COMPMID-3871: Create BatchNormalization SVE/SVE2

1. Decouple data type for NHWC
2. Add NHWC SVE support for BachNormalization

Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: I0383b969b555b429d9acebb4efa17ecba9429ea7
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4755
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>

8 files changed, 678 insertions, 145 deletions

diff --git a/Android.bp b/Android.bp
index 9f7f447fe3..1032950f3e 100644
--- a/Android.bp
+++ b/Android.bp
@@ -368,6 +368,10 @@ cc_library_static {
         "src/core/NEON/kernels/arm_gemm/quantized.cpp",
         "src/core/NEON/kernels/arm_gemm/rowsum_indirect_s8.cpp",
         "src/core/NEON/kernels/arm_gemm/rowsum_indirect_u8.cpp",
+        "src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp",
+        "src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp",
+        "src/core/NEON/kernels/batchnormalization/impl/SVE/fp16.cpp",
+        "src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp",
         "src/core/NEON/kernels/convolution/common/padding.cpp",
         "src/core/NEON/kernels/convolution/common/qasymm8.cpp",
         "src/core/NEON/kernels/convolution/common/qsymm8.cpp",
diff --git a/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp b/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp
index afb08e5d1c..b4cac74dc4 100644
--- a/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp
@@ -37,19 +37,77 @@
 #include "src/core/NEON/kernels/detail/NEActivationFunctionDetail.h"
 #include "src/core/NEON/wrapper/wrapper.h"
 
+#include "src/core/NEON/kernels/batchnormalization/impl/list.h"
+#include "src/core/common/Registrars.h"
+
 #include <map>
 
 namespace arm_compute
 {
 namespace
 {
+struct BatchNormalizationSelectorData
+{
+    DataType dt;
+};
+using BatchNormalizationSelectorPtr = std::add_pointer<bool(const BatchNormalizationSelectorData &data)>::type;
+using BatchNormalizationKernelPtr   = std::add_pointer<void(ITensor *, ITensor *, const ITensor *, const ITensor *, const ITensor *, const ITensor *,
+                                                            float, ActivationLayerInfo &, const Window &)>::type;
+
+struct BatchNormalizationKernel
+{
+    const char                         *name;
+    const BatchNormalizationSelectorPtr is_selected;
+    BatchNormalizationKernelPtr         ukernel;
+};
+
+static const BatchNormalizationKernel available_kernels[] =
+{
+#if defined(__ARM_FEATURE_SVE)
+    {
+        "fp16_sve_batch_normalization",
+        [](const BatchNormalizationSelectorData & data) { return data.dt == DataType::F16; },
+        REGISTER_FP16_SVE(arm_compute::cpu::fp16_sve_batch_normalization)
+    },
+    {
+        "f32_sve_batch_normalization",
+        [](const BatchNormalizationSelectorData & data) { return data.dt == DataType::F32; },
+        REGISTER_FP32_SVE(arm_compute::cpu::fp32_sve_batch_normalization)
+    },
+#else  /* !defined(__ARM_FEATURE_SVE) */
+    {
+        "fp16_neon_batch_normalization",
+        [](const BatchNormalizationSelectorData & data) { return data.dt == DataType::F16; },
+        REGISTER_FP16_NEON(arm_compute::cpu::fp16_neon_batch_normalization)
+    },
+    {
+        "f32_neon_batch_normalization",
+        [](const BatchNormalizationSelectorData & data) { return data.dt == DataType::F32; },
+        REGISTER_FP32_NEON(arm_compute::cpu::fp32_neon_batch_normalization)
+    },
+#endif /* !defined(__ARM_FEATURE_SVE) */
+};
+
+const BatchNormalizationKernel *get_implementation(const BatchNormalizationSelectorData &data)
+{
+    for(const auto &uk : available_kernels)
+    {
+        if(uk.is_selected(data))
+        {
+            return &uk;
+        }
+    }
+    return nullptr;
+}
+
 Status
 validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *mean, const ITensorInfo *var,
                    const ITensorInfo *beta, const ITensorInfo *gamma, float epsilon, ActivationLayerInfo act_info)
 {
     ARM_COMPUTE_UNUSED(epsilon);
-    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+
+    const auto *uk = get_implementation(BatchNormalizationSelectorData{ input->data_type() });
+    ARM_COMPUTE_RETURN_ERROR_ON(uk == nullptr || uk->ukernel == nullptr);
 
     if(act_info.enabled())
     {
@@ -83,27 +141,6 @@ validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const IT
 
