TAPA Basics

This section introduces the basic concepts of a TAPA program and the major steps of the TAPA compilation process

TAPA Program

TAPA dataflow programs decouple communication and computation. A TAPA program has two types of building blocks: streams and tasks.

• A stream is essentially a FIFO

• A task consumes input data from some streams, perform computation, then produces output data to some other streams.

• All tasks execute in parallel and communicate with each other through streams.

TAPA Compilation Process

The TAPA compilation process ultimately involves two steps.

• First, TAPA extracts each task and synthesize it independently using an existing HLS compilers (e.g., Vitis HLS).

• Second, TAPA generates optimized logic to stitch the RTL of each task together into the final accelerator.

Magic of TAPA

The core idea of TAPA is to synthesize each task using existing HLS tools and generate our own customized RTL to compose the tasks. This allows us to stand on the shoulder of existing HLS tools, take advantage of them and go higher.

• Speed. Since we explicitly decouples communication and computation, and then synthesize each task in parallel, we gain significant speedup compared to conventional HLS tools. In addition, we design both software and hardware simulator specifically optimized for our programming model and thus beat universal simulation tools.

• High Frequency. We add layout-aware optimization when generating glue logic to compose the individual tasks. Specifically, we pre-determine the approximate placement location of each task, and properly pipeline the communication structures between tasks based on their locations.

• Expressiveness. Since we generate all RTL for communication structures, we are free to define additional communication APIs. So far, we have implemented a set of highly flexbile APIs for stream interaction and external memory access, which are unavailable in other HLS tools.