ALPyNA: acceleration of loops in Python for novel architectures

Abstract

We present ALPyNA, an automatic loop parallelization framework for Python, which analyzes data dependences within nested loops and dynamically generates CUDA kernels for GPU execution. The ALPyNA system applies classical dependence analysis techniques to discover and exploit potential parallelism. The skeletal structure of the dependence graph is determined statically (if possible) or at runtime; this is combined with type and bounds information discovered at runtime, to auto-generate high-performance kernels for offload to GPU.We demonstrate speedups of up to 1000x relative to the native CPython interpreter across four array-intensive numerical Python benchmarks. Performance improvement is related to both iteration domain size and dependence graph complexity. Nevertheless, this approach promises to bring the benefits of manycore parallelism to application developers.

Python programmers have GPUs too: automatic Python loop parallelization with staged dependence analysis

“…This execution time includes the overhead of transferring data back and forth between host and GPU. A breakdown of analysis, compilation and execution times for four benchmarks is given in prior work [17] and shows that execution time for large domain sizes quickly dominates. Table 1 report minimum, maximum, and geometric mean speedup achieved by ALPyNA over the CPython interpreter.…”

“…We have already published a brief description and preliminary study of an earlier version of ALPyNA [17]. Since that report, we have extended the framework to handle a richer set of loop nest semantics.…”

“…The present paper gives detail regarding the motivation and implementation, as well as a more extensive empirical evaluation, using 12 benchmarks (cf. four in [17]); the majority of extra benchmarks require the richer set of Python features now supported in ALPyNA.…”

“…ALPyNA determines loop bounds statically by parsing the original code (Section 4.1) or dynamically using introspection within the context of the closure that represents the loop nest (Section 4.2). The bounds for each instantiation of a loop nest is determined by introspecting the evaluation of the bounds expression at runtime (as demonstrated in previous work [17]).…”

Pricing Python parallelism: a dynamic language cost model for heterogeneous platforms

“…Rather than require that developers have parallel programming expertise, our approach is to automatically parallelize loop nests in vanilla Python on GPUs. The ALPyNA framework has demonstrated significant reduction in runtimes of large and moderately complex loop nests [14,15] (Section 2). However offload overheads like kernel compilation and data transfer mean that offloading small loops is slower than CPU execution.…”