hpicgs / bhtsne Goto Github PK

use intrinsics for more efficient processing of vectorizable for loops

[--algorithmParameter value]* [-outputOption]* [--outputFile value] fileName
The order of parameters and options is not relevant.

• copy and adjust argumentParser from cppassist
• implement main
• test

refactor C-style C++ to modern C++ using features from newer standards

• tsne
• sptree
• vptree

write content or remove cmake target

<param> and set<param> getter and setter for every algorithm parameter
load<fileExtension> methods for each supported extension (e.g. legacyDat, csv, our binary)
run method; or better name
save<fileExtension> methods for each supported extension (e.g. svg, csv, [legacy]binary)

TODO:

• write tests
• implement interface
• write comments

• readme
• doxygen interface documentation
• latex manual? see #15
• improve perplexity explanation

add the data sets again but ignore them for the launchpad package

Many phases/steps from the algorithm perform an operation on all dimensions of a data point.
Using a library such as Vc, automatic vectorization is enforced.

enable package managers like apt/vcpkg/homebrew to simply install this project

The goal is to store at leat n * (n-1)/2 values and compute them at most.

Example:

• TSNE::computeSquaredEuclideanDistance

implement time measurement, so performance before and after refactoring/optimization can be compared

• measure legacy implementation
• measure current implementation

enable simple cross-platform setup for Windows, macOS and Ubuntu by creating a CMake project

figure out, whether using the gpu can improve performance, and if it does, implement that

E.g., by bypassing C++ std::getline and use the more C-like interface std::strtod (example: https://github.com/cginternals/unifiedmemory-demo/blob/master/source/unifiedmemory/Computation.cpp#L216).
The trade-off is the missing error handling, which you lack nonetheless.
Further, the parsed data point can be moved to the target row.

instead of passing an implicit sparse matrix around in 3 seperate vectors

Example:

• TSNE::computeGaussianPerplexity the three inner loops computing cur_P, sum_P, and H may get consolidated and vectorized.

All vectors shoould use aligned allocators and manual SSE / AVX code should use aligned load memory instructions.

add commandline option to write outData to disk on every n-th iteration, so "data developement" can be seen

To compute the new m_centerOfMass, a number of potential precision-losing operations are performed.

figure out how the algorithm can be parallelized and implement that

Example:

• TSNE::computeGaussianPerplexity while initializing obj_X

write tests, so mistakes are easily caught when refactoring and optimizing