Hi Alexei and Richy, Richard B. Kreckel schrieb:
Alexei Sheplyakov wrote:
The following transformations are done automagically now: (-x+y)*(x-y) -> -(x-y)^2 (3*x+y)/(6*x+2*y)^2 -> 1/4*(y+3*x)^(-1)
Please note: exam_series13() from check/exam_pseries.cpp fails due to this patch, because the expression 1/x*(-(1+x)/(1-x)) + (1+x)/x/(1-x) gets evaluated to zero.
This patch appears to come with a slight performance hit. On my machine, 14x14 Vandermonde matrix determinants degrades from 19s to 32s, Lewis-Wester test O1 from 7.5s to 11s, and Lewis-Wester test A from 0.055s to 0.22s. (I'm seriously surprised by the latter since it only involves integers, but it's reproducible.)
Looking at the code it appears to me like mul::eval recurses now. Wouldn't it be faster to collect the coefficients in one sweep? Well, I recognize that most terms will be trivial in the following recursions, but still.
Also, you write that "in some very unlucky event it can even loop forever". I don't quite understand. Do you have an example for this?
something new here? I kind of feel like doing a release soon, 1.3.8 and also maybe then a 1.4, but this seems to be a sort of issue to be solved first. Regards, Jens