Full runtime polyhedral optimizing loop transformations with the generation, instantiation, and scheduling of code‐bones

Abstract

Summary In this paper, we present a new runtime code generation technique for speculative loop optimization and parallelization. The main benefit of this technique, compared to previous approaches, is to enable advanced optimizing loop transformations at runtime with an acceptable time overhead. The loop transformations that may be applied are those handled by the polyhedral model. The proposed code generation strategy is based on the generation of “code‐bones” at compile‐time, which are parametrized code snippets either dedicated to speculation management or to computations of the original target program. These code‐bones are then instantiated and assembled at runtime to constitute the speculatively optimized code, as soon as an optimizing polyhedral transformation has been determined. Their granularity threshold is sufficient to apply any polyhedral transformation, while still enabling fast runtime code generation. This approach has been implemented in the speculative loop parallelizing framework APOLLO.

Memory Access Classification for Vertical Task Parallelism

“…The need of identification could be circumvented by using a framework such as Apollo [7] that allows to speculate on a particular pattern, and then to rollback if the expected pattern was not respected.…”

“…If not, a rollback to the previous correct state is performed and a sequential execution of the loop is carried out. For certain types of data accesses, namely affine or nearly affine functions over loop counts, polyhedral models can be used to provide race-free static task parallelism [6], and these approaches have also recently be used successfully to build runtime systems that are able to rollback execution if access violations occurred [7]. For programs with low effective parallelism, this can lead to a significant overhead due to rollback followed by sequential execution when many conflicts occur between threads.…”

ALPyNA: acceleration of loops in Python for novel architectures

“…Caamaño et al [4] describe the working of a runtime optimising polyhedral compiler called Apollo. To reduce complexity, they analyze small windows of LLVM-IR and statically generate variants of code for which dependences cannot be known until runtime.…”

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

“…Speculative code variants: Apollo [8] is a runtime optimizing polyhedral compiler. Polyhedral transformations are performed on small windows of LLVM-IR and variants for each transformation are generated with guard conditions.…”