Batchnorm inference instances, external API, client examples and gtests (#531)

* File renaming and class renaming for device element-wise operation

* Add batchnorm-infer instances, external API and client example

* Add batchnorm-infer profiler module and gtests

* Remove file device_elementwise_extension.hpp and move NormalizeInInfer operation to element_wise_operation.hpp

* Remove the using of class aliasing for DeviceElementwiseForBatchNormInfer

* Rename class and file due to conflict from device_elementwise_2d.hpp

* Fix namespace in batcnnorm_infer_nhwc client example

include/ck/tensor_operation/gpu/element/element_wise_operation.hpp

@@ -314,6 +314,40 @@ struct Normalize
     double epsilon_;
 };
 
+// used by BatchNorm inference
+// y = gamma * (x-mean) / sqrt(epsilon+variance) + beta
+// The data type of mean and variance is used as AccDataType
+struct NormalizeInInfer
+{
+    NormalizeInInfer(double epsilon = 1e-4) : epsilon_(epsilon) {}
+
+    template <typename T1, typename T2, typename T3, typename T4>
+    __host__ __device__ constexpr void operator()(T1& y,
+                                                  const T1& x,
+                                                  const T2& mean,
+                                                  const T2& variance,
+                                                  const T3& gamma,
+                                                  const T4& beta) const
+    {
+        static_assert(std::is_same<T2, float>::value || std::is_same<T2, double>::value,
+                      "Data type is not supported by this operation!");
+
+        using ck::type_convert;
+        using ck::math::sqrt;
+
+        T2 tmp_x, tmp_y;
+
+        tmp_x = type_convert<T2>(x);
+
+        tmp_y = ((tmp_x - mean) / sqrt(variance + type_convert<T2>(epsilon_))) *
+                    type_convert<T2>(gamma) +
+                type_convert<T2>(beta);
+        y = type_convert<T1>(tmp_y);
+    };
+
+    double epsilon_;
+};
+
 template <typename Y, typename X>
 struct UnaryTypeConvert;