|Re: [eigen] SelfAdjointView<> as LHS expression for optimizations?|
[ Thread Index |
| More lists.tuxfamily.org/eigen Archives
- To: eigen@xxxxxxxxxxxxxxxxxxx
- Subject: Re: [eigen] SelfAdjointView<> as LHS expression for optimizations?
- From: Gael Guennebaud <gael.guennebaud@xxxxxxxxx>
- Date: Mon, 8 Nov 2010 23:57:08 +0100
- Dkim-signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=gamma; h=domainkey-signature:received:mime-version:received:in-reply-to :references:from:date:message-id:subject:to:content-type :content-transfer-encoding; bh=YghRDB15ywzZQlpiA+wdRpxnb3zfPXNrUas/29Gctc0=; b=rIJKBVW4tELvswMxGdDhyvx8rD+ZNQ9MI8yYzccbMyM1paHIe1IKyPhm/dxgtNIH6I CdiJCAP4bTmsYk6fPh2P2KR8bRNs36NaKH19rBIT9Yl62l3ld5YjxZmsEyFKU1SjZwr0 wF8IwHei8zXoCN6pUkvILMnnhu8j7mi6ur0Us=
- Domainkey-signature: a=rsa-sha1; c=nofws; d=gmail.com; s=gamma; h=mime-version:in-reply-to:references:from:date:message-id:subject:to :content-type:content-transfer-encoding; b=WmjVoHz5TOXV2D9+js1eESNRMe+GgY9geR9FaDeZlyMAyWvFdPhz2TWKHRC3LIPjDg jngMh0A3NWwXmjWHp45twyOK5KAs29NKQyIntimvI3atV/TWr+fck832ZEFNsW3jFIOY G1CudK5k2VHA5h8dFuxBUIzp+Y7B4RYgeE0KA=
On Fri, Nov 5, 2010 at 8:49 PM, Jose Luis Blanco
> Hi all,
> The operation I have in mind is:
> RESULT = H * C * H.transpose()
> with C symmetric and squared.
> The point is that two "problems" (well... I have to admit, small
> ones!) occur with a naive Eigen expresion "H * C * H.transpose()":
> 1) Many multiplications could be saved by realizing that the result is
> exactly symmetric. I know the implications may be far reaching, but,
> could SelfAdjointView<> be extended so it can be used to store
> results? I mean, something like:
> Eigen::MatrixXd RESULT;
> RESULT.selfadjointView<Eigen::Lower>() =
> or an alternative that wouldn't require LHS = operator for SelfAdjointView<>:
> Eigen::MatrixXd RESULT =
> Any ideas about the difficulty of implementing any of these two ideas?
> Or... is there anything similar already done?
Good question. Actually I don't any particular optimizations beside
computing only half of the second product. Basically, this would be
implemented as follow:
Temp = H * C.selfadjointView<Lower>();
Result.triangularView<Lower> = Temp * H.adjoint();
The second line currently does not work, however this is very easy to
do by extending our internal selfadjoint product routine to compute
half of a general A*B product instead of handling AA* or A*A only.
Actually this routines cannot exploit the fact that both sides come
from the same matrix anyway....
Regarding the API we could simply let the user write:
Result.triangularView<Lower> = H * C.selfadjointView<Lower>() * H.adjoint();
For consistency reason,
Result.selfadjointView<Lower> = H * C.selfadjointView<Lower>() * H.adjoint();
would do the same.
> 2) Apart from efficiency, there is this numerical accuracy problem of
> the resulting matrix not being *exactly* symmetric... a*b*c != c*b*a
> for most architectures. Ok, the errors are tiny, but it may lead to
> inconsistencies that some time ago affected the Cholesky
> implementation I used to use... that's why I worry about all this in
> the first place!