Accelerate data flow signal processing applications in heterogeneous CPU/GPU systems

摘要

The method comprises: forming a virtual tile cluster having tiles, the tile comprising a processor and memory of a CPU device and a GPU device, and the tile of a GPU device further comprising a sub-processor; Forming a virtual integrated memory that can be accessed by the CPU and GPU devices; Receiving a job; And assigning a task to a tile of the virtual tile cluster according to a predefined rule. When a job is assigned to a tile of a GPU device, the method includes: broadcasting the job to a subprocessor of the tile using a GPU shuffle instruction; And dividing the data for the job and allocating the divided data to the subprocessor-each subprocessor executes a codelet using each divided data. The task is executed by at least one tile of the virtual tile cluster.