From 027d89091bfec53c700101ddfa9ab174a48923e1 Mon Sep 17 00:00:00 2001
From: "philblunsom@gmail.com"
 <philblunsom@gmail.com@ec762483-ff6d-05da-a07a-a48fb63a330f>
Date: Fri, 16 Jul 2010 04:50:35 +0000
Subject: working on mpi implementation.

git-svn-id: https://ws10smt.googlecode.com/svn/trunk@283 ec762483-ff6d-05da-a07a-a48fb63a330f
---
 gi/pyp-topics/src/Makefile.mpi          |   3 +-
 gi/pyp-topics/src/contexts_corpus.cc    |   4 +-
 gi/pyp-topics/src/contexts_corpus.hh    |   2 +
 gi/pyp-topics/src/makefile.depend       | 228 +++++++++++++++++++++++---
 gi/pyp-topics/src/mpi-pyp-topics.cc     | 148 ++++++++++-------
 gi/pyp-topics/src/mpi-pyp-topics.hh     |  17 +-
 gi/pyp-topics/src/mpi-pyp.hh            | 273 ++++++++++++++++++++++++++++++--
 gi/pyp-topics/src/mpi-train-contexts.cc |   2 +-
 gi/pyp-topics/src/pyp.hh                |   9 +-
 9 files changed, 582 insertions(+), 104 deletions(-)

(limited to 'gi')

diff --git a/gi/pyp-topics/src/Makefile.mpi b/gi/pyp-topics/src/Makefile.mpi
index 8c859881..b7b8a290 100644
--- a/gi/pyp-topics/src/Makefile.mpi
+++ b/gi/pyp-topics/src/Makefile.mpi
@@ -16,7 +16,8 @@ mpi-pyp-contexts-train: mpi-train-contexts.o $(local_objs)
 
 .PHONY: depend echo
 depend:
-	$(CXX) -MM $(CXXFLAGS) *.cc *.c | sed 's/^\(.*\.o:\)/obj\/\1/' > makefile.depend
+#$(CXX) -MM $(CXXFLAGS) *.cc *.c | sed 's/^\(.*\.o:\)/obj\/\1/' > makefile.depend
+	$(CXX) -MM $(CXXFLAGS) *.cc *.c > makefile.depend
 
 clean:
 	rm -f *.o
diff --git a/gi/pyp-topics/src/contexts_corpus.cc b/gi/pyp-topics/src/contexts_corpus.cc
index 26d5718a..1cf69429 100644
--- a/gi/pyp-topics/src/contexts_corpus.cc
+++ b/gi/pyp-topics/src/contexts_corpus.cc
@@ -28,7 +28,7 @@ void read_callback(const ContextsLexer::PhraseContextsType& new_contexts, void*
   Document* doc(new Document());
 
   //cout << "READ: " << new_contexts.phrase << "\t";
-  for (int i=0; i < new_contexts.counts.size(); ++i) {
+  for (int i=0; i < (int)new_contexts.counts.size(); ++i) {
     int cache_word_count = corpus_ptr->m_dict.max();
 
     //string context_str = corpus_ptr->m_dict.toString(new_contexts.contexts[i]);
@@ -101,7 +101,7 @@ void filter_callback(const ContextsLexer::PhraseContextsType& new_contexts, void
 
   map<string,int>* context_counts = (static_cast<map<string,int>*>(extra));
 
-  for (int i=0; i < new_contexts.counts.size(); ++i) {
+  for (int i=0; i < (int)new_contexts.counts.size(); ++i) {
     int context_index = new_contexts.counts.at(i).first;
     int count = new_contexts.counts.at(i).second;
     //int count = new_contexts.counts[i];
diff --git a/gi/pyp-topics/src/contexts_corpus.hh b/gi/pyp-topics/src/contexts_corpus.hh
index 66b71783..4d3d5669 100644
--- a/gi/pyp-topics/src/contexts_corpus.hh
+++ b/gi/pyp-topics/src/contexts_corpus.hh
@@ -63,6 +63,8 @@ public:
 
