COMPMID-2245: Extend NEFuseBatchNormalization to support DepthwiseConvolution weights

Change-Id: I2ee4aebfd69865290ed6c78dd17ff1299353317e
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1371
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

5 files changed, 603 insertions, 140 deletions

diff --git a/arm_compute/core/NEON/kernels/NEFuseBatchNormalizationKernel.h b/arm_compute/core/NEON/kernels/NEFuseBatchNormalizationKernel.h
index 97d6a9f55b..947f8e49c4 100644
--- a/arm_compute/core/NEON/kernels/NEFuseBatchNormalizationKernel.h
+++ b/arm_compute/core/NEON/kernels/NEFuseBatchNormalizationKernel.h
@@ -53,44 +53,50 @@ public:
     ~NEFuseBatchNormalizationKernel() = default;
     /** Set the source, destination of the kernel
      *
-     * @param[in]  conv_weights  Convolution layer weights tensor. Data type supported: F16/F32
-     * @param[in]  bn_mean       Batch normalization layer mean tensor. Same as @p conv_weights
-     * @param[in]  bn_var        Batch normalization layer variance tensor. Same as @p conv_weights
-     * @param[out] fused_weights Output fused weights tensor. Same as @p conv_weights
-     * @param[out] fused_bias    Output fused bias tensor. Same as @p conv_weights
-     * @param[in]  conv_bias     (Optional) Convolution layer bias tensor. Same as @p conv_weights
-     * @param[in]  bn_beta       (Optional) Batch normalization layer beta tensor. Same as @p conv_weights
-     * @param[in]  bn_gamma      (Optional) Batch normalization layer gamma tensor. Same as @p conv_weights
+     * @param[in]  input_weights Input weights tensor for convolution or depthwise convolution layer. Data type supported: F16/F32. Data layout supported: NCHW, NHWC
+     * @param[in]  bn_mean       Batch normalization layer mean tensor. Same as @p input_weights
+     * @param[in]  bn_var        Batch normalization layer variance tensor. Same as @p input_weights
+     * @param[out] fused_weights (Optional) Output fused weights tensor. It can be a nullptr in case of in-place computation. Same as @p input_weights
+     * @param[out] fused_bias    (Optional) Output fused bias tensor. It can be a nullptr in case of in-place computation and input_bias != nullptr. Same as @p input_weights
+     * @param[in]  input_bias    (Optional) Input bias tensor for convolution or depthwise convolution layer. It can be a nullptr in case the bias tensor is not required. Same as @p input_weights
+     * @param[in]  bn_beta       (Optional) Batch normalization layer beta tensor. It can be a nullptr in case the beta tensor is not required. Same as @p input_weights
+     *                           @note if nullptr, bn_beta is set to 0.0
+     * @param[in]  bn_gamma      (Optional) Batch normalization layer gamma tensor. It can be a nullptr in case the gamma tensor is not required. Same as @p input_weights
+     *                           @note if nullptr, bn_gamma is set to 1.0
      * @param[in]  epsilon       (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f.
+     * @param[in]  fbn_type      (Optional) Fused batch normalization type. Defaults to CONVOLUTION.
      */
-    void configure(const ITensor *conv_weights, const ITensor *bn_mean, const ITensor *bn_var, ITensor *fused_weights, ITensor *fused_bias,
-                   const ITensor *conv_bias = nullptr, const ITensor *bn_beta = nullptr, const ITensor *bn_gamma = nullptr,
-                   float epsilon = 0.001f);
+    void configure(const ITensor *input_weights, const ITensor *bn_mean, const ITensor *bn_var, ITensor *fused_weights, ITensor *fused_bias,
+                   const ITensor *input_bias = nullptr, const ITensor *bn_beta = nullptr, const ITensor *bn_gamma = nullptr,
+                   float epsilon = 0.001f, FuseBatchNormalizationType fbn_type = FuseBatchNormalizationType::CONVOLUTION);
     /** Static function to check if given info will lead to a valid configuration of @ref NEFuseBatchNormalizationKernel
      *
-     * @param[in] conv_weights  Convolution layer weights tensor. Data type supported: F16/F32
-     * @param[in] bn_mean       Batch normalization layer mean tensor. Same as @p conv_weights
-     * @param[in] bn_var        Batch normalization layer variance tensor. Same as @p conv_weights
-     * @param[in] fused_weights Output fused weights tensor. Same as @p conv_weights
-     * @param[in] fused_bias    Output fused bias tensor. Same as @p conv_weights
-     * @param[in] conv_bias     (Optional) Convolution layer bias tensor. Same as @p conv_weights
-     * @param[in] bn_beta       (Optional) Batch normalization layer beta tensor. Same as @p conv_weights
-     * @param[in] bn_gamma      (Optional) Batch normalization layer gamma tensor. Same as @p conv_weights
+     * @param[in] input_weights Input weights tensor info for convolution or depthwise convolution layer. Data type supported: F16/F32. Data layout supported: NCHW, NHWC
+     * @param[in] bn_mean       Batch normalization layer mean tensor info. Same as @p input_weights
+     * @param[in] bn_var        Batch normalization layer variance tensor info. Same as @p input_weights
+     * @param[in] fused_weights (Optional) Output fused weights tensor info. It can be a nullptr in case of in-place computation. Same as @p input_weights
+     * @param[in] fused_bias    (Optional) Output fused bias tensor info. It can be a nullptr in case of in-place computation and input_bias != nullptr. Same as @p input_weights
+     * @param[in] input_bias    (Optional) Input bias tensor info for convolution or depthwise convolution layer. It can be a nullptr in case the bias tensor is not required. Same as @p input_weights
+     * @param[in] bn_beta       (Optional) Batch normalization layer beta tensor info. It can be a nullptr in case the beta tensor is not required. Same as @p input_weights
+     *                          @note if nullptr, bn_beta is set to 0.0
+     * @param[in] bn_gamma      (Optional) Batch normalization layer gamma tensor info. It can be a nullptr in case the gamma tensor is not required. Same as @p input_weights
+     *                          @note if nullptr, bn_gamma is set to 1.0
      * @param[in] epsilon       (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f.
+     * @param[in] fbn_type      (Optional) Fused batch normalization type. Defaults to CONVOLUTION.
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *conv_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
+    static Status validate(const ITensorInfo *input_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
                            const ITensorInfo *fused_weights, const ITensorInfo *fused_bias,
-                           const ITensorInfo *conv_bias = nullptr, const ITensorInfo *bn_beta = nullptr, const ITensorInfo *bn_gamma = nullptr,
-                           float epsilon = 0.001f);
+                           const ITensorInfo *input_bias = nullptr, const ITensorInfo *bn_beta = nullptr, const ITensorInfo *bn_gamma = nullptr,
+                           float epsilon = 0.001f, FuseBatchNormalizationType fbn_type = FuseBatchNormalizationType::CONVOLUTION);
 
