COMPMID-1741: Implement NEFuseBatchNormalizationKernel

Change-Id: Ib3ba4b22804a94a1e8ef6d7076e28c2fc1cd2fa2
Reviewed-on: https://review.mlplatform.org/385
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Anthony Barbier <Anthony.barbier@arm.com>

7 files changed, 587 insertions, 8 deletions

diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index 49bba3a750..755e68a2e1 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -64,6 +64,7 @@
 #include "arm_compute/core/NEON/kernels/NEFillInnerBorderKernel.h"
 #include "arm_compute/core/NEON/kernels/NEFlattenLayerKernel.h"
 #include "arm_compute/core/NEON/kernels/NEFloorKernel.h"
+#include "arm_compute/core/NEON/kernels/NEFuseBatchNormalizationKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMAssemblyBaseKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMInterleave4x4Kernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.h"
diff --git a/arm_compute/core/NEON/kernels/NEFuseBatchNormalizationKernel.h b/arm_compute/core/NEON/kernels/NEFuseBatchNormalizationKernel.h
new file mode 100644
index 0000000000..4d5855e6f5
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEFuseBatchNormalizationKernel.h
@@ -0,0 +1,110 @@
+/*
+ * Copyright (c) 2018 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, INNEUDING 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 NEAIM, 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 __ARM_COMPUTE_NEFUSEBATCHNORMALIZATIONKERNEL_H__
+#define __ARM_COMPUTE_NEFUSEBATCHNORMALIZATIONKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+namespace arm_compute
+{
+// Forward declarations
+class ITensor;
+
+/** OpenNE kernel to fuse the batch normalization node to a preceding convolution node */
+class NEFuseBatchNormalizationKernel : public INEKernel
+{
+public:
+    const char *name() const override
+    {
+        return "NEFuseBatchNormalizationKernel";
+    }
+    /** Default constructor */
+    NEFuseBatchNormalizationKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEFuseBatchNormalizationKernel(const NEFuseBatchNormalizationKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEFuseBatchNormalizationKernel &operator=(const NEFuseBatchNormalizationKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEFuseBatchNormalizationKernel(NEFuseBatchNormalizationKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEFuseBatchNormalizationKernel &operator=(NEFuseBatchNormalizationKernel &&) = default;
+    /** Default destructor */
+    ~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]  epsilon       (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f.
+     */
+    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);
+    /** 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] epsilon       (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *conv_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);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    const ITensor *_conv_weights;
+    const ITensor *_conv_bias;
+    const ITensor *_bn_mean;
+    const ITensor *_bn_var;
+    const ITensor *_bn_gamma;
+    const ITensor *_bn_beta;
+    ITensor       *_fused_weights;
+    ITensor       *_fused_bias;
+    float          _epsilon;
+    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,
+                                       const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window);
+
+    FuseBatchNormFunction *_func;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_NEFUSEBATCHNORMALIZATIONKERNEL_H__ */
diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h
index 8873110982..9a023ec51f 100644
--- a/arm_compute/runtime/NEON/NEFunctions.h
+++ b/arm_compute/runtime/NEON/NEFunctions.h
@@ -63,6 +63,7 @@
 #include "arm_compute/runtime/NEON/functions/NEFlattenLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEFloor.h"
 #include "arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEFuseBatchNormalization.h"
 #include "arm_compute/runtime/NEON/functions/NEGEMM.h"
 #include "arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h"
 #include "arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h"
diff --git a/arm_compute/runtime/NEON/functions/NEFuseBatchNormalization.h b/arm_compute/runtime/NEON/functions/NEFuseBatchNormalization.h
new file mode 100644
index 0000000000..82ed329d86
--- /dev/null
+++ b/arm_compute/runtime/NEON/functions/NEFuseBatchNormalization.h
@@ -0,0 +1,94 @@
+/*
+ * Copyright (c) 2018 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, INNEUDING 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 NEAIM, 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 __ARM_COMPUTE_NEFUSEBATCHNORMALIZATION_H__
+#define __ARM_COMPUTE_NEFUSEBATCHNORMALIZATION_H__
+
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/kernels/NEFuseBatchNormalizationKernel.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/IFunction.h"
+
+namespace arm_compute
+{
+// Forward declarations
+class ITensor;
+
+/** Basic function to fuse the batch normalization node to a preceding convolution node */
+class NEFuseBatchNormalization : public IFunction
+{
+public:
+    /** Default constructor */
+    NEFuseBatchNormalization();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEFuseBatchNormalization(const NEFuseBatchNormalization &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEFuseBatchNormalization &operator=(const NEFuseBatchNormalization &) = delete;
+    /** Allow instances of this class to be moved */
+    NEFuseBatchNormalization(NEFuseBatchNormalization &&) = default;
+    /** Allow instances of this class to be moved */
+    NEFuseBatchNormalization &operator=(NEFuseBatchNormalization &&) = default;
+    /** Default destructor */
+    ~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]  epsilon       (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f.
+     */
+    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);
+    /** 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] epsilon       (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *conv_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);
+
+    // Inherited methods overridden:
+    void run() override;
+
+private:
+    NEFuseBatchNormalizationKernel _fuse_bn_kernel;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_NEFUSEBATCHNORMALIZATION_H__ */
diff --git a/src/core/NEON/kernels/NEFuseBatchNormalizationKernel.cpp b/src/core/NEON/kernels/NEFuseBatchNormalizationKernel.cpp
new file mode 100644
index 0000000000..25a0848977
--- /dev/null
+++ b/src/core/NEON/kernels/NEFuseBatchNormalizationKernel.cpp
@@ -0,0 +1,286 @@
+/*
+ * Copyright (c) 2018 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, INNEUDING 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 NEAIM, 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/NEON/kernels/NEFuseBatchNormalizationKernel.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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Window.h"
+
+#include "support/ToolchainSupport.h"
+
+#include "arm_compute/core/NEON/wrapper/wrapper.h"
+#include "utils/TypePrinter.h"
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *conv_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)
