COMPMID-1461 SSD support: Create NEON PriorBox

Change-Id: I99e1c3939cfea4b9cb0ddfa313706f31b213ca89

10 files changed, 673 insertions, 5 deletions

diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index 76eb5cb5f2..0cfde722d9 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -99,6 +99,7 @@
 #include "arm_compute/core/NEON/kernels/NEPermuteKernel.h"
 #include "arm_compute/core/NEON/kernels/NEPixelWiseMultiplicationKernel.h"
 #include "arm_compute/core/NEON/kernels/NEPoolingLayerKernel.h"
+#include "arm_compute/core/NEON/kernels/NEPriorBoxLayerKernel.h"
 #include "arm_compute/core/NEON/kernels/NEQuantizationLayerKernel.h"
 #include "arm_compute/core/NEON/kernels/NEROIPoolingLayerKernel.h"
 #include "arm_compute/core/NEON/kernels/NEReductionOperationKernel.h"
diff --git a/arm_compute/core/NEON/kernels/NEPriorBoxLayerKernel.h b/arm_compute/core/NEON/kernels/NEPriorBoxLayerKernel.h
new file mode 100644
index 0000000000..ac61b61142
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEPriorBoxLayerKernel.h
@@ -0,0 +1,105 @@
+/*
+ * 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, 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 __ARM_COMPUTE_NEPRIORBOXLAYERKERNEL_H__
+#define __ARM_COMPUTE_NEPRIORBOXLAYERKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Interface for the kernel to calculate prior boxes */
+class NEPriorBoxLayerKernel : public INEKernel
+{
+public:
+    const char *name() const override
+    {
+        return "NEPriorBoxLayerKernel";
+    }
+    /** Default constructor */
+    NEPriorBoxLayerKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEPriorBoxLayerKernel(const NEPriorBoxLayerKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEPriorBoxLayerKernel &operator=(const NEPriorBoxLayerKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEPriorBoxLayerKernel(NEPriorBoxLayerKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEPriorBoxLayerKernel &operator=(NEPriorBoxLayerKernel &&) = default;
+    /** Set the input and output tensors.
+     *
+     * @param[in]  input1 First source tensor. Data types supported: F32. Data layouts supported: NCHW/NHWC.
+     * @param[in]  input2 Second source tensor. Data types and layouts supported: same as @p input1
+     * @param[out] output Destination tensor. Output dimensions are [W * H * num_priors * 4, 2]. Data type supported: same as @p input
+     * @param[in]  info   Prior box layer info.
+     */
+    void configure(const ITensor *input1, const ITensor *input2, ITensor *output, const PriorBoxLayerInfo &info);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEPriorBoxLayerKernel
+     *
+     * @param[in] input1 First source tensor info. Data types supported: F32. Data layouts supported: NCHW/NHWC.
+     * @param[in] input2 Second source tensor info. Data types and layouts supported: same as @p input1
+     * @param[in] output Destination tensor info. Output dimensions are [W * H * num_priors * 4, 2]. Data type supported: same as @p input
+     * @param[in] info   Prior box layer info.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const PriorBoxLayerInfo &info);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    /** Stores the coordinates of the calculated prior boxes.
+     *
+     * @param[out] out        Output pointer.
+     * @param[in]  offset     Output offset to write to.
+     * @param[in]  center_x   Center pixel value on x-axis.
+     * @param[in]  center_y   Center pixel value on y-axis.
+     * @param[in]  box_width  Prior box width.
+     * @param[in]  box_height Prior box height.
+     * @param[in]  width      Input width.
+     * @param[in]  height     Input height.
+     */
+    template <DataLayout DL>
+    void store_coordinates(float *out, const int offset, const float center_x, const float center_y, const float box_width, const float box_height, const int width, const int height);
+    /** Function to calculate prior boxes.
+     *
+     * @param[in] window Input region on which to execute the kernel.
