Gael Guennebaud
e86dbd5255
fix apply Jacobi for complexes and add documentation for some *Jacobi* functions
2009-08-23 23:49:44 +02:00
Benoit Jacob
7bedf5e9cb
add initial, rough, full-pivoting RRQR decomposition
...
lots of room for improvement!
and add Gael a (c) line in Householder.h
2009-08-22 01:13:21 -04:00
Benoit Jacob
2f0b4e1abc
fix compilation with gcc 4.1. Indeed the path for recent gcc doesn't work with gcc 4.1, and looking at the implementation of vector in g++ 4.1, it was exactly our fallback case, so use that.
2009-08-21 12:16:37 -04:00
Benoit Jacob
72b002eab9
work around internal compiler error with gcc 4.1 and 4.2, reported on the forum
2009-08-20 12:19:15 -04:00
Gael Guennebaud
d56be9c128
* make HessenbergDecomposition uses the Householder module
...
* bugfix in ei_blas_traits for .conjugate().conjugate()
2009-08-17 17:41:01 +02:00
Gael Guennebaud
ff0f005d4c
change the make householder algorithm so that the remaining coefficient
...
is real, and make Tridiagonalization use it
2009-08-17 17:04:32 +02:00
Gael Guennebaud
e125c199bb
add EIGEN_TRANSFORM_PLUGIN
2009-08-17 09:16:41 +02:00
Gael Guennebaud
737bed19c1
make HouseholderQR uses the Householder module
2009-08-16 19:22:15 +02:00
Gael Guennebaud
fc9480cbb3
bugfix in compute_matrix_flags, optimization in LU,
...
improve doc, and workaround aliasing detection in MatrixBase_eval snippet
(not very nice but I don't know how to do it in a better way)
2009-08-16 10:55:10 +02:00
Benoit Jacob
ee982709d3
in all decs, make the compute() methods return *this
...
(implements feature request #18 )
2009-08-15 23:12:39 -04:00
Gael Guennebaud
65fe5f76fd
rename back MayAliasBit to EvalBeforeAssigningBit
2009-08-16 00:14:05 +02:00
Gael Guennebaud
f5f2b222a3
make SVD reuses applyJacobi
2009-08-16 00:02:36 +02:00
Gael Guennebaud
044dd0c1dd
revert previous change in Quaternion::setFromTwoVectors
2009-08-15 23:37:20 +02:00
Benoit Jacob
03c1e79f35
svd: sort in decreasing order, remove unused code
2009-08-15 19:20:48 -04:00
Gael Guennebaud
239ada95b7
add overloads of lazyAssign to detect common aliasing issue with
...
transpose and adjoint
2009-08-15 22:19:29 +02:00
Benoit Jacob
a3e6047c25
fix and improve docs
2009-08-15 15:29:44 -04:00
Gael Guennebaud
50c703f0c7
As proposed on the list:
...
- rename EvalBeforeAssignBit to MayAliasBit
- make .lazy() remove the MayAliasBit only, and mark it as deprecated
- add a NoAlias pseudo expression, and MatrixBase::noalias() function
Todo:
- we have to decide whether += and -= assume no aliasing by default ?
- once we agree on the API: update the Sparse module and the unit tests respectively.
2009-08-15 18:35:51 +02:00
Gael Guennebaud
13a8956188
bugfix in inner-product specialization,
...
compilation fix in stable norm,
optimize apply householder
2009-08-15 13:12:50 +02:00
Gael Guennebaud
7b60713e87
my previous fix was not very good
2009-08-15 11:52:50 +02:00
Gael Guennebaud
0da31a6e1d
bugfix and compilation fix in ProductBase
2009-08-15 10:55:11 +02:00
Gael Guennebaud
bff4238d15
fix setFromTwoVectors because of the change in sorting of the the singular values
2009-08-15 10:24:27 +02:00
Gael Guennebaud
109a4f650b
fix a couple of warnings
2009-08-15 10:20:01 +02:00
Benoit Jacob
fe4a86443f
fix warning
2009-08-14 20:16:04 -04:00
Benoit Jacob
a5f820b873
forgot to update this
2009-08-14 20:03:14 -04:00
Benoit Jacob
2f74801ca4
as discussed on list: default to align cols, reorganize parameters accordingly so that the default corresponds to 0 flag,
...
and implement FullPrecision output (non-default).
2009-08-14 16:31:42 -04:00
Benoit Jacob
16abc0ba7f
try to support 16 bit platforms... optimistic, but can't hurt
2009-08-14 15:49:14 -04:00
Benoit Jacob
22ae236d4e
machine_epsilon -> epsilon as wrapper around numeric_traits
2009-08-14 15:12:32 -04:00
Gael Guennebaud
6373c3cd00
oops bis, I forgot that SelfAdjointEigneSolver directly called the selector...