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
 
 private:
-    const ITensor *_conv_weights;
-    const ITensor *_conv_bias;
+    const ITensor *_input_weights;
+    const ITensor *_input_bias;
     const ITensor *_bn_mean;
     const ITensor *_bn_var;
     const ITensor *_bn_gamma;
@@ -101,7 +107,7 @@ private:
     bool           _run_in_place_weights;
     bool           _run_in_place_bias;
 
-    using FuseBatchNormFunction = void(const ITensor *conv_weights, const ITensor *conv_bias, ITensor *fused_weights, ITensor *fused_bias,
+    using FuseBatchNormFunction = void(const ITensor *input_weights, const ITensor *input_bias, ITensor *fused_weights, ITensor *fused_bias,
                                        const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window);
 
     FuseBatchNormFunction *_func;
diff --git a/arm_compute/runtime/NEON/functions/NEFuseBatchNormalization.h b/arm_compute/runtime/NEON/functions/NEFuseBatchNormalization.h
index 5e6286966a..3a2f6ccb6d 100644
--- a/arm_compute/runtime/NEON/functions/NEFuseBatchNormalization.h
+++ b/arm_compute/runtime/NEON/functions/NEFuseBatchNormalization.h
@@ -52,37 +52,43 @@ public:
     ~NEFuseBatchNormalization() = default;
     /** Set the input and output tensors.
      *
-     * @param[in]  conv_weights  Convolution layer weights tensor. Data type supported: F16/F32
-     * @param[in]  bn_mean       Batch normalization layer mean tensor. Same as @p conv_weights
-     * @param[in]  bn_var        Batch normalization layer variance tensor. Same as @p conv_weights
-     * @param[out] fused_weights Output fused weights tensor. Same as @p conv_weights
-     * @param[out] fused_bias    Output fused bias tensor. Same as @p conv_weights
-     * @param[in]  conv_bias     (Optional) Convolution layer bias tensor. Same as @p conv_weights
-     * @param[in]  bn_beta       (Optional) Batch normalization layer beta tensor. Same as @p conv_weights
-     * @param[in]  bn_gamma      (Optional) Batch normalization layer gamma tensor. Same as @p conv_weights
+     * @param[in]  input_weights Input weights tensor for convolution or depthwise convolution layer. Data type supported: F16/F32. Data layout supported: NCHW, NHWC
+     * @param[in]  bn_mean       Batch normalization layer mean tensor. Same as @p input_weights
+     * @param[in]  bn_var        Batch normalization layer variance tensor. Same as @p input_weights
+     * @param[out] fused_weights (Optional) Output fused weights tensor. It can be a nullptr in case of in-place computation. Same as @p input_weights
+     * @param[out] fused_bias    (Optional) Output fused bias tensor. It can be a nullptr in case of in-place computation and input_bias != nullptr. Same as @p input_weights
+     * @param[in]  input_bias    (Optional) Input bias tensor for convolution or depthwise convolution layer. It can be a nullptr in case the bias tensor is not required. Same as @p input_weights
+     * @param[in]  bn_beta       (Optional) Batch normalization layer beta tensor. It can be a nullptr in case the beta tensor is not required. Same as @p input_weights
+     *                           @note if nullptr, bn_beta is set to 0.0
+     * @param[in]  bn_gamma      (Optional) Batch normalization layer gamma tensor. It can be a nullptr in case the gamma tensor is not required. Same as @p input_weights
+     *                           @note if nullptr, bn_gamma is set to 1.0
      * @param[in]  epsilon       (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f.
+     * @param[in]  fbn_type      (Optional) Fused batch normalization type. Defaults to Convolution.
      */
-    void configure(const ITensor *conv_weights, const ITensor *bn_mean, const ITensor *bn_var, ITensor *fused_weights, ITensor *fused_bias,
-                   const ITensor *conv_bias = nullptr, const ITensor *bn_beta = nullptr, const ITensor *bn_gamma = nullptr,
-                   float epsilon = 0.001f);
+    void configure(const ITensor *input_weights, const ITensor *bn_mean, const ITensor *bn_var, ITensor *fused_weights, ITensor *fused_bias,
+                   const ITensor *input_bias = nullptr, const ITensor *bn_beta = nullptr, const ITensor *bn_gamma = nullptr,
+                   float epsilon = 0.001f, FuseBatchNormalizationType fbn_type = FuseBatchNormalizationType::CONVOLUTION);
     /** Static function to check if given info will lead to a valid configuration of @ref NEFuseBatchNormalization
      *
-     * @param[in] conv_weights  Convolution layer weights tensor. Data type supported: F16/F32
-     * @param[in] bn_mean       Batch normalization layer mean tensor. Same as @p conv_weights
-     * @param[in] bn_var        Batch normalization layer variance tensor. Same as @p conv_weights
-     * @param[in] fused_weights Output fused weights tensor. Same as @p conv_weights
-     * @param[in] fused_bias    Output fused bias tensor. Same as @p conv_weights
-     * @param[in] conv_bias     (Optional) Convolution layer bias tensor. Same as @p conv_weights
-     * @param[in] bn_beta       (Optional) Batch normalization layer beta tensor. Same as @p conv_weights
-     * @param[in] bn_gamma      (Optional) Batch normalization layer gamma tensor. Same as @p conv_weights
+     * @param[in] input_weights Input weights tensor info for convolution or depthwise convolution layer. Data type supported: F16/F32. Data layout supported: NCHW, NHWC
+     * @param[in] bn_mean       Batch normalization layer mean tensor info. Same as @p input_weights
+     * @param[in] bn_var        Batch normalization layer variance tensor info. Same as @p input_weights
+     * @param[in] fused_weights (Optional) Output fused weights tensor info. It can be a nullptr in case of in-place computation. Same as @p input_weights
+     * @param[in] fused_bias    (Optional) Output fused bias tensor info. It can be a nullptr in case of in-place computation and input_bias != nullptr. Same as @p input_weights
+     * @param[in] input_bias    (Optional) Input bias tensor info for convolution or depthwise convolution layer. It can be a nullptr in case the bias tensor is not required. Same as @p input_weights
+     * @param[in] bn_beta       (Optional) Batch normalization layer beta tensor info. It can be a nullptr in case the beta tensor is not required. Same as @p input_weights
+     *                          @note if nullptr, bn_beta is set to 0.0
+     * @param[in] bn_gamma      (Optional) Batch normalization layer gamma tensor info. It can be a nullptr in case the gamma tensor is not required. Same as @p input_weights
+     *                          @note if nullptr, bn_gamma is set to 1.0
      * @param[in] epsilon       (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f.
+     * @param[in] fbn_type      (Optional) Fused batch normalization type. Defaults to Convolution.
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *conv_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
+    static Status validate(const ITensorInfo *input_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
                            const ITensorInfo *fused_weights, const ITensorInfo *fused_bias,
-                           const ITensorInfo *conv_bias = nullptr, const ITensorInfo *bn_beta = nullptr, const ITensorInfo *bn_gamma = nullptr,
-                           float epsilon = 0.001f);
+                           const ITensorInfo *input_bias = nullptr, const ITensorInfo *bn_beta = nullptr, const ITensorInfo *bn_gamma = nullptr,
+                           float epsilon = 0.001f, FuseBatchNormalizationType fbn_type = FuseBatchNormalizationType::CONVOLUTION);
 
