Open Projects
LLVM Polly keeps here a list of open projects which each of themselves would
be a great contribution to Polly. All of these projects are meant to be self
contained and should take a newcomer around 3-4 months of work. The projects
we propose are all suitable as
Google Summer of
Code projects. In case you are interested in a Google Summer of code
project make sure to reach out via the Polly
mailing list early to
discuss your project proposal.
Integrate Polly with the LLVM vectorizers
Polly is not only a self-contained optimizer, but also provides a powerful
dependence and other program analyses. Currently, these analyses are only used
for our own optimizations. However, LLVM passes such as the loop vectorizer
would clearly benefit from having direct access to the available Polly
analyses. In this project, you would define in collaboration with the LLVM
community and considering existing dependence analysis interface a new
dependence analysis interface for Polly that allows passes to directly query
Polly analysis. Even though this project sounds straightforward at a first
glance, sorting out how to actually make this happen with the current and
the new pass managers, understanding how and when to invalidate the Polly
analysis and if dependence information can be computed on-demand make this
still a challenging project. If successful, this project may be a great way
to bring features of Polly to standard -O3 optimizations.
Register tiling to obtain fast BLAS kernels with Polly
Even though Polly is already able to speep up compute kernels significantly,
when comparing to the best BLAS routines we still are at least one order of
magnitude off. In this project you will investigate what is needed to close
this performance gap. Earlier investigations have shown that register tiling
is one important piece towards this goal. In combination with good tile size
models and some back-end work, this project is shooting to make common blas
operations, but also many non-blas kernels competitive with vendor math
libraries and outperforming the code icc/gcc currently generate.
Polly support for Julia - First steps
Julia is a new matlab style programming
language that provides C like performance for scientific computing. Even
though Julia also translates to LLVM-IR, parsing and optimizing Julia code
poses new challenges that currently prevent Polly from optimizing Julia
code despite the clear need for optimizations such as loop-tiling for Julia.
In this project you will -- starting from first proof-of-concept patches --
integrate Polly into Julia and ensure that Julia code can benefit from the
same high-level loop optimizations as todays C code already does. If time
permits, making Polly's recent bound-check elimination logic work in Julia
code would allow the optimization of Julia code, even if save out-of-bound
checking is used.
Interactive Polyhedral Web Calculator
At the core of Polly we use the isl math library. isl allows us to describe
loop transformations with relatively simple higher level operations while
still providing the full expressiveness of integer polyhedra. To understand
and describe the transformations we are performing it is often very convenient
to quickly script example transformations in a scripting language like python.
isl already comes with a python binding generator, with
pypyjs there is a python interpreter for the web and with emscriptem isl
itself can also be compiled to javascript. In this project you combine all
these components to obtain an interactive polyhedral web calculator, that uses
latest web technology to nicely illustrate the integer polyhedra you obtain.
