diff --git a/include/caffe/neuron_layers.hpp b/include/caffe/neuron_layers.hpp
index 323215134c7..aff58233e5c 100644
--- a/include/caffe/neuron_layers.hpp
+++ b/include/caffe/neuron_layers.hpp
@@ -734,7 +734,8 @@ class PReLULayer : public NeuronLayer<Dtype> {
       const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
 
   bool channel_shared_;
-  Blob<Dtype> multiplier_;  // dot multipler for backward computation of params
+  Blob<Dtype> multiplier_;  // dot multiplier for backward computation of params
+  Blob<Dtype> backward_buff_;  // temporary buffer for backward computation
   Blob<Dtype> bottom_memory_;  // memory for in-place computation
 };
 
diff --git a/src/caffe/layers/prelu_layer.cpp b/src/caffe/layers/prelu_layer.cpp
index 7119a274dd3..7a38f9fac80 100644
--- a/src/caffe/layers/prelu_layer.cpp
+++ b/src/caffe/layers/prelu_layer.cpp
@@ -45,7 +45,8 @@ void PReLULayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
 
   // Propagate gradients to the parameters (as directed by backward pass).
   this->param_propagate_down_.resize(this->blobs_.size(), true);
-  multiplier_.Reshape(vector<int>(1, bottom[0]->count() / bottom[0]->num()));
+  multiplier_.Reshape(vector<int>(1, bottom[0]->count(1)));
+  backward_buff_.Reshape(vector<int>(1, bottom[0]->count(1)));
   caffe_set(multiplier_.count(), Dtype(1), multiplier_.mutable_cpu_data());
 }
 
diff --git a/src/caffe/layers/prelu_layer.cu b/src/caffe/layers/prelu_layer.cu
index fd0eda5d191..dfa238d85bd 100644
--- a/src/caffe/layers/prelu_layer.cu
+++ b/src/caffe/layers/prelu_layer.cu
@@ -86,22 +86,22 @@ void PReLULayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
     int cdim = channels * dim;
     Dtype dsum = 0.;
     for (int n = 0; n < bottom[0]->num(); ++n) {
-      Dtype* temp_buff = multiplier_.mutable_gpu_diff();
       // compute element-wise diff
       // NOLINT_NEXT_LINE(whitespace/operators)
-      PReLUParamBackward<Dtype><<<CAFFE_GET_BLOCKS(count),
+      PReLUParamBackward<Dtype><<<CAFFE_GET_BLOCKS(cdim),
           CAFFE_CUDA_NUM_THREADS>>>(
           cdim, top_diff + top[0]->offset(n),
-          bottom_data + bottom[0]->offset(n), multiplier_.mutable_gpu_diff());
+          bottom_data + bottom[0]->offset(n),
+          backward_buff_.mutable_gpu_diff());
       CUDA_POST_KERNEL_CHECK;
       if (channel_shared_) {
         Dtype d;
-        caffe_gpu_dot<Dtype>(channels * dim, multiplier_.gpu_diff(),
+        caffe_gpu_dot<Dtype>(channels * dim, backward_buff_.gpu_diff(),
             multiplier_.gpu_data(), &d);
         dsum += d;
       } else {
         caffe_gpu_gemv<Dtype>(CblasNoTrans, channels, dim, 1.,
-            multiplier_.gpu_diff(), multiplier_.gpu_data(), 1.,
+            backward_buff_.gpu_diff(), multiplier_.gpu_data(), 1.,
             slope_diff);
       }
     }