Sumith : Optimized Awesomeness

GSoC Progress - Week 2


Hello, this post contains the second report of my GSoC progress. At one point, I had changed the deadline from Sundays to Fridays, but I seem to be running a week late on the post names. Will be corrected next week.

Progress

We decided that instead of coding up the Polynomial upfront, we try to speed up the expand2b benchmark i.e. try to nail speed at a lower level, then think of the design decisions and wrap it up into Polynomial class.
The plan was this:
* Add support for Piranha in SymEngine CMake
* Implement packing of exponents and a check function to ensure it fits, use this to fasten expand2b
* Use Piranha's integer and benchmark expand2b again

The faster hashtable was kept for later.

Report

PR 464 was merged.
Implements support for Piranha in CMake along with it's dependencies Boost and PTHREAD.
The above two dependencies come as separate CMake option as well. We feel that the Boost support can be improved, that can be done at a later stage.

PR 470 Speeding up the benchmark.
* The pack and check function were implemented
* Used std::valarray instead std::vector(inspired by issue 111) but the benchmark slowed down, hence change was not adopted
* Implemented functions poly2packed() and packed2poly(), for converting between the two representations
* Implemented function poly_mul2() for multiplying the packed polynomials
* Re-wrote expand2b to use packed, now expand2c

Very nice speedup was obtained from using the packed structure, a more detailed report can be found here.
But we are still far from Piranha and we have lots to do :)

Most of the week's time went to learning to link libraries and writing cmake files for my own projects so that I could figure what was happening in PR 464. Now I feel it was very easy a task and shouldn't have consumed the time it did.

Targets for Week 3

My aim is to get all the code and optimization, possible at this level, done by next week, so that we can start wrapping in the coming weeks.
* Use piranha::integer for coefficients, benchmark
* Implement switching between packed structure and tuple depending on whether exponents fit or not

If time permits, I want to implement a system to use int for small cofficients and switch to mpz_class when large.
We have lots to do to hit the speed that we expect.

That's all folks.
Au Revoir