| Re: [eigen] Eigen "views" |
[ Thread Index |
Date Index
| More lists.tuxfamily.org/eigen Archives
]
- To: eigen@xxxxxxxxxxxxxxxxxxx
- Subject: Re: [eigen] Eigen "views"
- From: Benoit Jacob <jacob.benoit.1@xxxxxxxxx>
- Date: Tue, 16 Aug 2011 11:03:00 -0400
- Dkim-signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=gamma; h=mime-version:in-reply-to:references:date:message-id:subject:from:to :content-type; bh=Y23bzwXfb66OPKDvyA778lRR3gAOX4+kIkM0W7JgMaI=; b=pKvkYkPqBOKvi0wIZ6xs57aqlm0HEgIMI0zJ3fNBgodhI2jhQV6osOce09cRTvmWO6 NfdQZCJ/G1yakXc1qnKTukcvKYI/5sCLFErIObim3iSLsdtnJb/SCq5/nYE7AmUuSswZ A5BAh9jID99zLUjqcal+bkGChF2xAIRjSLhTc=
2011/8/16 Benoit Jacob <jacob.benoit.1@xxxxxxxxx>:
> 2011/8/16 Jose Luis Blanco <joseluisblancoc@xxxxxxxxx>:
>> I would be also quite interesting in this feature, but... shouldn't it
>> come at the cost of losing vectorization optimizations?
>
> For sure, this feature would be tricky to vectorize. I'd not worry
> about this at first and just do a non-vectorized version. Then, a
> vectorized version could be made where the packet() method fetches the
> 4 (or 2) scalars independently and loads them into a xmm packet, but
> it could be that the loss of memory locality when iterating over
> packets
....or rather, just the fact that each packet load would have to do 4
(or 2) random memory accesses.
> offsets the benefits of vectorization. This would have to be
> benchmarked.
>
> Cheers,
> Benoit
>
>> Or are the inner workings of Eigen such smart that would dynamically
>> build SSE "packets", etc.?
>>
>> Cheers,
>> JL
>>
>> On Tue, Aug 16, 2011 at 3:10 PM, Christoph Hertzberg
>> <chtz@xxxxxxxxxxxxxxxxxxxxxxxx> wrote:
>>> On 16.08.2011 15:07, Christoph Hertzberg wrote:
>>>>
>>>> On 16.08.2011 08:48, Gael Guennebaud wrote:
>>>>>
>>>>> [...]
>>>>>
>>>>> m2 = m1(indices1,indices2);
>>>>> m2 = m1(Eigen::All,indices2);
>>>>> m2 = m1(indices1,Eigen::All);
>>>>
>>>> Yes, I think this looks as close as one can get to Matlab-syntax (not
>>>> that this is a requirement, but it's definitely not a bad thing).
>>>
>>> You could even make m1(i, indices) return a row-vector etc ...
>>>
>>>>
>>>> Christoph
>>>>
>>>
>>>
>>
>> --
>> ___________________________________________________________
>>
>> Dr. Jose-Luis Blanco-Claraco
>> Dpt. Ing. Civil, Mat. y Fabric - Phone: +34 951 952435
>> E.T.S.I. Industriales - Despacho 2.037
>> Universidad de Malaga - Campus Universitario de Teatinos
>> 29071 Malaga, Spain
>> https://sites.google.com/site/jlblancosite/
>> ___________________________________________________________
>>
>>
>>
>