2009-08-14 13:49:29 +02:00
Gael Guennebaud
8abec72259
oops forgot to remove the #include in Core
2009-08-14 09:49:33 +02:00
Gael Guennebaud
13e95f7f68
optimize "apply Jacobi" for small sizes, and move it to Jacobi.h
2009-08-14 00:17:14 +02:00
Benoit Jacob
f2536416da
* remove EIGEN_DONT_INLINE that harm performance for small sizes
...
* normalize left Jacobi rotations to avoid having to swap rows
* set precision to 2*machine_epsilon instead of machine_epsilon, we lose 1 bit of precision
but gain between 10% and 100% speed, plus reduce the risk that some day we hit a bad matrix
where it's impossible to approach machine precision
2009-08-13 14:56:39 -04:00
Benoit Jacob
76a3089a43
oops, don't set the precision to -1 !!
2009-08-13 09:56:53 -04:00
Benoit Jacob
13b191d94b
apply Koldo's workaround for MSVC bug
2009-08-13 09:53:47 -04:00
Gael Guennebaud
1b257a7620
add an optimized "apply in place a rotation in the plane",
...
and make Jacobi and SelfAdjointEigenSolver use it
=> ~ x1.75 speedup for JacobiSVD and x2 for SelfAdjointEigenSolver
2009-08-13 11:42:02 +02:00
Benoit Jacob
1d80f561ad
apply change discussed on the list :
...
* new default precision "-1" means use the current stream precision
* otherwise, save and restore the stream precision
2009-08-13 22:50:55 -04:00
Benoit Jacob
99802094e4
do without an empirical homemade formula that i wasn't comfortable about...
...
turns out it's not needed anymore and removing it seems to only increase the precision
2009-08-12 18:30:37 -04:00
Benoit Jacob
2b618a2c16
make jacobi SVD more robust after experimenting with very nasty matrices...
...
it turns out to be better to repeat the jacobi steps on a given (p,q) pair until it
is diagonal to machine precision, before going to the next (p,q) pair. it's also
an optimization as experiments show that in a majority of cases this allows to find out
that the (p,q) pair is already diagonal to machine precision.
2009-08-12 18:23:39 -04:00
Benoit Jacob
309d540d4a
add parentheses; hopefully this solves Koldos MSVC compilation issue...
2009-08-12 10:14:15 -04:00
Benoit Jacob
22d65d47d0
finally, the good approach was two-sided Jacobi. Indeed, it allows
...
to guarantee the precision of the output, which is very valuable.
Here, we guarantee that the diagonal matrix returned by the SVD is
actually diagonal, to machine precision.
Performance isn't bad at all at 50% of the current householder SVD
performance for a 200x200 matrix (no vectorization) and we have
lots of room for improvement.
2009-08-12 02:35:07 -04:00
Benoit Jacob
ce033ebdfe
add EIGEN_DEBUG_VAR
2009-08-11 16:12:34 -04:00
Gael Guennebaud
afbd73b5cd
overload operartor* with a ProductBase such that "scalar * (mat * mat)" is optimized
...
as one could naturally expect
2009-08-11 15:15:06 +02:00
Gael Guennebaud
a4f6642518
fix issue #36 (missing return *this in Rotation2D
2009-08-11 15:11:47 +02:00
Gael Guennebaud
ea884e6f48
remove #include Bidiagonalization, and add missing ";"
2009-08-11 15:08:03 +02:00
Gael Guennebaud
35b4077a5d
merge
2009-08-09 23:11:25 +02:00
Gael Guennebaud
ef55e7f4ce
make custom asm directive volatile
2009-08-09 23:09:46 +02:00
Benoit Jacob
216ee335ac
LinearVectorization: If the destination isn't aligned,
...
we have to do runtime checks and we don't unroll, so it's only good for large enough sizes
2009-08-09 22:19:12 +02:00
Benoit Jacob
1f1705868b
now you can #define EIGEN_DEBUG_ASSIGN, and all the values in ei_assign_traits are printed
2009-08-09 21:35:13 +02:00
Benoit Jacob
527557672a
disable the assembly for fast unaligned stores. indeed, there is a strange bug that is triggered by this code:
...
#include<Eigen/Core>
int main()
{
Eigen::Matrix4f m;
m <<1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16;
m.col(0).swap(m.col(1));
std::cout << m << std::endl;
}
when the fast unaligned stores are used, the column is copied instead of being swapped.
2009-08-09 20:49:55 +02:00
Benoit Jacob
8e08680119
don't depend on uninitialized value
2009-08-09 17:18:42 +02:00
Benoit Jacob
3ed83fa681
* add Jacobi transformations
...
* add Jacobi (Hestenes) SVD decomposition for square matrices
* add function for trivial Householder
2009-08-09 16:58:13 +02:00