     return Status{};
 }
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, ITensorInfo *mean, ITensorInfo *var, ITensorInfo *gamma, ITensorInfo *beta)
-{
-    ARM_COMPUTE_UNUSED(mean, var, gamma, beta);
-
-    // Configure kernel window
-    Window win = calculate_max_window(*input, Steps());
-
-    if(output != nullptr)
-    {
-        // Output auto initialization if not yet initialized
-        auto_init_if_empty(*output, *input->clone());
-
-        // NEBatchNormalizationLayerKernel doesn't need padding so update_window_and_padding() can be skipped
-        Coordinates coord;
-        coord.set_num_dimensions(output->num_dimensions());
-        output->set_valid_region(ValidRegion(coord, output->tensor_shape()));
-    }
-
-    return std::make_pair(Status{}, win);
-}
 } //namespace
 
 template <typename T, bool fused_activation, typename F>
@@ -212,102 +249,17 @@ void NEBatchNormalizationLayerKernel::batch_normalization_nchw(const Window &win
     input, output);
 }
 
-template <typename T, bool fused_activation, typename F>
-void NEBatchNormalizationLayerKernel::batch_normalization_nhwc(const Window &window)
-{
-    /** NEON vector tag type. */
-    using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>;
-
-    const int  window_step_x  = 16 / sizeof(T);
-    const auto window_start_x = static_cast<int>(window.x().start());
-    const auto window_end_x   = static_cast<int>(window.x().end());
-
-    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
-    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-    Iterator input(_input, win_collapsed);
-    Iterator output(_output, win_collapsed);
-
-    F activation_functor(_act_info);
-
-    const auto input_mean  = reinterpret_cast<const T *>(_mean->ptr_to_element(Coordinates(0, 0)));
-    const auto input_var   = reinterpret_cast<const T *>(_var->ptr_to_element(Coordinates(0, 0)));
-    const auto input_gamma = (_gamma != nullptr) ? reinterpret_cast<const T *>(_gamma->ptr_to_element(Coordinates(0, 0))) : nullptr;
-    const auto input_beta  = (_beta != nullptr) ? reinterpret_cast<const T *>(_beta->ptr_to_element(Coordinates(0, 0))) : nullptr;
-
-    const auto epsilon_vec = wrapper::vdup_n(static_cast<T>(_epsilon), ExactTagType{});
-    execute_window_loop(win_collapsed, [&](const Coordinates &)
-    {
-        const auto input_ptr  = reinterpret_cast<const T *>(input.ptr());
-        const auto output_ptr = reinterpret_cast<T *>(output.ptr());
-
-        // Perform core calculations using vector operations
-        int x = window_start_x;
-        for(; x <= (window_end_x - window_step_x); x += window_step_x)
-        {
-            // Conctruct vectors
-            const auto mean_vec  = wrapper::vloadq(input_mean + x);
-            const auto var_vec   = wrapper::vloadq(input_var + x);
-            const auto gamma_vec = (input_gamma != nullptr) ? wrapper::vloadq(input_gamma + x) : wrapper::vdup_n(static_cast<T>(1.f), ExactTagType{});
-            const auto beta_vec  = (input_beta != nullptr) ? wrapper::vloadq(input_beta + x) : wrapper::vdup_n(static_cast<T>(0.f), ExactTagType{});
-
-            // Calculate denominator
-            const auto denominator = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec));
-
-            // Calculate x bar
-            const auto numerator = wrapper::vsub(wrapper::vloadq(input_ptr + x), mean_vec);
-            const auto x_bar     = wrapper::vmul(numerator, denominator);
-            auto       res       = wrapper::vmla(beta_vec, x_bar, gamma_vec);
-
-            // Perform fused activation
-            if(fused_activation)
-            {
-                activation_functor(res);
-            }
-
-            // Store results
-            wrapper::vstore(output_ptr + x, res);
-        }
-
-        // Compute left-over elements
-        for(; x < window_end_x; ++x)
-        {
-            // Conctruct vectors
-            const T gamma = (input_gamma != nullptr) ? input_gamma[x] : 1.f;
-            const T beta  = (input_beta != nullptr) ? input_beta[x] : 0.f;
-
-            const T denominator = sqrt(input_var[x] + _epsilon);
-            const T numerator   = input_ptr[x] - input_mean[x];
-            const T x_bar       = numerator / denominator;
-            T       res         = beta + x_bar * gamma;
-
-            // Perform fused activation
-            if(fused_activation)
-            {
-                activation_functor(res);
-            }
-
-            // Store results
-            *reinterpret_cast<T *>(output_ptr + x) = res;
-        }
-    },
-    input, output);
-}
-
 void NEBatchNormalizationLayerKernel::configure_non_fused()
 {
