diff options
author | Benoit Jacob <jacob.benoit.1@gmail.com> | 2009-09-26 22:48:16 -0400 |
---|---|---|
committer | Benoit Jacob <jacob.benoit.1@gmail.com> | 2009-09-26 22:48:16 -0400 |
commit | 924b55e9a92cc30f6caf9e53ea6c5ec96f275dc3 (patch) | |
tree | 6edf89d39af1d38b65462f68e105fb30f23de975 /Eigen/src/Core/ReturnByValue.h | |
parent | e82ab8a5dd1974db041098f40e8df819e6569d16 (diff) |
when copying a ReturnByValue into a MatrixBase, first eval it to a PlainMatrixType.
This allows to limit the number of instantiations of the big template method evalTo.
Also allows to get rid of the dummy MatrixBase::resize().
See "TODO" comment.
Diffstat (limited to 'Eigen/src/Core/ReturnByValue.h')
-rw-r--r-- | Eigen/src/Core/ReturnByValue.h | 18 |
1 files changed, 16 insertions, 2 deletions
diff --git a/Eigen/src/Core/ReturnByValue.h b/Eigen/src/Core/ReturnByValue.h index 55652db48..297ed2456 100644 --- a/Eigen/src/Core/ReturnByValue.h +++ b/Eigen/src/Core/ReturnByValue.h @@ -65,8 +65,22 @@ template<typename Derived> template<typename OtherDerived> Derived& MatrixBase<Derived>::operator=(const ReturnByValue<OtherDerived>& other) { - other.evalTo(derived()); - return derived(); + // Here we evaluate to a temporary matrix tmp, which we then copy. The main purpose + // of this is to limit the number of instantiations of the template method evalTo<Destination>(): + // we only instantiate for PlainMatrixType. + // Notice that this behaviour is specific to this operator in MatrixBase. The corresponding operator in class Matrix + // does not evaluate into a temporary first. + // TODO find a way to avoid evaluating into a temporary in the cases that matter. At least Block<> matters + // for the implementation of blocked algorithms. + // Should we: + // - try a trick like for the products, where the destination is abstracted as an array with stride? + // - or just add an operator in class Block, so we get a separate instantiation there (bad) but at least not more + // than that, and at least that's easy to make work? + // - or, since here we're talking about a compromise between code size and performance, let the user choose? + // Not obvious: many users will never find out about this feature, and it's hard to find a good API. + PlainMatrixType tmp; + other.evalTo(tmp); + return derived() = tmp; } #endif // EIGEN_RETURNBYVALUE_H |