Scattering OMP parallelisation #243
Conversation
src/analysis.cpp
Outdated
| break; | ||
| } | ||
| } | ||
|
|
||
| ScatteringFunction::ScatteringFunction(const json &j, Space &spc) try : spc(spc) { | ||
| from_json(j); | ||
| name = "scatter"; | ||
| usecom = j.value("com", true); | ||
| use_com = j.value("com", true); | ||
| save_after_sample = j.value("stepsave", false); // save to disk for each sample |
There was a problem hiding this comment.
How come that this could come like this from the pull request #241? :-D Done. Documentation shall also contain ipbc. As ipbc has no meaning for debye scheme, shall we keep it as an independent attribute, or shall we have three different schemes?
There was a problem hiding this comment.
yes, debye is quite different and also requres dq, qmin, qmx while these are automatic w. explicit and ipbc that om the other hand require qmax. I would keep ipbc as an option in the explicit scheme and split debye to a separate analysis.
src/scatter.h
Outdated
| } | ||
| // as of January 2020 the std::transform_reduce is not implemented in libc++ | ||
| T sum_cos = 0; | ||
| for (auto &qr : qr_products) { |
There was a problem hiding this comment.
How well does Eigen fare for this operation? It's supposed to handle vectorisation on a number of architectures. I believe the operation here would by qr_products.sin().sum(). Note also that for the coming Eigen 4, iterators are supported.
There was a problem hiding this comment.
Btw. your sin/cos optimization may be applicable also for Ewald summation:
Line 127 in 58adfe9
There was a problem hiding this comment.
Update: on GCC9/MacOS, the separate cos/sin (ifdef GNUC) are slower
There was a problem hiding this comment.
The following gives a factor four speed-up on GCC and Clang:
Eigen::VectorXd qr_products;
qr_products.noalias() = positions * q;
T sum_sin = qr_products.array().sin().sum();
T sum_cos = qr_products.array().cos().sum();where positions must by a MatrixXd(N,3) object. Best to await a unit-test to see if it's doing the right thing, though :-)
Create a merge commit is probably the best strategy how to merge this pull request. |
OpenMP is enabled for loops in the sample methods that have O(N^2) complexity, where N is number of particles. Furthermore, the sample methods are rewritten to allow automatic vectorisation of trigonometric functions in GCC by using libmvec vector math library when --ffast-math is enabled. This speeds up the analysis by factor of 4 on modern processors.
Different implementations exhibit very different performance (factor of 5 observed) depending on the compiler and system environment. Add unit and performance tests.
Description
Refactor and paralellise scattering analysis
OpenMP is enabled for loops in the sample methods that have O(N^2) complexity, where N is number of particles.
Furthermore, the sample methods are rewritten to allow automatic vectorisation of trigonometric functions in GCC by using
libmvecvector math library when--ffast-mathis enabled. This speeds up the analysis by factor of 4 on modern processors.Checklist
Note
The pull request includes and thus obsoletes the pull request #241.