for code review 6

Author: sweetsky0901

paddle/operators/math/maxouting.cc

@@ -85,11 +85,9 @@ class MaxOutGradFunctor<platform::CPUPlace, T> {
           int output_idx = blen + clen + f;
           for (int g = 0; g < groups && continue_match; ++g) {
               int input_idx = input_idx0 + fea_size * g;
-              input_grad_data[input_idx] = 0;
               if (input_data[input_idx] == output_data[output_idx]) {
                 input_grad_data[input_idx] += output_grad_data[output_idx];
                 continue_match = false;
-                break;
               }
           }
         }

paddle/operators/math/maxouting.cu

@@ -69,8 +69,7 @@ __global__ void KernelMaxoutGrad(
         }
       }
       if (max_index != -1) {
-        // atomic add
-        platform::CudaAtomicAdd(input_grad + max_index, output_grad[index]);
+        input_grad[max_index] += output_grad[index];
       }
     }
 }

paddle/operators/maxout_op.cu.cc

@@ -15,13 +15,11 @@
 #include "paddle/operators/maxout_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(maxout, ops::MaxOutKernel<paddle::platform::GPUPlace,
-                       float>);
-REGISTER_OP_GPU_KERNEL(maxout, ops::MaxOutKernel<paddle::platform::GPUPlace,
-                       double>);
+REGISTER_OP_GPU_KERNEL(maxout,
+                       ops::MaxOutKernel<paddle::platform::GPUPlace, float>,
+                       ops::MaxOutKernel<paddle::platform::GPUPlace, double>);
 REGISTER_OP_GPU_KERNEL(maxout_grad,
                        ops::MaxOutGradKernel<paddle::platform::GPUPlace,
-                       float>);
-REGISTER_OP_GPU_KERNEL(maxout_grad,
+                        float>,
                        ops::MaxOutGradKernel<paddle::platform::GPUPlace,
-                       double>);
+                        double>);