     // Inherited methods overridden:
     void run() override;
diff --git a/src/core/NEON/kernels/NEFuseBatchNormalizationKernel.cpp b/src/core/NEON/kernels/NEFuseBatchNormalizationKernel.cpp
index d45e3ce56a..836e429aba 100644
--- a/src/core/NEON/kernels/NEFuseBatchNormalizationKernel.cpp
+++ b/src/core/NEON/kernels/NEFuseBatchNormalizationKernel.cpp
@@ -26,74 +26,86 @@
 #include "arm_compute/core/CPP/Validate.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/wrapper/wrapper.h"
 #include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/Window.h"
-
 #include "support/ToolchainSupport.h"
 
-#include "arm_compute/core/NEON/wrapper/wrapper.h"
 #include "utils/TypePrinter.h"
+#include <map>
+
 namespace arm_compute
 {
 namespace
 {
-Status validate_arguments(const ITensorInfo *conv_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
+Status validate_arguments(const ITensorInfo *input_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
                           const ITensorInfo *fused_weights, const ITensorInfo *fused_bias,
-                          const ITensorInfo *conv_bias, const ITensorInfo *bn_beta, const ITensorInfo *bn_gamma,
-                          float epsilon)
+                          const ITensorInfo *input_bias, const ITensorInfo *bn_beta, const ITensorInfo *bn_gamma,
+                          float epsilon, FuseBatchNormalizationType fbn_type)
 {
     ARM_COMPUTE_UNUSED(epsilon);
-    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(conv_weights);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(conv_weights, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input_weights, bn_mean, bn_var);
+    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input_weights);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input_weights, 1, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(bn_mean, bn_var);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(conv_weights, bn_mean, bn_var);
-
-    unsigned int kernels_idx = get_data_layout_dimension_index(conv_weights->data_layout(), DataLayoutDimension::BATCHES);
-    ARM_COMPUTE_RETURN_ERROR_ON(conv_weights->dimension(kernels_idx) != bn_mean->dimension(0));
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, bn_mean, bn_var);
+    ARM_COMPUTE_RETURN_ERROR_ON(input_bias == nullptr && fused_bias == nullptr);
+    ARM_COMPUTE_RETURN_ERROR_ON(bn_mean->num_dimensions() > 1);
 
+    if(fbn_type == FuseBatchNormalizationType::CONVOLUTION)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(input_weights->dimension(3) != bn_mean->dimension(0));
+    }
+    else
+    {
+        const size_t channel_idx = get_data_layout_dimension_index(input_weights->data_layout(), DataLayoutDimension::CHANNEL);
+        ARM_COMPUTE_RETURN_ERROR_ON(input_weights->dimension(channel_idx) != bn_mean->dimension(0));
+    }
     // Validate bias
-    if(conv_bias != nullptr)
+    if(input_bias != nullptr)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(bn_mean, conv_bias);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(conv_weights, conv_bias);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(bn_mean, input_bias);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, input_bias);
     }
     // Validate beta
     if(bn_beta != nullptr)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(bn_mean, bn_beta);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(conv_weights, bn_beta);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, bn_beta);
     }
     // Validate gamma
     if(bn_gamma != nullptr)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(bn_mean, bn_gamma);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(conv_weights, bn_gamma);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, bn_gamma);
     }
 
     // Validate output weights
     if(fused_weights != nullptr && fused_weights->total_size() != 0)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(conv_weights, fused_weights);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(conv_weights, fused_weights);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(conv_weights, fused_weights);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input_weights, fused_weights);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input_weights, fused_weights);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, fused_weights);
     }
     // Validate output bias
     if(fused_bias != nullptr && fused_bias->total_size() != 0)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(bn_mean, fused_bias);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(conv_weights, fused_bias);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, fused_bias);
     }
 
     return Status{};
 }
 
-template <typename ScalarType, int size>
-void fused_batch_normmalization(const ITensor *conv_weights, const ITensor *conv_bias, ITensor *fused_weights, ITensor *fused_bias,
-                                const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window)
+template <typename VectorType>
+void fused_batch_normalization_conv(const ITensor *conv_weights, const ITensor *conv_bias, ITensor *fused_weights, ITensor *fused_bias,
+                                    const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window)
 {
-    using ExactTagType = typename wrapper::traits::neon_vector<ScalarType, size>::tag_type;
+    using ScalarType   = typename VectorType::scalar_type;
+    const int size     = 16 / conv_weights->info()->element_size();
+    using ExactTagType = typename VectorType::tag_type;
 
     const bool run_in_place_weights = (fused_weights == nullptr) || (fused_weights == conv_weights);
     const bool run_in_place_bias    = (fused_bias == nullptr) || (conv_bias != nullptr && fused_bias == conv_bias);
@@ -112,8 +124,6 @@ void fused_batch_normmalization(const ITensor *conv_weights, const ITensor *conv
     const auto conv_bias_in  = (conv_bias != nullptr ? reinterpret_cast<ScalarType *>(conv_bias->ptr_to_element(Coordinates(0, 0))) : nullptr);
     auto       conv_bias_out = (run_in_place_bias ? conv_bias_in : reinterpret_cast<ScalarType *>(fused_bias->ptr_to_element(Coordinates(0, 0))));
 
