Section: New Results

Multithreaded code generation for process networks

Participants : Loïc Besnard, Thierry Gautier.

As part of an in-depth comparison of process models, we have recently revisited the relation between the model of asynchronous dataflow represented by Kahn Process Networks (KPNs) and that of synchronous dataflow represented by the polychronous model of computation. In particular, we have precisely described in which conditions polychronous programs can be seen as KPNs. In this context, we have considered different cases of process networks, including so-called “polyendochronous processes”. Under some conditions expressed by clock equation systems, (networks of) processes exhibiting polyhierarchies of clocks are polyendochronous and, as compositions of endochronous processes, may be seen as KPNs.

Based on this characterization, we have developed in the open-source Polychrony toolset a new strategy of code generation for such (polyendochronous) process networks. Typically, after the clock calculus, a program $P$ is organized as a composition of processes, $P=\left(\right|P1\left|P2\right|...\left|Pn\right|)$, each one structured around a clock tree. When $P$ is characterized as polyendochronous, it contains generally clock constraints such as $\mathrm{𝐶𝑙𝑘}1=\mathrm{𝐶𝑙𝑘}2$, with $\mathrm{𝐶𝑙𝑘}1$ being a clock in the subtree corresponding to $P1$ and $\mathrm{𝐶𝑙𝑘}2$ a clock in the subtree corresponding to $P2$.

Such a constraint induces a synchronization between two parts ($P1$, $P2$) of the program when $\mathrm{𝐶𝑙𝑘}1$ or $\mathrm{𝐶𝑙𝑘}2$ occurs. The principle of the code generation for polyendochronous processes is based on the existing distributed code generation, but with the additional resynchronization of parts of the application induced by the constraints on clocks ($\mathrm{𝐶𝑙𝑘}1$, $\mathrm{𝐶𝑙𝑘}2$) not placed in the same clock trees. For distributed code generation, it is considered that each (clock) hierarchy will run on a specific processor. In this case, the purpose is mainly to partition the application, and the processors will be virtual ones.

The code generation of each partition consists in the definition of several tasks: one task per cluster (a cluster being a subpart that may be executed as soon as its inputs are available, without any communication with the external world); one task per input/output of the partition; one task for the cluster of state variables; one task that manages the steps. Synchronization between these tasks is obtained by semaphores (one semaphore per task). This code generation technique for the class of networks called “polyendochronous processes” has been added in the Polychrony toolset (http://polychrony.inria.fr) and a paper describing the comparison of process models is currently in submission.