"Hi , I had recently installed DS5 and ran the matric multiplication application from SVE examples , since I had chosen the N1 as the FVP environment, I observed the vector length was 8 x 64 bits(since application had float 64 as datatype) = 512 bits. I am exploring more on how to force to use the predicate register to utilize the 2048 bits in SVE2 platforms "
Hi there, I have moved your question to the Architectures and Processors forum. Many thanks.
Did you mean the N1 or V1?
Either way, SVE (and SVE2) is a scalable vector instruction set, supporting vector lengths from 128 to 20248 bits. A given processor will have a maximum supported vector length, for example 256 bits. You can use vector lengths below the maximum, but you can't above what the hardware supports.
Software can written to be vector length agnostic, meaning that it will exploit whatever the configure (or maximum) vector length of the thing it is running on is.
If you wanted to experiment with wider length lengths, the AEM (Architecture Envelope Model) might be useful.
Thanks for replying , I meant for N1 only , i was not aware of this AEM , is this an FVP within in the DS5 ? , my intention is to experiment 2048 bits vector length in FVP environment using DS5 since i dont have the hardware, kindly point to on which Neoverse FVP model in DS5 would help for the experiment
My knowledge of DS-5/Arm DS is a little out of date I'm afraid. But I know you can download the AEM standalone from this page:
You want the model called "Armv-A Base RevC AEM FVP". This is a "generic" model, rather than a model of a specific processor. That means you have greater freedom in configuring it. You will need to set the desired max vector length using a parameter when you launch the model (the parameter is called "veclen").More info here:
DS-5 is a very old toolchain. It has long been superceded by Arm Development Studio, which includes Neoverse N1 (and N2) FVPs and examples:https://developer.arm.com/Tools%20and%20Software/Arm%20Development%20Studio