Buffer-accessor model vs unified shared memory

Questions

  • When should I use the buffer-accessor model and when is USM more appropriate?

Objectives

  • Learn the trade-offs of the buffer-accessor and USM models.

In episodes Data management with buffers and accessors and Data management with unified shared memory, we saw what the current SYCL standard has to offer in terms of data management models. Some of the exercises showed how to implement the same task using both models. What is the “best” strategy? The answer is, unsurprinsingly, “it depends”. The question should rather be: would you like the runtime to manage data transfers or do you want to be fully in control? The following chart offers a guide to inform a decision between the two models.

../_images/data-management-strategy.svg

Choosing between the USM and buffer-accessor model in SYCL. The fundamental question is how comfortable we are with the runtime managing data movement for us.

The chart does not take into consideration whether or not we are working with an established codebase. If that is the case, USM, which is a pointer-based approach, might provide an easier porting path to SYCL.

Note

Why can’t we use both models in the same application? The current standard does not describe how to achieve interoperability between buffers and USM-allocated memory. This might lead to unnecessary copies and degraded performance and the even more unfortunate consequence that, once a data management model is chosen, the application is effectively locked into it forever. Future SYCL standards might remedy this situation. In the meantime, hipSYCL does offer buffer-USM interoperability as an extension to the standard: https://hipsycl.github.io/hipsycl/extension/hipsycl-091-buffer-usm-interop/

Porting the heat equation mini-app to SYCL

As a last assignment for this workshop, we’ll re-write our port of the heat equation mini-app in ref:heat-equation using USM.

Todo

WRITE THE USM PORT!