+     */
+    template <DataLayout DL>
+    void calculate_prior_boxes(const Window &window);
+    /** Common signature for all the specialised PriorBox functions
+     *
+     * @param[in] window Input region on which to execute the kernel..
+     */
+    using PriorBoxFunction = void (NEPriorBoxLayerKernel::*)(const Window &window);
+    PriorBoxFunction  _func;
+    const ITensor    *_input1;
+    const ITensor    *_input2;
+    ITensor          *_output;
+    PriorBoxLayerInfo _info;
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_NEPRIORBOXLAYERKERNEL_H__ */
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index 1c9571463b..fb277584fd 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -847,7 +847,7 @@ public:
     /** Constructor
      *
      * @param[in] min_sizes     Min sizes vector.
-     * @param[in] variances     Variances vector. Size must be equal to 4.
+     * @param[in] variances     Variances vector.
      * @param[in] offset        Offset value.
      * @param[in] flip          (Optional) Flip the aspect ratios.
      * @param[in] clip          (Optional) Clip coordinates so that they're within [0,1].
@@ -864,7 +864,7 @@ public:
           _flip(flip),
           _clip(clip),
           _max_sizes(max_sizes),
-          _aspect_ratios(aspect_ratios),
+          _aspect_ratios(),
           _img_size(img_size),
           _steps(steps)
     {
diff --git a/arm_compute/runtime/CL/functions/CLPriorBoxLayer.h b/arm_compute/runtime/CL/functions/CLPriorBoxLayer.h
index 2376cd3fc6..6adf7f9dd1 100644
--- a/arm_compute/runtime/CL/functions/CLPriorBoxLayer.h
+++ b/arm_compute/runtime/CL/functions/CLPriorBoxLayer.h
@@ -24,8 +24,6 @@
 #ifndef __ARM_COMPUTE_CLPRIORBOXLAYER_H__
 #define __ARM_COMPUTE_CLPRIORBOXLAYER_H__
 
-#include "arm_compute/runtime/IFunction.h"
-
 #include "arm_compute/core/CL/kernels/CLPriorBoxLayerKernel.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/CL/ICLSimpleFunction.h"
diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h
index 555d7d7551..d9c8355178 100644
--- a/arm_compute/runtime/NEON/NEFunctions.h
+++ b/arm_compute/runtime/NEON/NEFunctions.h
@@ -99,6 +99,7 @@
 #include "arm_compute/runtime/NEON/functions/NEPhase.h"
 #include "arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h"
 #include "arm_compute/runtime/NEON/functions/NEPoolingLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEPriorBoxLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEQuantizationLayer.h"
 #include "arm_compute/runtime/NEON/functions/NERNNLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEROIPoolingLayer.h"
diff --git a/arm_compute/runtime/NEON/functions/NEPriorBoxLayer.h b/arm_compute/runtime/NEON/functions/NEPriorBoxLayer.h
new file mode 100644
index 0000000000..34ba39d960
--- /dev/null
+++ b/arm_compute/runtime/NEON/functions/NEPriorBoxLayer.h
@@ -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, 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 __ARM_COMPUTE_NEPRIORBOXLAYER_H__
+#define __ARM_COMPUTE_NEPRIORBOXLAYER_H__
+
+#include "arm_compute/core/NEON/kernels/NEPriorBoxLayerKernel.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/INESimpleFunction.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Basic function to run @ref NEPriorBoxLayerKernel. */
+class NEPriorBoxLayer : public INESimpleFunction
+{
+public:
+    /** Set the input and output tensors.
+     *
+     * @param[in]  input1 First source tensor. Data types supported: F32. Data layouts supported: NCHW/NHWC.
+     * @param[in]  input2 Second source tensor. Data types and layouts supported: same as @p input1
+     * @param[out] output Destination tensor. Output dimensions are [W * H * num_priors * 4, 2]. Data type supported: same as @p input
+     * @param[in]  info   Prior box layer info.
