[GitHub] eric-haibin-lin commented on a change in pull request #11004: Only allocate cudnn-rnn dropout memory if dropout p > 0 and acquire descriptors during initialization

[GitBox <gi...@apache.org>]

eric-haibin-lin commented on a change in pull request #11004: Only allocate cudnn-rnn dropout memory if dropout p > 0 and acquire descriptors during initialization
URL: https://github.com/apache/incubator-mxnet/pull/11004#discussion_r189451435
 
 

 ##########
 File path: src/operator/cudnn_rnn-inl.h
 ##########
 @@ -76,37 +76,56 @@ class CuDNNRNNOp : public Operator{
       param_.lstm_q_ = true;
     else
       param_.lstm_q_ = false;
+
+    // Create descriptors
+    CUDNN_CALL(cudnnCreateTensorDescriptor(&hx_desc_));
+    CUDNN_CALL(cudnnCreateTensorDescriptor(&cx_desc_));
+    CUDNN_CALL(cudnnCreateTensorDescriptor(&hy_desc_));
+    CUDNN_CALL(cudnnCreateTensorDescriptor(&cy_desc_));
+    CUDNN_CALL(cudnnCreateTensorDescriptor(&dhx_desc_));
+    CUDNN_CALL(cudnnCreateTensorDescriptor(&dcx_desc_));
+    CUDNN_CALL(cudnnCreateTensorDescriptor(&dhy_desc_));
+    CUDNN_CALL(cudnnCreateTensorDescriptor(&dcy_desc_));
+
+    CUDNN_CALL(cudnnCreateFilterDescriptor(&w_desc_));
+    CUDNN_CALL(cudnnCreateFilterDescriptor(&dw_desc_));
+
+    CUDNN_CALL(cudnnCreateRNNDescriptor(&rnn_desc_));
+    CUDNN_CALL(cudnnCreateDropoutDescriptor(&dropout_desc_));
   }
 
   ~CuDNNRNNOp() {
+    CUDNN_CALL(cudnnDestroyTensorDescriptor(hx_desc_));
+    CUDNN_CALL(cudnnDestroyTensorDescriptor(cx_desc_));
+    CUDNN_CALL(cudnnDestroyTensorDescriptor(hy_desc_));
+    CUDNN_CALL(cudnnDestroyTensorDescriptor(cy_desc_));
+    CUDNN_CALL(cudnnDestroyTensorDescriptor(dhx_desc_));
+    CUDNN_CALL(cudnnDestroyTensorDescriptor(dcx_desc_));
+    CUDNN_CALL(cudnnDestroyTensorDescriptor(dhy_desc_));
+    CUDNN_CALL(cudnnDestroyTensorDescriptor(dcy_desc_));
+
+    CUDNN_CALL(cudnnDestroyFilterDescriptor(w_desc_));
+    CUDNN_CALL(cudnnDestroyFilterDescriptor(dw_desc_));
+    CUDNN_CALL(cudnnDestroyRNNDescriptor(rnn_desc_));
+    CUDNN_CALL(cudnnDestroyDropoutDescriptor(dropout_desc_));
+
     if (init_cudnn_) {
       for (size_t i = 0; i < x_desc_vec_.size(); ++i) {
         CUDNN_CALL(cudnnDestroyTensorDescriptor(x_desc_vec_[i]));
         CUDNN_CALL(cudnnDestroyTensorDescriptor(y_desc_vec_[i]));
         CUDNN_CALL(cudnnDestroyTensorDescriptor(dx_desc_vec_[i]));
         CUDNN_CALL(cudnnDestroyTensorDescriptor(dy_desc_vec_[i]));
+        init_cudnn_ = false;
 
 Review comment:
   why moved inside the loop?

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services