Automatic End-to-End Joint Optimization for Kernel Compilation on DSPs
DescriptionDigital signal processors (DSPs) are extensively applied in more compute-heavy embedded sensing applications.
The compilation for DSP kernels is a static process, relied heavily on vendor-made compiler and optimization by experts.
Hand-crafted methods suffer from heavy burden on programmers, while state-of-the-art automatic compilation methods focus more on a certain aspect,requiring empirical adjustments by programmers in the actual scenario.

We propose JOKer, an automatic end-to-end multi-level joint optimizer for kernel compilation on DSPs.
JOKer integrates means of optimization at core steps in compiling process and automatically tunes compilation configurations for hiding the details of optimization workload and provides an end-to-end workflow.
Event Type
Research Manuscript
TimeThursday, July 13th3:40pm - 3:55pm PDT
Location3006, 3rd Floor
Embedded Systems
Embedded Software