     std::vector<std::string> context2string(const WordID& id) const {
       std::vector<std::string> res;
+      assert (id >= 0);
+      std::cerr << m_dict.Convert(id) << std::endl;
       m_dict.AsVector(id, &res);
       return res;
     }
diff --git a/gi/pyp-topics/src/makefile.depend b/gi/pyp-topics/src/makefile.depend
index 88bab79e..88bc73c1 100644
--- a/gi/pyp-topics/src/makefile.depend
+++ b/gi/pyp-topics/src/makefile.depend
@@ -1,4 +1,4 @@
-obj/contexts_corpus.o: contexts_corpus.cc contexts_corpus.hh \
+contexts_corpus.o: contexts_corpus.cc contexts_corpus.hh \
   /Users/pblunsom/packages/include/boost/ptr_container/ptr_vector.hpp \
   /Users/pblunsom/packages/include/boost/ptr_container/ptr_sequence_adapter.hpp \
   /Users/pblunsom/packages/include/boost/ptr_container/detail/reversible_ptr_container.hpp \
@@ -432,7 +432,7 @@ obj/contexts_corpus.o: contexts_corpus.cc contexts_corpus.hh \
   /Users/pblunsom/packages/include/boost/type_traits/add_cv.hpp \
   /Users/pblunsom/packages/include/boost/type_traits/remove_volatile.hpp \
   /Users/pblunsom/packages/include/boost/type_traits/function_traits.hpp
-obj/contexts_lexer.o: contexts_lexer.cc contexts_lexer.h \
+contexts_lexer.o: contexts_lexer.cc contexts_lexer.h \
   ../../../decoder/dict.h \
   /Users/pblunsom/packages/include/boost/functional/hash.hpp \
   /Users/pblunsom/packages/include/boost/functional/hash/hash.hpp \
@@ -463,7 +463,7 @@ obj/contexts_lexer.o: contexts_lexer.cc contexts_lexer.h \
   /Users/pblunsom/packages/include/boost/detail/container_fwd.hpp \
   ../../../decoder/wordid.h ../../../decoder/filelib.h \
   ../../../decoder/gzstream.h
-obj/corpus.o: corpus.cc corpus.hh \
+corpus.o: corpus.cc corpus.hh \
   /Users/pblunsom/packages/include/boost/shared_ptr.hpp \
   /Users/pblunsom/packages/include/boost/smart_ptr/shared_ptr.hpp \
   /Users/pblunsom/packages/include/boost/config.hpp \
@@ -874,8 +874,8 @@ obj/corpus.o: corpus.cc corpus.hh \
   /Users/pblunsom/packages/include/boost/detail/is_incrementable.hpp \
   /Users/pblunsom/packages/include/boost/ptr_container/detail/void_ptr_iterator.hpp \
   gzstream.hh
-obj/gzstream.o: gzstream.cc gzstream.hh
-obj/mpi-pyp-topics.o: mpi-pyp-topics.cc \
+gzstream.o: gzstream.cc gzstream.hh
+mpi-pyp-topics.o: mpi-pyp-topics.cc \
   /Users/pblunsom/packages/include/boost/mpi/communicator.hpp \
   /Users/pblunsom/packages/include/boost/mpi/config.hpp \
   /Users/pblunsom/packages/include/boost/config.hpp \
@@ -1448,8 +1448,108 @@ obj/mpi-pyp-topics.o: mpi-pyp-topics.cc \
   /Users/pblunsom/packages/include/boost/random/detail/const_mod.hpp \
   /Users/pblunsom/packages/include/boost/random/detail/seed.hpp \
   /Users/pblunsom/packages/include/boost/mpi/environment.hpp mpi-pyp.hh \
-  log_add.h slice-sampler.h mt19937ar.h corpus.hh
-obj/mpi-train-contexts.o: mpi-train-contexts.cc \
+  /Users/pblunsom/packages/include/boost/tuple/tuple.hpp \
+  /Users/pblunsom/packages/include/boost/ref.hpp \
+  /Users/pblunsom/packages/include/boost/tuple/detail/tuple_basic.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/cv_traits.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/add_volatile.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/add_cv.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/remove_volatile.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/function_traits.hpp \
+  /Users/pblunsom/packages/include/boost/serialization/map.hpp \
+  /Users/pblunsom/packages/include/boost/serialization/utility.hpp \
+  /Users/pblunsom/packages/include/boost/serialization/collections_save_imp.hpp \
+  /Users/pblunsom/packages/include/boost/serialization/collections_load_imp.hpp \
+  /Users/pblunsom/packages/include/boost/serialization/detail/stack_constructor.hpp \
+  /Users/pblunsom/packages/include/boost/aligned_storage.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/type_with_alignment.hpp \
+  /Users/pblunsom/packages/include/boost/mpi.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives/all_gather.hpp \
+  /Users/pblunsom/packages/include/boost/serialization/vector.hpp \
+  /Users/pblunsom/packages/include/boost/serialization/collection_traits.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives/broadcast.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives_fwd.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives/gather.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives/all_reduce.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives/reduce.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/detail/computation_tree.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/operations.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives/all_to_all.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives/scatter.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives/scan.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/graph_communicator.hpp \
+  /Users/pblunsom/packages/include/boost/graph/graph_traits.hpp \
+  /Users/pblunsom/packages/include/boost/pending/property.hpp \
+  /Users/pblunsom/packages/include/boost/pending/detail/property.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/same_traits.hpp \
+  /Users/pblunsom/packages/include/boost/graph/properties.hpp \
+  /Users/pblunsom/packages/include/boost/property_map/property_map.hpp \
+  /Users/pblunsom/packages/include/boost/pending/cstddef.hpp \
+  /Users/pblunsom/packages/include/boost/concept_check.hpp \
+  /Users/pblunsom/packages/include/boost/concept/assert.hpp \
+  /Users/pblunsom/packages/include/boost/concept/detail/general.hpp \
+  /Users/pblunsom/packages/include/boost/concept/detail/has_constraints.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/conversion_traits.hpp \
+  /Users/pblunsom/packages/include/boost/concept/usage.hpp \
+  /Users/pblunsom/packages/include/boost/concept/detail/concept_def.hpp \
+  /Users/pblunsom/packages/include/boost/concept/detail/concept_undef.hpp \
+  /Users/pblunsom/packages/include/boost/concept_archetype.hpp \
+  /Users/pblunsom/packages/include/boost/property_map/vector_property_map.hpp \
+  /Users/pblunsom/packages/include/boost/graph/property_maps/constant_property_map.hpp \
+  /Users/pblunsom/packages/include/boost/graph/property_maps/null_property_map.hpp \
+  /Users/pblunsom/packages/include/boost/iterator/counting_iterator.hpp \
+  /Users/pblunsom/packages/include/boost/detail/numeric_traits.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/has_nothrow_assign.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/has_trivial_assign.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/has_nothrow_constructor.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/has_trivial_constructor.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/has_nothrow_copy.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/has_nothrow_destructor.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/has_trivial_destructor.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/has_virtual_destructor.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/is_signed.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/is_unsigned.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/is_compound.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/is_floating_point.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/is_member_object_pointer.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/is_object.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/is_stateless.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/rank.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/extent.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/remove_all_extents.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/function_traits.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/aligned_storage.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/floating_point_promotion.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/integral_promotion.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/promote.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/integral_promotion.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/floating_point_promotion.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/make_unsigned.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/is_signed.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/is_unsigned.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/make_signed.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/decay.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/is_complex.hpp \
+  /Users/pblunsom/packages/include/boost/detail/select_type.hpp \
+  /Users/pblunsom/packages/include/boost/graph/iteration_macros.hpp \
+  /Users/pblunsom/packages/include/boost/shared_array.hpp \
+  /Users/pblunsom/packages/include/boost/smart_ptr/shared_array.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/group.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/intercommunicator.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/nonblocking.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/skeleton_and_content.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/detail/forward_skeleton_iarchive.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/detail/forward_skeleton_oarchive.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/detail/ignore_iprimitive.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/detail/ignore_oprimitive.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/detail/content_oarchive.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/detail/broadcast_sc.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/detail/communicator_sc.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/timer.hpp pyp.hh \
+  slice-sampler.h log_add.h mt19937ar.h corpus.hh
+mpi-train-contexts.o: mpi-train-contexts.cc \
   /Users/pblunsom/packages/include/boost/program_options/parsers.hpp \
   /Users/pblunsom/packages/include/boost/program_options/config.hpp \
   /Users/pblunsom/packages/include/boost/config.hpp \
@@ -2064,8 +2164,100 @@ obj/mpi-train-contexts.o: mpi-train-contexts.cc \
   /Users/pblunsom/packages/include/boost/random/linear_congruential.hpp \
   /Users/pblunsom/packages/include/boost/random/detail/const_mod.hpp \
   /Users/pblunsom/packages/include/boost/random/detail/seed.hpp \
-  mpi-pyp.hh log_add.h slice-sampler.h mt19937ar.h corpus.hh \
-  contexts_corpus.hh contexts_lexer.h ../../../decoder/dict.h \
+  mpi-pyp.hh /Users/pblunsom/packages/include/boost/tuple/tuple.hpp \
+  /Users/pblunsom/packages/include/boost/tuple/detail/tuple_basic.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/cv_traits.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/add_cv.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/remove_volatile.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/function_traits.hpp \
