Other Data Path Allocators

Before discussing the specifics of the EMUCS data path allocator, several other data path allocators will be described briefly. McFarland [McFarlandSS] provides a tutorial on High-Level Synthesis in which he defines two classes of data path allocators iterative/constructive and global allocation. Iterative/constructive techniques bind one element at a time, while global allocation techniques find simultaneous solutions to a number of bindings at one time. Examples of each of these techniques are described below. 

The CAMAD system [Peng86] views synthesis as a series of optimizations. Optimizations are iteratively applied to the data-flow to do such things as merging operators together so that they share hardware, or serializing parallel operators so that they can be merged. The choice of which optimizations to apply and the order in which to apply them is determined by hardware costs. 

In the ADAM system, functional units are chosen by MAHA [Parker86]. In this system, functional units are allocated concurrently with the creation of the schedule. Functional unit allocation falls into the category of iterative techniques. Functional units are allocated in a first-come-first-served basis, with operators sharing functional units when possible. MAHA allocates additional functional units as they become necessary to the implementation of the schedule. Registers are chosen later using a global allocation technique implemented by REAL [Kurdahi87], another component of the ADAM system. 

DAA [Kowalski84] is a rule-based expert system approach to synthesis. The choice of functional units, however, is determined algorithmically by BUD [McFarland86a], which estimates area and performance for several design options, and chooses the best. The DAA program chooses registers and an appropriate interconnection scheme for the design. 

---

A regular computation-intensive signal flow—as occurs in algebraic analysis, Altering, or format conversion— is combined with nested branches and multiple data-dependent loops. This explicit irregularity complicates not only the controller synthesis but also other synthesis tasks, Uke scheduling and data-path allocation. These tasks need to deal with control-flow hierarchy explicitly, which is an important aspect of our approach (see section 3). 

First, an ASAP schedule is constructed, assuming infinite resources, and one cycle per operation. Then optimizations are applied, moving operations to other control steps to reduce the maximum number of operations of each type in any one control step, and grouping operations into functional units so as to have a minimum number of functional units. Uien the scheduler traverses the control step schedule, passing the operations in each control step to the data path allocator. The data path allocator tries to bind those operations using heuristics if it fails, the scheduler tries to delay operations until later control steps, and if that also fails, the user is notified that the resource constraints should be increased. 

It is one implication of this approach that more care needs to be taken with the selection of landmark points than is presently the case. The allocation of a landmark halfway between two others, or in the middle of a surface path bounded by some curves, adds information to a data-set even if the landmark is located exactly where a spline driven by the remaining information would place it. It adds, precisely, the information that that location was observed with only digitizing error, so points nearby have a prediction error that is lower simply by virtue of the additional data. Where landmarks are widely spaced, the deformations predicted by, say, a thin-plate spline are far less reliable than where landmarks are precise regardless of the statistics of the landmark locations themselves. 

---