Currently we have a dependency cycle between `graph` and `graph_parallel`. It's not the only one, but it's notable with the fact that both libraries are buildable and we therefore can't ignore it as we can ignore cycles between header-only libraries that at release time become a single directory. In the particular, this impedes efforts to generate CMake config files. Under one scheme I'm currently working on, for instance, Boost.Python can't be installed because it tries to install Graph, which tries to install GraphParallel, which tries to install Graph. The cycle between these two is an interesting one. boost/graph/breadth_first_search.hpp does: #ifdef BOOST_GRAPH_USE_MPI # include <boost/graph/distributed/breadth_first_search.hpp> #endif at its end, and boost/graph/distributed/breadth_first_search.hpp does: #ifndef BOOST_GRAPH_USE_MPI #error "Parallel BGL files should not be included unless <boost/graph/use_mpi.hpp> has been included" #endif #include <boost/graph/breadth_first_search.hpp> at its very start. So these headers include each other. One would naively think that only the parallel header should include the normal one, but one is not well versed in graph library architectures, so one might quite possibly be missing something, or a lot. Comments?
On 10/5/2018 8:18 PM, Peter Dimov via Boost wrote:
Currently we have a dependency cycle between `graph` and `graph_parallel`. It's not the only one, but it's notable with the fact that both libraries are buildable and we therefore can't ignore it as we can ignore cycles between header-only libraries that at release time become a single directory.
In the particular, this impedes efforts to generate CMake config files. Under one scheme I'm currently working on, for instance, Boost.Python can't be installed because it tries to install Graph, which tries to install GraphParallel, which tries to install Graph.
The cycle between these two is an interesting one. boost/graph/breadth_first_search.hpp does:
#ifdef BOOST_GRAPH_USE_MPI # include <boost/graph/distributed/breadth_first_search.hpp> #endif
at its end, and boost/graph/distributed/breadth_first_search.hpp does:
#ifndef BOOST_GRAPH_USE_MPI #error "Parallel BGL files should not be included unless <boost/graph/use_mpi.hpp> has been included" #endif
#include <boost/graph/breadth_first_search.hpp>
at its very start. So these headers include each other.
One would naively think that only the parallel header should include the normal one, but one is not well versed in graph library architectures, so one might quite possibly be missing something, or a lot.
I believe you are correct here. Boost Graph should not depend on Boost Graph Parallel in any situation. Boost Graph Parallel is a library end-users should be able to use as they choose and Boost Graph itself should be usable without ever bringing in the Parallel code.
Comments?
On Fri, Oct 05, 2018 at 10:43:23PM -0400, Edward Diener via Boost wrote:
On 10/5/2018 8:18 PM, Peter Dimov via Boost wrote:
Currently we have a dependency cycle between `graph` and `graph_parallel`. It's not the only one, but it's notable with the fact that both libraries are buildable and we therefore can't ignore it as we can ignore cycles between header-only libraries that at release time become a single directory.
In the particular, this impedes efforts to generate CMake config files. Under one scheme I'm currently working on, for instance, Boost.Python can't be installed because it tries to install Graph, which tries to install GraphParallel, which tries to install Graph.
The cycle between these two is an interesting one. boost/graph/breadth_first_search.hpp does:
#ifdef BOOST_GRAPH_USE_MPI # include <boost/graph/distributed/breadth_first_search.hpp> #endif
at its end, and boost/graph/distributed/breadth_first_search.hpp does:
#ifndef BOOST_GRAPH_USE_MPI #error "Parallel BGL files should not be included unless <boost/graph/use_mpi.hpp> has been included" #endif
#include <boost/graph/breadth_first_search.hpp>
at its very start. So these headers include each other.
One would naively think that only the parallel header should include the normal one, but one is not well versed in graph library architectures, so one might quite possibly be missing something, or a lot.
I believe you are correct here. Boost Graph should not depend on Boost Graph Parallel in any situation. Boost Graph Parallel is a library end-users should be able to use as they choose and Boost Graph itself should be usable without ever bringing in the Parallel code.
I concur. I worked with these two libraries a few years ago, and my recollection is Boost Graph was implemented first and has no dependency on Boost Graph Parallel. At least for my use cases, this was true. Karen. -- Karen Shaeffer The subconscious mind is driven by your deeply Neuralscape Services held beliefs -- not your deeply held desires.
On Sat, 6 Oct 2018, Peter Dimov via Boost wrote:
Currently we have a dependency cycle between `graph` and `graph_parallel`. It's not the only one, but it's notable with the fact that both libraries are buildable and we therefore can't ignore it as we can ignore cycles between header-only libraries that at release time become a single directory.
In the particular, this impedes efforts to generate CMake config files. Under one scheme I'm currently working on, for instance, Boost.Python can't be installed because it tries to install Graph, which tries to install GraphParallel, which tries to install Graph.
The cycle between these two is an interesting one. boost/graph/breadth_first_search.hpp does:
#ifdef BOOST_GRAPH_USE_MPI # include <boost/graph/distributed/breadth_first_search.hpp> #endif
at its end, and boost/graph/distributed/breadth_first_search.hpp does:
#ifndef BOOST_GRAPH_USE_MPI #error "Parallel BGL files should not be included unless <boost/graph/use_mpi.hpp> has been included" #endif
#include <boost/graph/breadth_first_search.hpp>
at its very start. So these headers include each other.
One would naively think that only the parallel header should include the normal one, but one is not well versed in graph library architectures, so one might quite possibly be missing something, or a lot.
Comments?
I don't know why you are calling this a dependency. The file declares the sequential stuff, and if the parallel version is available/requested (BOOST_GRAPH_USE_MPI) makes it available as well. So you don't need the parallel version to use the sequential one. -- Marc Glisse
Marc Glisse wrote:
On Sat, 6 Oct 2018, Peter Dimov via Boost wrote:
Currently we have a dependency cycle between `graph` and `graph_parallel`.
...
I don't know why you are calling this a dependency.
Boostdep is calling it a dependency, not me. I could hand-edit the dependency file each time after it's generated, in principle, or even script that, but I am not a fan of adding even more special cases to our process. And in either case, when BOOST_GRAPH_USE_MPI is defined, both headers _include each other_.
participants (4)
-
Edward Diener -
Karen Shaeffer -
Marc Glisse -
Peter Dimov