+     */
+    void configure(const ITensor *input1, const ITensor *input2, ITensor *output, const PriorBoxLayerInfo &info);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEPriorBoxLayer
+     *
+     * @param[in] input1 First source tensor info. Data types supported: F32. Data layouts supported: NCHW/NHWC.
+     * @param[in] input2 Second source tensor info. Data types and layouts supported: same as @p input1
+     * @param[in] output Destination tensor info. Output dimensions are [W * H * num_priors * 4, 2]. Data type supported: same as @p input
+     * @param[in] info   Prior box layer info.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const PriorBoxLayerInfo &info);
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_NEPRIORBOXLAYER_H__ */
diff --git a/src/core/NEON/kernels/NEPriorBoxLayerKernel.cpp b/src/core/NEON/kernels/NEPriorBoxLayerKernel.cpp
new file mode 100644
index 0000000000..2f6317921b
--- /dev/null
+++ b/src/core/NEON/kernels/NEPriorBoxLayerKernel.cpp
@@ -0,0 +1,358 @@
+/*
+ * 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, 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/NEON/kernels/NEPriorBoxLayerKernel.h"
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+
+#include <arm_neon.h>
+#include <cstdint>
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const PriorBoxLayerInfo &info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input1, input2);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input1, input2);
+
+    // Check variances
+    const int var_size = info.variances().size();
+    if(var_size > 1)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(var_size != 4, "Must provide 4 variance values");
+        for(int i = 0; i < var_size; ++i)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MSG(var_size <= 0, "Must be greater than 0");
+        }
+    }
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.steps()[0] < 0.f, "Step x should be greater or equal to 0");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.steps()[1] < 0.f, "Step y should be greater or equal to 0");
+
+    if(!info.max_sizes().empty())
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.max_sizes().size() != info.min_sizes().size(), "Max and min sizes dimensions should match");
+    }
+
+    for(unsigned int i = 0; i < info.max_sizes().size(); ++i)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.max_sizes()[i] < info.min_sizes()[i], "Max size should be greater than min size");
+    }
+
+    if(output != nullptr && output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(get_data_layout_dimension_index(input1->data_layout(), DataLayoutDimension::HEIGHT)) != 2);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input1, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output, const PriorBoxLayerInfo &info)
+{
+    ARM_COMPUTE_UNUSED(input2);
+
+    Window win            = {};
+    bool   window_changed = false;
+    switch(input1->data_layout())
+    {
+        case DataLayout::NCHW:
+        {
+            const int          num_priors                        = info.aspect_ratios().size() * info.min_sizes().size() + info.max_sizes().size();
+            const unsigned int num_elems_processed_per_iteration = 4 * num_priors;
+            win                                                  = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
+            AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
+            window_changed = update_window_and_padding(win, output_access);
+            break;
+        }
+        case DataLayout::NHWC:
+        {
+            win = calculate_max_window(*output, Steps());
+            break;
+        }
+        default:
+            ARM_COMPUTE_ERROR("Not implemented");
+    };
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
+
+NEPriorBoxLayerKernel::NEPriorBoxLayerKernel()
+    : _func(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr), _info()
+{
+}
+
+template <DataLayout DL>
+void NEPriorBoxLayerKernel::store_coordinates(float *out, const int offset, const float center_x, const float center_y, const float box_width, const float box_height, const int width,
+                                              const int height)
+{
+    float xmin = (center_x - box_width / 2.f) / width;
+    float ymin = (center_y - box_height / 2.f) / height;