-    const bool is_nhwc = _input->info()->data_layout() == DataLayout::NHWC;
     switch(_input->info()->data_type())
     {
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
         case DataType::F16:
-            _func = (is_nhwc) ? &NEBatchNormalizationLayerKernel::batch_normalization_nhwc<float16_t, false, detail::dummy<float16_t, 8>> :
-                    &NEBatchNormalizationLayerKernel::batch_normalization_nchw<float16_t, false, detail::dummy<float16_t, 8>>;
+            _func = &NEBatchNormalizationLayerKernel::batch_normalization_nchw<float16_t, false, detail::dummy<float16_t, 8>>;
             break;
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
         case DataType::F32:
-            _func = (is_nhwc) ? &NEBatchNormalizationLayerKernel::batch_normalization_nhwc<float, false, detail::dummy<float, 4>> :
-                    &NEBatchNormalizationLayerKernel::batch_normalization_nchw<float, false, detail::dummy<float, 4>>;
+            _func = &NEBatchNormalizationLayerKernel::batch_normalization_nchw<float, false, detail::dummy<float, 4>>;
             break;
         default:
             ARM_COMPUTE_ERROR("Element size not supported");
@@ -324,13 +276,6 @@ void NEBatchNormalizationLayerKernel::configure_fused()
         { ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, &NEBatchNormalizationLayerKernel::batch_normalization_nchw<float, true, detail::brelu<float, 4>> },
         { ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, &NEBatchNormalizationLayerKernel::batch_normalization_nchw<float, true, detail::lubrelu<float, 4>> }
     };
-    // NHWC Fused Batched Normalization with activation functions : FP32
-    static std::map<ActivationLayerInfo::ActivationFunction, BatchNormFunctionPtr> bn_fused_map_f32_nhwc =
-    {
-        { ActivationLayerInfo::ActivationFunction::RELU, &NEBatchNormalizationLayerKernel::batch_normalization_nhwc<float, true, detail::relu<float, 4>> },
-        { ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, &NEBatchNormalizationLayerKernel::batch_normalization_nhwc<float, true, detail::brelu<float, 4>> },
-        { ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, &NEBatchNormalizationLayerKernel::batch_normalization_nhwc<float, true, detail::lubrelu<float, 4>> }
-    };
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
     // NCHW Fused Batched Normalization with activation functions : FP16
     static std::map<ActivationLayerInfo::ActivationFunction, BatchNormFunctionPtr> bn_fused_map_f16_nchw =
@@ -339,24 +284,17 @@ void NEBatchNormalizationLayerKernel::configure_fused()
         { ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, &NEBatchNormalizationLayerKernel::batch_normalization_nchw<float16_t, true, detail::brelu<float16_t, 8>> },
         { ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, &NEBatchNormalizationLayerKernel::batch_normalization_nchw<float16_t, true, detail::lubrelu<float16_t, 8>> }
     };
-    // NHWC Fused Batched Normalization with activation functions : FP16
-    static std::map<ActivationLayerInfo::ActivationFunction, BatchNormFunctionPtr> bn_fused_map_f16_nhwc =
-    {
-        { ActivationLayerInfo::ActivationFunction::RELU, &NEBatchNormalizationLayerKernel::batch_normalization_nhwc<float16_t, true, detail::relu<float16_t, 8>> },
-        { ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, &NEBatchNormalizationLayerKernel::batch_normalization_nhwc<float16_t, true, detail::brelu<float16_t, 8>> },
-        { ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, &NEBatchNormalizationLayerKernel::batch_normalization_nhwc<float16_t, true, detail::lubrelu<float16_t, 8>> }
-    };
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 
     switch(_input->info()->data_type())
     {
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
         case DataType::F16:
-            _func = (_input->info()->data_layout() == DataLayout::NHWC) ? bn_fused_map_f16_nhwc[_act_info.activation()] : bn_fused_map_f16_nchw[_act_info.activation()];
+            _func = bn_fused_map_f16_nchw[_act_info.activation()];
             break;
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
         case DataType::F32:
-            _func = (_input->info()->data_layout() == DataLayout::NHWC) ? bn_fused_map_f32_nhwc[_act_info.activation()] : bn_fused_map_f32_nchw[_act_info.activation()];
+            _func = bn_fused_map_f32_nchw[_act_info.activation()];
             break;
         default:
             ARM_COMPUTE_ERROR("Element size not supported");
@@ -398,20 +336,32 @@ void NEBatchNormalizationLayerKernel::configure(ITensor *input, ITensor *output,
     }
 