-    int slice = -1;
-
     const auto input_mean  = reinterpret_cast<const ScalarType *>(bn_mean->ptr_to_element(Coordinates(0, 0)));
     const auto input_var   = reinterpret_cast<const ScalarType *>(bn_var->ptr_to_element(Coordinates(0, 0)));
     const auto input_gamma = (bn_gamma != nullptr) ? reinterpret_cast<const ScalarType *>(bn_gamma->ptr_to_element(Coordinates(0, 0))) : nullptr;
@@ -133,45 +143,38 @@ void fused_batch_normmalization(const ITensor *conv_weights, const ITensor *conv
     auto conv_bias_in_scalar = ScalarType(0.0);
     execute_window_loop(win, [&](const Coordinates & id)
     {
-        if(slice != id[3])
+        var = input_var[id[3]];
+        if(input_gamma != nullptr)
         {
-            slice = id[3];
-            mean  = input_mean[slice];
-            var   = input_var[slice];
-            gamma = ScalarType(1.0);
-            beta  = ScalarType(0.0);
+            gamma = input_gamma[id[3]];
+        }
 
-            // Construct vectors
-            mean_vec = wrapper::vdup_n(mean, ExactTagType{});
-            var_vec  = wrapper::vdup_n(var, ExactTagType{});
-            if(input_gamma != nullptr)
-            {
-                gamma     = input_gamma[slice];
-                gamma_vec = wrapper::vdup_n(gamma, ExactTagType{});
-            }
+        if((id[0] == 0) && (id[1] == 0) && (id[2] == 0))
+        {
             if(input_beta != nullptr)
             {
-                beta     = input_beta[slice];
+                beta     = input_beta[id[3]];
                 beta_vec = wrapper::vdup_n(beta, ExactTagType{});
             }
+
+            // Construct vectors
+            mean     = input_mean[id[3]];
+            mean_vec = wrapper::vdup_n(mean, ExactTagType{});
+
             if(conv_bias_in != nullptr)
             {
-                conv_bias_in_scalar = conv_bias_in[slice];
-            }
-            else
-            {
-                conv_bias_in_scalar = ScalarType(0);
+                conv_bias_in_scalar = conv_bias_in[id[3]];
             }
-
-            conv_bias_in_scalar  = (conv_bias_in_scalar - mean) / sqrt(var + ScalarType(epsilon));
-            conv_bias_in_scalar  = (conv_bias_in_scalar * gamma) + beta;
-            conv_bias_out[slice] = conv_bias_in_scalar;
-            rvar_vec             = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec));
+            auto conv_bias_tmp_scalar = (conv_bias_in_scalar - mean) / std::sqrt(var + ScalarType(epsilon));
+            conv_bias_out[id[3]]      = (conv_bias_tmp_scalar * gamma) + beta;
         }
 
         int  x              = window_start_x;
         auto conv_w_in_ptr  = reinterpret_cast<const ScalarType *>(conv_w_in.ptr());
         auto conv_w_out_ptr = reinterpret_cast<ScalarType *>(conv_w_out.ptr());
+        var_vec             = wrapper::vdup_n(var, ExactTagType{});
+        gamma_vec           = wrapper::vdup_n(gamma, ExactTagType{});
+        rvar_vec            = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec));
 
