inference acc: add the round to even handling.

Author: vera121

src/caffe/layers/conv_layer.cu

@@ -32,6 +32,7 @@ void ConvolutionLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
     const int count_t = top[i]->count();
     if (output_scale != Dtype(1)) {
       caffe_gpu_scal(count_t, output_scale, top_data);
+      caffe_gpu_round(count_t, top_data);
     }
     //CUSTOMIZATION-->
   }

src/caffe/layers/deconv_layer.cu

@@ -31,6 +31,7 @@ void DeconvolutionLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
     const int count_t = top[i]->count();
     if (output_scale != Dtype(1)) {
       caffe_gpu_scal(count_t, output_scale, top_data);
+      caffe_gpu_round(count_t, top_data);
     }
     //CUSTOMIZATION-->
   }

src/caffe/layers/inner_product_layer.cu

@@ -38,6 +38,7 @@ void InnerProductLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
     const int count_t = top[0]->count();
     if (output_scale_ != Dtype(1)) {
       caffe_gpu_scal(count_t, output_scale_, top_data);
+      caffe_gpu_round(count_t, top_data);
     }
   //CUSTOMIZATION-->
 }

src/caffe/util/math_functions.cu

@@ -163,7 +163,7 @@ void caffe_gpu_scale<double>(const int n, const double alpha, const double *x,
 template <typename Dtype>
 __global__ void round_kernel(const int n, Dtype* y) {
   CUDA_KERNEL_LOOP(index, n) {
-    y[index] = round(y[index]);
+    y[index] = rint(y[index]);
   }
 }