+  /Users/pblunsom/packages/include/boost/serialization/map.hpp \
+  /Users/pblunsom/packages/include/boost/serialization/utility.hpp \
+  /Users/pblunsom/packages/include/boost/serialization/collections_save_imp.hpp \
+  /Users/pblunsom/packages/include/boost/serialization/collections_load_imp.hpp \
+  /Users/pblunsom/packages/include/boost/serialization/detail/stack_constructor.hpp \
+  /Users/pblunsom/packages/include/boost/aligned_storage.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/type_with_alignment.hpp \
+  /Users/pblunsom/packages/include/boost/mpi.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives/all_gather.hpp \
+  /Users/pblunsom/packages/include/boost/serialization/vector.hpp \
+  /Users/pblunsom/packages/include/boost/serialization/collection_traits.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives/broadcast.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives_fwd.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives/gather.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives/all_reduce.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives/reduce.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/detail/computation_tree.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/operations.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives/all_to_all.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives/scatter.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/collectives/scan.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/graph_communicator.hpp \
+  /Users/pblunsom/packages/include/boost/graph/graph_traits.hpp \
+  /Users/pblunsom/packages/include/boost/pending/property.hpp \
+  /Users/pblunsom/packages/include/boost/pending/detail/property.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/same_traits.hpp \
+  /Users/pblunsom/packages/include/boost/graph/properties.hpp \
+  /Users/pblunsom/packages/include/boost/property_map/property_map.hpp \
+  /Users/pblunsom/packages/include/boost/pending/cstddef.hpp \
+  /Users/pblunsom/packages/include/boost/concept_check.hpp \
+  /Users/pblunsom/packages/include/boost/concept/assert.hpp \
+  /Users/pblunsom/packages/include/boost/concept/detail/general.hpp \
+  /Users/pblunsom/packages/include/boost/concept/detail/has_constraints.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/conversion_traits.hpp \
+  /Users/pblunsom/packages/include/boost/concept/usage.hpp \
+  /Users/pblunsom/packages/include/boost/concept/detail/concept_def.hpp \
+  /Users/pblunsom/packages/include/boost/concept/detail/concept_undef.hpp \
+  /Users/pblunsom/packages/include/boost/concept_archetype.hpp \
+  /Users/pblunsom/packages/include/boost/property_map/vector_property_map.hpp \
+  /Users/pblunsom/packages/include/boost/graph/property_maps/constant_property_map.hpp \
+  /Users/pblunsom/packages/include/boost/graph/property_maps/null_property_map.hpp \
+  /Users/pblunsom/packages/include/boost/iterator/counting_iterator.hpp \
+  /Users/pblunsom/packages/include/boost/detail/numeric_traits.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/has_nothrow_assign.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/has_trivial_assign.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/has_nothrow_constructor.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/has_trivial_constructor.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/has_nothrow_copy.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/has_nothrow_destructor.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/has_virtual_destructor.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/is_compound.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/is_floating_point.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/is_member_object_pointer.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/is_object.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/is_stateless.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/rank.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/extent.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/remove_all_extents.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/function_traits.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/aligned_storage.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/floating_point_promotion.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/integral_promotion.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/promote.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/integral_promotion.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/floating_point_promotion.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/make_signed.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/decay.hpp \
+  /Users/pblunsom/packages/include/boost/type_traits/is_complex.hpp \
+  /Users/pblunsom/packages/include/boost/detail/select_type.hpp \
+  /Users/pblunsom/packages/include/boost/graph/iteration_macros.hpp \
+  /Users/pblunsom/packages/include/boost/shared_array.hpp \
+  /Users/pblunsom/packages/include/boost/smart_ptr/shared_array.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/group.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/intercommunicator.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/nonblocking.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/skeleton_and_content.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/detail/forward_skeleton_iarchive.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/detail/forward_skeleton_oarchive.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/detail/ignore_iprimitive.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/detail/ignore_oprimitive.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/detail/content_oarchive.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/detail/broadcast_sc.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/detail/communicator_sc.hpp \
+  /Users/pblunsom/packages/include/boost/mpi/timer.hpp pyp.hh \
+  slice-sampler.h log_add.h mt19937ar.h corpus.hh contexts_corpus.hh \
+  contexts_lexer.h ../../../decoder/dict.h \
   /Users/pblunsom/packages/include/boost/functional/hash.hpp \
   /Users/pblunsom/packages/include/boost/functional/hash/hash.hpp \
   /Users/pblunsom/packages/include/boost/functional/hash/hash_fwd.hpp \
@@ -2078,7 +2270,7 @@ obj/mpi-train-contexts.o: mpi-train-contexts.cc \
   /Users/pblunsom/packages/include/boost/functional/hash/extensions.hpp \
   /Users/pblunsom/packages/include/boost/detail/container_fwd.hpp \
   ../../../decoder/wordid.h gzstream.hh
-obj/pyp-topics.o: pyp-topics.cc timing.h clock_gettime_stub.c pyp-topics.hh \
+pyp-topics.o: pyp-topics.cc timing.h clock_gettime_stub.c pyp-topics.hh \
   /Users/pblunsom/packages/include/boost/ptr_container/ptr_vector.hpp \
   /Users/pblunsom/packages/include/boost/ptr_container/ptr_sequence_adapter.hpp \
   /Users/pblunsom/packages/include/boost/ptr_container/detail/reversible_ptr_container.hpp \
@@ -2484,7 +2676,7 @@ obj/pyp-topics.o: pyp-topics.cc timing.h clock_gettime_stub.c pyp-topics.hh \
   /Users/pblunsom/packages/include/boost/random/linear_congruential.hpp \
   /Users/pblunsom/packages/include/boost/random/detail/const_mod.hpp \
   /Users/pblunsom/packages/include/boost/random/detail/seed.hpp pyp.hh \
-  log_add.h slice-sampler.h mt19937ar.h corpus.hh \
+  slice-sampler.h log_add.h mt19937ar.h corpus.hh \
   /Users/pblunsom/packages/include/boost/shared_ptr.hpp \
   /Users/pblunsom/packages/include/boost/smart_ptr/shared_ptr.hpp \
   /Users/pblunsom/packages/include/boost/config/no_tr1/memory.hpp \
@@ -2658,7 +2850,7 @@ obj/pyp-topics.o: pyp-topics.cc timing.h clock_gettime_stub.c pyp-topics.hh \
   /Users/pblunsom/packages/include/boost/smart_ptr/scoped_ptr.hpp \
   /Users/pblunsom/packages/include/boost/type_traits/is_fundamental.hpp \
   /Users/pblunsom/packages/include/boost/thread/condition.hpp
-obj/train-contexts.o: train-contexts.cc \
+train-contexts.o: train-contexts.cc \
   /Users/pblunsom/packages/include/boost/program_options/parsers.hpp \
   /Users/pblunsom/packages/include/boost/program_options/config.hpp \
   /Users/pblunsom/packages/include/boost/config.hpp \
@@ -3140,7 +3332,7 @@ obj/train-contexts.o: train-contexts.cc \
   /Users/pblunsom/packages/include/boost/random/linear_congruential.hpp \
   /Users/pblunsom/packages/include/boost/random/detail/const_mod.hpp \
   /Users/pblunsom/packages/include/boost/random/detail/seed.hpp pyp.hh \
-  log_add.h slice-sampler.h mt19937ar.h corpus.hh workers.hh \
+  slice-sampler.h log_add.h mt19937ar.h corpus.hh workers.hh \
   /Users/pblunsom/packages/include/boost/bind.hpp \
   /Users/pblunsom/packages/include/boost/bind/bind.hpp \
   /Users/pblunsom/packages/include/boost/is_placeholder.hpp \
@@ -3275,7 +3467,7 @@ obj/train-contexts.o: train-contexts.cc \
   /Users/pblunsom/packages/include/boost/functional/hash/extensions.hpp \
   /Users/pblunsom/packages/include/boost/detail/container_fwd.hpp \
   ../../../decoder/wordid.h gzstream.hh
-obj/train.o: train.cc \
+train.o: train.cc \
   /Users/pblunsom/packages/include/boost/program_options/parsers.hpp \
   /Users/pblunsom/packages/include/boost/program_options/config.hpp \
   /Users/pblunsom/packages/include/boost/config.hpp \
@@ -3757,7 +3949,7 @@ obj/train.o: train.cc \
   /Users/pblunsom/packages/include/boost/random/linear_congruential.hpp \
   /Users/pblunsom/packages/include/boost/random/detail/const_mod.hpp \
   /Users/pblunsom/packages/include/boost/random/detail/seed.hpp pyp.hh \
-  log_add.h slice-sampler.h mt19937ar.h corpus.hh workers.hh \
+  slice-sampler.h log_add.h mt19937ar.h corpus.hh workers.hh \
   /Users/pblunsom/packages/include/boost/bind.hpp \
   /Users/pblunsom/packages/include/boost/bind/bind.hpp \
   /Users/pblunsom/packages/include/boost/is_placeholder.hpp \
@@ -3892,6 +4084,6 @@ obj/train.o: train.cc \
   /Users/pblunsom/packages/include/boost/functional/hash/extensions.hpp \
   /Users/pblunsom/packages/include/boost/detail/container_fwd.hpp \
   ../../../decoder/wordid.h gzstream.hh
-obj/clock_gettime_stub.o: clock_gettime_stub.c
-obj/gammadist.o: gammadist.c gammadist.h mt19937ar.h
-obj/mt19937ar.o: mt19937ar.c mt19937ar.h
+clock_gettime_stub.o: clock_gettime_stub.c
+gammadist.o: gammadist.c gammadist.h mt19937ar.h
+mt19937ar.o: mt19937ar.c mt19937ar.h
diff --git a/gi/pyp-topics/src/mpi-pyp-topics.cc b/gi/pyp-topics/src/mpi-pyp-topics.cc
index 2ad28278..4525302e 100644
--- a/gi/pyp-topics/src/mpi-pyp-topics.cc
+++ b/gi/pyp-topics/src/mpi-pyp-topics.cc
@@ -4,7 +4,7 @@
 #include "mpi-pyp-topics.hh"
 