         for(; x <= (window_end_x - window_step_x); x += window_step_x)
         {
@@ -186,28 +189,245 @@ void fused_batch_normmalization(const ITensor *conv_weights, const ITensor *conv
         // Compute left-over elements
         for(; x < window_end_x; ++x)
         {
-            *(conv_w_out_ptr + x) = *(conv_w_in_ptr + x) / sqrt(var + ScalarType(epsilon)) * gamma;
+            *(conv_w_out_ptr + x) = *(conv_w_in_ptr + x) / std::sqrt(var + ScalarType(epsilon)) * gamma;
         }
     },
     conv_w_in, conv_w_out);
 }
+
+template <typename VectorType>
+void fused_batch_normalization_dwc_nhwc(const ITensor *dwc_weights, const ITensor *dwc_bias, ITensor *fused_weights, ITensor *fused_bias,
+                                        const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window)
+{
+    using ScalarType   = typename VectorType::scalar_type;
+    const int size     = 16 / dwc_weights->info()->element_size();
+    using ExactTagType = typename VectorType::tag_type;
+
+    const bool run_in_place_weights = (fused_weights == nullptr) || (fused_weights == dwc_weights);
+    const bool run_in_place_bias    = (fused_bias == nullptr) || (dwc_bias != nullptr && fused_bias == dwc_bias);
+
+    // Set build options
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const int  window_step_x  = size;
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    Iterator dwc_w_in(dwc_weights, win);
+    Iterator dwc_w_out(run_in_place_weights ? dwc_weights : fused_weights, win);
+
+    const auto dwc_bias_in  = (dwc_bias != nullptr ? reinterpret_cast<ScalarType *>(dwc_bias->ptr_to_element(Coordinates(0, 0))) : nullptr);
+    auto       dwc_bias_out = (run_in_place_bias ? dwc_bias_in : reinterpret_cast<ScalarType *>(fused_bias->ptr_to_element(Coordinates(0, 0))));
+
+    const auto input_mean  = reinterpret_cast<const ScalarType *>(bn_mean->ptr_to_element(Coordinates(0, 0)));
+    const auto input_var   = reinterpret_cast<const ScalarType *>(bn_var->ptr_to_element(Coordinates(0, 0)));
+    const auto input_gamma = (bn_gamma != nullptr) ? reinterpret_cast<const ScalarType *>(bn_gamma->ptr_to_element(Coordinates(0, 0))) : nullptr;
+    const auto input_beta  = (bn_beta != nullptr) ? reinterpret_cast<const ScalarType *>(bn_beta->ptr_to_element(Coordinates(0, 0))) : nullptr;
+
+    auto       mean_vec     = wrapper::vdup_n(ScalarType(0), ExactTagType{});
+    auto       var_vec      = wrapper::vdup_n(ScalarType(0), ExactTagType{});
+    auto       gamma_vec    = wrapper::vdup_n(ScalarType(1), ExactTagType{});
+    auto       beta_vec     = wrapper::vdup_n(ScalarType(0), ExactTagType{});
+    auto       rvar_vec     = wrapper::vdup_n(ScalarType(0), ExactTagType{});
+    auto       dwc_bias_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{});
+    const auto epsilon_vec  = wrapper::vdup_n(ScalarType(epsilon), ExactTagType{});
+
+    auto gamma              = ScalarType(1.0);
+    auto beta               = ScalarType(0.0);
+    auto dwc_bias_in_scalar = ScalarType(0);
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        int x = window_start_x;
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
+        {
+            var_vec = wrapper::vloadq(input_var + x);
+            if(input_gamma != nullptr)
+            {
+                gamma_vec = wrapper::vloadq(input_gamma + x);
+            }
+
+            if((id[2] == 0) && (id[1] == 0))
+            {
+                mean_vec = wrapper::vloadq(input_mean + x);
+
+                // Construct vectors
+                if(input_beta != nullptr)
+                {
+                    beta_vec = wrapper::vloadq(input_beta + x);
+                }
+
+                if(dwc_bias_in != nullptr)
+                {
+                    dwc_bias_vec = wrapper::vloadq(dwc_bias_in + x);
+                }
+
+                auto dwc_bias_tmp_vec = wrapper::vmul(wrapper::vsub(dwc_bias_vec, mean_vec), wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec)));
+                dwc_bias_tmp_vec      = wrapper::vadd(wrapper::vmul(dwc_bias_tmp_vec, gamma_vec), beta_vec);
+                wrapper::vstore(dwc_bias_out + x, dwc_bias_tmp_vec);
+            }
+
+            auto dwc_w_in_ptr  = reinterpret_cast<const ScalarType *>(dwc_w_in.ptr());
+            auto dwc_w_out_ptr = reinterpret_cast<ScalarType *>(dwc_w_out.ptr());
+
+            auto wn  = wrapper::vloadq(dwc_w_in_ptr + x);
+            rvar_vec = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec));
+            wn       = wrapper::vmul(wn, rvar_vec);
+            wn       = wrapper::vmul(wn, gamma_vec);
+
+            // Store results
+            wrapper::vstore(dwc_w_out_ptr + x, wn);
+        }
+
+        // Compute left-over elements
+        for(; x < window_end_x; ++x)
+        {
+            auto var = input_var[x];
+            if(input_gamma != nullptr)
+            {
+                gamma = input_gamma[x];
+            }
+
+            if(id[2] == 0 && id[1] == 0)
+            {
+                auto mean = input_mean[x];
+                if(input_beta != nullptr)
+                {
+                    beta = input_beta[x];
+                }
+                if(dwc_bias_in != nullptr)
+                {
+                    dwc_bias_in_scalar = dwc_bias_in[x];
+                }
+
+                auto dwc_bias_tmp_scalar = (dwc_bias_in_scalar - mean) / std::sqrt(var + ScalarType(epsilon));
+                dwc_bias_out[x]          = (dwc_bias_tmp_scalar * gamma) + beta;
+            }
+
+            const auto dwc_w_in_ptr  = reinterpret_cast<const ScalarType *>(dwc_w_in.ptr());
+            auto       dwc_w_out_ptr = reinterpret_cast<ScalarType *>(dwc_w_out.ptr());
+
+            *(dwc_w_out_ptr + x) = *(dwc_w_in_ptr + x) / std::sqrt(var + ScalarType(epsilon)) * gamma;
+        }
+    },
+    dwc_w_in, dwc_w_out);
+}
+
+template <typename VectorType>
+void fused_batch_normalization_dwc_nchw(const ITensor *dwc_weights, const ITensor *dwc_bias, ITensor *fused_weights, ITensor *fused_bias,
+                                        const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window)
+{
+    using ScalarType   = typename VectorType::scalar_type;
+    const int size     = 16 / dwc_weights->info()->element_size();
+    using ExactTagType = typename VectorType::tag_type;
+
+    const bool run_in_place_weights = (fused_weights == nullptr) || (fused_weights == dwc_weights);
+    const bool run_in_place_bias    = (fused_bias == nullptr) || (dwc_bias != nullptr && fused_bias == dwc_bias);
+
+    // Set build options
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    const int  window_step_x  = size;
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    Iterator dwc_w_in(dwc_weights, win);
+    Iterator dwc_w_out(run_in_place_weights ? dwc_weights : fused_weights, win);
+
+    const auto dwc_bias_in  = (dwc_bias != nullptr ? reinterpret_cast<ScalarType *>(dwc_bias->ptr_to_element(Coordinates(0, 0))) : nullptr);
+    auto       dwc_bias_out = (run_in_place_bias ? dwc_bias_in : reinterpret_cast<ScalarType *>(fused_bias->ptr_to_element(Coordinates(0, 0))));
+
+    const auto input_mean  = reinterpret_cast<const ScalarType *>(bn_mean->ptr_to_element(Coordinates(0, 0)));
+    const auto input_var   = reinterpret_cast<const ScalarType *>(bn_var->ptr_to_element(Coordinates(0, 0)));
+    const auto input_gamma = (bn_gamma != nullptr) ? reinterpret_cast<const ScalarType *>(bn_gamma->ptr_to_element(Coordinates(0, 0))) : nullptr;
+    const auto input_beta  = (bn_beta != nullptr) ? reinterpret_cast<const ScalarType *>(bn_beta->ptr_to_element(Coordinates(0, 0))) : nullptr;
+
+    auto       mean_vec    = wrapper::vdup_n(ScalarType(0), ExactTagType{});
+    auto       var_vec     = wrapper::vdup_n(ScalarType(0), ExactTagType{});
+    auto       gamma_vec   = wrapper::vdup_n(ScalarType(1), ExactTagType{});
+    auto       beta_vec    = wrapper::vdup_n(ScalarType(0), ExactTagType{});
+    auto       rvar_vec    = wrapper::vdup_n(ScalarType(0), ExactTagType{});
+    const auto epsilon_vec = wrapper::vdup_n(ScalarType(epsilon), ExactTagType{});
+
+    auto mean               = ScalarType(0.0);
+    auto var                = ScalarType(0.0);
+    auto gamma              = ScalarType(1.0);
+    auto beta               = ScalarType(0.0);
+    auto dwc_bias_in_scalar = ScalarType(0.0);
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        var = input_var[id[2]];
+        if(input_gamma != nullptr)
+        {
+            gamma = input_gamma[id[2]];
+        }
+
+        if(id[1] == 0)
+        {
+            mean = input_mean[id[2]];
+
+            // Construct vectors
+            mean_vec = wrapper::vdup_n(mean, ExactTagType{});
+            if(input_beta != nullptr)
+            {
+                beta     = input_beta[id[2]];
+                beta_vec = wrapper::vdup_n(beta, ExactTagType{});
+            }
+
+            if(dwc_bias_in != nullptr)
+            {
+                dwc_bias_in_scalar = dwc_bias_in[id[2]];
+            }
+
+            auto dwc_bias_tmp_scalar = (dwc_bias_in_scalar - mean) / std::sqrt(var + ScalarType(epsilon));
+            dwc_bias_out[id[2]]      = (dwc_bias_tmp_scalar * gamma) + beta;
+        }
+
+        int  x             = window_start_x;
+        auto dwc_w_in_ptr  = reinterpret_cast<const ScalarType *>(dwc_w_in.ptr());
+        auto dwc_w_out_ptr = reinterpret_cast<ScalarType *>(dwc_w_out.ptr());
+        var_vec            = wrapper::vdup_n(var, ExactTagType{});
+        gamma_vec          = wrapper::vdup_n(gamma, ExactTagType{});
+        rvar_vec           = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec));
+
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
+        {
+            auto wn = wrapper::vloadq(dwc_w_in_ptr + x);
+            wn      = wrapper::vmul(wn, rvar_vec);
+            wn      = wrapper::vmul(wn, gamma_vec);
+
+            // Store results
+            wrapper::vstore(dwc_w_out_ptr + x, wn);
+        }
+
+        // Compute left-over elements
+        for(; x < window_end_x; ++x)
+        {
+            *(dwc_w_out_ptr + x) = *(dwc_w_in_ptr + x) / std::sqrt(var + ScalarType(epsilon)) * gamma;
+        }
+    },
+    dwc_w_in, dwc_w_out);
+}
+
 } // namespace
 
