now includes the MPI rank in each log msg; bug fix: fixed a few compilation problems on Windows

src/common/logging.cpp

@@ -122,6 +122,15 @@ void createLoggers(const marian::Config* options) {
 #endif
 }
 
+// modify the log pattern for the "general" logger to include the MPI rank
+// This is called upon initializing MPI. It is needed to associated error messages to ranks.
+void switchtoMultinodeLogging(std::string nodeIdStr) {
+  Logger log = spdlog::get("general");
+  if (log)
+    log->set_pattern("[%Y-%m-%d %T " + nodeIdStr + "] %v");
+}
+
+
 namespace marian {
   void noinline logCallStack(size_t skipLevels)
   {

src/common/logging.h

@@ -110,3 +110,4 @@ void checkedLog(std::string logger, std::string level, Args... args) {
 }
 
 void createLoggers(const marian::Config* options = nullptr);
+void switchtoMultinodeLogging(std::string nodeIdStr);

src/graph/node_operators_binary.h

@@ -490,7 +490,7 @@ struct RowsNodeOp : public NaryNodeOp {
 struct SelectNodeOp : public NaryNodeOp {
   SelectNodeOp(Expr a, Expr indices, int axis)
       : NaryNodeOp({a, indices}, newShape(a, axis, indices->shape().elements())),
-        axis_{a->shape().axis(axis)} {
+        axis_(a->shape().axis(axis)) {
     matchOrAbort<IndexType>(indices->value_type());
   }
 

src/layers/word2vec_reader.h

@@ -56,7 +56,7 @@ public:
     embs.reserve(dimVoc * dimEmb);
 
     // Populate output vector with embedding
-    for(size_t word = 0; word < (size_t)dimVoc; ++word) {
+    for(Word word = 0; word < (Word)dimVoc; ++word) {
       // For words not occuring in the file use uniform distribution
       if(word2vec.find(word) == word2vec.end()) {
         auto randVals = randomEmbeddings(dimVoc, dimEmb);

src/models/hardatt.h

@@ -293,7 +293,7 @@ public:
         if(specialSymbols_.count(embIdx[i])) {
           stateHardAtt->getAttentionIndices()[i]++;
           if(stateHardAtt->getAttentionIndices()[i] >= dimSrcWords)
-            stateHardAtt->getAttentionIndices()[i] = dimSrcWords - 1;
+            stateHardAtt->getAttentionIndices()[i] = (unsigned int)dimSrcWords - 1;
         }
     }
   }

src/models/transformer_factory.h

@@ -13,4 +13,6 @@ Ptr<EncoderBase> NewEncoderTransformer(Ptr<Options> options);
 Ptr<DecoderBase> NewDecoderTransformer(Ptr<Options> options);
 }  // namespace marian
 
+#ifndef _MSC_VER
 #include "models/transformer.h"
+#endif

src/tensors/cpu/tensor_operators.cpp

@@ -517,7 +517,7 @@ void Insert(Tensor out,
 
   int length = inShape.elements();
   functional::Array<int, functional::Shape::size()> dims;
-  int axisCPU = axis + functional::Shape::size() - out->shape().size();
+  int axisCPU = (int)(axis + functional::Shape::size() - out->shape().size());
 
   for(int index = 0; index < length; ++index) {
     inShape.dims(index, dims);

src/training/communicator.cpp

@@ -80,14 +80,16 @@ public:
     MPI_Comm_size(MPI_COMM_WORLD, &comm_world_size_);
     MPI_Comm_rank(MPI_COMM_WORLD, &my_rank_);
 
+    // patch logging pattern to include the MPI process info
+    switchtoMultinodeLogging(std::to_string(MPIWrapper::myMPIRank()));
+
     // log hostnames in order, and test
-    // @TODO: We call ourselves here. Not sure if that is properly allowed.
     for (size_t r = 0; r < numMPIProcesses(); r++) {
       MPIWrapper::barrier();
       if (r == MPIWrapper::myMPIRank())
-        LOG(info, "[mpi] initialized as {}", MPIWrapper::idStr());
+        LOG(info, "[mpi] initialized {} processes", MPIWrapper::numMPIProcesses());
+      MPIWrapper::barrier();
     }
-    MPIWrapper::barrier();
   }
 
