The recently invented thick control flow (TCF) model packs together an
unbounded number of fibers, thread-like computational entities, flowing
through the same control path. This promises to simplify parallel
programming by partially eliminating looping and artificial thread
arithmetics. In this paper we outline an architecture for efficiently
executing programs written for the TCF model. It features scalable
latency hiding via replication of instructions, radical synchronization
cost reduction via a wave-based synchronization mechanism, and improved
low-level parallelism exploitation via chaining of functional units.
Replication of instructions is supported by a dynamic
multithreading-like mechanism, which saves the fiber-wise data into
special replicated register blocks. The architecture facilitates
programmers with compact, unbounded notation of fibers and groups of
them together with strong synchronous shared memory algorithmics.
According to evaluations, the architecture is able to efficiently handle
workloads featuring computational elements with the same control flow,
independently of the number of elements. In its turn, this pays out as
improved performance and lower power consumption due to elimination of
redundant parts of computation and machinery.