 NEFuseBatchNormalizationKernel::NEFuseBatchNormalizationKernel()
-    : _conv_weights(nullptr), _conv_bias(nullptr), _bn_mean(nullptr), _bn_var(nullptr), _bn_gamma(nullptr), _bn_beta(nullptr), _fused_weights(nullptr), _fused_bias(nullptr), _epsilon(),
+    : _input_weights(nullptr), _input_bias(nullptr), _bn_mean(nullptr), _bn_var(nullptr), _bn_gamma(nullptr), _bn_beta(nullptr), _fused_weights(nullptr), _fused_bias(nullptr), _epsilon(),
       _run_in_place_weights(false), _run_in_place_bias(false), _func(nullptr)
 {
 }
 
-void NEFuseBatchNormalizationKernel::configure(const ITensor *conv_weights, const ITensor *bn_mean, const ITensor *bn_var,
+void NEFuseBatchNormalizationKernel::configure(const ITensor *input_weights, const ITensor *bn_mean, const ITensor *bn_var,
                                                ITensor *fused_weights, ITensor *fused_bias,
-                                               const ITensor *conv_bias, const ITensor *bn_beta, const ITensor *bn_gamma,
-                                               float epsilon)
+                                               const ITensor *input_bias, const ITensor *bn_beta, const ITensor *bn_gamma,
+                                               float epsilon, FuseBatchNormalizationType fbn_type)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(conv_weights, bn_mean, bn_var);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input_weights, bn_mean, bn_var);
 
-    _conv_weights  = conv_weights;
-    _conv_bias     = conv_bias;
+    _input_weights = input_weights;
+    _input_bias    = input_bias;
     _bn_mean       = bn_mean;
     _bn_var        = bn_var;
     _bn_beta       = bn_beta;
@@ -216,15 +436,15 @@ void NEFuseBatchNormalizationKernel::configure(const ITensor *conv_weights, cons
     _fused_bias    = fused_bias;
     _epsilon       = epsilon;
 
-    _run_in_place_weights = (fused_weights == nullptr) || (fused_weights == conv_weights);
-    _run_in_place_bias    = (fused_bias == nullptr) || (conv_bias != nullptr && fused_bias == conv_bias);
+    _run_in_place_weights = (fused_weights == nullptr) || (fused_weights == input_weights);
+    _run_in_place_bias    = (fused_bias == nullptr) || (input_bias != nullptr && fused_bias == input_bias);
 
     // Auto initialize outputs
     if(_fused_weights != nullptr)
     {
         // Output tensor auto initialization if not yet initialized
-        auto_init_if_empty(*_fused_weights->info(), *_conv_weights->info()->clone());
-        fused_weights->info()->set_valid_region(conv_weights->info()->valid_region());
+        auto_init_if_empty(*_fused_weights->info(), *_input_weights->info()->clone());
+        fused_weights->info()->set_valid_region(input_weights->info()->valid_region());
     }
     if(_fused_bias != nullptr)
     {
@@ -234,42 +454,53 @@ void NEFuseBatchNormalizationKernel::configure(const ITensor *conv_weights, cons
     }
 
     // Validate arguments
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(conv_weights->info(), bn_mean->info(), bn_var->info(),
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input_weights->info(), bn_mean->info(), bn_var->info(),
                                                   (fused_weights != nullptr) ? fused_weights->info() : nullptr,
                                                   (fused_bias != nullptr) ? fused_bias->info() : nullptr,
-                                                  (conv_bias != nullptr) ? conv_bias->info() : nullptr,
+                                                  (input_bias != nullptr) ? input_bias->info() : nullptr,
                                                   (bn_beta != nullptr) ? bn_beta->info() : nullptr,
                                                   (bn_gamma != nullptr) ? bn_gamma->info() : nullptr,
-                                                  epsilon));
+                                                  epsilon, fbn_type));
 
     // Configure kernel window
-    Window win = calculate_max_window(*conv_weights->info());
+    Window win = calculate_max_window(*input_weights->info());
     INEKernel::configure(win);
 
-    // Configure function to run based on different data types
-    const DataType data_type = _conv_weights->info()->data_type();
-    switch(data_type)
+    // Configure function
+    static std::map<std::string, FuseBatchNormFunction *> map_function =
     {
-        case DataType::F32:
-            _func = &fused_batch_normmalization<float, 4>;
-            break;
+        { "fused_batch_normalization_conv_NHWC_F32", &fused_batch_normalization_conv<wrapper::traits::neon_vector<float, 4>> },
+        { "fused_batch_normalization_conv_NCHW_F32", &fused_batch_normalization_conv<wrapper::traits::neon_vector<float, 4>> },
+        { "fused_batch_normalization_dwc_NHWC_F32", &fused_batch_normalization_dwc_nhwc<wrapper::traits::neon_vector<float, 4>> },
+        { "fused_batch_normalization_dwc_NCHW_F32", &fused_batch_normalization_dwc_nchw<wrapper::traits::neon_vector<float, 4>> },
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        case DataType::F16:
-            _func = &fused_batch_normmalization<float16_t, 8>;
-            break;
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        default:
-            ARM_COMPUTE_ERROR("Not Supported");
-            break;
+        { "fused_batch_normalization_conv_NHWC_F16", &fused_batch_normalization_conv<wrapper::traits::neon_vector<float16_t, 8>> },
+        { "fused_batch_normalization_conv_NCHW_F16", &fused_batch_normalization_conv<wrapper::traits::neon_vector<float16_t, 8>> },
+        { "fused_batch_normalization_dwc_NHWC_F16", &fused_batch_normalization_dwc_nhwc<wrapper::traits::neon_vector<float16_t, 8>> },
+        { "fused_batch_normalization_dwc_NCHW_F16", &fused_batch_normalization_dwc_nchw<wrapper::traits::neon_vector<float16_t, 8>> },
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+    };
+
+    std::string function_to_call("fused_batch_normalization_");
+    function_to_call += fbn_type == FuseBatchNormalizationType::CONVOLUTION ? "conv_" : "dwc_";
+    function_to_call += string_from_data_layout(_input_weights->info()->data_layout());
+    function_to_call += "_";
+    function_to_call += string_from_data_type(_input_weights->info()->data_type());
+
+    auto it = map_function.find(function_to_call);
+
+    if(it != map_function.end())
+    {
+        _func = it->second;
     }
 }
 