 //#include <boost/date_time/posix_time/posix_time_types.hpp>
-void PYPTopics::sample_corpus(const Corpus& corpus, int samples,
+void MPIPYPTopics::sample_corpus(const Corpus& corpus, int samples,
                               int freq_cutoff_start, int freq_cutoff_end,
                               int freq_cutoff_interval,
                               int max_contexts_per_document) {
@@ -23,33 +23,33 @@ void PYPTopics::sample_corpus(const Corpus& corpus, int samples,
   }
   int local_documents = m_mpi_end - m_mpi_start;
 
-
   if (!m_backoff.get()) {
     m_word_pyps.clear();
-    m_word_pyps.push_back(PYPs());
+    m_word_pyps.push_back(MPIPYPs());
   }
 
   if (m_am_root) std::cerr << "\n Training with " << m_word_pyps.size()-1 << " backoff level"
-    << (m_word_pyps.size()==2 ? ":" : "s:") << std::endl;
+    << (m_word_pyps.size()>1 ? ":" : "s:") << std::endl;
 
-  for (int i=0; i<(int)m_word_pyps.size(); ++i)
-  {
+  for (int i=0; i<(int)m_word_pyps.size(); ++i) {
     m_word_pyps.at(i).reserve(m_num_topics);
     for (int j=0; j<m_num_topics; ++j)
-      m_word_pyps.at(i).push_back(new PYP<int>(0.5, 1.0));
+      m_word_pyps.at(i).push_back(new MPIPYP<int>(0.5, 1.0));
   }
   if (m_am_root) std::cerr << std::endl;
 
-  m_document_pyps.reserve(corpus.num_documents());
-  for (int j=0; j<corpus.num_documents(); ++j)
+  m_document_pyps.reserve(local_documents);
+  //m_document_pyps.reserve(corpus.num_documents());
+  //for (int j=0; j<corpus.num_documents(); ++j)
+  for (int j=0; j<local_documents; ++j)
     m_document_pyps.push_back(new PYP<int>(0.5, 1.0));
 
   m_topic_p0 = 1.0/m_num_topics;
   m_term_p0 = 1.0/corpus.num_types();
   m_backoff_p0 = 1.0/corpus.num_documents();
 
-  if (m_am_root) std::cerr << " Documents: " << corpus.num_documents() << " Terms: "
-    << corpus.num_types() << std::endl;
+  if (m_am_root) std::cerr << " Documents: " << corpus.num_documents() << "(" 
+    << local_documents << ")" << " Terms: " << corpus.num_types() << std::endl;
 
   int frequency_cutoff = freq_cutoff_start;
   if (m_am_root) std::cerr << " Context frequency cutoff set to " << frequency_cutoff << std::endl;
@@ -57,13 +57,16 @@ void PYPTopics::sample_corpus(const Corpus& corpus, int samples,
   timer.Reset();
   // Initialisation pass
   int document_id=0, topic_counter=0;
-  for (Corpus::const_iterator corpusIt=corpus.begin();
-       corpusIt != corpus.end(); ++corpusIt, ++document_id) {
-    m_corpus_topics.push_back(DocumentTopics(corpusIt->size(), 0));
+  for (int i=0; i<local_documents; ++i) {
+    document_id = i+m_mpi_start;
+
+  //for (Corpus::const_iterator corpusIt=corpus.begin();
+  //     corpusIt != corpus.end(); ++corpusIt, ++document_id) {
+    m_corpus_topics.push_back(DocumentTopics(corpus.at(document_id).size(), 0));
 
     int term_index=0;
-    for (Document::const_iterator docIt=corpusIt->begin();
-         docIt != corpusIt->end(); ++docIt, ++term_index) {
+    for (Document::const_iterator docIt=corpus.at(document_id).begin();
+         docIt != corpus.at(document_id).end(); ++docIt, ++term_index) {
       topic_counter++;
       Term term = *docIt;
 
@@ -80,21 +83,41 @@ void PYPTopics::sample_corpus(const Corpus& corpus, int samples,
 
         if (m_use_topic_pyp) {
           F p0 = m_topic_pyp.prob(new_topic, m_topic_p0);
-          int table_delta = m_document_pyps[document_id].increment(new_topic, p0);
+          int table_delta = m_document_pyps.at(i).increment(new_topic, p0);
           if (table_delta)
             m_topic_pyp.increment(new_topic, m_topic_p0);
         }
-        else m_document_pyps[document_id].increment(new_topic, m_topic_p0);
+        else m_document_pyps.at(i).increment(new_topic, m_topic_p0);
       }
 
-      m_corpus_topics[document_id][term_index] = new_topic;
+      m_corpus_topics.at(i).at(term_index) = new_topic;
+    }
+  }
+
+  // Synchronise the topic->word counds across the processes.
+  for (std::vector<MPIPYPs>::iterator levelIt=m_word_pyps.begin();
+       levelIt != m_word_pyps.end(); ++levelIt) {
+    for (MPIPYPs::iterator pypIt=levelIt->begin();
+         pypIt != levelIt->end(); ++pypIt) {
+      if (!m_am_root) boost::mpi::communicator().barrier();
+      std::cerr << "Before Sync Process " << m_rank << ":";
+      pypIt->debug_info(std::cerr); std::cerr << std::endl;
+      if (m_am_root) boost::mpi::communicator().barrier();
+
+      pypIt->synchronise();
+
+      if (!m_am_root) boost::mpi::communicator().barrier();
+      std::cerr << "After Sync Process " << m_rank << ":";
+      pypIt->debug_info(std::cerr); std::cerr << std::endl;
+      if (m_am_root) boost::mpi::communicator().barrier();
     }
   }
+
   if (m_am_root) std::cerr << "  Initialized in " << timer.Elapsed() << " seconds\n";
 
   int* randomDocIndices = new int[local_documents];
   for (int i = 0; i < local_documents; ++i)
-	  randomDocIndices[i] = i+m_mpi_start;
+	  randomDocIndices[i] = i;
 
