new code

source/sample/transformer/T2TPredictor.cpp

@@ -206,6 +206,7 @@ void T2TPredictor::Predict(T2TStateBundle * next, XTensor * encoding, XTensor *
     /* the decoder output of the last position */
     /* the decoder output of the last position */
     decodingStep = CopyIndexed(decoding, decoding.order - 2, selectSrc, selectTgt);
     decodingStep = CopyIndexed(decoding, decoding.order - 2, selectSrc, selectTgt);
 
 
+
     /* generate the output probabilities */
     /* generate the output probabilities */
     m->outputLayer->Make(decodingStep, output);
     m->outputLayer->Make(decodingStep, output);
     
     

source/sample/transformer/T2TSearch.cpp

@@ -158,26 +158,27 @@ void T2TSearch::Score(T2TStateBundle * prev, T2TStateBundle * beam)
     _SumDim(&prob, &probPathPrev, &score, 0);
     _SumDim(&prob, &probPathPrev, &score, 0);
 
 
     InitTensor(&len, &lenPrev);
     InitTensor(&len, &lenPrev);
+    InitTensor(&lp, &lenPrev);
 
 
-    _ScaleAndShift(&lenPrev, &len, 1.0F, 1.0F);
+    //_ScaleAndShift(&lenPrev, &len, 1.0F, 1.0F);
 
 
     /* the GNMT-like length penalty */
     /* the GNMT-like length penalty */
-    lp = T2TLengthPenalizer::GNMT(len, alpha);
+    //lp = T2TLengthPenalizer::GNMT(len, alpha);
 
 
     lp.Reshape(lp.unitNum);
     lp.Reshape(lp.unitNum);
 
 
     /* score = log-prob/lp */
     /* score = log-prob/lp */
-    _DivDim(&score, &lp, &score, 0);
+    //_DivDim(&score, &lp, &score, 0);
 
 
     InitTensor(&mask, &prev->endMark);
     InitTensor(&mask, &prev->endMark);
-    CopyValues(prev->endMark, mask);
-    _ScaleAndShiftMe(&mask, -1e9F);
+    //CopyValues(prev->endMark, mask);
+    //_ScaleAndShiftMe(&mask, -1e9F);
     
     
     mask.Reshape(mask.unitNum);
     mask.Reshape(mask.unitNum);
 
 
     /* mask the completed hypotheses so that they cannot 
     /* mask the completed hypotheses so that they cannot 
        be involved in further sorting and beam search. */
        be involved in further sorting and beam search. */
-    _SumDim(&score, &mask, &score, 0);
+    //_SumDim(&score, &mask, &score, 0);
     
     
     prob.Reshape(order, dims);
     prob.Reshape(order, dims);
     score.Reshape(order, dims);
     score.Reshape(order, dims);
@@ -227,9 +228,9 @@ void T2TSearch::Generate(T2TStateBundle * beam)
     score.Reshape(order, dimsBeam);
     score.Reshape(order, dimsBeam);
     
     
     /* keep the most promissing candidates in the beam */
     /* keep the most promissing candidates in the beam */
-    TopK(score, scoreTopK, index, -1, beamSize);
+    //TopK(score, scoreTopK, index, -1, beamSize);
     
     
-    CopyValues(scoreTopK, preID);
+    CopyValues(index, preID);
     
     
     int sizeVocab = score.GetDim(-1);
     int sizeVocab = score.GetDim(-1);
     
     

source/tensor/core/arithmetic/SumDim.cu

@@ -84,15 +84,18 @@ void KernelAddWithCol(T * a, T * b, T * c, int rowNum, int colNum, int blockSize
     int colIndex = blockDim.x * blockIdx.x + threadIdx.x;
     int colIndex = blockDim.x * blockIdx.x + threadIdx.x;
     int row = blockDim.y * blockIdx.y + threadIdx.y;
     int row = blockDim.y * blockIdx.y + threadIdx.y;
 
 
-    int col = colIndex % colNum;
-    int block = colIndex / colNum;
+    int col = colIndex % blockSize;
+    int block = colIndex / blockSize;
 
 
     if(row >= rowNum || block >= blockNum)
     if(row >= rowNum || block >= blockNum)
         return;
         return;
 
 
-    if(threadIdx.x == 0)
+    if(threadIdx.x == 0){
+        printf("(%d %d) ", row, block);
         bv[threadIdx.y] = b[row];
         bv[threadIdx.y] = b[row];
+    }
 
 
+    /*
     __syncthreads();
     __syncthreads();
 
 
     int offset = block * blockSize + row * colNum + col;
     int offset = block * blockSize + row * colNum + col;
@@ -100,7 +103,7 @@ void KernelAddWithCol(T * a, T * b, T * c, int rowNum, int colNum, int blockSize
     if(betaFired)
     if(betaFired)
         c[offset] = a[offset] + bv[threadIdx.y] * beta;
         c[offset] = a[offset] + bv[threadIdx.y] * beta;
     else
     else
-        c[offset] = a[offset] + bv[threadIdx.y];
+        c[offset] = a[offset] + bv[threadIdx.y];*/
 }
 }
 
