Updated code to work with the requirement of each matrix owning its storage in order to resize.

Source/SequenceTrainingLib/parallelforwardbackward.cpp

@@ -14,6 +14,7 @@
 #include "cudalatticeops.h"
 #include <numeric> // for debug
 #include "cudalib.h"
+#include <memory>
 
 #define TWO_CHANNEL // [v-hansu]
 using namespace msra::cuda;
@@ -623,16 +624,28 @@ struct parallelstateimpl
         if (errorsignalgpustorage->GetNumRows() != 0 && errorsignalgpustorage->GetNumRows() != errorsignal.rows())
             throw ::logic_error("gpumatrixstorage->rows() shall be fixed once allocated");
         if (errorsignalgpustorage->GetNumCols() < errorsignal.cols())
+        {
+            // Note: This is required because otherwise errorsignalgpustorage will be a view of the storage object in
+            // errorsignalgpustorage, and thuse it can't resize. This is perhaps not the optimal way to do this, but
+            // how else? Why do these two matrices exist? Why not just one?
+            errorsignalgpu = nullptr;
             errorsignalgpustorage->Resize(errorsignal.rows(), errorsignal.cols());
-        *errorsignalgpu = errorsignalgpustorage->ColumnSlice(0, errorsignal.cols());
+        }
+        //*errorsignalgpu = errorsignalgpustorage->ColumnSlice(0, errorsignal.cols());
+        errorsignalgpu = make_unique<Microsoft::MSR::CNTK::Matrix<float>>(errorsignalgpustorage->ColumnSlice(0, errorsignal.cols()));
 
         if (cacheerrsignalneg)
         {
             if (errorsignalneggpustorage->GetNumRows() != 0 && errorsignalneggpustorage->GetNumRows() != errorsignal.rows())
                 throw ::logic_error("gpumatrixstorage->rows() shall be fixed once allocated");
             if (errorsignalneggpustorage->GetNumCols() < errorsignal.cols())
+            {
+                // Same as above.
+                errorsignalneggpu = nullptr;
                 errorsignalneggpustorage->Resize(errorsignal.rows(), errorsignal.cols());
-            *errorsignalneggpu = errorsignalneggpustorage->ColumnSlice(0, errorsignal.cols());
+            }
+            //*errorsignalneggpu = errorsignalneggpustorage->ColumnSlice(0, errorsignal.cols());
+            errorsignalneggpu = make_unique<Microsoft::MSR::CNTK::Matrix<float>>(errorsignalneggpustorage->ColumnSlice(0, errorsignal.cols()));
         }
     }