+    float xmax = (center_x + box_width / 2.f) / width;
+    float ymax = (center_y + box_height / 2.f) / height;
+
+    switch(DL)
+    {
+        case DataLayout::NCHW:
+        {
+            float32x4_t vec_elements = { xmin, ymin, xmax, ymax };
+            if(_info.clip())
+            {
+                static const float32x4_t CONST_0 = vdupq_n_f32(0.f);
+                static const float32x4_t CONST_1 = vdupq_n_f32(1.f);
+                vec_elements                     = vmaxq_f32(vminq_f32(vec_elements, CONST_1), CONST_0);
+            }
+            vst1q_f32(out + offset, vec_elements);
+        }
+        break;
+        case DataLayout::NHWC:
+        {
+            const int output_offset = _output->info()->strides_in_bytes()[1] / _output->info()->element_size();
+            if(_info.clip())
+            {
+                xmin = std::min(std::max(xmin, 0.f), 1.f);
+                ymin = std::min(std::max(ymin, 0.f), 1.f);
+                xmax = std::min(std::max(xmax, 0.f), 1.f);
+                ymax = std::min(std::max(ymax, 0.f), 1.f);
+            }
+
+            *(out + output_offset * offset)       = xmin;
+            *(out + output_offset * (offset + 1)) = ymin;
+            *(out + output_offset * (offset + 2)) = xmax;
+            *(out + output_offset * (offset + 3)) = ymax;
+        }
+        break;
+        default:
+            ARM_COMPUTE_ERROR("Not implemented");
+    }
+}
+
+template <DataLayout DL>
+void NEPriorBoxLayerKernel::calculate_prior_boxes(const Window &window)
+{
+    const int num_priors = _info.aspect_ratios().size() * _info.min_sizes().size() + _info.max_sizes().size();
+
+    const int width_idx  = get_data_layout_dimension_index(DL, DataLayoutDimension::WIDTH);
+    const int height_idx = get_data_layout_dimension_index(DL, DataLayoutDimension::HEIGHT);
+
+    const int layer_width  = _input1->info()->dimension(width_idx);
+    const int layer_height = _input1->info()->dimension(height_idx);
+
+    int img_width  = _info.img_size().x;
+    int img_height = _info.img_size().y;
+    if(img_width == 0 || img_height == 0)
+    {
+        img_width  = _input2->info()->dimension(width_idx);
+        img_height = _input2->info()->dimension(height_idx);
+    }
+
+    float step_x = _info.steps()[0];
+    float step_y = _info.steps()[1];
+    if(step_x == 0.f || step_y == 0.f)
+    {
+        step_x = static_cast<float>(img_width) / layer_width;
+        step_y = static_cast<float>(img_height) / layer_height;
+    }
+
+    Window slice = {};
+
+    switch(DL)
+    {
+        case DataLayout::NCHW:
+            slice = window.first_slice_window_2D();
+            slice.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), 2));
+            break;
+        case DataLayout::NHWC:
+            slice = window.first_slice_window_3D();
+            slice.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), 4 * num_priors));
+            slice.set(Window::DimZ, Window::Dimension(0, _output->info()->dimension(2), 2));
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Not implemented");
+    }
+
+    Iterator output(_output, slice);
+    execute_window_loop(slice, [&](const Coordinates & id)
+    {
+        float center_x = 0;
+        float center_y = 0;
+        int   idx      = 0;
+        switch(DL)
+        {
+            case DataLayout::NCHW:
+                idx      = id.x() / (4 * num_priors);
+                center_x = (static_cast<float>(idx % layer_width) + _info.offset()) * step_x;
+                center_y = (static_cast<float>(idx / layer_width) + _info.offset()) * step_y;
+                break;
+            case DataLayout::NHWC:
+                idx      = id.y() / (4 * num_priors);
+                center_x = (static_cast<float>(idx % layer_width) + _info.offset()) * step_x;
+                center_y = (static_cast<float>(idx / layer_width) + _info.offset()) * step_y;
+                break;
+            default:
+                ARM_COMPUTE_ERROR("Not implemented");
+        }
+
+        float box_width;
+        float box_height;
+        int   offset = 0;
+
+        auto out = reinterpret_cast<float *>(output.ptr());
+        for(unsigned int i = 0; i < _info.min_sizes().size(); ++i)
+        {
+            const float min_size = _info.min_sizes().at(i);
+            box_width            = min_size;
