Reduce number of MIOpen handles to 1 per GPU in convolution layers

include/caffe/util/cudnn.hpp

@@ -3,6 +3,9 @@
 #ifdef USE_ACCMI
 
 #ifdef USE_MIOPEN
+#include <mutex>
+#include <unordered_map>
+
 #include <miopen/miopen.h>
 
 #include "caffe/common.hpp"
@@ -41,6 +44,48 @@ namespace caffe {
 
 namespace miopen {
 
+class miopenHandleMap {
+private:
+  miopenHandleMap() : map_(), lock_() {}
+public:
+  static miopenHandleMap& getInstance() {
+    static miopenHandleMap instance_{};
+    return instance_;
+  }
+
+  ~miopenHandleMap() {
+    for (auto iter : map_) {
+      miopenDestroy(iter.second);
+    }
+    map_.clear();
+  }
+
+  miopenHandle_t getHandle(int device) {
+    miopenHandle_t ret = nullptr;
+    std::lock_guard<std::mutex> l(lock_);
+    if (map_.find(device) != map_.end()) {
+      ret = map_[device];
+    }
+    return ret;
+  }
+
+  bool setHandle(int device, miopenHandle_t handle) {
+    bool ret = false;
+    std::lock_guard<std::mutex> l(lock_);
+    if (map_.find(device) != map_.end()) {
+      LOG(FATAL) << "Duplicated MIOpen handle for device: " << device;
+    } else {
+      map_[device] = handle;
+      ret = true;
+    }
+    return ret;
+  }
+
+private:
+  std::unordered_map<int, miopenHandle_t> map_;
+  std::mutex lock_;
+};
+
 template <typename Dtype> class dataType;
 template<> class dataType<float>  {
  public:

src/caffe/layers/cudnn_conv_layer.cpp

@@ -13,7 +13,7 @@ CuDNNConvolutionLayer<Dtype>::CuDNNConvolutionLayer(const LayerParameter& param)
     : ConvolutionLayer<Dtype>(param), handles_setup_(false),
       fwd_algo_(), bwd_weight_algo_(), bwd_data_algo_(),
       workspace_fwd_sizes_(), workspace_bwd_filter_sizes_(), workspace_bwd_data_sizes_(),
-      workspace() { }
+      workspace(), handle_(nullptr) { }
 
 /**
  * TODO(dox) explain cuDNN interface
@@ -57,7 +57,18 @@ void CuDNNConvolutionLayer<Dtype>::LayerSetUp(
 
   for (int g = 0; g < this->group_ * WORKSPACE_PER_GROUP; g++) {
 #ifdef USE_MIOPEN
-    MIOPEN_CHECK(miopenCreateWithStream(&handle_, nullptr));
+    int device;
+    HIP_CHECK(hipGetDevice(&device));
+
+    auto& hmap = caffe::miopen::miopenHandleMap::getInstance();
+    handle_ = hmap.getHandle(device);
+    if (handle_ == nullptr) {
+      DLOG(INFO) << "Creating MIOpen handle on device: " << device;
+      MIOPEN_CHECK(miopenCreateWithStream(&handle_, nullptr));
+      hmap.setHandle(device, handle_);
+    } else {
+      DLOG(INFO) << "Get MIOpen handle from cache on device: " << device;
+    }
 #endif
 
     workspace[g] = NULL;
@@ -435,10 +446,6 @@ CuDNNConvolutionLayer<Dtype>::~CuDNNConvolutionLayer() {
 #endif
 #endif
 
-#ifdef USE_MIOPEN
-    miopenDestroy(handle_);
-#endif
-
   hipFree(workspaceData);
 }