Summation based benchmark for calculators

[pier4r]

(12-23-2017 01:56 PM)AlexFekken Wrote:  I did not just mean "calling the same functions", but "calling the same functions with almost exactly the same (perhaps even the same) argument". I would say that any proper benchmark should call whatever functions it calls with a decent spread of the inputs. Using atan for (mainly) large values and then passing the almost constant result into sin (near an extremum!) pretty much garantees the exact opposite of that.

Just my 2c of course...

You also have a point. But for the objective of the test (not much accuracy, rather timing) I would suppose that whatever spread of the inputs would be more or less similar. If you have better compositions of functions (trig, exp/ln/, power, add/sub/mul/div ) let me know! It may be an interesting input to learn a thing or two.

At first I was exploring the hyperbolic functions (also because I never used them) but those may (a) be not so commonly implemented and (b) less often used.

Edit. Meanwhile I forced the sharp 506w to do a similar tasks like the summation using the integral calculation, although the result is off due to fixed coefficents.

The integral calculator in the sharp el-506w is done with the Simpson's rule:


This means that with a careful choice of the input parameters I can hope that the function is evaluated almost exactly in the 1000 points of the summation written above.
In particular:
a = 1
b = 1000
n = 500
N = 1000
h = 0.999
(if someone finds better parameters, please tell me!)

I do not know how the summation is done, but I assume that there are two internal registers for the sum of the "even" steps and the "odd" steps that then gets multiplied by 4 and 2. The computation is done in 4m 58s .