Design Entry Clause Samples

Design Entry. Modern HLS tools like AutoESL and PICO (semi-)automatically leverage a wide range of compiler optimization techniques such as common subexpression elimina- tion and loop unrolling, and computer architecture techniques such as pipelining and resource sharing to improve cost and performance aspects of a design. For some of these techniques, the effectiveness is highly dependent on the structure of the appli- cation. Therefore, the decision when and how to apply a particular technique often has to be made by the designer. Some techniques can be applied or controlled with a tool pragma, while other techniques must be reflected in the way the algorithm is de- scribed. In this section we describe the techniques applied for the M-RVD QRD block of the sphere decoder application. In particular, we have applied a combination of time division multiplexing, loop unrolling, array partitioning, and case-specific opti- mizations. All of these techniques have been applied by modifying C code only such that a different architectural interpretation is obtained, while the functional interpre- tation is preserved. For designs without feedback, an HLS tool is generally able to instantiate regis- ters freely to increase clock frequency and throughput. However, in pipelines that are part of a feedback loop, registers cannot be inserted freely without introduc- ing pipeline stalls. Hence, feedback loops, also known as recurrences, in a design are the key limiter of throughput [Pap91]. For example, Figure 6.3 shows the high level structure of the 8×8 M-RVD QRD loop nest. Although there are several recur- rences in the application, the critical recurrence in this code occurs when the result We observe several characteristics of this design. First, the code accurately reflects the order in which data is processed in the reference design. Second, the TDM refac- toring is expressed entirely at the C code level. This means it can be seamlessly ported to any HLS tool that supports the used C constructs, such as multi-dimensional ar- rays. Third, the number of datasets to iterate over, that is, the TDM depth, cannot be determined without knowing the sizes of the critical recurrences. Although AutoESL does not compute the number of datasets automatically, the HLS process does an- alyze the source code for recurrences and reports to the designer where recurrences are not satisfied. The designer can use this information in a subsequent AutoESL run. In the sphere decoder application, since 3...