|
From: | Brian Padalino |
Subject: | Re: [Discuss-gnuradio] PS3/Cell BE platform |
Date: | Thu, 16 Nov 2006 10:24:31 -0500 |
On 11/16/06, Jason <address@hidden> wrote:
Legal nitpicks aside, the SPE will apparently take in an array of four single-precision floats (128bits total in one register), and multiply or add it to another array of four floats, all in one instruction. This would massively reduce the number of instructions for, say, multiplying two 4x4 matrices together. Then toss in the fact that there are 7 SPE's, each moving at 3.2GHz, each with 256K of DMA accessible local memory. Sorry, I'm drooling. :)
From my understanding, each SPE has 128 registers each with a width of
128 bits. Your program and data that you are currently working on needs to all fit within the local store memory of 256k - so you need to be very aware of where your data is and where it is going. You can then have one SPE dedicated to streaming FFT's to another SPE which is doing some kind of correlation/filtering then followed by whatever else you want the other SPE's to do. It should also be noted that DMA from one SPE to another is not all equal - some combinations work faster than others, but if I remember correctly, you can't choose which SPE you assign programs to - it just chooses one given the interface from IBM. The IBM dev kit has a complete simulator in it that should be able to do everything you might want to do with a Cell, just without the great performance. Brian
[Prev in Thread] | Current Thread | [Next in Thread] |