[llvm-dev] [GSoC 2016] Improvement of vectorization process in Polly - Mid term report

Tobias Grosser via llvm-dev llvm-dev at lists.llvm.org
Wed Jun 22 10:13:30 PDT 2016


On 06/21/2016 09:34 AM, Roman Gareev wrote:
> Dear LLVM contributors,
> 
> Over the last weeks, I have been working on implementation of tilings
> and interchangings of specific loops based on the algorithm used to
> optimize matrix multiplication and described in "Analytical Modeling
> is Enough for High Performance BLIS" [1]. It corresponds to the
> original timeline.
> 
> We outlined the following steps:
> 
> 1. Determination of statements that contain matrix multiplication (a
> corresponding patch [2] is accepted)
> 
> 2. Creation of a micro-kernel (a corresponding patch [3] is accepted)
> 
> 3. Creation of a macro-kernel (a corresponding patch [4] is under review)
> 
> For further details, please see my blog post
> http://romangareev.blogspot.ru/2016/06/gsoc-2016-report-i.html
> 
> During the vacation stated in the timeline, I will work on completion
> of the step 3.  The next steps are implementation of the packing
> transformation and implementation of a determination of vectorization
> factors and profitability of vectorization based on the LLVM
> vectorization cost model.

Thank you Roman for the update and the nice writeup. I am looking
forward to see progress on the packing transformation and the cost
model.

Best,
Tobias


More information about the llvm-dev mailing list