-Status NEFuseBatchNormalizationKernel::validate(const ITensorInfo *conv_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
+Status NEFuseBatchNormalizationKernel::validate(const ITensorInfo *input_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
                                                 const ITensorInfo *fused_weights, const ITensorInfo *fused_bias,
-                                                const ITensorInfo *conv_bias, const ITensorInfo *bn_beta, const ITensorInfo *bn_gamma,
-                                                float epsilon)
+                                                const ITensorInfo *input_bias, const ITensorInfo *bn_beta, const ITensorInfo *bn_gamma,
+                                                float epsilon, FuseBatchNormalizationType fbn_type)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(conv_weights, bn_mean, bn_var, fused_weights, fused_bias, conv_bias, bn_beta, bn_gamma, epsilon));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input_weights, bn_mean, bn_var, fused_weights, fused_bias, input_bias, bn_beta, bn_gamma, epsilon, fbn_type));
     return Status{};
 }
 
@@ -278,6 +509,6 @@ void NEFuseBatchNormalizationKernel::run(const Window &window, const ThreadInfo
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
-    (*_func)(_conv_weights, _conv_bias, _fused_weights, _fused_bias, _bn_mean, _bn_var, _bn_beta, _bn_gamma, _epsilon, window);
+    (*_func)(_input_weights, _input_bias, _fused_weights, _fused_bias, _bn_mean, _bn_var, _bn_beta, _bn_gamma, _epsilon, window);
 }
 } // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEFuseBatchNormalization.cpp b/src/runtime/NEON/functions/NEFuseBatchNormalization.cpp
index dc48731c4d..68dc159f75 100644
--- a/src/runtime/NEON/functions/NEFuseBatchNormalization.cpp
+++ b/src/runtime/NEON/functions/NEFuseBatchNormalization.cpp
@@ -36,20 +36,20 @@ NEFuseBatchNormalization::NEFuseBatchNormalization()
 {
 }
 
-void NEFuseBatchNormalization::configure(const ITensor *conv_weights, const ITensor *bn_mean, const ITensor *bn_var,
+void NEFuseBatchNormalization::configure(const ITensor *input_weights, const ITensor *bn_mean, const ITensor *bn_var,
                                          ITensor *fused_weights, ITensor *fused_bias,
-                                         const ITensor *conv_bias, const ITensor *bn_beta, const ITensor *bn_gamma,
-                                         float epsilon)
+                                         const ITensor *input_bias, const ITensor *bn_beta, const ITensor *bn_gamma,
+                                         float epsilon, FuseBatchNormalizationType fbn_type)
 {
-    _fuse_bn_kernel.configure(conv_weights, bn_mean, bn_var, fused_weights, fused_bias, conv_bias, bn_beta, bn_gamma, epsilon);
+    _fuse_bn_kernel.configure(input_weights, bn_mean, bn_var, fused_weights, fused_bias, input_bias, bn_beta, bn_gamma, epsilon, fbn_type);
 }
 
-Status NEFuseBatchNormalization::validate(const ITensorInfo *conv_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
+Status NEFuseBatchNormalization::validate(const ITensorInfo *input_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
                                           const ITensorInfo *fused_weights, const ITensorInfo *fused_bias,
-                                          const ITensorInfo *conv_bias, const ITensorInfo *bn_beta, const ITensorInfo *bn_gamma,
-                                          float epsilon)
+                                          const ITensorInfo *input_bias, const ITensorInfo *bn_beta, const ITensorInfo *bn_gamma,
+                                          float epsilon, FuseBatchNormalizationType fbn_type)
 {
-    return NEFuseBatchNormalizationKernel::validate(conv_weights, bn_mean, bn_var, fused_weights, fused_bias, conv_bias, bn_beta, bn_gamma, epsilon);
+    return NEFuseBatchNormalizationKernel::validate(input_weights, bn_mean, bn_var, fused_weights, fused_bias, input_bias, bn_beta, bn_gamma, epsilon, fbn_type);
 }
 
