Re: scheduling techniques

Nadjiba Sahraoui wrote:
>
> I'm investigating techniques on fine-grain scheduling of loops given
> constraints on:
>
> 1- number of registers
> 2- number of load/store /cycle
> 3- number of functional units
>
> I have to experiment these techniques for a convolution.
>
> If you have information about papers, reports, tools, methods or sites
> let me know please asap.
>


The real problem here is that the number of scheduling techniques is
too enormous to consider them all. There must be dozens of different
techniques for scheduling a loop, from trace scheduling to the
countless variations on the theme of modulo scheduling. You really
need to narrow down the number of techniques that you want to look at.


For a survey of the major software pipelining techniques, you
should look at Allan et al's survey of software pipelining.


@ARTICLE{Allen+95,
                AUTHOR = {Vicki H. Allan and Reese B. Jones and
                                              Randall M. Lee and Stephen J. Allan},
                TITLE = {Software Pipelining},
                BOOKTITLE = {ACM Computing Surveys},
                ORGANIZATION= {ACM},
                MONTH = {Sept},
                YEAR = 1995}


Although very good, this survey is really just the tip of the iceberg
when it comes to software pipelining.


If you decide to go with modulo scheduling (as many people do), you
may want to try to get a copy of MOST (Modulo Scheduling Testbed).
This is a research testbed that includes implementations of many of
the more popular modulo scheduling techniques. I don't think it is
publicly available, but someone out there probably knows where you can
get it from.


David.

Return to the comp.compilers page.
Search the comp.compilers archives again.