 
 /*
 /*

source/tensor/core/movement/CopyValues.cpp

@@ -41,7 +41,8 @@ void _CopyValues(const XTensor * s, XTensor * t, XStream * stream)
     CheckNTErrors(s != NULL && t != NULL, "The input tensor and output tensor must be nonempty!");
     CheckNTErrors(s != NULL && t != NULL, "The input tensor and output tensor must be nonempty!");
     CheckNTErrors(s->data != NULL, "Cannot copy an empty data array!");
     CheckNTErrors(s->data != NULL, "Cannot copy an empty data array!");
     CheckNTErrors(t->data != NULL, "Cannot copy to an empty data array!");
     CheckNTErrors(t->data != NULL, "Cannot copy to an empty data array!");
-    CheckNTErrors(s->unitNum == t->unitNum, "Unmatched data item number!");
+    CheckNTErrors(s->unitSize == t->unitSize, "Incompatible data types in value copy.");
+    CheckNTErrors(s->unitNum == t->unitNum, "The data items are be the same.");
 
 
     if ((s->dataType == X_FLOAT16 && t->dataType == X_FLOAT) ||
     if ((s->dataType == X_FLOAT16 && t->dataType == X_FLOAT) ||
         (s->dataType == X_FLOAT && t->dataType == X_FLOAT16)) {
         (s->dataType == X_FLOAT && t->dataType == X_FLOAT16)) {
@@ -90,9 +91,9 @@ void _CopyValues(const XTensor * s, const int sBeg, const int sLen, XTensor * t,
     CheckNTErrors(s != NULL && t != NULL, "The input tensor and output tensor must be nonempty!");
     CheckNTErrors(s != NULL && t != NULL, "The input tensor and output tensor must be nonempty!");
     CheckNTErrors(s->data != NULL, "Cannot copy an empty data array!");
     CheckNTErrors(s->data != NULL, "Cannot copy an empty data array!");
     CheckNTErrors(t->data != NULL, "Cannot copy to an empty data array!");
     CheckNTErrors(t->data != NULL, "Cannot copy to an empty data array!");
-    CheckNTErrors(s->unitSize == t->unitSize, "The input tensors must be of the same unit size!");
-    CheckNTErrors(s->order > sBeg && sBeg >= 0 && sLen <= s->unitNum, "Wrong segment on the source side");
-    CheckNTErrors(t->order > tBeg && tBeg >= 0, "Wrong segment on the target side");
+    CheckNTErrors(s->unitSize == t->unitSize, "Incompatible data types in value copy.");
+    CheckNTErrors(sBeg >= 0 && sBeg + sLen <= s->unitNum, "Wrong segment of the source data array");
+    CheckNTErrors(tBeg >= 0 && tBeg + sLen <= t->unitNum, "Wrong segment of the target data array");
 
 
     if (!s->isSparse && !t->isSparse) {
     if (!s->isSparse && !t->isSparse) {
         XMemCopy((char*)t->data + tBeg * t->unitSize, t->devID,
         XMemCopy((char*)t->data + tBeg * t->unitSize, t->devID,

source/tensor/core/movement/CopyValues.cu

@@ -37,10 +37,11 @@ copy a range of elements from a source vector to a target vector
 */
 */
 void _CudaCopyValues(const XTensor * s, XTensor * t, XStream * stream)
 void _CudaCopyValues(const XTensor * s, XTensor * t, XStream * stream)
 {
 {
-    CheckNTErrors((s != NULL && t != NULL), "The input tensor and output tensor must be nonempty!");
+    CheckNTErrors(s != NULL && t != NULL, "The input tensor and output tensor must be nonempty!");
     CheckNTErrors(s->dataType == t->dataType, "Unmatched data type!");
     CheckNTErrors(s->dataType == t->dataType, "Unmatched data type!");
-    CheckNTErrors((s->unitSize == t->unitSize), "Incompatible vectors in value copy.");
-    CheckNTErrors((s->denseRatio <= s->denseRatio), "Incompatible vectors in value copy.");
+    CheckNTErrors(s->unitSize == t->unitSize, "Incompatible data types in value copy.");
+    CheckNTErrors(s->unitNum == t->unitNum, "The data items are be the same.");
+    CheckNTErrors(s->denseRatio <= t->denseRatio, "Incompatible vectors in value copy.");
 
 
     /* dense -> dense */
     /* dense -> dense */
     if (!s->isSparse && !t->isSparse) {
     if (!s->isSparse && !t->isSparse) {