     // Configure activation function to run
-    if(_act_info.enabled())
+    const bool is_nchw = _input->info()->data_layout() == DataLayout::NCHW;
+    if(is_nchw)
     {
-        configure_fused();
-    }
-    else
-    {
-        configure_non_fused();
+        if(_act_info.enabled())
+        {
+            configure_fused();
+        }
+        else
+        {
+            configure_non_fused();
+        }
     }
 
     // Configure kernel window
-    auto win_config = validate_and_configure_window(input->info(), (run_in_place) ? nullptr : output->info(), mean->info(), var->info(), (gamma != nullptr) ? gamma->info() : nullptr,
-                                                    (beta != nullptr) ? beta->info() : nullptr);
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-    INEKernel::configure(win_config.second);
+    Window win = calculate_max_window(*input->info(), Steps());
+    INEKernel::configure(win);
+
+    if(output != nullptr)
+    {
+        // Output auto initialization if not yet initialized
+        auto_init_if_empty(*output->info(), *input->info()->clone());
+
+        Coordinates coord;
+        coord.set_num_dimensions(output->info()->num_dimensions());
+        output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
+    }
 }
 
 Status NEBatchNormalizationLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output,
@@ -420,9 +370,6 @@ Status NEBatchNormalizationLayerKernel::validate(const ITensorInfo *input, const
                                                  float epsilon, ActivationLayerInfo act_info)
 {
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, mean, var, beta, gamma, epsilon, act_info));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output ? output->clone().get() : nullptr, mean->clone().get(), var->clone().get(),
-                                                              (gamma != nullptr) ? gamma->clone().get() : nullptr, (beta != nullptr) ? beta->clone().get() : nullptr)
-                                .first);
 
     return Status{};
 }
@@ -432,8 +379,17 @@ void NEBatchNormalizationLayerKernel::run(const Window &window, const ThreadInfo
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
-    ARM_COMPUTE_ERROR_ON(_func == nullptr);
+    ARM_COMPUTE_ERROR_ON(_func == nullptr && _input->info()->data_layout() == DataLayout::NCHW);
 
