[30.08.2020] Report published on the final compilation infrastructure
MNEMOSENE has recently published a report describing the final version of the parallelizing compiler for the Computation-in-Memory (CIM) architecture developed in the project. Created by MNEMOSENE consortium partner INRIA, the compiler allows for the compilation of high-level representations of critical operations of target applications for the CIM architecture with a simplified and extensible compilation flow.
The end-to-end compilation flow and orchestration of accelerated functions with host-code and the CIM run-time were verified with multiple implementations of a HAR application executed on the project’s simulator.
The use of multi-level intermediate representation (MLIR) allows for a more robust identification of performance-critical operations for offloading, simplified transformations and a simplified code generation scheme based on common MLIR and Low Level Virtual Machine (LLVM) infrastructure.
To read the full report, please click on the following link.
The end-to-end compilation flow and orchestration of accelerated functions with host-code and the CIM run-time were verified with multiple implementations of a HAR application executed on the project’s simulator.
The use of multi-level intermediate representation (MLIR) allows for a more robust identification of performance-critical operations for offloading, simplified transformations and a simplified code generation scheme based on common MLIR and Low Level Virtual Machine (LLVM) infrastructure.
To read the full report, please click on the following link.