* somehow the NICE_RANDOM stuff wasn't being used anymore and
tests were sometimes failing again. Fixed by #including Eigen/Array
instead of cherry-picking just Random.h.
* little fixes in the unaligned assert page
* add Transform::operator= taking rotation.
An old remnant was left commented out. Why was it disabled?
* slight optimization in operator= taking translation
* slight optimization (perhaps) in the new memory assertion
for this very nasty bug (unaligned member in dynamically allocated struct)
that our friends at Krita just encountered:
http://bugs.kde.org/show_bug.cgi?id=177133
CCBUG:177133
- in matrix-matrix product, static assert on the two scalar types to be the same.
- Similarly in CwiseBinaryOp. POTENTIALLY CONTROVERSIAL: we don't allow anymore binary
ops to take two different scalar types. The functors that we defined take two args
of the same type anyway; also we still allow the return type to be different.
Again the reason is that different scalar types are incompatible with vectorization.
Better have the user realize explicitly what mixing different numeric types costs him
in terms of performance.
See comment in CwiseBinaryOp constructor.
- This allowed to fix a little mistake in test/regression.cpp, mixing float and double
- Remove redundant semicolon (;) after static asserts
* add a LDL^T factorization with solver using code from T. Davis's LDL
library (LPGL2.1+)
* various bug fixes in trianfular solver, matrix product, etc.
* improve cmake files for the supported libraries
* split the sparse unit test
* etc.
Some naming questions:
- for "extend" we could also think of: "expand", "union", "add"
- same for "clamp": "crop", "intersect"
- same for "contains": "isInside", "intersect"
=> ah "intersect" is conflicting, so that eliminates this one !
* remove the automatic resizing feature of operator =
* add function Matrix::set() to be used when the previous
behavior is wanted
* the default constructor of dynamic-size matrices now
creates a "null" matrix (data=0, rows = cols = 0)
instead of a 1x1 matrix
* fix UnixX typos ;)
as described on the wiki (one map per N column)
Here's some bench results for the 4 currently supported map impl:
std::map => 18.3385 (581 MB)
gnu::hash_map => 6.52574 (555 MB)
google::dense => 2.87982 (315 MB)
google::sparse => 15.7441 (165 MB)
This is the time is second (and memory consumption) to insert/lookup
10 million of coeffs with random coords inside a 10000^2 matrix,
with one map per packet of 64 columns => google::dense really rocks !
Note I use for the key value the index of the column in the packet (between 0 and 63)
times the number of rows and I used the default hash function.... so maybe there is
room for improvement here....
solver from suitesparse (as cholmod). It seems to be even faster
than SuperLU and it was much simpler to interface ! Well,
the factorization is faster, but for the solve part, SuperLU is
quite faster. On the other hand the solve part represents only a
fraction of the whole procedure. Moreover, I bench random matrices
that does not represents real cases, and I'm not sure at all
I use both libraries with their best settings !
* rename Cholesky to LLT
* rename CholeskyWithoutSquareRoot to LDLT
* rename MatrixBase::cholesky() to llt()
* rename MatrixBase::choleskyNoSqrt() to ldlt()
* make {LLT,LDLT}::solve() API consistent with other modules
Note that we are going to keep a source compatibility untill the next beta release.
E.g., the "old" Cholesky* classes, etc are still available for some time.
To be clear, Eigen beta2 should be (hopefully) source compatible with beta1,
and so beta2 will contain all the deprecated API of beta1. Those features marked
as deprecated will be removed in beta3 (or in the final 2.0 if there is no beta 3 !).
Also includes various updated in sparse Cholesky.
* several fixes (transpose, matrix product, etc...)
* Added a basic cholesky factorization
* Added a low level hybrid dense/sparse vector class
to help writing code involving intensive read/write
in a fixed vector. It is currently used to implement
the matrix product itself as well as in the Cholesky
factorization.
However, for matrices larger than 5, it seems there is constantly a quite large error for a very
few coefficients. I don't what's going on, but that's certainely not due to numerical issues only.
(also note that the test with the pseudo eigenvectors fails the same way)
* eigenvectors => pseudoEigenvectors
* added pseudoEigenvalueMatrix
* clear the documentation
* added respective unit test
Still missing: a proper eigenvectors() function.
based on the former.
* opengl_demo: makes IcoSphere better (vertices are instanciated only once) and
removed the generation of a big geometry for the fancy spheres...
* add a WithAlignedOperatorNew class with overloaded operator new
* make Matrix (and Quaternion, Transform, Hyperplane, etc.) use it
if needed such that "*(new Vector4) = xpr" does not failed anymore.
* Please: make sure your classes having fixed size Eigen's vector
or matrice attributes inherit WithAlignedOperatorNew
* add a ei_new_allocator STL memory allocator to use with STL containers.
This allocator really calls operator new on your types (unlike GCC's
new_allocator). Example:
std::vector<Vector4f> data(10);
will segfault if the vectorization is enabled, instead use:
std::vector<Vector4f,ei_new_allocator<Vector4f> > data(10);
NOTE: you only have to worry if you deal with fixed-size matrix types
with "sizeof(matrix_type)%16==0"...
few bits left of the comma and for floating-point types will never return zero.
This replaces the custom functions in test/main.h, so one does not anymore need
to think about that when writing tests.
