[MIOPEN] Optimization: reduce HIP module API calls in Reshape()

Author: whchung

include/caffe/layers/cudnn_conv_layer.hpp

@@ -59,6 +59,8 @@ class CuDNNConvolutionLayer : public ConvolutionLayer<Dtype> {
   mlopenTensorDescriptor_t              bias_desc_;
   mlopenTensorDescriptor_t              filter_desc_;
   vector<mlopenConvolutionDescriptor_t> conv_descs_;
+
+  int N_, C_, W_, H_;
 #endif
 
 #ifdef USE_CUDNN

src/caffe/layers/cudnn_conv_layer.cpp

@@ -146,6 +146,10 @@ void CuDNNConvolutionLayer<Dtype>::LayerSetUp(
 #endif
   }
 
+#ifdef USE_MIOPEN
+  N_ = C_ = W_ = H_ = 0;
+#endif
+
   handles_setup_ = true;
 }
 
@@ -185,6 +189,25 @@ void CuDNNConvolutionLayer<Dtype>::Reshape(
   size_t workspace_limit_bytes = 8*1024*1024;
 #endif
 
+#ifdef USE_MIOPEN
+  bool doReshape = false;
+  if ((N_ != this->num_) ||
+      (C_ != this->channels_ / this->group_) ||
+      (H_ != height) ||
+      (W_ != width)) {
+    doReshape = true;
+    //LOG(INFO) << "doReshape\n";
+    N_ = this->num_;
+    C_ = this->channels_ / this->group_;
+    H_ = height;
+    W_ = width;
+  } else {
+    //LOG(INFO) << "NOT doReshape\n";
+  }
+
+  if (doReshape) {
+#endif
+
   for (int i = 0; i < bottom.size(); i++) {
 #ifdef USE_MIOPEN
     miopen::setTensor4dDesc<Dtype>(&bottom_descs_[i],
@@ -370,6 +393,10 @@ void CuDNNConvolutionLayer<Dtype>::Reshape(
 #endif
   }
 
+#ifdef USE_MIOPEN
+  } // if (doReshape)
+#endif
+
   // reduce over all workspace sizes to get a maximum to allocate / reallocate
   size_t total_workspace_fwd = 0;
   size_t total_workspace_bwd_data = 0;