   virtual size_t myMPIRank()        const override { return (size_t)my_rank_; };

src/training/communicator.h

@@ -81,7 +81,11 @@ public:
     v.resize(vecLen);
     bCast(v.data(), v.size(), getDataType(v.data()), rootRank, comm);
   }
-  std::string idStr() { // helper to identify the node in logs
+  std::string idStr() const { // helper to identify the node in logs
+    return hostnameAndProcessId() + " MPI rank " + std::to_string(myMPIRank()) + " out of " + std::to_string(numMPIProcesses());
+  }
+protected:
+  static std::string hostnameAndProcessId() { // helper to get hostname:pid
 #ifdef _WIN32
     std::string hostname = getenv("COMPUTERNAME");
     auto processId = GetCurrentProcessId();
@@ -93,7 +97,7 @@ public:
     }();
     auto processId = getpid();
 #endif
-    return hostname + ":" + std::to_string(processId) + " MPI rank " + std::to_string(myMPIRank()) + " out of " + std::to_string(numMPIProcesses());
+    return hostname + ":" + std::to_string(processId);
   }
 };
 

src/training/graph_group_singleton.cpp

@@ -17,7 +17,6 @@ void SingletonGraph::execute(Ptr<data::Batch> batch) {
   graph_->backward();
 
   // Get batch stats
-  size_t batch_words = batch->wordsTrg();
   opt_->update(graph_);
 
   if(mvAvg_) {

vs/Marian.vcxproj

@@ -123,6 +123,8 @@
     <ClInclude Include="..\src\command\marian_conv.cpp" />
     <ClCompile Include="..\src\command\marian_main.cpp" />
     <ClCompile Include="..\src\common\binary.cpp" />
+    <ClCompile Include="..\src\common\cli_wrapper.cpp" />
+    <ClCompile Include="..\src\common\config_validator.cpp" />
     <ClCompile Include="..\src\common\io.cpp" />
     <ClCompile Include="..\src\common\utils.cpp" />
     <ClCompile Include="..\src\common\logging.cpp" />
@@ -202,9 +204,15 @@
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="..\src\common\binary.h" />
+    <ClInclude Include="..\src\common\cli_helper.h" />
+    <ClInclude Include="..\src\common\cli_wrapper.h" />
+    <ClInclude Include="..\src\common\config_validator.h" />
+    <ClInclude Include="..\src\common\filesystem.h" />
+    <ClInclude Include="..\src\common\hash.h" />
     <ClInclude Include="..\src\common\io.h" />
     <ClInclude Include="..\src\common\io_item.h" />
     <ClInclude Include="..\src\common\timer.h" />
+    <ClInclude Include="..\src\common\types.h" />
     <ClInclude Include="..\src\layers\loss.h" />
     <ClInclude Include="..\src\layers\weight.h" />
     <ClInclude Include="..\src\marian.h" />

vs/Marian.vcxproj.filters

@@ -247,6 +247,12 @@
     <ClCompile Include="..\src\3rd_party\ExceptionWithCallStack.cpp">
       <Filter>3rd_party</Filter>
     </ClCompile>
+    <ClCompile Include="..\src\common\cli_wrapper.cpp">
+      <Filter>common</Filter>
+    </ClCompile>
+    <ClCompile Include="..\src\common\config_validator.cpp">
+      <Filter>common</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="..\src\marian.h" />
@@ -1043,6 +1049,24 @@
     <ClInclude Include="..\src\common\timer.h">
       <Filter>common</Filter>
     </ClInclude>
+    <ClInclude Include="..\src\common\cli_helper.h">
+      <Filter>common</Filter>
+    </ClInclude>
+    <ClInclude Include="..\src\common\cli_wrapper.h">
+      <Filter>common</Filter>
+    </ClInclude>
+    <ClInclude Include="..\src\common\config_validator.h">
+      <Filter>common</Filter>
+    </ClInclude>
+    <ClInclude Include="..\src\common\filesystem.h">
+      <Filter>common</Filter>
+    </ClInclude>
+    <ClInclude Include="..\src\common\hash.h">
+      <Filter>common</Filter>
+    </ClInclude>
+    <ClInclude Include="..\src\common\types.h">
+      <Filter>common</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="3rd_party">