bug fix: BatchLUSequenceReader would move features and labels to the …

…CPU, causing downstream issues. Now it keeps the location, so that it works with GPU now
qingsong99 · Nov 30, 2015 · 95bf296 · 95bf296
1 parent e647197
commit 95bf296
Show file tree

Hide file tree

Showing 2 changed files with 11 additions and 6 deletions.
diff --git a/DataReader/LUSequenceReader/LUSequenceReader.cpp b/DataReader/LUSequenceReader/LUSequenceReader.cpp
@@ -890,6 +890,7 @@ bool BatchLUSequenceReader<ElemType>::GetMinibatch(std::map<std::wstring, Matrix
             }
         }
 
+        locObs.SetPreferredDeviceId(features.GetDeviceId());    // needed, otherwise SetValue() below will inherit CPUDEVICE a as target
         features.SetValue(locObs);
 
         lablsize = GetLabelOutput(matrices, m_labelInfo[labelInfoOut], actualmbsize);
@@ -917,11 +918,12 @@ size_t BatchLUSequenceReader<ElemType>::GetLabelOutput(std::map<std::wstring,
     Matrix<ElemType>* labels = matrices[m_labelsName[labelInfoOut]];
     if (labels == nullptr) return 0;
 
-    DEVICEID_TYPE device = labels->GetDeviceId();
-
     labels->Resize(labelInfo.dim, actualmbsize);
     labels->SetValue(0);
-    labels->TransferFromDeviceToDevice(device, CPUDEVICE, true);
+
+    // build it on the CPU side
+    DEVICEID_TYPE deviceId = labels->GetDeviceId();
+    labels->TransferFromDeviceToDevice(deviceId, CPUDEVICE, true);
 
     size_t nbrLabl = 0;
     for (size_t j = 0; j < actualmbsize; ++j)
@@ -955,6 +957,9 @@ size_t BatchLUSequenceReader<ElemType>::GetLabelOutput(std::map<std::wstring,
         nbrLabl++;
     }
 
+    // move it back to GPU if that's where it was before
+    labels->TransferFromDeviceToDevice(CPUDEVICE, deviceId, true);
+
     return nbrLabl;
 }
 

diff --git a/Math/Math/Matrix.h b/Math/Math/Matrix.h
@@ -121,12 +121,12 @@ namespace Microsoft { namespace MSR { namespace CNTK {
         void SetDataLocation(CurrentDataLocation location, MatrixType type=UNDETERMINED) const;
 
     public:
-        MatrixType GetMatrixType() const {return m_matrixType;};
+        MatrixType GetMatrixType() const { return m_matrixType; }
         MatrixFormat GetFormat() const { return m_baseMatrix->GetFormat(); }
-        bool OwnBuffer() const {return m_baseMatrix->OwnBuffer();}
+        bool OwnBuffer() const { return m_baseMatrix->OwnBuffer(); }
         int GetDeviceId() const; //-1 if CPU, otherwise GPU CUDA device id
         DEVICEID_TYPE GetPreferredDeviceId() const { return m_preferredDeviceId; }; //-1 if CPU, otherwise GPU CUDA device id
-        void SetPreferredDeviceId(DEVICEID_TYPE preferredDeviceId){ if (m_preferredDeviceId != preferredDeviceId) m_preferredDeviceId = preferredDeviceId; }
+        void SetPreferredDeviceId(DEVICEID_TYPE preferredDeviceId) { m_preferredDeviceId = preferredDeviceId; }
         //Moves matrix from device id_from to device with id_to. 
         //If emptyTransfer=true, then no data is ever moved, just corresponding GPU/CPU matrices are deleted and then created using empty constructor
         void TransferFromDeviceToDevice(int id_from, int id_to, bool ismoved = false,/*if false then keep source and set location to BOTH*/ bool emptyTransfer = false, bool updatePreferredDevice = true) const;