This allow code factorization and generic template specialization
of functions
* added any_rotation * {Translation,Scaling,Transform} products methods
* rewrite of the actually broken ToRoationMatrix helper class to
a global ei_toRotationMatrix function.
NonAffine, Affine (default), contains NoShear, contains NoScaling
that allows significant speed improvements. If you like it, this concept could be applied to
Transform::extractRotation (or to a more advanced decomposition function) and to Hyperplane::transformed()
and maybe to some other places... e.g., I think a Transform::normalMatrix() function would not harm and
warn user that the transformation of normals is not that trivial (I saw this mistake much too often)
* handling Quaternion, AngleAxis and Rotation2D, 2 options here:
1- make all of them inheriting a common base class Rotation such that we can
have a single version of operator* for all the rotation type (they all get converted to a matrix)
2- write a version for all type (so 3 rotations types * 3 for Transform,Translation and Scaling)
* real documentation
- the coefficients are stored in a single vector
- added transformation methods
- removed Line* typedef since in 2D this is really an hyperplane
and not really a line...
- HyperPlane => Hyperplane
and various cleaning in Altivec code. Altivec vectorization have been re-enabled
in CoreDeclaration
* added copy constructors in non empty functors because I observed weird behavior with
std::complex<>
AngleAxis*Vector products were wrong because they returned the product
_expression_
toRotationMatrix()*other;
and toRotationMatrix() died before that expression would be later
evaluated. Here it would not have been practical to NestByValue as this
is a whole matrix. So, let them simply evaluate and return the result by
value.
The geometry.cpp unit-test only checked for compatibility between
various rotations, it didn't check the correctness of the rotations
themselves. That's why this bug escaped us. So, this commit checks that
the rotations produced by AngleAxis have all the expected properties.
Since the compatibility with the other rotations is already checked,
this should validate them as well.
* added a meta.cpp unit test
* EIGEN_TUNE_FOR_L2_CACHE_SIZE now represents L2 block size in Bytes (whence the ei_meta_sqrt...)
* added a CustomizeEigen.dox page
* added a TOC to QuickStartGuide.dox
* replaced the Flags template parameter of Matrix by StorageOrder
and move it back to the 4th position such that we don't have to
worry about the two Max* template parameters
* extended EIGEN_USING_MATRIX_TYPEDEFS with the ei_* math functions
* bugfix in Dot unroller
* added special random generator for the unit tests and reduced the tolerance threshold by an order of magnitude
this fixes issues with sum.cpp but other tests still failed sometimes, this have to be carefully checked...
asm("...") from the code while fixing MSVC compat (so your changes crossed
one another).
- move the pragma warning to CoreDeclarations, it's the right place to do early
platform checks.
CCMAIL:ps_ml@gmx.de
IoFormat OctaveFmt(4, AlignCols, ", ", ";\n", "", "", "[", "]");
cout << mat.format(OctaveFmt);
The first "4" is the precision.
Documentation missing.
* Some compilation fixes
* add some explanations in the typedefs page
* expand a bit the new QuickStartGuide. Some placeholders (not a pb since
it's not even yet linked to from other pages). The point I want to make is
that it's super important to have fully compilable short programs (even
with compile instructions for the first one) not just small snippets, at
least at the beginning. Let's start with examples of compilable programs.
- the decompostion code has been adfapted from JAMA
- handles non square matrices of size MxN with M>=N
- does not work for complex matrices
- includes a solver where the parts corresponding to zero singular values are set to zero
- 33 new snippets
- unfuck doxygen output in Cwise (issues with function macros)
- more see-also links from outside, making Cwise more discoverable
* rename matrixNorm() to operatorNorm(). There are many matrix norms
(the L2 is another one) but only one is called the operator norm.
Risk of confusion with keyword operator is not too scary after all.
* remove the cast operators in the Geometry module: they are replaced by constructors
and new operator= in Matrix
* extended the operations supported by Rotation2D
* rewrite in solveTriangular:
- merge the Upper and Lower specializations
- big optimization of the path for row-major triangular matrices
I don't see any reason not to allow it, it doesn't add much code, and
it makes porting from eigen1 easier.
*expand tests/basicstuff to first test coefficient access methods
* fix .normalized() so that Random().normalized() works; since the return
type became complicated to write down i just let it return an actual
vector, perhaps not optimal.
* add Sparse/CMakeLists.txt. I suppose that it was intentional that it
didn't have CMakeLists, but in <=2.0 releases I'll just manually remove
Sparse.
- added a MapBase base xpr on top of which Map and the specialization
of Block are implemented
- MapBase forces both aligned loads (and aligned stores, see below) in expressions
such as "x.block(...) += other_expr"
* Significant vectorization improvement:
- added a AlignedBit flag meaning the first coeff/packet is aligned,
this allows to not generate extra code to deal with the first unaligned part
- removed all unaligned stores when no unrolling
- removed unaligned loads in Sum when the input as the DirectAccessBit flag
* Some code simplification in CacheFriendly product
* Some minor documentation improvements
- removes much code
- 2.5x faster (even though LU uses complete pivoting contrary to what
inverse used to do!)
- there _were_ numeric stability problems with the partial pivoting
approach of inverse(), with 200x200 matrices inversion failed almost
half of the times (overflow). Now these problems are solved thanks to
complete pivoting.
* remove some useless stuff in LU