qualcomm/fastrpc

FastRPC is a userspace library that enables remote procedure calls between a CPU and DSP for offloading compute-intensive tasks to specialized processors.

The tool solves the problem of efficiently executing algorithms on digital signal processors, which can perform certain workloads more compute- and power-efficiently than CPUs through specialized resources like HVX. FastRPC hides the complexity of cross-processor communication by using an IDL-based interface definition approach. Developers define function interfaces in IDL files, which the QAIC compiler processes to generate marshaling code (stubs on the CPU side and skeletons on the DSP side) along with necessary headers. The framework handles parameter serialization, kernel-level transport via rpmsg, and remote invocation dispatch, allowing applications to call DSP functions as if they were local procedures.

FastRPC is suited for teams developing on Qualcomm platforms who have algorithms benefiting from DSP execution, particularly those leveraging HVX vector processing capabilities. The framework is most appropriate for applications where offloading large processing tasks to the DSP provides meaningful performance or power gains. It requires integration into both CPU and DSP build pipelines, with stubs and headers linked into CPU executables and skeletons and headers linked into DSP libraries. The tool is specific to Qualcomm's heterogeneous processor architecture and is not a general-purpose RPC framework.

The project maintains active engagement with its codebase through regular updates and demonstrates responsiveness to issues and pull requests. Development activity shows consistent attention to bug fixes and feature improvements across the FastRPC stack components. The maintainers provide clear documentation of the architecture and workflow, indicating a commitment to helping developers understand the framework's internals. The project accepts external contributions and maintains organized issue tracking, suggesting an open development model.