+{
+    ARM_COMPUTE_UNUSED(epsilon);
+    //ARM_COMPUTE_RETURN_ERROR_ON_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_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));
+
+    // Validate bias
+    if(conv_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);
+    }
+    // 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);
+    }
+    // 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);
+    }
+
+    // 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);
+    }
+    // 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);
+    }
+
+    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)
+{
+    using ExactTagType = typename wrapper::traits::neon_vector<ScalarType, size>::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);
+
+    // 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 conv_w_in(conv_weights, win);
+    Iterator conv_w_out(run_in_place_weights ? conv_weights : fused_weights, win);
+
+    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;
+    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 conv_bias_in_scalar = ScalarType(0.0);
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        if(slice != id[3])
+        {
+            slice = id[3];
+            mean  = input_mean[slice];
+            var   = input_var[slice];
+            gamma = ScalarType(1.0);
+            beta  = ScalarType(0.0);
+
+            // 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(input_beta != nullptr)
+            {
+                beta     = input_beta[slice];
+                beta_vec = wrapper::vdup_n(beta, 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_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));
+        }
+
+        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());
+
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
+        {
+            auto wn = wrapper::vloadq(conv_w_in_ptr + x);
+            wn      = wrapper::vmul(wn, rvar_vec);
+            wn      = wrapper::vmul(wn, gamma_vec);
+
+            // Store results
+            wrapper::vstore(conv_w_out_ptr + x, wn);
+        }
+
+        // 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_in, conv_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(),
+      _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,
+                                               ITensor *fused_weights, ITensor *fused_bias,
+                                               const ITensor *conv_bias, const ITensor *bn_beta, const ITensor *bn_gamma,
+                                               float epsilon)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(conv_weights, bn_mean, bn_var);
+
+    _conv_weights  = conv_weights;
+    _conv_bias     = conv_bias;
+    _bn_mean       = bn_mean;
+    _bn_var        = bn_var;
+    _bn_beta       = bn_beta;
+    _bn_gamma      = bn_gamma;
+    _fused_weights = fused_weights;
+    _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);
+
+    // 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());
+    }
+    if(_fused_bias != nullptr)
+    {
+        // Output tensor auto initialization if not yet initialized
+        auto_init_if_empty(*_fused_bias->info(), *_bn_mean->info()->clone());
+        _fused_bias->info()->set_valid_region(bn_mean->info()->valid_region());
+    }
+
+    // Validate arguments
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(conv_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,
+                                                  (bn_beta != nullptr) ? bn_beta->info() : nullptr,
+                                                  (bn_gamma != nullptr) ? bn_gamma->info() : nullptr,
+                                                  epsilon));
+
+    // Configure kernel window
+    Window win = calculate_max_window(*conv_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)
+    {
+        case DataType::F32:
+            _func = &fused_batch_normmalization<float, 4>;
+            break;
+#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;
+    }
+}
+
+Status NEFuseBatchNormalizationKernel::validate(const ITensorInfo *conv_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)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(conv_weights, bn_mean, bn_var, fused_weights, fused_bias, conv_bias, bn_beta, bn_gamma, epsilon));
+    return Status{};
+}
+
+void NEFuseBatchNormalizationKernel::run(const Window &window, const ThreadInfo &info)
+{
+    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);
+}
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEFuseBatchNormalization.cpp b/src/runtime/NEON/functions/NEFuseBatchNormalization.cpp
new file mode 100644
index 0000000000..7f17a15cba
--- /dev/null
+++ b/src/runtime/NEON/functions/NEFuseBatchNormalization.cpp
@@ -0,0 +1,59 @@
+/*
+ * Copyright (c) 2018 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, INNEUDING 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 NEAIM, 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 "arm_compute/core/Error.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/NEScheduler.h"
+
+namespace arm_compute
+{
+NEFuseBatchNormalization::NEFuseBatchNormalization()
+    : _fuse_bn_kernel()
+{
+}
+
+void NEFuseBatchNormalization::configure(const ITensor *conv_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)
+{
+    _fuse_bn_kernel.configure(conv_weights, bn_mean, bn_var, fused_weights, fused_bias, conv_bias, bn_beta, bn_gamma, epsilon);
+}
+
+Status NEFuseBatchNormalization::validate(const ITensorInfo *conv_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)
+{
+    return NEFuseBatchNormalizationKernel::validate(conv_weights, bn_mean, bn_var, fused_weights, fused_bias, conv_bias, bn_beta, bn_gamma, epsilon);
+}
+
+void NEFuseBatchNormalization::run()
+{
+    NEScheduler::get().schedule(&_fuse_bn_kernel, Window::DimY);
+}
+} // namespace arm_compute
diff --git a/tests/validation/NEON/BatchNormalizationLayer.cpp b/tests/validation/NEON/BatchNormalizationLayer.cpp
index ca13d26495..155f6c336e 100644
--- a/tests/validation/NEON/BatchNormalizationLayer.cpp
+++ b/tests/validation/NEON/BatchNormalizationLayer.cpp
@@ -23,18 +23,22 @@
  */
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/NEON/functions/NEBatchNormalizationLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEConvolutionLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEFuseBatchNormalization.h"
 #include "arm_compute/runtime/Tensor.h"
 #include "arm_compute/runtime/TensorAllocator.h"
 #include "tests/NEON/Accessor.h"
 #include "tests/PaddingCalculator.h"
 #include "tests/datasets/RandomBatchNormalizationLayerDataset.h"
 #include "tests/datasets/ShapeDatasets.h"
+#include "tests/datasets/SmallConvolutionLayerDataset.h"
 #include "tests/framework/Asserts.h"
 #include "tests/framework/Macros.h"
 #include "tests/framework/datasets/Datasets.h"
 #include "tests/validation/Helpers.h"
 #include "tests/validation/Validation.h"
 #include "tests/validation/fixtures/BatchNormalizationLayerFixture.h"
+#include "tests/validation/fixtures/BatchNormalizationLayerFusionFixture.h"
 
