On the Limits of Automatic Processor Specialisation by Mapping Dataflow Sections on Ad-hoc Functional Units

Many available systems-on-chip embedded processors can be specialised for a given application-domain by adding ad-hoc functional units. These functional units can be used to map clusters of elementary arithmetic or logic operations (sections of the dataflow graph); more complex hardware addons could attack the control flow (e.g., map loops onto sequential functional units). For automatic processor specialisation, the two strategies imply different methods and have a different complexity. Before incurring in the complexities of control flow, this paper attacks the question that appears so far unanswered: how much can one improve the performance of an embedded processor on specific algorithms by automatically mapping only dataflow sections of code on special functional units? The basic scope for performance improvement is assessed and broken out in different sources: hardware parallelism; avoided quantisation of instructions in an integral number of cycles; and simplification of the logic due to constants. The scope for speedup is increased through additional manual optimisations (amenable to automation): bit-width analysis and arithmetic optimisation. Finally, ILP techniques such as loop unrolling and predication are used to increase the size of the basic blocks and give more scope for the sources of improvement. The results show that significant improvements in speed (up to 6 times) can be targeted without necessarily mapping the control flow onto hardware.