Merge pull request #261 from MyPandaShaoxiang/developing

fix normalize

Author: LittleMaer

saber/funcs/impl/cuda/base/cuda_c/saber_normalize.cu

@@ -9,7 +9,7 @@ template <typename Dtype, bool has_scale, bool shared>
 __global__ void normalize_kernel_no_across_spatial(const int size_in_channel, const int n,const int channels, \
     const Dtype* scale, const Dtype* bottom_data, Dtype* top_data, const float eps, const int p){
 
-    CUDA_KERNEL_LOOP(index, size_in_channel){
+    CUDA_KERNEL_LOOP(index, size_in_channel * n){
         float sqr_sum = 0.f;
         int num_index=index/size_in_channel;
         int index_in_channel=index%size_in_channel;
@@ -26,17 +26,16 @@ __global__ void normalize_kernel_no_across_spatial(const int size_in_channel, co
         if (p == 1) {
             norm = 1.f / (sqr_sum + eps);
         } else {
-            norm = 1.f / (sqrtf(sqr_sum) + eps);
+            norm = 1.f / sqrtf(sqr_sum + eps);
         }
-        Dtype has_scale_norm=scale[0]*norm;
         for (int i = 0; i < channels; ++i) {
             if (has_scale) {
                 if (shared) {
                     top_data[data_index + i * size_in_channel] = \
-                        bottom_data[data_index + i * size_in_channel] * scale[0] * has_scale_norm;
+                        bottom_data[data_index + i * size_in_channel] * scale[0] * norm;
                 } else {
                     top_data[data_index + i * size_in_channel] = \
-                        bottom_data[data_index + i * size_in_channel] * scale[i] * has_scale_norm;
+                        bottom_data[data_index + i * size_in_channel] * scale[i] * norm;
                 }
             } else {
                 top_data[data_index + i * size_in_channel] = \