 namespace arm_compute
 {
@@ -44,16 +48,21 @@ namespace validation
 {
 namespace
 {
-constexpr AbsoluteTolerance<float> tolerance_f32(0.00001f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
+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 */
+#endif                                                   // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 const auto act_infos = framework::dataset::make("ActivationInfo",
 {
     ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU),
     ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f),
     ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f),
 });
+const auto common_fusion_dataset = combine(combine(combine(framework::dataset::make("UseBias", { false, true }),
+                                                           framework::dataset::make("UseBeta", { false, true })),
+                                                   framework::dataset::make("UseGamma", { false, true })),
+                                           framework::dataset::make("Epsilon", { 0.001f }));
 } // namespace
 
 TEST_SUITE(NEON)
@@ -150,7 +159,7 @@ FIXTURE_DATA_TEST_CASE(Random, NEBatchNormalizationLayerFixture<float>, framewor
                                                                                                                    framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
 {
     // Validate output
-    validate(Accessor(_target), _reference, tolerance_f32, 0);
+    validate(Accessor(_target), _reference, abs_tolerance_f32, 0);
 }
 TEST_SUITE_END()
 
@@ -166,12 +175,31 @@ FIXTURE_DATA_TEST_CASE(Random, NEBatchNormalizationLayerFixture<half>, framework
     // Validate output
     validate(Accessor(_target), _reference, tolerance_f16, 0);
 }
-TEST_SUITE_END()
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
-TEST_SUITE_END()
+TEST_SUITE_END() // FP16
+#endif           /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+TEST_SUITE_END() // Float
 
-TEST_SUITE_END()
-TEST_SUITE_END()
+TEST_SUITE_END() // BatchNormalizationLayer
+
+TEST_SUITE(BatchNormalizationLayerFusion)
+template <typename T>
+using NEBatchNormalizationLayerFusionFixture = BatchNormalizationLayerFusionValidationFixture<Tensor, Accessor, NEConvolutionLayer, NEFuseBatchNormalization, T>;
+
+TEST_SUITE(Float)
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEBatchNormalizationLayerFusionFixture<float>, framework::DatasetMode::PRECOMMIT,
+                       combine(combine(combine(datasets::SmallConvolutionLayerDataset(), common_fusion_dataset),
+                                       framework::dataset::make("DataType", DataType::F32)),
+                               framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+}
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // Float
+
+TEST_SUITE_END() // BatchNormalizationLayerFusion
+TEST_SUITE_END() // NEON
 } // namespace validation
 } // namespace test
 } // namespace arm_compute