-    (this->*_func)(window);
+    const bool is_nchw = _input->info()->data_layout() == DataLayout::NCHW;
+    if(is_nchw)
+    {
+        (this->*_func)(window);
+    }
+    else
+    {
+        const auto *uk = get_implementation(BatchNormalizationSelectorData{ _input->info()->data_type() });
+        uk->ukernel(_input, _output, _mean, _var, _beta, _gamma, _epsilon, _act_info, window);
+    }
 }
 } // namespace arm_compute
diff --git a/src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp
new file mode 100644
index 0000000000..dfadef34f7
--- /dev/null
+++ b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp
@@ -0,0 +1,146 @@
+/*
+ * Copyright (c) 2020 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/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/kernels/detail/NEActivationFunctionDetail.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/StdTypes.h"
+#include "src/core/common/Validate.h"
+
+#include <arm_neon.h>
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
+namespace arm_compute
+{
+namespace
+{
+using BatchNomalizationPtr = void (*)(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
+                                      float epsilon, ActivationLayerInfo &act_info, const Window &window);
+
+template <typename T>
+void batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
+                         float epsilon, ActivationLayerInfo &act_info, const Window &window)
+{
+    /** NEON vector tag type. */
+    using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<float16_t, wrapper::traits::BitWidth::W128>;
+
+    const int  window_step_x  = 8;
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(src, win_collapsed);
+    Iterator output(dst, win_collapsed);
+
+    const auto input_mean  = reinterpret_cast<const float16_t *>(mean->ptr_to_element(Coordinates(0, 0)));
+    const auto input_var   = reinterpret_cast<const float16_t *>(var->ptr_to_element(Coordinates(0, 0)));
+    const auto input_gamma = (gamma != nullptr) ? reinterpret_cast<const float16_t *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr;
+    const auto input_beta  = (beta != nullptr) ? reinterpret_cast<const float16_t *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr;
+
+    T activation_functor(act_info);
+
+    const auto epsilon_vec = wrapper::vdup_n(static_cast<float16_t>(epsilon), ExactTagType{});
+    execute_window_loop(win_collapsed, [&](const Coordinates &)
+    {
+        const auto input_ptr  = reinterpret_cast<const float16_t *>(input.ptr());
+        const auto output_ptr = reinterpret_cast<float16_t *>(output.ptr());
+
+        // Perform core calculations using vector operations
+        int x = window_start_x;
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
+        {
+            // Conctruct vectors
+            const auto mean_vec  = wrapper::vloadq(input_mean + x);
+            const auto var_vec   = wrapper::vloadq(input_var + x);
+            const auto gamma_vec = (input_gamma != nullptr) ? wrapper::vloadq(input_gamma + x) : wrapper::vdup_n(static_cast<float16_t>(1.f), ExactTagType{});
+            const auto beta_vec  = (input_beta != nullptr) ? wrapper::vloadq(input_beta + x) : wrapper::vdup_n(static_cast<float16_t>(0.f), ExactTagType{});
+
+            // Calculate denominator
+            const auto denominator = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec));
+
+            // Calculate x bar
+            const auto numerator = wrapper::vsub(wrapper::vloadq(input_ptr + x), mean_vec);
+            const auto x_bar     = wrapper::vmul(numerator, denominator);
+            auto       res       = wrapper::vmla(beta_vec, x_bar, gamma_vec);
+
+            // Perform fused activation
+            if(act_info.enabled())
+            {
+                activation_functor(res);
+            }
+
+            // Store results
+            wrapper::vstore(output_ptr + x, res);
+        }
+
+        // Compute left-over elements
+        for(; x < window_end_x; ++x)
+        {
+            // Conctruct vectors
+            const float16_t gamma = (input_gamma != nullptr) ? input_gamma[x] : 1.f;
+            const float16_t beta  = (input_beta != nullptr) ? input_beta[x] : 0.f;
+
+            const float16_t denominator = sqrt(input_var[x] + epsilon);
+            const float16_t numerator   = input_ptr[x] - input_mean[x];
+            const float16_t x_bar       = numerator / denominator;
+            float16_t       res         = beta + x_bar * gamma;
+
+            // Perform fused activation
+            if(act_info.enabled())
+            {
+                activation_functor(res);
+            }
+
+            // Store results
+            *reinterpret_cast<float16_t *>(output_ptr + x) = res;
+        }
+    },
+    input, output);
+}
+
+// Fused Batched Normalization with activation functions
+static std::map<ActivationLayerInfo::ActivationFunction, BatchNomalizationPtr> fused_map =
+{
+    { ActivationLayerInfo::ActivationFunction::RELU, &batch_normalization<detail::relu<float16_t, 8>> },
+    { ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, &batch_normalization<detail::brelu<float16_t, 8>> },
+    { ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, &batch_normalization<detail::lubrelu<float16_t, 8>> }
+};
+}
+namespace cpu
+{
+void fp16_neon_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
+                                   float epsilon, ActivationLayerInfo &act_info, const Window &window)
+{
+    fused_map[act_info.activation()](src, dst, mean, var, beta, gamma, epsilon, act_info, window);
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
\ No newline at end of file
