Updates due to JuliaDiff/TaylorSeries.jl#361

[PerezHz]

This PR contains updates which go together with JuliaDiff/TaylorSeries.jl#361

EDIT: this PR depends also on JuliaDiff/TaylorSeries.jl#364

[PerezHz]

Tests were actually passing modulo coverage, except for the last commit, but cannot reproduce the issue locally, so... I'm re-running tests and see if it's maybe a CI issue

[PerezHz]

After re-running tests, they are passing except for Julia 1.6 on the macos platform... does look like a CI issue?

[lbenet]

After re-running tests, they are passing except for Julia 1.6 on the macos platform... does look like a CI issue?

Also windows in Julia 1.6 is failing. I doubt it is a CI issue, it should be deterministic, right?

[PerezHz]

Also windows in Julia 1.6 is failing. I doubt it is a CI issue, it should be deterministic, right?

So you're saying maybe it's an initialization issue?

[PerezHz]

Windows and MacOS failures on Julia 1.6 look pretty consistent, whereas all the other versions and platforms tests passing also look consistent. Probably it's then a sqrt issue which is somehow only visible on Julia 1.6 on Windows and MacOS but not on Ubuntu?

[PerezHz]

Update: managed to reproduce the error on julia 1.6 (macos)

[PerezHz]

Turns out the issue in julia1.6 can be traced back to rounding off issues in powers of Float64:

julia> 0.2029886550675951^3 # integer power
 0.008364024538422981

julia> 0.2029886550675951^3.0 # float power
 0.00836402453842298

[PerezHz]

Pushed a fix to TaylorSeries...

[PerezHz]

... et voilà!

[PerezHz]

In a nutshell, since some allocations were incurred in the preamble of parsed jetcoeffs! methods, in one of the last commits I removed the preamble in favor of using mutating methods for 0-th order, modifying the recursion loop accordingly. For power operations such as x = r^3 the preamble did a float conversion x = Taylor1(constant_term(r) ^ float(constant_term(3)), order). The equivalent conversion in pow! was missing, and hence different results were obtained by parsed and non-parsed jetcoeffs! methods for Julia 1.6 (and my guess is that this issue somehow did not make it to the official Linux x64 binaries, so the issue did not appear there).

[lbenet]

In a nutshell, since some allocations were incurred in the preamble of parsed jetcoeffs! methods, in one of the last commits I removed the preamble in favor of using mutating methods for 0-th order, modifying the recursion loop accordingly. For power operations such as x = r^3 the preamble did a float conversion x = Taylor1(constant_term(r) ^ float(constant_term(3)), order). The equivalent conversion in pow! was missing, and hence different results were obtained by parsed and non-parsed jetcoeffs! methods for Julia 1.6 (and my guess is that this issue somehow did not make it to the official Linux x64 binaries, so the issue did not appear there).

Thanks for checking and solving this!

[lbenet]

Registration of Taylor series v0.18.0 is on its way... 10-15 minutes are required 😄

[lbenet]

v0.18.0 is released!

[PerezHz]

v0.18.0 is released!

Many thanks! I'll re-trigger tests here!

[PerezHz]

With TaylorSeries v0.18.0 released this is now ready for review

[lbenet]

With TaylorSeries v0.18.0 released this is now ready for review

I'll try to review it this afternoon... sorry for the delay.

[lbenet]

Thanks a lot @PerezHz. My only suggestion is to bump a new (minor or patch?) version. Aside from that, LGTM!

[PerezHz]

Thanks a lot @PerezHz. My only suggestion is to bump a new (minor or patch?) version. Aside from that, LGTM!

Alright, bumped the patch version and will merge afterwards, thanks!