Implementing a big.LITTLE architecture with architecturally diverse cores
Bryce Snell · 2024
In processors with heterogeneous instruction sets, an unsupported instruction forces a thread, function, or process to be moved to a core that can support the instruction or the unsupported instruction must be emulated through software. Current methods greatly reduce performance due to slow algorithms, lost work, many memory operations, or untrained caches. This work proposes a new method that allows an unsupported instruction to be directly inserted into the pipeline of the processor with proper support. Using various workloads this paper demonstrates that such a solution can provide significant execution time improvements over both emulation and thread migration solutions. Furthermore, this design provided a second advantage by demonstrating reductions in the total energy required to run these workloads.