Optimizing Data Movement - 2021.2 English

Vitis Unified Software Platform Documentation: Application Acceleration Development (UG1393)

Document ID
English (United States)
Release Date
2021.2 English
Figure 1. Optimizing Data Movement Flow

In the OpenCL execution model, all data is transferred from the host main memory to the global device memory first, and then from the global device memory to the kernel for computation. The computation results are written back from the kernel to the global device memory, and lastly from the global memory to the host main memory. A key factor in determining strategies for kernel data movement optimization is understanding how data can be efficiently moved around between different level of memories maximizing the efficient use of bandwidth on all the memory interfaces.

During data movement optimization, it is important to isolate data transfer code from computation code because inefficiency in computation might cause stalls in data movement. You should focus on modifying the data transfer logic in the host and kernel code during this optimization step. The goal is to maximize the system level data throughput by maximizing data transfer bandwidth and device global memory bandwidth usage. It usually takes multiple iterations of running software emulation, hardware emulation, as well as execution in hardware to achieve optimum performance.