better control of the strategy of the merge method

source/tensor/XTensor.h

@@ -49,6 +49,8 @@ struct XLink;
 #define USE_BATCHED_STRIDED_MAT_MUL
 #define USE_BATCHED_STRIDED_MAT_MUL
 #define MIN_TENSOR_SPLIT_NUM 0
 #define MIN_TENSOR_SPLIT_NUM 0
 #define MIN_TENSOR_SPLIT_LIST_NUM 1024
 #define MIN_TENSOR_SPLIT_LIST_NUM 1024
+#define MIN_TENSOR_MERGE_NUM 0
+#define MIN_TENSOR_MERGE_LIST_NUM 1024
 #define MIN_TENSOR_CAT_NUM 8
 #define MIN_TENSOR_CAT_NUM 8
 
 
 /* computation flags */
 /* computation flags */

source/tensor/core/shape/Merge.cpp

@@ -94,7 +94,7 @@ void _Merge(const XTensor * s, XTensor * t, int whereToMerge, int leadingDim)
     gridSize = blockNum;
     gridSize = blockNum;
     gridNum = s->unitNum / (blockSize * blockNum);
     gridNum = s->unitNum / (blockSize * blockNum);
 
 
-    if (mergedNum * gridNum <= MIN_TENSOR_SPLIT_NUM) {
+    if (mergedNum * gridNum <= MIN_TENSOR_MERGE_NUM) {
         int sPitch = blockSize * s->unitSize;
         int sPitch = blockSize * s->unitSize;
         int tPtich = blockSize * mergedNum * t->unitSize;
         int tPtich = blockSize * mergedNum * t->unitSize;
         int mSize = blockSize * t->unitSize;
         int mSize = blockSize * t->unitSize;
@@ -253,7 +253,7 @@ void _Merge(const XList * smalls, XTensor * big, int whereToMerge)
     gridNum = s0->unitNum / (blockSize * blockNum);
     gridNum = s0->unitNum / (blockSize * blockNum);
 
 
     /* merging with fewer data copy operations */
     /* merging with fewer data copy operations */
-    if (mergedNum * gridNum <= MIN_TENSOR_SPLIT_LIST_NUM) {
+    if (mergedNum * gridNum <= MIN_TENSOR_MERGE_LIST_NUM) {
         int sPitch = blockSize * s0->unitSize;
         int sPitch = blockSize * s0->unitSize;
         int tPtich = blockSize * mergedNum * big->unitSize;
         int tPtich = blockSize * mergedNum * big->unitSize;
         int mSize = blockSize * big->unitSize;
         int mSize = blockSize * big->unitSize;