Re: [Starpu-devel] Scheduling accelerators

[Muhammad Usama Asad <muhammad-usama.asad@s2012.tu-chemnitz.de>]

Hey Samuel,

Thanks for your reply. The SoC has 2kbyes local memory. This problem  
is one of the leading research area these days. For Automotive  
industry, the Texas Instruments SoC TDA2X is facing the same  
scheduling issues how to get the maximum out of accelerators. The team  
here is actually interested in a automated way to automatically  
partition vision algorithms and map tasks to processors.

I am therefore assigned this task as part of my masters Thesis. I will  
appreciate your teams precious input. Can you recommend some  
researchers group or experts for partitioning ?

Probably for mapping and scheduling I will be using StarPU and will be  
getting your guidance in this regard.

I am highly grateful for your support.

Thanks & regards,
Muhammad






Zitat von Samuel Thibault <samuel.thibault@inria.fr>:

> Muhammad Usama Asad, on Tue 12 Jan 2016 18:43:36 +0100, wrote:
> > So its not well enough to use for Embedded SoCs. Actually this SoC consists
> > of vision accelerators, can you recommend any other scheduler or software
> > framework for scheduling accelerators ?
>
> The problem is quite general actually.  A PhD student in our team
> studied what a dynamic scheduler can afford in an embedded environment,
> and it may not be much.  Again, how much local memory does it have? Is
> it 256KiB, 2MiB, 32MiB?  General-purpose dynamic scheduling would
> usually needs rather 32MiB than even 2MiB.
>
> Samuel