Re: [eigen] 32 byte alignment for avx

[ Thread Index | Date Index | More Archives ]

sure, at some point this information will have to be carried out by
the packet_traits class, that's the plan.


On Sun, Nov 27, 2011 at 4:16 PM, Mark Borgerding <mark@xxxxxxxxxxxxxx> wrote:
> On 11/26/2011 05:50 PM, Benoit Jacob wrote:
>> We never align to more than 16 bytes, so requiring higher alignment than
>> that is wrong.
> Has anyone talked about aligning to 32 byte boundaries for AVX SIMD?
> For those who don't know: Starting with Sandy Bridge, intel processors have
> a new flavor of SIMD: AVX, which uses 256 bit registers.  This allows us to
> do the same operation to 8 floats at once, vs. 4 floats at once using SSE
> thru SSE4.2.  The instruction set is structured such that future processor
> generations may use 512 bits or beyond.
> I've done a bit for my work and the speedups are impressive.  Not quite the
> 2x that the increase in register width would suggest, but often 30-60%
> speedups on CPU-intensive problems.
> -- Mark

Mail converted by MHonArc 2.6.19+