Fusion of loops for parallelism and locality

Abstract

Loop fusion improves data locality and reduces synchronization in data-parallel applications. However, loop fusion is not always legal. Even when legal, fusion may introduce loop-carried dependences which prevent parallelism. In addition, performance losses result from cache conflicts in fused loops. In this paper, we present new techniques to: 1) allow fusion of loop nests in the presence of fusion-preventing dependences, 2) maintain parallelism and allow the parallel execution of fused loops with minimal synchronization, and 3) eliminate cache conflicts in fused loops. We describe algorithms for implementing these techniques in compilers. The techniques are evaluated on a 56-processor KSR2 multiprocessor and on a 16-processor Convex SPP-1000 multiprocessor. The results demonstrate performance improvements for both kernels and complete applications. The results also indicate that careful evaluation of the profitability of fusion is necessary as more processors are used.

A Survey of Loop Parallelization: Models, Approaches, and Recent Developments

“…This transformation is not only help for parallelization, but also contributes benefits to other aspects, such as reducing loop bound testing, promoting data locality etc [19]. Manjikian et al [20] presented a typical approach, which integrates both parallelism and memory locality optimization. Darte [19] discussed the complexity of loop transformation, which is help for knowing the ultimate potentialities of this method.…”

A Comparison of Compiler Tiling Algorithms

“…Mitchel et al study interactions between tiling for multiple objectives at once [16]. In addition to tiling, researchers working on locality optimizations have considered both computation-reordering transformations such as loop permutation [9,17,25] and loop fission/fusion [15,17]. Scalar replacement replaces array references with scalars, reducing the number of memory references if the compiler later puts the scalar in a register [3].…”

A Performance Study for Iterative Stencil Loops on GPUs with Ghost Zone Optimizations

“…However, these techniques do not consider the ghost zone technique that reduces the inter-tile communication. Although loop fusion [25] and time skewing [36] are able to generate tiles that can execute concurrently with improved locality, they cannot eliminate the communication between concurrent tiles if more than one stencil loops are fused into one tile. This enforces bulk-synchronous systems, such as NVIDIA's Tesla architecture, to frequently synchronizes computation among different PEs, which eventually penalizes performance.…”