trunk,cudamatrix: bugfix, we should use GPU memory caching via CuDevice for CuArray too...

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3309 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

src/cudamatrix/cu-array-inl.h

@@ -50,7 +50,11 @@ void CuArray<T>::Resize(MatrixIndexT dim, MatrixResizeType resize_type) {
   
 #if HAVE_CUDA == 1
   if (CuDevice::Instantiate().Enabled()) {
-    CU_SAFE_CALL(cudaMalloc((void**)&data_, dim*sizeof(T)));
+    Timer tim;
+    this->data_ = static_cast<T*>(CuDevice::Instantiate().Malloc(dim * sizeof(T)));
+    this->dim_ = dim;
+    if (resize_type == kSetZero) this->SetZero();
+    CuDevice::Instantiate().AccuProfile("CuArray::Resize", tim.Elapsed());    
   } else
 #endif
   {
@@ -75,7 +79,7 @@ void CuArray<T>::Destroy() {
 #if HAVE_CUDA == 1
   if (CuDevice::Instantiate().Enabled()) { 
     if (data_ != NULL) {
-      CU_SAFE_CALL(cudaFree(data_));
+      CuDevice::Instantiate().Free(this->data_);
     }
   } else
 #endif