   // Sampling phase
   for (int curr_sample=0; curr_sample < samples; ++curr_sample) {
@@ -110,8 +133,8 @@ void PYPTopics::sample_corpus(const Corpus& corpus, int samples,
     // Randomize the corpus indexing array
     int tmp;
     int processed_terms=0;
-    for (int i = local_documents-1; i > 0; --i) {
-        //i+1 since j \in [0,i] but rnd() \in [0,1)
+    for (int i = (local_documents-1); i > 0; --i) {
+      //i+1 since j \in [0,i] but rnd() \in [0,1)
     	int j = (int)(rnd() * (i+1));
       assert(j >= 0 && j <= i);
      	tmp = randomDocIndices[i];
@@ -120,15 +143,17 @@ void PYPTopics::sample_corpus(const Corpus& corpus, int samples,
     }
 
     // for each document in the corpus
-    int document_id;
-    for (int i=0; i<local_documents; ++i) {
-    	document_id = randomDocIndices[i];
+    for (int rand_doc=0; rand_doc<local_documents; ++rand_doc) {
+    	int doc_index = randomDocIndices[rand_doc];
+    	int document_id = doc_index + m_mpi_start;
+      const Document& doc = corpus.at(document_id);
 
       // for each term in the document
       int term_index=0;
-      Document::const_iterator docEnd = corpus.at(document_id).end();
-      for (Document::const_iterator docIt=corpus.at(document_id).begin();
+      Document::const_iterator docEnd = doc.end();
+      for (Document::const_iterator docIt=doc.begin();
            docIt != docEnd; ++docIt, ++term_index) {
+
         if (max_contexts_per_document && term_index > max_contexts_per_document)
           break;
         
@@ -140,36 +165,49 @@ void PYPTopics::sample_corpus(const Corpus& corpus, int samples,
         processed_terms++;
 
         // remove the prevous topic from the PYPs
-        int current_topic = m_corpus_topics[document_id][term_index];
+        int current_topic = m_corpus_topics.at(doc_index).at(term_index);
         // a negative label mean that term hasn't been sampled yet
         if (current_topic >= 0) {
           decrement(term, current_topic);
 
-          int table_delta = m_document_pyps[document_id].decrement(current_topic);
+          int table_delta = m_document_pyps.at(doc_index).decrement(current_topic);
           if (m_use_topic_pyp && table_delta < 0)
             m_topic_pyp.decrement(current_topic);
         }
 
         // sample a new_topic
-        int new_topic = sample(document_id, term);
+        int new_topic = sample(doc_index, term);
 
         // add the new topic to the PYPs
-        m_corpus_topics[document_id][term_index] = new_topic;
+        m_corpus_topics.at(doc_index).at(term_index) = new_topic;
         increment(term, new_topic);
 
         if (m_use_topic_pyp) {
           F p0 = m_topic_pyp.prob(new_topic, m_topic_p0);
-          int table_delta = m_document_pyps[document_id].increment(new_topic, p0);
+          int table_delta = m_document_pyps.at(doc_index).increment(new_topic, p0);
           if (table_delta)
             m_topic_pyp.increment(new_topic, m_topic_p0);
         }
-        else m_document_pyps[document_id].increment(new_topic, m_topic_p0);
+        else m_document_pyps.at(doc_index).increment(new_topic, m_topic_p0);
       }
       if (document_id && document_id % 10000 == 0) {
         if (m_am_root) std::cerr << "."; std::cerr.flush();
       }
     }
     m_world.barrier();
+    // Synchronise the topic->word counds across the processes.
+    for (std::vector<MPIPYPs>::iterator levelIt=m_word_pyps.begin();
+         levelIt != m_word_pyps.end(); ++levelIt) {
+      for (MPIPYPs::iterator pypIt=levelIt->begin();
+           pypIt != levelIt->end(); ++pypIt) {
+        std::cerr << "Before Sync Process " << m_rank << ":";
+        pypIt->debug_info(std::cerr); std::cerr << std::endl;
+        pypIt->synchronise();
+        std::cerr << "After Sync Process " << m_rank << ":";
+        pypIt->debug_info(std::cerr); std::cerr << std::endl;
+      }
+    }
+
     if (m_am_root) std::cerr << " ||| sampled " << processed_terms << " terms.";
 
     if (curr_sample != 0 && curr_sample % 10 == 0) {
@@ -179,9 +217,9 @@ void PYPTopics::sample_corpus(const Corpus& corpus, int samples,
 
       // resample the hyperparamters
       F log_p=0.0;
-      for (std::vector<PYPs>::iterator levelIt=m_word_pyps.begin();
+      for (std::vector<MPIPYPs>::iterator levelIt=m_word_pyps.begin();
            levelIt != m_word_pyps.end(); ++levelIt) {
-        for (PYPs::iterator pypIt=levelIt->begin();
+        for (MPIPYPs::iterator pypIt=levelIt->begin();
              pypIt != levelIt->end(); ++pypIt) {
           pypIt->resample_prior();
           log_p += pypIt->log_restaurant_prob();
@@ -206,7 +244,7 @@ void PYPTopics::sample_corpus(const Corpus& corpus, int samples,
       int k=0;
       if (m_am_root) std::cerr << "Topics distribution: ";
       std::cerr.precision(2);
-      for (PYPs::iterator pypIt=m_word_pyps.front().begin();
+      for (MPIPYPs::iterator pypIt=m_word_pyps.front().begin();
            pypIt != m_word_pyps.front().end(); ++pypIt, ++k) {
         if (m_am_root && k % 5 == 0) std::cerr << std::endl << '\t';
         if (m_am_root) std::cerr << "<" << k << ":" << pypIt->num_customers() << ","
@@ -220,8 +258,8 @@ void PYPTopics::sample_corpus(const Corpus& corpus, int samples,
 }
 
 
-void PYPTopics::decrement(const Term& term, int topic, int level) {
-  //std::cerr << "PYPTopics::decrement(" << term << "," << topic << "," << level << ")" << std::endl;
+void MPIPYPTopics::decrement(const Term& term, int topic, int level) {
+  //std::cerr << "MPIPYPTopics::decrement(" << term << "," << topic << "," << level << ")" << std::endl;
   m_word_pyps.at(level).at(topic).decrement(term);
   if (m_backoff.get()) {
     Term backoff_term = (*m_backoff)[term];
@@ -230,8 +268,8 @@ void PYPTopics::decrement(const Term& term, int topic, int level) {
   }
 }
 
-void PYPTopics::increment(const Term& term, int topic, int level) {
-  //std::cerr << "PYPTopics::increment(" << term << "," << topic << "," << level << ")" << std::endl;
+void MPIPYPTopics::increment(const Term& term, int topic, int level) {
+  //std::cerr << "MPIPYPTopics::increment(" << term << "," << topic << "," << level << ")" << std::endl;
   m_word_pyps.at(level).at(topic).increment(term, word_pyps_p0(term, topic, level));
 
   if (m_backoff.get()) {
@@ -241,7 +279,7 @@ void PYPTopics::increment(const Term& term, int topic, int level) {
   }
 }
 
-int PYPTopics::sample(const DocumentId& doc, const Term& term) {
+int MPIPYPTopics::sample(const DocumentId& doc, const Term& term) {
   // First pass: collect probs
   F sum=0.0;
   std::vector<F> sums;
@@ -252,7 +290,7 @@ int PYPTopics::sample(const DocumentId& doc, const Term& term) {
     if (m_use_topic_pyp) topic_prob = m_topic_pyp.prob(k, m_topic_p0);
 
     //F p_k_d = m_document_pyps[doc].prob(k, topic_prob);
-    F p_k_d = m_document_pyps[doc].unnormalised_prob(k, topic_prob);
+    F p_k_d = m_document_pyps.at(doc).unnormalised_prob(k, topic_prob);
 
     sum += (p_w_k*p_k_d);
     sums.push_back(sum);
@@ -266,9 +304,9 @@ int PYPTopics::sample(const DocumentId& doc, const Term& term) {
   assert(false);
 }
 
