A model for fusion and code motion in an automatic parallelizing compiler

Abstract

Loop fusion has been studied extensively, but in a manner isolated from other transformations. This was mainly due to the lack of a powerful intermediate representation for application of compositions of high-level transformations. Fusion presents strong interactions with parallelism and locality. Currently, there exist no models to determine good fusion structures integrated with all components of an autoparallelizing compiler. This is also one of the reasons why all the benefits of optimization and automatic parallelization of long sequences of loop nests spanning hundreds of lines of code have never been explored.We present a fusion model in an integrated automatic parallelization framework that simultaneously optimizes for hardware prefetch stream buffer utilization, locality, and parallelism. Characterizing the legal space of fusion structures in the polyhedral compiler framework is not difficult. However, incorporating useful optimization criteria into such a legal space to pick good fusion structures is very hard. The model we propose captures utilization of hardware prefetch streams, loss of parallelism, as well as constraints imposed by privatization and code expansion into a single convex optimization space. The model scales very well to program sections spanning hundreds of lines of code. It has been implemented into the polyhedral pass of the IBM XL optimizing compiler. Experimental results demonstrate its effectiveness in finding good fusion structures for codes including SPEC benchmarks and large applications. An improvement ranging from 5% to nearly a factor of 2.75× is obtained over the current production compiler optimizer on these benchmarks.

PolyGLoT: A Polyhedral Loop Transformation Framework for a Graphical Dataflow Language

“…Finally, the proof for Theorem 5 also implies a loop-carried input for a particular variable access is created on a loop only when all the accesses to a variable inside the loop-nest are read accesses. Therefore, a flow path from a loopinvariant source vertex to a loop-carried input never exists, ensuring characteristic (6).…”

“…There are many compilers, both proprietary and open-source which now use the polyhedral compiler framework [12,15,6,16]. Research in this area, however, has predominantly focused on imperative languages such as C, C++, and Fortran.…”

Optimizing memory hierarchy allocation with loop transformations for high-level synthesis

“…The work in the first category focuses on improving data locality and date reuse by code transformation, especially loop transformation [2][3][4][5][6][7][8][9][10][11]. By changing the accessing order of array references in the loop nests, the co-located references become temporally "closer", which means a smaller data reuse buffer.…”

“…To improve data locality, iteration distances between dependent array instances are formulated in the objective function. Beside affine transformation, loop fusion/distribution, code motion and tiling for imperfectly nests loops have also been studied in recent work [8][9][10][11]. However, these models [2][3][4][5][6][7][8][9][10][11] use simple platform-independent objective functions, which cannot accurately model the impact of memory hierarchy allocation in hardware synthesis.…”

Code generation for parallel execution of a class of irregular loops on distributed memory systems

“…Automatic parallelization and locality optimization of affine loop nests have been addressed for shared-memory multiprocessors and GPUs with good success [4], [7], [8], [16], [29], [30]. However, many large-scale simulation applications must be executed in a distributed-memory environment, using irregular or sparse computations where the control-flow and array-access patterns are data-dependent.…”