Add 5 arguments mul!

[amontoison]

No description provided.

[coveralls]

Coverage increased (+0.02%) to 96.349% when pulling 8459f58 on amontoison:mul! into 174c648 on JuliaSmoothOptimizers:master.

[codecov]

Codecov Report

Merging #146 into master will increase coverage by 0.02%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #146      +/-   ##
==========================================
+ Coverage   96.32%   96.34%   +0.02%     
==========================================
  Files          14       14              
  Lines         625      630       +5     
==========================================
+ Hits          602      607       +5     
  Misses         23       23              

Impacted Files	Coverage Δ
src/operations.jl	100.00% <100.00%> (ø)
src/lsr1.jl	98.55% <0.00%> (+0.04%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 560071b...8459f58. Read the comment docs.

[abelsiqueira]

Is this a 1.3 thing? If so, we should also update the version because we drop support for Julia. I'm just not sure if it should be major or minor.

[amontoison]

Yes, it's a feature of Julia 1.3. I think release 1.1 for LinearOperators is appropriate, no?

[dpo]

Does this assume that op Is able to directly update y? That's how it works with matrices I think.

[amontoison]

No, one vector is allocated (or already preallocated) for op * x product and after y is updated.

[dpo]

Maybe operators should be defined by the 5-arg mul!. What do you think?

[amontoison]

Yes, it's a good idea.

[dpo]

Frankly, I have to say that I strongly dislike the mul!() notation. The first and foremost reason why this package was created is to avoid such notation and be able to simply write A * x. If this notation propagates into Krylov.jl, it defeats the entire purpose.

[abelsiqueira]

I think it's just for compatibility in other packages, in case someone decides to use it. Since matrices will support them, we might as well have it.
We should prevent using it in our packages, especially because It doesn't seem to improve speed in our case.
By the way, I just realized that we don't need to drop < 1.3, since defining 5-arg mul! shouldn't break anything, right @amontoison?

[dpo]

Compatibility with what though? Has there been demand for this? If it still allocates a vector, what is the advantage?

[amontoison]

Yes, you're right Abel, we don't need to drop < 1.3. I think it's still relevant to have it like the 3-arguments mul!. I have something in mind for Krylov.jl that use it without it's explicit form with a macro.

[dpo]

I continue to think that it would be better to define operators with a 5-arg prod!() instead of the prod() that we are using now, so that we can simply write A * x. No macro needed.

[dpo]

Let's put this on hold temporarily.

[amontoison]

superseded by #180