Skip to Main content Skip to Navigation
Reports

Kernel Offloading with Optimized Remote Accesses

Christophe Alias 1 Alain Darte 1 Alexandru Plesco 1
1 COMPSYS - Compilation and embedded computing systems
Inria Grenoble - Rhône-Alpes, LIP - Laboratoire de l'Informatique du Parallélisme
Abstract : The use of hardware accelerators, e.g., with GPGPUs or customized circuits using FPGAs, are particularly interesting for accelerating data- and compute-intensive applications. However, to get high performance when offloading computation kernels to such an acceleration platform, it is mandatory to restructure the application code and to generate adequate communication mechanisms to the external memory, i.e., for remote accesses. In a previous report, we showed, in the context of the high-level synthesis (HLS) of hardware accelerators, how to derive an optimized organization for an accelerator communicating to an external DDR memory. Loop tiling is used to enable block communications, suitable for DDR memories. Pipelined communication processes are generated to overlap communications and computations, thereby hiding some latencies, in a way similar to double buffering. Finally, data reuse among tiles is exploited to avoid remote accesses when data are already available in the local memory. The goal of this report is to show how to specify the sets of data to be read from and to be written to the external memory so as to reduce communications and reuse data as much as possible in the accelerator. The main difficulty in such a specification arises when some data are (re)defined in the accelerator, and even worse in the case of approximations, when some data may be redefined but this is not sure. We show that techniques based on parametric polyhedral optimizations can be used to generate the sets of data to be loaded (resp. stored) just before (resp. after) each tile. An interesting feature is that the case where no approximation is needed appears nicely as a particular case of the general procedure.
Complete list of metadatas

Cited literature [4 references]  Display  Hide  Download

https://hal.inria.fr/inria-00611179
Contributor : Alain Darte <>
Submitted on : Monday, July 25, 2011 - 5:30:29 PM
Last modification on : Wednesday, November 20, 2019 - 2:38:25 AM
Long-term archiving on: : Monday, November 12, 2012 - 11:30:08 AM

File

RR-7697.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : inria-00611179, version 1

Collections

Citation

Christophe Alias, Alain Darte, Alexandru Plesco. Kernel Offloading with Optimized Remote Accesses. [Research Report] RR-7697, INRIA. 2011, pp.29. ⟨inria-00611179⟩

Share

Metrics

Record views

373

Files downloads

160