Embedded Systems

Implementation of the Pico-CNN Deep Learning Inference Framework in OpenCL for Embedded Devices

Bear­beitet von S. Brandt.

Bach­e­lor’s The­sis / Mas­ter’s The­sis / Stu­dent Re­search Pro­ject

Ab­stract

In­fer­ence of deep learn­ing al­go­rithms on em­bed­ded/edge de­vices is a very ac­tive area of re­search in acad­e­mia and in­dus­try. How­ever, pop­u­lar deep learn­ing frame­works are not suited for in­fer­ence on em­bed­ded/edge de­vices. In order to meet this de­mand the Chair for Em­bed­ded Sys­tems de­vel­oped the open source deep learn­ing in­fer­ence frame­work Pico-CNN [1] (https://​github.​com/​ekut-​es/​pico-​cnn) which is com­pletely writ­ten in C, op­ti­mized for em­bed­ded/edge de­vices, and is not de­pen­dent on third party li­braries.

In this stu­dent re­search pro­ject an OpenCL [2] vari­ant of Pico-CNN should be im­ple­mented and op­ti­mized to uti­lized em­bed­ded CPUs and GPUs.

Re­quire­ments

  • C/C++
  • Deep Learn­ing
  • Linux (op­tional)

Ref­er­ences

[1] K. Lübeck and O. Bring­mann, “A Het­ero­ge­neous and Re­con­fig­urable Em­bed­ded Ar­chi­tec­ture for En­ergy-Ef­fi­cient Ex­e­cu­tion of Con­vo­lu­tional Neural Net­works,” in Ar­chi­tec­ture of Com­put­ing Sys­tems – ARCS 2019, pp. 267–280 (Copen­hagen, Den­mark).

[2] Wikipedia, OpenCL: https://​en.​wikipedia.​org/​wiki/​OpenCL

Con­tact

Lübeck, Kon­stan­tin

Jung, Alexan­der

Bring­mann, Oliver