(x) up to O(x199)O(x^{199}), but in Flint you might be only doing series up to O(x99)O(x^{99}).

Ralf Stephan
Nov 05 2015 19:10
Ah that is true!
Ondřej Čertík
Nov 05 2015 19:35
@rwst would you mind rerunning the benchmarks with the correct series size? When we say length N, let's just make it mean O(xN)O(x^N), I think that's the most intuitive.
I just submitted a benchmark when double precision coefficients are used: rwst/series-benchmark#1
And it runs at about 88% of theoretical peak performance on my computer.
Ondřej Čertík
Nov 05 2015 19:47
@bluescarni we've discussed this already in private email, but since others might also be interested in this: what's nice is that in Piranha, you can also switch to double precision coefficients, and thus comparing against my Fortran code then tells you the overall slowdown of the Piranha's machinery, compared to the theoretical maximum (that can be essentially attained with the above code). And double precision coefficients might actually be useful in some applications.