 void NEFuseBatchNormalization::run()
diff --git a/tests/validation/NEON/FuseBatchNormalization.cpp b/tests/validation/NEON/FuseBatchNormalization.cpp
new file mode 100644
index 0000000000..1a8f928c6e
--- /dev/null
+++ b/tests/validation/NEON/FuseBatchNormalization.cpp
@@ -0,0 +1,220 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/runtime/NEON/functions/NEFuseBatchNormalization.h"
+#include "tests/Globals.h"
+#include "tests/NEON/Accessor.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/FuseBatchNormalizationFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+AbsoluteTolerance<float> absolute_tolerance_f32(0.001f);
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+AbsoluteTolerance<float> absolute_tolerance_f16(0.2f);
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+} // namespace
+
+template <typename T>
+using NEFuseBatchNormalizationConvFixture = FuseBatchNormalizationFixture<Tensor, Accessor, NEFuseBatchNormalization, 4, T>;
+template <typename T>
+using NEFuseBatchNormalizationDWCFixture = FuseBatchNormalizationFixture<Tensor, Accessor, NEFuseBatchNormalization, 3, T>;
+
+// *INDENT-OFF*
+// clang-format off
+
+/** Shapes to test - Precommit */
+const auto shape_conv_values_precommit = concat(datasets::Small4DShapes(), datasets::Small3DShapes());
+
+/** Shapes to test - Nightly */
+const auto shape_conv_values_nightly = concat(datasets::Large4DShapes(), datasets::Large3DShapes());
+
+/** Data layout to test */
+const auto data_layout_values = framework::dataset::make("DataLayout", { DataLayout::NHWC, DataLayout::NCHW });
+
+/** In-place flags to test */
+const auto in_place_values = framework::dataset::make("InPlace", { true, false });
+
+/** With bias flags to test */
+const auto with_bias_values = framework::dataset::make("WithBias", { true, false });
+
+/** With gamma flags to test */
+const auto with_gamma_values = framework::dataset::make("WithGamma", { true, false });
+
+/** With beta flags to test */
+const auto with_beta_values = framework::dataset::make("WithBeta", { true, false });
+
+TEST_SUITE(NEON)
+TEST_SUITE(FuseBatchNormalization)
+TEST_SUITE(Convolution)
+TEST_SUITE(Float)
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEFuseBatchNormalizationConvFixture<float>, framework::DatasetMode::PRECOMMIT,
+                                        combine(combine(combine(combine(combine(combine(
+                                                        shape_conv_values_precommit,
+                                                        framework::dataset::make("DataType", { DataType::F32 })),
+                                                        data_layout_values),
+                                                        in_place_values),
+                                                        with_bias_values),
+                                                        with_gamma_values),
+                                                        with_beta_values))
+{
+    // Validate outputs
+    validate(Accessor(_target_w), _reference_w, absolute_tolerance_f32);
+    validate(Accessor(_target_b), _reference_b, absolute_tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, NEFuseBatchNormalizationConvFixture<float>, framework::DatasetMode::NIGHTLY,
+                                        combine(combine(combine(combine(combine(combine(
+                                                        shape_conv_values_nightly,
+                                                        framework::dataset::make("DataType", { DataType::F32 })),
+                                                        data_layout_values),
+                                                        in_place_values),
+                                                        with_bias_values),
+                                                        with_gamma_values),
+                                                        with_beta_values))
+{
+    // Validate outputs
+    validate(Accessor(_target_w), _reference_w, absolute_tolerance_f32);
+    validate(Accessor(_target_b), _reference_b, absolute_tolerance_f32);
+}
+TEST_SUITE_END() // FP32
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEFuseBatchNormalizationConvFixture<half>, framework::DatasetMode::PRECOMMIT,
+                                        combine(combine(combine(combine(combine(combine(
+                                                        shape_conv_values_precommit,
+                                                        framework::dataset::make("DataType", { DataType::F16 })),
+                                                        data_layout_values),
+                                                        in_place_values),
+                                                        with_bias_values),
+                                                        with_gamma_values),
+                                                        with_beta_values))
+{
+    // Validate outputs
+    validate(Accessor(_target_w), _reference_w, absolute_tolerance_f16);
+    validate(Accessor(_target_b), _reference_b, absolute_tolerance_f16);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, NEFuseBatchNormalizationConvFixture<half>, framework::DatasetMode::NIGHTLY,
+                                        combine(combine(combine(combine(combine(combine(
+                                                        shape_conv_values_nightly,
+                                                        framework::dataset::make("DataType", { DataType::F16 })),
+                                                        data_layout_values),
+                                                        in_place_values),
+                                                        with_bias_values),
+                                                        with_gamma_values),
+                                                        with_beta_values))
+{
+    // Validate outputs
+    validate(Accessor(_target_w), _reference_w, absolute_tolerance_f16);
+    validate(Accessor(_target_b), _reference_b, absolute_tolerance_f16);
+}
+TEST_SUITE_END() // FP16
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+TEST_SUITE_END() // Float
+TEST_SUITE_END() // Convolution
+TEST_SUITE(DepthwiseConvolution)
+TEST_SUITE(Float)
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEFuseBatchNormalizationDWCFixture<float>, framework::DatasetMode::PRECOMMIT,
+                                        combine(combine(combine(combine(combine(combine(
+                                                        datasets::Small3DShapes(),
+                                                        framework::dataset::make("DataType", { DataType::F32 })),
+                                                        data_layout_values),
+                                                        in_place_values),
+                                                        with_bias_values),
+                                                        with_gamma_values),
+                                                        with_beta_values))
+{
+    // Validate outputs
+    validate(Accessor(_target_w), _reference_w, absolute_tolerance_f32);
+    validate(Accessor(_target_b), _reference_b, absolute_tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, NEFuseBatchNormalizationDWCFixture<float>, framework::DatasetMode::NIGHTLY,
+                                        combine(combine(combine(combine(combine(combine(
+                                                        datasets::Large3DShapes(),
+                                                        framework::dataset::make("DataType", { DataType::F32 })),
+                                                        data_layout_values),
+                                                        in_place_values),
+                                                        with_bias_values),
+                                                        with_gamma_values),
+                                                        with_beta_values))
+{
+    // Validate outputs
+    validate(Accessor(_target_w), _reference_w, absolute_tolerance_f32);
+    validate(Accessor(_target_b), _reference_b, absolute_tolerance_f32);
+}
+
+TEST_SUITE_END() // FP32
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEFuseBatchNormalizationDWCFixture<half>, framework::DatasetMode::PRECOMMIT,
+                                        combine(combine(combine(combine(combine(combine(
+                                                        datasets::Small3DShapes(),
+                                                        framework::dataset::make("DataType", { DataType::F16 })),
+                                                        data_layout_values),
+                                                        in_place_values),
+                                                        with_bias_values),
+                                                        with_gamma_values),
+                                                        with_beta_values))
+{
+    // Validate outputs
+    validate(Accessor(_target_w), _reference_w, absolute_tolerance_f16);
+    validate(Accessor(_target_b), _reference_b, absolute_tolerance_f16);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, NEFuseBatchNormalizationDWCFixture<half>, framework::DatasetMode::NIGHTLY,
+                                        combine(combine(combine(combine(combine(combine(
+                                                        datasets::Large3DShapes(),
+                                                        framework::dataset::make("DataType", { DataType::F16 })),
+                                                        data_layout_values),
+                                                        in_place_values),
+                                                        with_bias_values),
+                                                        with_gamma_values),
+                                                        with_beta_values))
+{
+    // Validate outputs
+    validate(Accessor(_target_w), _reference_w, absolute_tolerance_f16);
+    validate(Accessor(_target_b), _reference_b, absolute_tolerance_f16);
+}
+
+TEST_SUITE_END() // FP16
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+TEST_SUITE_END() // Float
+TEST_SUITE_END() // DepthwiseConvolution
+TEST_SUITE_END() // FuseBatchNormalization
+TEST_SUITE_END() // NEON
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
\ No newline at end of file