better CPU code

source/tensor/core/arithmetic/MatrixMulBatched.cpp

@@ -84,17 +84,25 @@ void _MatrixMulBatched(const XTensor * a, MATRIX_TRANS_TYPE transposedA,
     XList * aList = new XList(10);
     XList * aList = new XList(10);
     XList * bList = new XList(10);
     XList * bList = new XList(10);
     XList * cList = new XList(10);
     XList * cList = new XList(10);
-    int aDimSize[2] = { -a->dimSizeRDI[1], a->dimSizeRDI[0] };
-    int bDimSize[2] = { -b->dimSizeRDI[1], b->dimSizeRDI[0] };
-    int cDimSize[2] = { -c->dimSizeRDI[1], c->dimSizeRDI[0] };
+    int aDimSize[2] = {-a->dimSizeRDI[1], a->dimSizeRDI[0]};
+    int bDimSize[2] = {-b->dimSizeRDI[1], b->dimSizeRDI[0]};
+    int cDimSize[2] = {-c->dimSizeRDI[1], c->dimSizeRDI[0]};
+
+    XTensor * tensorBuf = new XTensor[blockNum * 3];
+    XTensor * aBuf = tensorBuf;
+    XTensor * bBuf = tensorBuf + blockNum;
+    XTensor * cBuf = tensorBuf + blockNum * 2;
 
 
     for (int p = 0; p < blockNum; p++) {
     for (int p = 0; p < blockNum; p++) {
         void * ap = (char*)a->data + aRealBlockSize * p;
         void * ap = (char*)a->data + aRealBlockSize * p;
         void * bp = (char*)b->data + bRealBlockSize * p;
         void * bp = (char*)b->data + bRealBlockSize * p;
         void * cp = (char*)c->data + cRealBlockSize * p;
         void * cp = (char*)c->data + cRealBlockSize * p;
-        XTensor * ai = NewTensor(2, aDimSize, a->dataType, a->denseRatio, a->devID, a->mem);
-        XTensor * bi = NewTensor(2, bDimSize, b->dataType, b->denseRatio, b->devID, b->mem);
-        XTensor * ci = NewTensor(2, cDimSize, c->dataType, c->denseRatio, c->devID, c->mem);
+        XTensor * ai = aBuf + p;
+        XTensor * bi = bBuf + p;
+        XTensor * ci = cBuf + p;
+        InitTensor(ai, 2, aDimSize, a->dataType, a->denseRatio, a->devID, a->mem);
+        InitTensor(bi, 2, bDimSize, b->dataType, b->denseRatio, b->devID, b->mem);
+        InitTensor(ci, 2, cDimSize, c->dataType, c->denseRatio, c->devID, c->mem);
         ai->data = ap;
         ai->data = ap;
         bi->data = bp;
         bi->data = bp;
         ci->data = cp;
         ci->data = cp;
@@ -132,22 +140,21 @@ void _MatrixMulBatched(const XTensor * a, MATRIX_TRANS_TYPE transposedA,
 
 
     for (int i = 0; i < aList->count; i++) {
     for (int i = 0; i < aList->count; i++) {
         XTensor * ai = (XTensor*)aList->GetItem(i);
         XTensor * ai = (XTensor*)aList->GetItem(i);
-        ai->data = NULL;
-        delete ai;
+        ai->data = NULL;;
     }
     }
 
 
     for (int i = 0; i < bList->count; i++) {
     for (int i = 0; i < bList->count; i++) {
         XTensor * bi = (XTensor*)bList->GetItem(i);
         XTensor * bi = (XTensor*)bList->GetItem(i);
         bi->data = NULL;
         bi->data = NULL;
-        delete bi;
     }
     }
 
 
     for (int i = 0; i < cList->count; i++) {
     for (int i = 0; i < cList->count; i++) {
         XTensor * ci = (XTensor*)cList->GetItem(i);
         XTensor * ci = (XTensor*)cList->GetItem(i);
         ci->data = NULL;
         ci->data = NULL;
-        delete ci;
     }
     }
 
 
+    delete[] tensorBuf;
+
     delete aList;
     delete aList;
     delete bList;
     delete bList;
     delete cList;
     delete cList;