fixed a bug in GPUMatrix::RowElementMultiplyWith and RowElementDivideBy. Changed ComputationNetwork.h to support loading nodes with number of children larger than 3.

.gitignore

@@ -5,6 +5,7 @@
 *.suo
 *.user
 *.sln.docstates
+*.orig
 
 # Build results
 

MachineLearning/cn/CompositeComputationNode.h

@@ -2135,7 +2135,6 @@ namespace Microsoft { namespace MSR { namespace CNTK {
             else  //all samples share the same prior
                 posterior.ColumnElementMultiplyWith(prior);  
 
-
             //compute GMM log-likelihood
             Matrix<ElemType>::Multiply(ConstOnes(1, numComponent, posterior.GetDeviceId()), false, posterior, false, functionValues);  //functionValues <-- total likelihood
             posterior.RowElementDivideBy(functionValues); //posterior <-- per-comp likelihood / total likelihood

MachineLearning/cn/ComputationNetwork.h

@@ -336,7 +336,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                     }
 
                     ComputationNodePtr nodePtr = GetNodeFromName(nodeName);
-                    ComputationNodePtr childNodePtr0, childNodePtr1, childNodePtr2;
+                    ComputationNodePtr childNodePtr0, childNodePtr1, childNodePtr2, childNodePtr3, childNodePtr4;
                     switch (numChildren)
                     {
                     case 1:
@@ -354,6 +354,21 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                         childNodePtr2 = GetNodeFromName(childrenNames[2]);
                         nodePtr->AttachInputs(childNodePtr0, childNodePtr1, childNodePtr2);
                         break;
+                    case 4:
+                        childNodePtr0 = GetNodeFromName(childrenNames[0]);
+                        childNodePtr1 = GetNodeFromName(childrenNames[1]);
+                        childNodePtr2 = GetNodeFromName(childrenNames[2]);
+                        childNodePtr3 = GetNodeFromName(childrenNames[3]);
+                        nodePtr->AttachInputs(childNodePtr0, childNodePtr1, childNodePtr2, childNodePtr3);
+                        break;
+                    case 5:
+                        childNodePtr0 = GetNodeFromName(childrenNames[0]);
+                        childNodePtr1 = GetNodeFromName(childrenNames[1]);
+                        childNodePtr2 = GetNodeFromName(childrenNames[2]);
+                        childNodePtr3 = GetNodeFromName(childrenNames[3]);
+                        childNodePtr4 = GetNodeFromName(childrenNames[4]);
+                        nodePtr->AttachInputs(childNodePtr0, childNodePtr1, childNodePtr2, childNodePtr3, childNodePtr4);
+                        break;
                     default:
                         throw std::logic_error("Invalid number of children.");
                     }

Math/Math/GPUMatrix.cu

@@ -1371,9 +1371,9 @@ namespace Microsoft { namespace MSR { namespace CNTK {
         if (!(a.GetNumRows() == 1 && a.GetNumCols() == GetNumCols()))
             throw std::invalid_argument("RowElementMultiplyWith: The input matrix should be a row vector and match [this]'s columns.");
 
-        long N=(long)a.GetNumRows();
+        long N = (long)this->GetNumRows();
-        long M=(long)this->GetNumCols();        
+        long M = (long)a.GetNumCols();
-        int blocksPerGrid =(int)ceil(1.0*N/threadsPerBlock);  
+        int blocksPerGrid = (int)ceil(1.0*M / threadsPerBlock);
         a.PrepareDevice();
         cudaEvent_t done = nullptr;
         if (do_sync)    CUDA_CALL(cudaEventCreate(&done));        
@@ -1394,9 +1394,9 @@ namespace Microsoft { namespace MSR { namespace CNTK {
         if (!(a.GetNumRows() == 1 && a.GetNumCols() == GetNumCols()))
             throw std::invalid_argument("RowElementDivideBy: The input matrix should be a row vector and match [this]'s columns.");
 
-        long N = (long)a.GetNumRows();
+        long N = (long)this->GetNumRows();
-        long M = (long)this->GetNumCols();
+        long M = (long)a.GetNumCols();
-        int blocksPerGrid = (int)ceil(1.0*N / threadsPerBlock);
+        int blocksPerGrid = (int)ceil(1.0*M / threadsPerBlock);
         a.PrepareDevice();
         cudaEvent_t done = nullptr;
         if (do_sync)    CUDA_CALL(cudaEventCreate(&done));
@@ -1417,9 +1417,9 @@ namespace Microsoft { namespace MSR { namespace CNTK {
         if (!(a.GetNumRows() == GetNumRows() && a.GetNumCols() == 1))
             throw std::invalid_argument("ColumnElementDivideBy: The input matrix should be a col vector and match [this]'s rows.");
 
-        long N=(long)a.GetNumRows();
+        long N = (long)a.GetNumRows();
-        long M=(long)this->GetNumCols();        
+        long M = (long)this->GetNumCols();
-        int blocksPerGrid =(int)ceil(1.0*N/threadsPerBlock);  
+        int blocksPerGrid = (int)ceil(1.0*N / threadsPerBlock);
         a.PrepareDevice();
         cudaEvent_t done = nullptr;
         if (do_sync)    CUDA_CALL(cudaEventCreate(&done));        

Math/Math/Matrix.cpp

@@ -1737,7 +1737,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
         if (a.IsEmpty() || IsEmpty())
             throw std::logic_error("RowElementMultiplyWith: Matrix is empty.");
 
-        if (!(a.GetNumRows() == GetNumRows() && a.GetNumCols() == 1))
+        if (!(a.GetNumCols() == GetNumCols() && a.GetNumRows() == 1))
             throw std::invalid_argument("RowElementMultiplyWith: The input matrix should be a row vector and match [this]'s columns.");
 
         //WARNING: a and this must have same type
@@ -1763,7 +1763,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
         if (a.IsEmpty() || IsEmpty())
             throw std::logic_error("RowElementDivideBy: Matrix is empty.");
 
-        if (!(a.GetNumRows() == GetNumRows() && a.GetNumCols() == 1))
+        if (!(a.GetNumCols() == GetNumCols() && a.GetNumRows() == 1))
             throw std::invalid_argument("RowElementDivideBy: The input matrix should be a row vector and match [this]'s columns.");
 
         //WARNING: a and this must have same type