A Compiler for Parallel Exeuction of Numerical Python Programs on ...

the element and its 4 neighbors and writes the result to the corresponding element in amatrix <strong>of</strong> the same dimensions. The kernel is highly parallel as each element is processedindependently but the amount <strong>of</strong> computation per point is very small. The compiler wasable to successfully carry out the array access analysis and was able to generate the GPUcode <strong>for</strong> this benchmark. However, the data transfer overhead considerably outweighed thecomputation time savings <strong>of</strong> the GPU. There<strong>for</strong>e, the benchmark ran slower when usinga GPU compared to OpenMP. The results are reported in Table 7.9. For each case, theminimum, maximum and mean execution times are presented.Table 7.9: Execution time <strong>for</strong> 5-point stencil benchmark (milliseconds)Problem size 1024 2048 3072 4096Serial Time min 10.8 43.3 97.5 175max 10.9 43.7 98.2 176.6mean 10.8 43.4 97.7 175.8OpenMP 1 thread min 10.1 46 70 150max 10.26 47 71.24 152mean 10.17 46.3 70.4 151OpenMP 4 threads min 5 24 47.3 58.1max 5 24 47.8 58.5mean 5 24 47.5 58.2GPU Total (Opt) min 65 98.8 123.1 1010max 65.8 99.6 124.9 1016mean 65.3 99.1 123.7 1014GPU Only (Opt) min 0.39 0.89 2.8 25max 0.39 0.89 2.8 26mean 0.39 0.89 2.8 25.5GPU Total (No Opt) min 35.3 63.1 112.1 1000.3max 35.6 63.7 112.7 1006mean 35.4 63.25 112.3 1002GPU Only (No Opt) min 0.59 2.0 4.2 50max 0.59 2.0 4.2 52mean 0.59 2.0 4.2 517.5 RPES benchmarkRPES benchmark is a <strong>Python</strong> adaption <strong>of</strong> the benchmark from Parboil suite. The benchmarkinvolves indirect memory loads and triangular loops. Un<strong>Python</strong> determined that aGPU version cannot be generated and there<strong>for</strong>e did not generate calls to the JIT compiler.There<strong>for</strong>e, the per<strong>for</strong>mance did not change when the GPU was enabled because theGPU was never used and the JIT compiler was not called. The compiler only generatedan OpenMP version <strong>of</strong> the benchmark. This benchmark illustrates that the programmercan safely add GPU parallel annotations without any fear <strong>of</strong> errors in the case <strong>of</strong> compilerlimitations.The benchmark was only tested with default parameters provided by the benchmark58