GDC09: parallel load, store and other ops with Larrabee

By Michael J Huelskoetter (90 posts) on August 14, 2009 at 5:24 am

“You must finish what you start.” This is my today's credo regarding GDC09 as I've blogged only yesterday about one of the planned tech sessions referring to Larrabee programming. And today I'm publishing the second session's preview.

This tech talk is being called "SIMD programming with Larrabee: Second Glance at the New Instructions in action" and has to do with register based programming for Larrabee software. Moreover Intel's Steve Hughes and Steve McCalla will immerse deeply into the topic and talk about the fact that ...

… every Larrabee CPU core has its own 32 vector registers which are 16 bits wide. These registers can be used for parallel executable operations. This results in 512 bit wide SIMD registers per processor core.

… Larrabee based programming provides two methods to perform vector operations: SOA (Structure of Arrays) and AOS (Array of Structure). Both make sense under certain conditions and Steve & Steve will demonstrate in detail, which method is the better one. This will be shown with the help of easy math examples like matrix calculation.

… Predication uses eight 16-bit wide registers to execute vector based comparative operations in parallel. Predication can be used for typical x86 loops either.

… Gather and Scatter are also important if you talk about Larrabee based coding. Gather corresponds to a load instruction which is executed simultaneously on 16 floating point registers. And Scatter means the reverse way, namely the simultaneous saving of these 512 bit wide vectors. Besides, we will also learn that Gather and Scatter are only limited by the cache speed.

… there is already a C ++ Larrabee protoype library which uses only one header and no .lib or .dll files anymore. And we'll hear also that debugging will be easier and that existing platforms will be supported.

So, there are many good reasons to mark your calender. You either join the session directly at GDC'09 on Tuesday morning from 10:10 to 11:00 am CET. Or you stop and follow this blog to get all the information in a dedicated article as we will video blog what Steve & Steve are talking about - and why Larrabee will be important for many developers in the future

Categories: Events, Game Development, Graphics & Media, Parallel Programming
Tags: ,

For more complete information about compiler optimizations, see our Optimization Notice.

Comments (1)

August 14, 2009 11:40 AM PDT


omnia b7610		 "which are 16 bits wide." I don't fully understand this sentence. What does that exactly mean?

Trackbacks (4)


Leave a comment  

To obtain technical support, please go to Software Support.
Name (required)*

Email (required; will not be displayed on this page)*

Your URL (optional)


Comment*