Subject: Re: [boost] Boost SIMD beta release
From: Dave Abrahams (dave_at_[hidden])
Date: 2012-12-23 21:55:33
on Sun Dec 23 2012, Mathias Gaunard <mathias.gaunard-AT-ens-lyon.org> wrote:
> On 23/12/12 13:41, Mathias Gaunard wrote:
>> Technical problem: for each segment, the type of the iterator needs to
>> be the same. The code for incrementing or dereferencing the iterator is
>> therefore the same.
>> This gives us two solutions:
>> - Do a branch in the dereferencing function of the iterator to load
>> normal data and trailing data differently (at every iteration, not ok --
>> this approach also works with normal iterators)
>> - Pre-emptively build the last vector of the data and copy it into the
>> iterator (which implies unnecessary loads, stores and cache misses and
>> makes the iterator expensive to copy), and consider that as an
>> additional segment of contiguous memory.
> The above only stands for input iterators. I forgot mentioning output ones.
> The first solution still works for output iterators, but the second
> one clearly does not.
> The only way we would have to make efficient segmented output
> iterators for this case would be to allow iterator of different
> segments to be of different types, which would require making the
> number of segments per level a compile-time constant.
Certainly not. You can do one runtime dispatch per top-level algorithm
call, depending on segment alignment, just as the Austern paper
does... and just as your approach would have to do AFAICT.
-- Dave Abrahams BoostPro Computing Software Development Training http://www.boostpro.com Clang/LLVM/EDG Compilers C++ Boost
Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk