Stream Initialization

A stream in a DFG must have certain characteristics specified by the algorithm: the data type of the stream element, and the dimensions explicitly spanned by the stream.

Figure 2: Example Stream Initializations

A fragment of a stream may also be specified as having a pre-defined value -- one known when the program was written. This is used to define the constants and stream initial conditions used by an algorithm. Some examples are shown in Fig. 2, using the Q language described in Appendix A².

The first example declaration is for a stream, One, with no defined dimensions -- a scalar, or a single stream element with a defined value of one (and a non-existent location). It will have a value of one for all stream elements in a stream fragment demanded of it. The second declaration, Line, is for a stream with two dimensions in which all stream elements with a position of zero along the x dimension are defined and have a value of one. The third one, Plate, specifies a four dimensional stream, with a two dimensional boundary condition specified in a three dimensional space. Pre-defined stream fragments will be ``replicated indefinitely'' as needed along all stream dimensions not specified in their definition (such as y in Line, or t in Plate) by a runtime mechanism.

Footnotes

... A²

There are more elegant ways of declaring these stream initializations in the Q language. Fig. 2 illustrates the canonical method actually defined by the instruction set architecture.