Subject: Re: [boost] performance of a linear algebra/matrix library
From: DE (satan66613_at_[hidden])
Date: 2010-05-11 14:04:25
on 11.05.2010 at 20:20
joel falcou wrote :
> DE wrote:
>> here a question arises:
>> since a compiler is able to generate very fast code involving simd
>> instructions is one supposed to provide simd-enabled implementation of
>> a generic library? personally i think now that it is worthless
> Yeah sure, let's look how icc is able to vectorize arbitrary large
> arithmetic function
> (like cos, sqrt etc that don't have a 1-1 SIMD intrinsic mapping).
i think i agree
> Other think. If you
> work on dynamic memory, the compiler won't be doing any vectorization as
> the memory
> may not be aligned properly.
icc11 does it for dynamically allocated memory both for C and C++ code
i think it's thanks to 'new' which allocates memory 8 byte aligned
> Add to that loop with dependencies that human know how to rewirte (but
> not compiler),
> arbitrary function support, SoA support for thing like complex or other
> ADS ...
what's SoA? and ADS?
> Oh and, if your code as right, then both would have been inlined. That's
> what happen in our library
> (which lead us to rmeove the unroll settings)
don't get it
Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk