that eigen2 owes all its performance to nonstandard restrict keyword.
well, this can also improve portability in case some compiler doesn't have __restrict.
* EIGEN_ALIGN and EIGEN_DONT_ALIGN and the corresponding logic in Macros.h
(instead of using EIGEN_ARCH_WANTS_ALIGNMENT)
* The body of ei_aligned_malloc and ei_aligned_free
The reason for this backporting is that a user complained that with eigen 2.0 he got a warning at Memory.h:81 that the return value of posix_memalign was not used, and that function was declared with an attribute warn_unused_result.
Looking at this, it seemed that the body of this function was already overly complicated, and fixing this warning made it even worse, while the devel branch had a much simpler body and didn't suffer from that problem.
Then it was necessary to update ei_aligned_free too, and to backport EIGEN_ALIGN.
Inch' Allah....
df9dfa1455
This is what is needed to make Step (in KDE/kdeedu) build.
The rest of Eigen (outside of Sparse) is unaffected except for a few trivial changes that were needed.
calling this 2.0.3, will tag if no problem.
EIGEN_WORK_AROUND_QT_BUG_CALLING_WRONG_OPERATOR_NEW_FIXED_IN_QT_4_5
is the right way to go for allowing placement new on a class having
overloaded operator new. Qt 4.5 won't add the :: prefix. (I still do not
understand how you can distinghish a placement new from an overloaded
operator new taking an allocator as argument...)
* try to be clever in matrix ctors and operator=: be lazy when we can, always allow
to copy rowvector into columnvector, check the template parameters,
try to factor the code better
* add missing copy ctor in UnalignedType
* fix bug in the traits of DiagonalProduct
* renaming: EIGEN_TUNE_FOR_CPU_CACHE_SIZE
* update the dox a little
That means a lot of features which were available for sparse matrices
via the dense (and super slow) implemention are no longer available.
All features which make sense for sparse matrices (aka can be implemented efficiently) will be
implemented soon, but don't expect to see an API as rich as for the dense path.
Other changes:
* no block(), row(), col() anymore.
* instead use .innerVector() to get a col or row vector of a matrix.
* .segment(), start(), end() will be back soon, not sure for block()
* faster cwise product
only used as fallback for now, needs benchmarking.
also notice that some malloc() impls do waste memory to keep track of alignment
and other stuff (check msdn's page on malloc).
* expand test_dynalloc to cover low level aligned alloc funcs. Remove the old
#ifdef EIGEN_VECTORIZE...
* rewrite the logic choosing an aligned alloc, some new stuff:
* malloc() already aligned on freebsd and windows x64 (plus apple already)
* _mm_malloc() used only if EIGEN_VECTORIZE
* posix_memalign: correct detection according to man page (not necessarily
linux specific), don't attempt to declare it if the platform didn't declare it
(there had to be a reason why it didn't declare it, right?)
ei_aligned_malloc now really behaves like a malloc
(untyped, doesn't call ctor)
ei_aligned_new is the typed variant calling ctor
EIGEN_MAKE_ALIGNED_OPERATOR_NEW now takes the class name as parameter