diff --git a/src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp
new file mode 100644
index 0000000000..a24f7f624a
--- /dev/null
+++ b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp
@@ -0,0 +1,143 @@
+/*
+ * Copyright (c) 2020 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/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/kernels/detail/NEActivationFunctionDetail.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/StdTypes.h"
+#include "src/core/common/Validate.h"
+
+#include <arm_neon.h>
+#include <cmath>
+#include <cstddef>
+
+namespace arm_compute
+{
+namespace
+{
+using BatchNomalizationPtr = void (*)(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
+                                      float epsilon, ActivationLayerInfo &act_info, const Window &window);
+
+template <typename T>
+void batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
+                         float epsilon, ActivationLayerInfo &act_info, const Window &window)
+{
+    /** NEON vector tag type. */
+    using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<float, wrapper::traits::BitWidth::W128>;
+
+    const int  window_step_x  = 4;
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(src, win_collapsed);
+    Iterator output(dst, win_collapsed);
+
+    const auto input_mean  = reinterpret_cast<const float *>(mean->ptr_to_element(Coordinates(0, 0)));
+    const auto input_var   = reinterpret_cast<const float *>(var->ptr_to_element(Coordinates(0, 0)));
+    const auto input_gamma = (gamma != nullptr) ? reinterpret_cast<const float *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr;
+    const auto input_beta  = (beta != nullptr) ? reinterpret_cast<const float *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr;
+
+    T activation_functor(act_info);
+
+    const auto epsilon_vec = wrapper::vdup_n(static_cast<float>(epsilon), ExactTagType{});
+    execute_window_loop(win_collapsed, [&](const Coordinates &)
+    {
+        const auto input_ptr  = reinterpret_cast<const float *>(input.ptr());
+        const auto output_ptr = reinterpret_cast<float *>(output.ptr());
+
+        // Perform core calculations using vector operations
+        int x = window_start_x;
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
+        {
+            // Conctruct vectors
+            const auto mean_vec  = wrapper::vloadq(input_mean + x);
+            const auto var_vec   = wrapper::vloadq(input_var + x);
+            const auto gamma_vec = (input_gamma != nullptr) ? wrapper::vloadq(input_gamma + x) : wrapper::vdup_n(static_cast<float>(1.f), ExactTagType{});
+            const auto beta_vec  = (input_beta != nullptr) ? wrapper::vloadq(input_beta + x) : wrapper::vdup_n(static_cast<float>(0.f), ExactTagType{});
+
+            // Calculate denominator
+            const auto denominator = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec));
+
+            // Calculate x bar
+            const auto numerator = wrapper::vsub(wrapper::vloadq(input_ptr + x), mean_vec);
+            const auto x_bar     = wrapper::vmul(numerator, denominator);
+            auto       res       = wrapper::vmla(beta_vec, x_bar, gamma_vec);
+
+            // Perform fused activation
+            if(act_info.enabled())
+            {
+                activation_functor(res);
+            }
+
+            // Store results
+            wrapper::vstore(output_ptr + x, res);
+        }
+
+        // Compute left-over elements
+        for(; x < window_end_x; ++x)
+        {
+            // Conctruct vectors
+            const float gamma = (input_gamma != nullptr) ? input_gamma[x] : 1.f;
+            const float beta  = (input_beta != nullptr) ? input_beta[x] : 0.f;
+
+            const float denominator = sqrt(input_var[x] + epsilon);
+            const float numerator   = input_ptr[x] - input_mean[x];
+            const float x_bar       = numerator / denominator;
+            float       res         = beta + x_bar * gamma;
+
+            // Perform fused activation
+            if(act_info.enabled())
+            {
+                activation_functor(res);
+            }
+
+            // Store results
+            *reinterpret_cast<float *>(output_ptr + x) = res;
+        }
+    },
+    input, output);
+}
+
+// Fused Batched Normalization with activation functions
+static std::map<ActivationLayerInfo::ActivationFunction, BatchNomalizationPtr> fused_map =
+{
+    { ActivationLayerInfo::ActivationFunction::RELU, &batch_normalization<detail::relu<float, 4>> },
+    { ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, &batch_normalization<detail::brelu<float, 4>> },
+    { ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, &batch_normalization<detail::lubrelu<float, 4>> }
+};
+}
+namespace cpu
+{
+void fp32_neon_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
+                                   float epsilon, ActivationLayerInfo &act_info, const Window &window)
+{
+    fused_map[act_info.activation()](src, dst, mean, var, beta, gamma, epsilon, act_info, window);
+}
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/core/NEON/kernels/batchnormalization/impl/SVE/fp16.cpp b/src/core/NEON/kernels/batchnormalization/impl/SVE/fp16.cpp
new file mode 100644
index 0000000000..00326ffc8d
--- /dev/null
+++ b/src/core/NEON/kernels/batchnormalization/impl/SVE/fp16.cpp
@@ -0,0 +1,119 @@
+/*
+ * Copyright (c) 2020 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/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/SVEMath.h"
+#include "src/core/common/StdTypes.h"
+#include "src/core/common/Validate.h"
+
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+void fp16_sve_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
