Set bottom_diff to 0 and remove Async memcopy

diff --git a/src/caffe/layers/pooling_layer.cu b/src/caffe/layers/pooling_layer.cu
index 95a754f4a1753242d38029d4a11ee1bc3fd8f41e..77126e6b1207276472da6165ae8dd410b82bbc56 100644 (file)
--- a/src/caffe/layers/pooling_layer.cu
+++ b/src/caffe/layers/pooling_layer.cu
@@ -327,6 +327,7 @@ void PoolingLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
     // Since we have the mask we only need count top_diff
     count = top[0]->count(); 
     // NOLINT_NEXT_LINE(whitespace/operators)
+    caffe_gpu_memset(count,Dtype(0.),bottom_diff);
     MaxPoolBackward<Dtype><<<CAFFE_GET_BLOCKS(count), CAFFE_CUDA_NUM_THREADS>>>(
         count, top_diff, top[0]->num(), channels_,
         height_, width_, pooled_height_, pooled_width_,