-PYPTopics::F PYPTopics::word_pyps_p0(const Term& term, int topic, int level) const {
+MPIPYPTopics::F MPIPYPTopics::word_pyps_p0(const Term& term, int topic, int level) const {
   //for (int i=0; i<level+1; ++i) std::cerr << "  ";
-  //std::cerr << "PYPTopics::word_pyps_p0(" << term << "," << topic << "," << level << ")" << std::endl;
+  //std::cerr << "MPIPYPTopics::word_pyps_p0(" << term << "," << topic << "," << level << ")" << std::endl;
 
   F p0 = m_term_p0;
   if (m_backoff.get()) {
@@ -283,24 +321,24 @@ PYPTopics::F PYPTopics::word_pyps_p0(const Term& term, int topic, int level) con
       p0 = m_term_p0;
   }
   //for (int i=0; i<level+1; ++i) std::cerr << "  ";
-  //std::cerr << "PYPTopics::word_pyps_p0(" << term << "," << topic << "," << level << ") = " << p0 << std::endl;
+  //std::cerr << "MPIPYPTopics::word_pyps_p0(" << term << "," << topic << "," << level << ") = " << p0 << std::endl;
   return p0;
 }
 
-PYPTopics::F PYPTopics::prob(const Term& term, int topic, int level) const {
+MPIPYPTopics::F MPIPYPTopics::prob(const Term& term, int topic, int level) const {
   //for (int i=0; i<level+1; ++i) std::cerr << "  ";
-  //std::cerr << "PYPTopics::prob(" << term << "," << topic << "," << level << " " << factor << ")" << std::endl;
+  //std::cerr << "MPIPYPTopics::prob(" << term << "," << topic << "," << level << " " << factor << ")" << std::endl;
 
   F p0 = word_pyps_p0(term, topic, level);
   F p_w_k = m_word_pyps.at(level).at(topic).prob(term, p0);
 
   //for (int i=0; i<level+1; ++i) std::cerr << "  ";
-  //std::cerr << "PYPTopics::prob(" << term << "," << topic << "," << level << ") = " << p_w_k << std::endl;
+  //std::cerr << "MPIPYPTopics::prob(" << term << "," << topic << "," << level << ") = " << p_w_k << std::endl;
 
   return p_w_k;
 }
 
-int PYPTopics::max_topic() const {
+int MPIPYPTopics::max_topic() const {
   if (!m_use_topic_pyp)
     return -1;
 
@@ -317,8 +355,8 @@ int PYPTopics::max_topic() const {
   return current_topic;
 }
 
-int PYPTopics::max(const DocumentId& doc) const {
-  //std::cerr << "PYPTopics::max(" << doc << "," << term << ")" << std::endl;
+int MPIPYPTopics::max(const DocumentId& doc) const {
+  //std::cerr << "MPIPYPTopics::max(" << doc << "," << term << ")" << std::endl;
   // collect probs
   F current_max=0.0;
   int current_topic=-1;
@@ -342,8 +380,8 @@ int PYPTopics::max(const DocumentId& doc) const {
   return current_topic;
 }
 
-int PYPTopics::max(const DocumentId& doc, const Term& term) const {
-  //std::cerr << "PYPTopics::max(" << doc << "," << term << ")" << std::endl;
+int MPIPYPTopics::max(const DocumentId& doc, const Term& term) const {
+  //std::cerr << "MPIPYPTopics::max(" << doc << "," << term << ")" << std::endl;
   // collect probs
   F current_max=0.0;
   int current_topic=-1;
@@ -368,7 +406,7 @@ int PYPTopics::max(const DocumentId& doc, const Term& term) const {
   return current_topic;
 }
 
-std::ostream& PYPTopics::print_document_topics(std::ostream& out) const {
+std::ostream& MPIPYPTopics::print_document_topics(std::ostream& out) const {
   for (CorpusTopics::const_iterator corpusIt=m_corpus_topics.begin();
        corpusIt != m_corpus_topics.end(); ++corpusIt) {
     int term_index=0;
@@ -382,7 +420,7 @@ std::ostream& PYPTopics::print_document_topics(std::ostream& out) const {
   return out;
 }
 
-std::ostream& PYPTopics::print_topic_terms(std::ostream& out) const {
+std::ostream& MPIPYPTopics::print_topic_terms(std::ostream& out) const {
   for (PYPs::const_iterator pypsIt=m_word_pyps.front().begin();
        pypsIt != m_word_pyps.front().end(); ++pypsIt) {
     int term_index=0;
diff --git a/gi/pyp-topics/src/mpi-pyp-topics.hh b/gi/pyp-topics/src/mpi-pyp-topics.hh
index 5da35d82..a85a776d 100644
--- a/gi/pyp-topics/src/mpi-pyp-topics.hh
+++ b/gi/pyp-topics/src/mpi-pyp-topics.hh
@@ -1,5 +1,5 @@
-#ifndef PYP_TOPICS_HH
-#define PYP_TOPICS_HH
+#ifndef MPI_PYP_TOPICS_HH
+#define MPI_PYP_TOPICS_HH
 
 #include <vector>
 #include <iostream>
@@ -14,14 +14,14 @@
 #include "mpi-pyp.hh"
 #include "corpus.hh"
 
-class PYPTopics {
+class MPIPYPTopics {
 public:
   typedef std::vector<int> DocumentTopics;
   typedef std::vector<DocumentTopics> CorpusTopics;
   typedef double F;
 
 public:
-  PYPTopics(int num_topics, bool use_topic_pyp=false, unsigned long seed = 0) 
+  MPIPYPTopics(int num_topics, bool use_topic_pyp=false, unsigned long seed = 0) 
     : m_num_topics(num_topics), m_word_pyps(1), 
     m_topic_pyp(0.5,1.0), m_use_topic_pyp(use_topic_pyp),
     m_seed(seed),
@@ -47,12 +47,12 @@ public:
     m_backoff.reset(new TermBackoff);
     m_backoff->read(filename);
     m_word_pyps.clear();
-    m_word_pyps.resize(m_backoff->order(), PYPs());
+    m_word_pyps.resize(m_backoff->order(), MPIPYPs());
   }
   void set_backoff(TermBackoffPtr backoff) {
     m_backoff = backoff;
     m_word_pyps.clear();
-    m_word_pyps.resize(m_backoff->order(), PYPs());
+    m_word_pyps.resize(m_backoff->order(), MPIPYPs());
   }
 
   F prob(const Term& term, int topic, int level=0) const;
@@ -70,9 +70,10 @@ private:
 
   CorpusTopics m_corpus_topics;
   typedef boost::ptr_vector< PYP<int> > PYPs;
+  typedef boost::ptr_vector< MPIPYP<int> > MPIPYPs;
   PYPs m_document_pyps;
-  std::vector<PYPs> m_word_pyps;
-  PYP<int> m_topic_pyp;
+  std::vector<MPIPYPs> m_word_pyps;
+  MPIPYP<int> m_topic_pyp;
   bool m_use_topic_pyp;
 
   unsigned long m_seed;
diff --git a/gi/pyp-topics/src/mpi-pyp.hh b/gi/pyp-topics/src/mpi-pyp.hh
index 58be7c5c..65358d20 100644
--- a/gi/pyp-topics/src/mpi-pyp.hh
+++ b/gi/pyp-topics/src/mpi-pyp.hh
@@ -1,5 +1,5 @@
-#ifndef _pyp_hh
-#define _pyp_hh
+#ifndef _mpipyp_hh
+#define _mpipyp_hh
 
