Re: [eigen] Problem with LU and Cholesky inversion

Hi List,

bug fixed in revision 870935. As I said the bug was in fact a shortcoming in solveTriangular() and the current solution is to evaluate the right-hand side to a column major matrix if it is not already the case. So regarding performance, it is (currently) better to use col-major matrices for the right hand side if possible.

Later I plane to implement much more efficient triangular solvers when the rhs is not a vector (currently it processes the rhs one column at once)

cheers,
gael.

On Fri, Oct 10, 2008 at 8:53 PM, Gael Guennebaud <gael.guennebaud@xxxxxxxxx> wrote:

Hi all,

as I said the problem is in solveTriangular(), there is a FIXME. In fact, it assumes the right hand side is column major. If it is not we should copy it to a column major one (allows much faster code), but I don't have time for that right now....

cheers,
gael.

On Fri, Oct 10, 2008 at 5:22 PM, Timothy Hunter <tjhunter@xxxxxxxxxxxx> wrote:

At one point, I was trying to figure out what was wrong in the
.solve() method of the cholesky decomposition and I decomposed the
solveTriangular(diag inverse(solveTriangular())) to check each step.
It had an assert failure about the row major bit.
Let me check if I can get this code back.

On Fri, Oct 10, 2008 at 7:35 AM, Benoît Jacob <jacob@xxxxxxxxxxxxxxx> wrote:
> The spamming continues...
>
> The LU::solve() actually does evaluate its argument b into a matrix c and uses
> only c from that point on... in particular it calls solveTriangular on c
> only... so I really don't understand why it has trouble.
>
> Moreover I added this test in test/lu.cpp:
>
> // unit-test for a bad bug that escaped us for a long time:
> // make sure that LU::solve() works well when passed an _expression_
> m4 = (m1*m2).transpose();
> lu.solve(m4.transpose(), &m2);
> VERIFY_IS_APPROX(m4.transpose(), m1*m2);
>
> This test succeeds ! This is strange as it seems the same as Timothy's code
> which makes LU::solve() fail.
>
> I'm puzzled!
> Benoit
>
>
> On Friday 10 October 2008 15:20:21 Benoît Jacob wrote:
>> Hm I read too fast: indeed we also have the same problem in LU.... which is
>> my baby i.e. my fault.
>>
>> Still not decided whether the right approach is to evaluate or to
>> nest-by-value...
>>
>> Cheers,
>> Benoit
>>
>> On Friday 10 October 2008 15:04:01 Benoît Jacob wrote:
>> > On Friday 10 October 2008 14:58:18 Benoît Jacob wrote:
>> > > Did you forget a nestByValue() ?
>> >
>> > Replying to myself: hm, rather a .eval() of course as a solver typically
>> > uses each coeff many times.
>> >
>> > By the way: the solve() API in Cholesky is not homogeneous with the other
>> > solvers, it returns by value while the other solvers have a C-style API.
>> > Can you harmonize that by the way?
>> >
>> > Cheers,
>> > Benoit
>> >
>> > ---
>>
>> ---
>
>
>
> ---
>
>