+                                  float epsilon, ActivationLayerInfo &act_info, const Window &window)
+{
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(src, win_collapsed);
+    Iterator output(dst, win_collapsed);
+
+    const auto input_mean  = reinterpret_cast<const float16_t *>(mean->ptr_to_element(Coordinates(0, 0)));
+    const auto input_var   = reinterpret_cast<const float16_t *>(var->ptr_to_element(Coordinates(0, 0)));
+    const auto input_gamma = (gamma != nullptr) ? reinterpret_cast<const float16_t *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr;
+    const auto input_beta  = (beta != nullptr) ? reinterpret_cast<const float16_t *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr;
+
+    const auto epsilon_vec = svdup_n_f16(epsilon);
+    const auto const_1     = svdup_n_f16(1.f);
+    const auto const_0     = svdup_n_f16(0.f);
+    const auto va          = svdup_n_f16(act_info.a());
+    const auto vb          = svdup_n_f16(act_info.b());
+    execute_window_loop(win_collapsed, [&](const Coordinates &)
+    {
+        const auto input_ptr  = reinterpret_cast<const float16_t *>(input.ptr());
+        const auto output_ptr = reinterpret_cast<float16_t *>(output.ptr());
+
+        // Compute S elements per iteration
+        int      x  = window_start_x;
+        svbool_t pg = svwhilelt_b16(x, window_end_x);
+        do
+        {
+            // Conctruct vectors
+            const auto mean_vec  = svld1_f16(pg, input_mean + x);
+            const auto var_vec   = svld1_f16(pg, input_var + x);
+            const auto gamma_vec = (input_gamma != nullptr) ? svld1_f16(pg, input_gamma + x) : const_1;
+            const auto beta_vec  = (input_beta != nullptr) ? svld1_f16(pg, input_beta + x) : const_0;
+
+            // Calculate denominator
+            const auto tmp         = svadd_f16_z(pg, var_vec, epsilon_vec);
+            auto       denominator = svrsqrte_f16(tmp);
+            denominator            = svmul_f16_z(pg, svrsqrts_f16(svmul_f16_z(pg, tmp, denominator), denominator), denominator);
+            denominator            = svmul_f16_z(pg, svrsqrts_f16(svmul_f16_z(pg, tmp, denominator), denominator), denominator);
+
+            // Calculate x bar
+            const auto numerator = svsub_f16_z(pg, svld1_f16(pg, input_ptr + x), mean_vec);
+            const auto x_bar     = svmul_f16_z(pg, numerator, denominator);
+            auto       res       = svmla_f16_z(pg, beta_vec, x_bar, gamma_vec);
+
+            // Perform fused activation
+            if(act_info.enabled())
+            {
+                if(act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU)
+                {
+                    res = svmax_f16_z(pg, const_0, res);
+                }
+                else if(act_info.activation() == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
+                {
+                    res = svmin_f16_z(pg, va, svmax_f16_z(pg, const_0, res));
+                }
+                else if(act_info.activation() == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+                {
+                    res = svmin_f16_z(pg, va, svmax_f16_z(pg, vb, res));
+                }
+            }
+
+            // Store results
+            svst1_f16(pg, output_ptr + x, res);
+
+            x += svcntw();
+            pg = svwhilelt_b16(x, window_end_x);
+        }
+        while(svptest_any(svptrue_b16(), pg));
+    },
+    input, output);
+}
+} // namespace cpu
+} // namespace arm_compute
+#endif // __ARM_FEATURE_SVE
diff --git a/src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp b/src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp
new file mode 100644
index 0000000000..317befd61e
--- /dev/null
+++ b/src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp
@@ -0,0 +1,119 @@
+/*
+ * Copyright (c) 2020 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/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/SVEMath.h"
+#include "src/core/common/StdTypes.h"
+#include "src/core/common/Validate.h"
+
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+void fp32_sve_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
+                                  float epsilon, ActivationLayerInfo &act_info, const Window &window)
+{
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(src, win_collapsed);
+    Iterator output(dst, win_collapsed);
+
+    const auto input_mean  = reinterpret_cast<const float *>(mean->ptr_to_element(Coordinates(0, 0)));
+    const auto input_var   = reinterpret_cast<const float *>(var->ptr_to_element(Coordinates(0, 0)));
+    const auto input_gamma = (gamma != nullptr) ? reinterpret_cast<const float *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr;
+    const auto input_beta  = (beta != nullptr) ? reinterpret_cast<const float *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr;
+
+    const auto epsilon_vec = svdup_n_f32(epsilon);
+    const auto const_1     = svdup_n_f32(1.f);
+    const auto const_0     = svdup_n_f32(0.f);
+    const auto va          = svdup_n_f32(act_info.a());
+    const auto vb          = svdup_n_f32(act_info.b());
+    execute_window_loop(win_collapsed, [&](const Coordinates &)
+    {
+        const auto input_ptr  = reinterpret_cast<const float *>(input.ptr());
+        const auto output_ptr = reinterpret_cast<float *>(output.ptr());
+
+        // Compute S elements per iteration
+        int      x  = window_start_x;
+        svbool_t pg = svwhilelt_b32(x, window_end_x);
+        do
+        {
+            // Conctruct vectors
+            const auto mean_vec  = svld1_f32(pg, input_mean + x);