 #include <math.h>
 #include <map>
@@ -9,11 +9,15 @@
 #include <boost/random/uniform_real.hpp>
 #include <boost/random/variate_generator.hpp>
 #include <boost/random/mersenne_twister.hpp>
+#include <boost/tuple/tuple.hpp>
+#include <boost/serialization/map.hpp>
+#include <boost/mpi.hpp>
+#include <boost/mpi/environment.hpp>
+#include <boost/mpi/communicator.hpp>
+#include <boost/mpi/operations.hpp>
 
-#include "pyp.h"
-#include "log_add.h"
-#include "slice-sampler.h"
-#include "mt19937ar.h"
+
+#include "pyp.hh"
 
 //
 // Pitman-Yor process with customer and table tracking
@@ -28,25 +32,104 @@ public:
   virtual int decrement(Dish d);
 
   void clear();
+  void reset_deltas();
 
-  void reset_deltas() { m_count_delta.clear(); }
+  void synchronise();
 
 private:
   typedef std::map<Dish, int> dish_delta_type;
-  typedef std::map<Dish, TableCounter> table_delta_type;
+  typedef std::map<Dish, typename PYP<Dish,Hash>::TableCounter> table_delta_type;
 
   dish_delta_type m_count_delta;
   table_delta_type m_table_delta;
 };
 
 template <typename Dish, typename Hash>
-MPIPYP<Dish,Hash>::MPIPYP(double a, double b, Hash)
-: PYP(a, b, Hash) {}
+MPIPYP<Dish,Hash>::MPIPYP(double a, double b, Hash h)
+: PYP<Dish,Hash>(a, b, 0, h) {}
 
 template <typename Dish, typename Hash>
 int 
 MPIPYP<Dish,Hash>::increment(Dish dish, double p0) {
-  int delta = PYP<Dish,Hash>::increment(dish, p0);
+  int delta = 0;
+  int table_joined=-1;
+  typename PYP<Dish,Hash>::TableCounter &tc = PYP<Dish,Hash>::_dish_tables[dish];
+
+  // seated on a new or existing table?
+  int c = PYP<Dish,Hash>::count(dish); 
+  int t = PYP<Dish,Hash>::num_tables(dish); 
+  int T = PYP<Dish,Hash>::num_tables();
+  double& a = PYP<Dish,Hash>::_a;
+  double& b = PYP<Dish,Hash>::_b;
+  double pshare = (c > 0) ? (c - a*t) : 0.0;
+  double pnew = (b + a*T) * p0;
+  assert (pshare >= 0.0);
+
+  if (mt_genrand_res53() < pnew / (pshare + pnew)) {
+    // assign to a new table
+    tc.tables += 1;
+    tc.table_histogram[1] += 1;
+    PYP<Dish,Hash>::_total_tables += 1;
+    delta = 1;
+  }
+  else {
+    // randomly assign to an existing table
+    // remove constant denominator from inner loop
+    double r = mt_genrand_res53() * (c - a*t);
+    for (std::map<int,int>::iterator
+         hit = tc.table_histogram.begin();
+         hit != tc.table_histogram.end(); ++hit) {
+      r -= ((hit->first - a) * hit->second);
+      if (r <= 0) {
+        tc.table_histogram[hit->first+1] += 1;
+        hit->second -= 1;
+        if (hit->second == 0)
+          tc.table_histogram.erase(hit);
+        table_joined = hit->first+1;
+        break;
+      }
+    }
+    if (r > 0) {
+      std::cerr << r << " " << c << " " << a << " " << t << std::endl;
+      assert(false);
+    }
+    delta = 0;
+  }
+
+  std::tr1::unordered_map<Dish,int,Hash>::operator[](dish) += 1;
+  //google::sparse_hash_map<Dish,int,Hash>::operator[](dish) += 1;
+  PYP<Dish,Hash>::_total_customers += 1;
+
+  // MPI Delta handling
+  // track the customer entering
+  typename dish_delta_type::iterator customer_it; 
+  bool customer_insert_result; 
+  boost::tie(customer_it, customer_insert_result) 
+    = m_count_delta.insert(std::make_pair(dish,0)); 
+
+  customer_it->second += 1;
+  if (customer_it->second == 0)
+    m_count_delta.erase(customer_it);
+
+  // increment the histogram bar for the table joined
+  if (!delta) {
+    assert (table_joined >= 0);
+    std::map<int,int> &histogram = m_table_delta[dish].table_histogram;
+    typename std::map<int,int>::iterator table_it; bool table_insert_result; 
+    boost::tie(table_it, table_insert_result) = histogram.insert(std::make_pair(table_joined,0)); 
+    table_it->second += 1;
+    if (table_it->second == 0) histogram.erase(table_it);
+
+    // decrement the histogram bar for the table left 
+    boost::tie(table_it, table_insert_result) = histogram.insert(std::make_pair(table_joined-1,0)); 
+    table_it->second -= 1;
+    if (table_it->second == 0) histogram.erase(table_it);
+  }
+  else {
+    typename PYP<Dish,Hash>::TableCounter &delta_tc = m_table_delta[dish];
+    delta_tc.tables += 1;
+    delta_tc.table_histogram[1] += 1;
+  }
 
   return delta;
 }
@@ -55,15 +138,177 @@ template <typename Dish, typename Hash>
 int 
 MPIPYP<Dish,Hash>::decrement(Dish dish)
 {
-  int delta = PYP<Dish,Hash>::decrement(dish);
+  typename std::tr1::unordered_map<Dish, int>::iterator dcit = find(dish);
+  //typename google::sparse_hash_map<Dish, int>::iterator dcit = find(dish);
+  if (dcit == PYP<Dish,Hash>::end()) {
+    std::cerr << dish << std::endl;
+    assert(false);
+  } 
+
+  int delta = 0, table_left=-1;
+
+  typename std::tr1::unordered_map<Dish, typename PYP<Dish,Hash>::TableCounter>::iterator dtit 
+    = PYP<Dish,Hash>::_dish_tables.find(dish);
+  //typename google::sparse_hash_map<Dish, TableCounter>::iterator dtit = _dish_tables.find(dish);
+  if (dtit == PYP<Dish,Hash>::_dish_tables.end()) {
+    std::cerr << dish << std::endl;
+    assert(false);
+  } 
+  typename PYP<Dish,Hash>::TableCounter &tc = dtit->second;
+
+  double r = mt_genrand_res53() * PYP<Dish,Hash>::count(dish);
+  for (std::map<int,int>::iterator hit = tc.table_histogram.begin();
+       hit != tc.table_histogram.end(); ++hit) {
+    r -= (hit->first * hit->second);
+    if (r <= 0) {
+      table_left = hit->first;
+      if (hit->first > 1) {
+        tc.table_histogram[hit->first-1] += 1;
+      }
+      else {
+        delta = -1;
+        tc.tables -= 1;
+        PYP<Dish,Hash>::_total_tables -= 1;
+      }
+
+      hit->second -= 1;
+      if (hit->second == 0) tc.table_histogram.erase(hit);
+      break;
+    }
+  }
+  if (r > 0) {
+    std::cerr << r << " " << PYP<Dish,Hash>::count(dish) << " " << PYP<Dish,Hash>::_a << " " 
+      << PYP<Dish,Hash>::num_tables(dish) << std::endl;
+    assert(false);
+  }
+
+  // remove the customer
+  dcit->second -= 1;
+  PYP<Dish,Hash>::_total_customers -= 1;
+  assert(dcit->second >= 0);
+  if (dcit->second == 0) {
+    PYP<Dish,Hash>::erase(dcit);
+    PYP<Dish,Hash>::_dish_tables.erase(dtit);
+  }
+
+  typename dish_delta_type::iterator it; 
+  bool insert_result; 
+  boost::tie(it, insert_result) = m_count_delta.insert(std::make_pair(dish,0)); 
+
+  it->second -= 1;
+
+  if (it->second == 0)
+    m_count_delta.erase(it);
+
+  assert (table_left >= 0);
+  typename PYP<Dish,Hash>::TableCounter& delta_tc = m_table_delta[dish];
+  if (table_left > 1)
+    delta_tc.table_histogram[table_left-1] += 1;
+  else delta_tc.tables -= 1;
+
+  std::map<int,int>::iterator tit = delta_tc.table_histogram.find(table_left);
+  //assert (tit != delta_tc.table_histogram.end());
+  tit->second -= 1;
+  if (tit->second == 0) delta_tc.table_histogram.erase(tit);
+
   return delta;
 }
 
