- a thin layer, using a bit boost::mpi (for the communicator mostly ...)
along the lines discussed in #12.
- implemented reduce, allreduce, bcast for arrays, simple scalars,
and any custom type that support boost serialization.
- Custom types : the operations are done recursively on members.
No change is needed in the class to use this mpi routine, as long as
serialize function is defined.
- For arrays of basic types (int, double...), a direct call to MPI C API, which works also for views
(as long as they are contiguous).
- For arrays of more complex types, we revert to boost::mpi.
- Added a simple test.
- Work still in progress :
- missing a simple scatter/gather for the arrays
- need more tests & API thinking.
- dispatch array code to array lib
- reduce is "sum" only, but do we need more.
Now the tail have a fixed size. It actually makes everything simpler. I took
order_min = -1 and order_max = 8. This makes the tails compatible with the
previous implementation. However we might want to change this to something like
-10, 10 so that they are self-contained. This commit should also fix issue #11.
- to use already a few c14 convenience details :
-> polymorphic std::plus, e.g.
boost::mpi::reduce (world, A,C, std::c14::plus<>(),0);
this plus determine the type by itself ...
-> errors on the type can be very cryptic on the gf.
-> add std::c14::make_unique
(equivalent of make_shared for unique_ptr).
- The previous version of the * operator for matrix was too clever.
It was giving a lazy object and then rewriting C = A *B into gemm (a,A,B,0,C).
The pb was in case of aliasing : when e.g. C = A, or is a part of A.
gemm is not correct that case, and as a result generic code like
a = a *b
may not be correct in matrix case, which is unacceptable.
- So we revert to a simple * operator for matrix
that does immediate computation.
Same thing for matrix* vector
- we also suppress a_x_ty class.
-> for M = a * b,
when M is a matrix, there is no overhead due to move assignment
-> however, when M is a view, there is an additionnal copy.
-Correctness comes first, hence the fix.
However, if one wants more speed and one can guarantee that
there is no aliasing possible, then one has to write a direct gemm call.
-> det_manip class was adapted, since in that case, we can show there
no alias, and we want the speed gain, so the * ops where replaced
by direct blas call (using the array blas interface).
-> also gemm, gemv, ger were overloaded in the case the return
matrix/vector (i.e. last parameter of the function) is not an lvalue,
but a temporary view created on the fly.
- Order of inclusion of Python include path and nunpy python include path
reversed.
--> on Ubuntu 12.04, there is a link numpy in /usr/include/python2.7
which was *before* the numpy include path, i.e. in this case
the new version of numpy installed by virtualenv.
By reversing the order of the include, the numpy arrayobject.h
file of virtualenv is found first, the Python.h is found anyway
in the correct system directory (does not change with new numpy version).
O. Parcollet
gf: security in the case beta<0 added in the mesh construction
gf: inline added in slice
test/triqs/gf: test of on_mesh() added
gfs: scalar for two-real_times
test/triqs/gf/ renamed in gfs, test gf_retw.cpp completed
gfs: evaluator homogeneised
two_times: evaluator corrected
test/triqs/gf/ renamed in gfs, test gf_retw.cpp completed
+ Correction after rebase
Fix a test : gf_re_im_freq_time
There is an issue with the last point.
To be fixed.
- change : all objects are by default
stored now by reference, not by copy any more.
Unless the trait force_copy_in_expr is true.
- rvalue refs are moved into the tree
- simplifies a lot the writing of lazy method, objects.
- added a macro for methods
- tests ok. Further check needed to control absence of copies...
- improved documentation
