Section: New Results

Interfacing MAQAO and BOAST Frameworks for Kernel Autotuning on ARM Platforms

In Project MontBlanc 2's deliverable D5.11  [13] we presented the integration of STORM's MAQAO software (a binary-level code analysis framework) with BOAST (an automatic performance tuning framework for meta-programming and optimizing computing kernels) developed at LIG's NANOSIM. From source meta-kernels written in the RUBY language, BOAST generates multiple versions, in various target languages, optionally applying optimization transformations and strategies, and exploring the space of compiler flags combinations, to discover the most effective kernel tuning parameters. MAQAO offers a scriptable framework to disassemble kernel binaries, explore binary instruction flows, register-level data dependencies, program control structures, to patch, re-assemble and instrument kernel binaries for tracing data access patterns, and to process them from custom analyzers written in the LUA language. This integration work built on the complementarity of these two environments by enabling MAQAO to process binary kernels generated by BOAST, and lead developers in a guided tuning cycle.