 template <typename Dish, typename Hash>
 void 
-MPIPYP<Dish,Hash>::clear()
-{
+MPIPYP<Dish,Hash>::clear() {
   PYP<Dish,Hash>::clear();
+  reset_deltas();
+}
+
+template <typename Dish, typename Hash>
+void 
+MPIPYP<Dish,Hash>::reset_deltas() { 
+  m_count_delta.clear(); 
+  m_table_delta.clear();
+}
+
+template <typename Dish>
+struct sum_maps {
+  typedef std::map<Dish,int> map_type;
+  map_type& operator() (map_type& l, map_type const & r) const {
+    for (typename map_type::const_iterator it=r.begin(); it != r.end(); it++)
+      l[it->first] += it->second;
+    return l;
+  }
+};
+
+// Needed Boost definitions
+namespace boost { 
+  namespace mpi {
+    template <>
+    struct is_commutative< sum_maps<int>, std::map<int,int> > : mpl::true_ {};
+  }
+
+  namespace serialization {
+    template<class Archive>
+    void serialize(Archive & ar, PYP<int>::TableCounter& t, const unsigned int version) {
+      ar & t.table_histogram;
+      ar & t.tables;
+    }
+
+  } // namespace serialization
+} // namespace boost
+
+
+template <typename Dish, typename Hash>
+void 
+MPIPYP<Dish,Hash>::synchronise() {
+  boost::mpi::communicator world; 
+  int rank = world.rank(), size = world.size();
+
+  // communicate the customer count deltas
+  dish_delta_type global_dish_delta; // the “merged” map
+  boost::mpi::all_reduce(world, m_count_delta, global_dish_delta, sum_maps<Dish>());
+
+  // update this restaurant
+  for (typename dish_delta_type::const_iterator it=global_dish_delta.begin(); 
+       it != global_dish_delta.end(); ++it) {
+    std::tr1::unordered_map<Dish,int,Hash>::operator[](it->first) += (it->second - m_count_delta[it->first]);
+    PYP<Dish,Hash>::_total_customers += (it->second - m_count_delta[it->first]);
+    //std::cerr << "Process " << rank << " adding " <<  (it->second - m_count_delta[it->first]) << " customers." << std::endl;
+  }
+
+  // communicate the table count deltas
+//  for (int process = 0; process < size; ++process) {
+//    if (rank == process) {
+//      // broadcast deltas
+//      std::cerr << " -- Rank " << rank << " broadcasting -- " << std::endl;
+//
+//      boost::mpi::broadcast(world, m_table_delta, process);
+//
+//      std::cerr << " -- Rank " << rank << " done broadcasting -- " << std::endl;
+//    }
+//    else {
+//      std::cerr << " -- Rank " << rank << " receiving -- " << std::endl;
+//      // receive deltas
+//      table_delta_type recv_table_delta;
+//
+//      boost::mpi::broadcast(world, recv_table_delta, process);
+//
+//      std::cerr << " -- Rank " << rank << " done receiving -- " << std::endl;
+//
+//      for (typename table_delta_type::const_iterator dish_it=recv_table_delta.begin(); 
+//           dish_it != recv_table_delta.end(); ++dish_it) {
+//        typename PYP<Dish,Hash>::TableCounter &tc = PYP<Dish,Hash>::_dish_tables[dish_it->first];
+//
+//        for (std::map<int,int>::const_iterator it=dish_it->second.table_histogram.begin(); 
+//             it != dish_it->second.table_histogram.end(); ++it) {
+//          tc.table_histogram[it->first] += it->second;
+//        }
+//        tc.tables += dish_it->second.tables;
+//        PYP<Dish,Hash>::_total_tables += dish_it->second.tables;
+//      }
+//    }
+//  }
+//  std::cerr << " -- Done Reducing -- " << std::endl;
+
+  reset_deltas();
 }
 
 #endif
diff --git a/gi/pyp-topics/src/mpi-train-contexts.cc b/gi/pyp-topics/src/mpi-train-contexts.cc
index 956ce123..0651ecac 100644
--- a/gi/pyp-topics/src/mpi-train-contexts.cc
+++ b/gi/pyp-topics/src/mpi-train-contexts.cc
@@ -86,7 +86,7 @@ int main(int argc, char **argv)
 
   // seed the random number generator: 0 = automatic, specify value otherwise
   unsigned long seed = 0; 
-  PYPTopics model(vm["topics"].as<int>(), vm.count("hierarchical-topics"), seed);
+  MPIPYPTopics model(vm["topics"].as<int>(), vm.count("hierarchical-topics"), seed);
 
   // read the data
   BackoffGenerator* backoff_gen=0;
diff --git a/gi/pyp-topics/src/pyp.hh b/gi/pyp-topics/src/pyp.hh
index 26f6ab2e..84decb0f 100644
--- a/gi/pyp-topics/src/pyp.hh
+++ b/gi/pyp-topics/src/pyp.hh
@@ -1,6 +1,7 @@
 #ifndef _pyp_hh
 #define _pyp_hh
 
+#include "slice-sampler.h"
 #include <math.h>
 #include <map>
 #include <tr1/unordered_map>
@@ -11,7 +12,6 @@
 #include <boost/random/mersenne_twister.hpp>
 
 #include "log_add.h"
-#include "slice-sampler.h"
 #include "mt19937ar.h"
 
 //
@@ -63,7 +63,7 @@ public:
   double b() const { return _b; }
   void set_b(double b) { _b = b; }
 
-  void clear();
+  virtual void clear();
   std::ostream& debug_info(std::ostream& os) const;
 
   double log_restaurant_prob() const;
@@ -75,13 +75,12 @@ public:
   void resample_prior_a();
   void resample_prior_b();
 
-private:
+protected:
   double _a, _b; // parameters of the Pitman-Yor distribution
   double _a_beta_a, _a_beta_b; // parameters of Beta prior on a
   double _b_gamma_s, _b_gamma_c; // parameters of Gamma prior on b
 
-  struct TableCounter
-  {
+  struct TableCounter {
     TableCounter() : tables(0) {};
     int tables;
     std::map<int, int> table_histogram; // num customers at table -> number tables
-- 
cgit v1.2.3