|
|
|
| e-Pub |
Previous | 
Home
Section: New Results
The XFOR Programming Structure
Participants : Imen Fassi, Philippe Clauss, Cédric Bastoul.
We have proposed a new programming control structure called “xfor” or “multifor”, providing users a way to schedule explicitly the statements of a loop nest, and take advantage of optimization and parallelization opportunities that are not easily attainable using the standard programming structures, or using automatic optimizing compilers [19] . This is the PhD work of Imen Fassi, who started her work in 2013 and who defended her thesis November the 27th, 2015 [12] .
It has been shown that xfor programs often reach better performance than programs optimized by fully automatic polyhedral compilers like Pluto [29] . It has also been shown that different versions of codes may perform very differently, although their memory behaviors are very similar. By analyzing further the origins of such performance differences, we noticed five important gaps in the currently adopted and well-established code optimization strategies [18] , [19] : insufficient data locality optimization, excess of conditional branches in the generated code, too verbose code with too many machine instructions, data locality optimization resulting in processor stalls, and finally missed vectorization opportunities.
To ease and extend the usage of the XFOR structure, we have developed:
-
Xfor-Wizard, which is a programming environment for XFOR programs, assisting users in writing XFOR codes and applying optimizing transformations. Automatic dependence analysis and comparisons against a referential code (XFOR-loops or classic for-loops) are achieved to order to help the user in ensuring semantic correctness of the written code.
-
XFORGEN, which is a tool to automatically generate an XFOR code that is equivalent to for-loops that have been automatically transformed using a static polyhedral compiler. The generated XFOR code exhibits the parameters of the transformations that have been applied and thus can be modified for further optimizations.