+            const auto var_vec   = svld1_f32(pg, input_var + x);
+            const auto gamma_vec = (input_gamma != nullptr) ? svld1_f32(pg, input_gamma + x) : const_1;
+            const auto beta_vec  = (input_beta != nullptr) ? svld1_f32(pg, input_beta + x) : const_0;
+
+            // Calculate denominator
+            const auto tmp         = svadd_f32_z(pg, var_vec, epsilon_vec);
+            auto       denominator = svrsqrte_f32(tmp);
+            denominator            = svmul_f32_z(pg, svrsqrts_f32(svmul_f32_z(pg, tmp, denominator), denominator), denominator);
+            denominator            = svmul_f32_z(pg, svrsqrts_f32(svmul_f32_z(pg, tmp, denominator), denominator), denominator);
+
+            // Calculate x bar
+            const auto numerator = svsub_f32_z(pg, svld1_f32(pg, input_ptr + x), mean_vec);
+            const auto x_bar     = svmul_f32_z(pg, numerator, denominator);
+            auto       res       = svmla_f32_z(pg, beta_vec, x_bar, gamma_vec);
+
+            // Perform fused activation
+            if(act_info.enabled())
+            {
+                if(act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU)
+                {
+                    res = svmax_f32_z(pg, const_0, res);
+                }
+                else if(act_info.activation() == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
+                {
+                    res = svmin_f32_z(pg, va, svmax_f32_z(pg, const_0, res));
+                }
+                else if(act_info.activation() == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+                {
+                    res = svmin_f32_z(pg, va, svmax_f32_z(pg, vb, res));
+                }
+            }
+
+            // Store results
+            svst1_f32(pg, output_ptr + x, res);
+
+            x += svcntw();
+            pg = svwhilelt_b32(x, window_end_x);
+        }
+        while(svptest_any(svptrue_b32(), pg));
+    },
+    input, output);
+}
+} // namespace cpu
+} // namespace arm_compute
+#endif // __ARM_FEATURE_SVE
diff --git a/src/core/NEON/kernels/batchnormalization/impl/list.h b/src/core/NEON/kernels/batchnormalization/impl/list.h
new file mode 100644
index 0000000000..8e0ea36f5a
--- /dev/null
+++ b/src/core/NEON/kernels/batchnormalization/impl/list.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2020 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.
+ */
+#ifndef SRC_CORE_NEON_KERNELS_BATCH_NORMALIZATION_LIST_H
+#define SRC_CORE_NEON_KERNELS_BATCH_NORMALIZATION_LIST_H
+
+namespace arm_compute
+{
+namespace cpu
+{
+#define DECLARE_BATCH_NORMALIZATION_KERNEL(func_name)                                                                              \
+    void func_name(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, \
+                   float epsilon, ActivationLayerInfo &act_info, const Window &window)
+
+DECLARE_BATCH_NORMALIZATION_KERNEL(fp16_neon_batch_normalization);
+DECLARE_BATCH_NORMALIZATION_KERNEL(fp16_sve_batch_normalization);
+DECLARE_BATCH_NORMALIZATION_KERNEL(fp32_neon_batch_normalization);
+DECLARE_BATCH_NORMALIZATION_KERNEL(fp32_sve_batch_normalization);
+
+#undef DECLARE_ACTIVATION_KERNEL
+} // namespace cpu
+} // namespace arm_compute
+
+#endif /* SRC_CORE_NEON_KERNELS_BATCH_NORMALIZATION_LIST_H */
diff --git a/tests/validation/NEON/BatchNormalizationLayer.cpp b/tests/validation/NEON/BatchNormalizationLayer.cpp
index 067c5bb198..b24357f8ad 100644
--- a/tests/validation/NEON/BatchNormalizationLayer.cpp
+++ b/tests/validation/NEON/BatchNormalizationLayer.cpp
@@ -51,8 +51,10 @@ namespace
 RelativeTolerance<float>           rel_tolerance_f32(0.05f);   /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
 constexpr AbsoluteTolerance<float> abs_tolerance_f32(0.0001f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-constexpr AbsoluteTolerance<float> tolerance_f16(0.01f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */
-#endif                                                   // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+RelativeTolerance<float>           rel_tolerance_f16(0.05f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */
+constexpr AbsoluteTolerance<float> abs_tolerance_f16(0.01f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */
+#endif                                                       // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
 const auto act_infos = framework::dataset::make("ActivationInfo",
 {
     ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU),
@@ -148,7 +150,7 @@ FIXTURE_DATA_TEST_CASE(RandomSmall, NEBatchNormalizationLayerFixture<half>, fram
                                                                                                                        framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
 {
     // Validate output
-    validate(Accessor(_target), _reference, tolerance_f16, 0);
+    validate(Accessor(_target), _reference, abs_tolerance_f16, 0);
 }
 
 FIXTURE_DATA_TEST_CASE(RandomLarge, NEBatchNormalizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::LargeRandomBatchNormalizationLayerDataset(),
@@ -159,7 +161,7 @@ FIXTURE_DATA_TEST_CASE(RandomLarge, NEBatchNormalizationLayerFixture<half>, fram
                                                                                                                        framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
 {
     // Validate output
-    validate(Accessor(_target), _reference, tolerance_f16, 0);
+    validate(Accessor(_target), _reference, abs_tolerance_f16, 0);
 }
 TEST_SUITE_END() // FP16
 #endif           /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */