ClassBasedCrossEntropyWithSoftmaxNode now taking advantage of optimiz…

…ed CPU-side element access

Source/ComputationNetworkLib/TrainingNodes.h

@@ -712,12 +712,8 @@ template class NoiseContrastiveEstimationNode<double>;
 template <class ElemType>
 class ClassBasedCrossEntropyWithSoftmaxNode : public ComputationNodeNonLooping /*ComputationNode*/<ElemType>, public NumInputs<4>
 {
-    typedef ComputationNodeNonLooping<ElemType> Base;
-    UsingComputationNodeMembersBoilerplate;
-    static const std::wstring TypeName()
-    {
-        return L"ClassBasedCrossEntropyWithSoftmax";
-    }
+    typedef ComputationNodeNonLooping<ElemType> Base; UsingComputationNodeMembersBoilerplate;
+    static const std::wstring TypeName() { return L"ClassBasedCrossEntropyWithSoftmax"; }
 
     // our inputs
     static const size_t LABELDATA = 0;
@@ -850,9 +846,8 @@ class ClassBasedCrossEntropyWithSoftmaxNode : public ComputationNodeNonLooping /
     // -sum(left_i * log(softmax_i(right)))
     virtual void /*ComputationNodeNonLooping::*/ ForwardPropNonLooping() override
     {
-        if (Input(LABELDATA)->Value().GetDeviceId() != CPUDEVICE)
-            LogicError("ClassBasedCrossEntropyWithSoftmax (ForwardPropNonLooping()): The label matrix is not using CPU device. This will make computation slow, even though the label data is probably saved on GPU. Because of the external loop over time with explicit class id retrieved from the label matrix, the computation will be very slow if the label matrix is saved on GPU. However, this is only a constraint for label matrix and other matrices such as data are suggested to reside on GPU. ");
-        // TODO: Get the label matrix into location=Both state.
+        // get the label matrix to CPU, ideally in location=BOTH state
+        Input(LABELDATA)->Value().TransferToDeviceIfNotThere(CPUDEVICE, /*ismoved =*/ false/*means: BOTH state OK*/, /*emptyTransfer =*/ false, /*updatePreferredDevice =*/ false);
 
         auto& functionValues = Value();
 
@@ -868,11 +863,11 @@ class ClassBasedCrossEntropyWithSoftmaxNode : public ComputationNodeNonLooping /
         m_totalNbrWords = ForColumnsWithClass([](size_t /*s*/, size_t /*t*/, const FrameRange& /*fr*/, size_t y_t, size_t /*c_t*/, size_t /*sz*/, size_t lft_bnd, size_t nbr_wrd)
         {
             if (nbr_wrd == 0)
-                LogicError("ClassBasedCrossEntropyWithSoftmax: Encountered a class of size 0. This sample seems to lack an NoInput flag.");
+                LogicError("ClassBasedCrossEntropyWithSoftmax: Encountered a class of size 0.");
             if (y_t < lft_bnd || y_t >= lft_bnd + nbr_wrd)
                 LogicError("ClassBasedCrossEntropyWithSoftmax: Word index out of bounds of class-member index range (word not a class member).");
         });
-        // m_totalNbrWords = total size of concatenated vector
+        // now m_totalNbrWords = total size of concatenated vector
 
         // buffer to hold the concatenated class-conditioned prob vectors
         m_softMax.Resize(1, m_totalNbrWords);

Source/Math/Matrix.cpp

@@ -256,7 +256,7 @@ void Matrix<ElemType>::SetDataLocation(CurrentDataLocation location, MatrixType
         m_baseMatrix = ((m_currentDataLocation == CurrentDataLocation::CPU) ? (BaseMatrix<ElemType>*) m_CPUSparseMatrix : (BaseMatrix<ElemType>*) m_GPUSparseMatrix);
 
     // sanity check
-    if (!m_baseMatrix)
+    if (!m_baseMatrix && m_matrixType != MatrixType::UNDETERMINED)
         LogicError("SetDataLocation: new m_baseMatrix must not be NULL.");
 }
 
@@ -777,7 +777,7 @@ void Matrix<ElemType>::CopySection(size_t numRows, size_t numCols, ElemType* dst
 template <class ElemType>
 Matrix<ElemType> Matrix<ElemType>::ColumnSlice(size_t startColumn, size_t numCols) const
 {
-    int devId = GetDeviceId(); // BUGBUG: Must work off current location
+    int devId = GetDeviceId();
 
     Matrix<ElemType> slice(matrixFlagDontOwnBuffer, (DEVICEID_TYPE) devId); // this already creates pointers
 
@@ -1075,7 +1075,7 @@ const ElemType Matrix<ElemType>::operator()(const size_t row, const size_t col)
 // WARNING: This function is very slow for GPUs since it requires copying values between CPUs and GPUs.
 // In addition, if ColumnSlice is used after this function but before the values are copied back to GPU
 // the operation will fail since the memory is not managed by the slice.
-// If you don't need to modify the values, please make sure to call the const version above.
+// If you don't need to modify the values, to call the const version above, or GetValue(row,col) which does that for you unambiguously.
 // TODO: Can we remove this, and have users use SetValue() instead? To avoid this potential error?
 template <class ElemType>
 ElemType& Matrix<ElemType>::operator()(const size_t row, const size_t col)