Quite interesting indeed. Thanks for sharing.
--Tim
On Tuesday, September 30, 2014 04:52:22 AM Hans W Borchers wrote:
> Sorry for reviving an old thread.
>
> I got interested in timings for vectorized vs. unvectorized versions of
> functions in MATLAB. So I wrote MATLAB versions of the `trapz` functions
> mentioned in this thread, and measured their running times by applying the
> `timeit` macro several hundred times. The results did surprise me:
>
> trapz1 (vectorized) : ~ 35 µs
> trapz2 (unvectorized): ~ 12 µs -- per loop
>
> for evaluating the trapezoidal rule on `x=linspace(0,pi); y=sin(x)`.
>
> I know one should not draw conclusions from such simple benchmarks. Still,
> does
> this comparison imply the old myth to "
> *vectorize MATLAB functions to make them faster*" is not true any more?
>
> The JIT compiler that MATLAB is equipped with is not as powerful as Julia's
> LLVM-based compiler, but it seems to be at work. Have others had similar
> experiences, or am I on a wrong path here?
>
> On Wednesday, April 23, 2014 11:20:50 PM UTC+2, Tomas Lycken wrote:
> > On Wednesday, April 23, 2014 11:10:15 PM UTC+2, Cameron McBride wrote:
> >> Or you can use the non-vectorized version and save the overhead of the
> >> temporary arrays being created by the addition and multiplication steps.
> >
> > There's really no way I can hide that I learnt scientific computing in
> > Matlab, is there? :P
> >
> >> On Wed, Apr 23, 2014 at 7:52 AM, Tomas Lycken <[email protected]>
> >>
> >> wrote:
> >>> The trapezoidal rule (http://en.wikipedia.org/wiki/Trapezoidal_rule)
> >>> would probably be almost trivial to implement.
> >>>
> >>> function trapz{T<:Real}(x::Vector{T}, y::Vector{T})
> >>>
> >>> if (length(y) != length(x))
> >>>
> >>> error("Vectors must be of same length")
> >>>
> >>> end
> >>> sum( (x[2:end] .- x[1:end-1]).*(y[2:end].+y[1:end-1]) ) / 2
> >>>
> >>> end
> >>>
> >>> x = [0:0.01:pi]
> >>> y = sin(x)
> >>>
> >>> trapz(x,y) # 1.9999820650436642
> >>>
> >>> This, of course, only currently works on vectors of real numbers, but
> >>> it's easy to extend it if you want.
> >>>
> >>> And there might be more accurate methods as well, of course (see e.g.
> >>> http://en.wikipedia.org/wiki/Simpson%27s_rule) but this one's often
> >>> "good enough".
> >>>
> >>> // T
> >>>
> >>> On Wednesday, April 23, 2014 8:43:48 AM UTC+2, Evgeny Shevchenko wrote:
> >>>> Hi, John.
> >>>> No, I didn't. I didn't find it and it seems to be not what i need:
> >>>>
> >>>> "no method quadgk(Array{Float64,1}, Array{Float64,1})"
> >>>>
> >>>> quadgk(f,a,b,...) expects a function as its first argument but I mean
> >>>> the case when y = f(x), but i don't have f, e.g. obtained experimental
> >>>> data, so x and y are 1-D arrays of floats.
> >>>>
> >>>>
> >>>> On Tue, Apr 22, 2014 at 7:49 PM, John Myles White <[email protected]
> >>>>
> >>>> > wrote:
> >>>>> Have you tried the quadgk function?
> >>>>>
> >>>>> -- John
> >>>>>
> >>>>> On Apr 22, 2014, at 7:32 AM, Evgeny Shevchenko <[email protected]> wrote:
> >>>>>
> >>>>> Hi
> >>>>>
> >>>>> Is there a package for numeric integration of obtained arrays, say x
> >>>>> and y? Couldn't find one, the led me to use @pyimport and numpy.trapz.
> >>>>>
> >>>>> --
> >>>>> pupoque@IRC