+            box_height           = min_size;
+            store_coordinates<DL>(out, offset, center_x, center_y, box_width, box_height, img_width, img_height);
+            offset += 4;
+
+            if(!_info.max_sizes().empty())
+            {
+                const float max_size = _info.max_sizes().at(i);
+                box_width            = std::sqrt(min_size * max_size);
+                box_height           = box_width;
+
+                store_coordinates<DL>(out, offset, center_x, center_y, box_width, box_height, img_width, img_height);
+                offset += 4;
+            }
+
+            // rest of priors
+            for(auto ar : _info.aspect_ratios())
+            {
+                if(fabs(ar - 1.) < 1e-6)
+                {
+                    continue;
+                }
+
+                box_width  = min_size * sqrt(ar);
+                box_height = min_size / sqrt(ar);
+
+                store_coordinates<DL>(out, offset, center_x, center_y, box_width, box_height, img_width, img_height);
+                offset += 4;
+            }
+        }
+
+        // set the variance
+        switch(DL)
+        {
+            case DataLayout::NCHW:
+            {
+                out = reinterpret_cast<float *>(_output->ptr_to_element(Coordinates(id.x(), 1)));
+                float32x4_t var;
+                if(_info.variances().size() == 1)
+                {
+                    var = vdupq_n_f32(_info.variances().at(0));
+                }
+                else
+                {
+                    const float32x4_t vars = { _info.variances().at(0), _info.variances().at(1), _info.variances().at(2), _info.variances().at(3) };
+                    var                    = vars;
+                }
+                for(int i = 0; i < num_priors; ++i)
+                {
+                    vst1q_f32(out + 4 * i, var);
+                }
+            }
+            break;
+            case DataLayout::NHWC:
+            {
+                for(int i = 0; i < num_priors; ++i)
+                {
+                    const int  prior_offset = 4 * i;
+                    const bool single_var   = _info.variances().size() == 1;
+                    *(reinterpret_cast<float *>(_output->ptr_to_element(Coordinates(0, id.y() + prior_offset + 0, 1)))) = _info.variances().at(0);
+                    *(reinterpret_cast<float *>(_output->ptr_to_element(Coordinates(0, id.y() + prior_offset + 1, 1)))) = single_var ? _info.variances().at(0) : _info.variances().at(1);
+                    *(reinterpret_cast<float *>(_output->ptr_to_element(Coordinates(0, id.y() + prior_offset + 2, 1)))) = single_var ? _info.variances().at(0) : _info.variances().at(2);
+                    *(reinterpret_cast<float *>(_output->ptr_to_element(Coordinates(0, id.y() + prior_offset + 3, 1)))) = single_var ? _info.variances().at(0) : _info.variances().at(3);
+                }
+            }
+            break;
+            default:
+                ARM_COMPUTE_ERROR("Not implemented");
+        }
+
+    },
+    output);
+}
+
+void NEPriorBoxLayerKernel::configure(const ITensor *input1, const ITensor *input2, ITensor *output, const PriorBoxLayerInfo &info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input1->info(), input2->info(), output->info(), info));
+
+    _input1 = input1;
+    _input2 = input2;
+    _info   = info;
+    _output = output;
+
+    switch(input1->info()->data_layout())
+    {
+        case DataLayout::NCHW:
+        {
+            _func = &NEPriorBoxLayerKernel::calculate_prior_boxes<DataLayout::NCHW>;
+            break;
+        }
+        case DataLayout::NHWC:
+        {
+            _func = &NEPriorBoxLayerKernel::calculate_prior_boxes<DataLayout::NHWC>;
+            break;
+        }
+        default:
+            ARM_COMPUTE_ERROR("Not implemented.");
+    }
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input1->info(), input2->info(), output->info(), info);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    INEKernel::configure(win_config.second);
+}
+
+Status NEPriorBoxLayerKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const PriorBoxLayerInfo &info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input1, input2, output, info));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input1->clone().get(), input2->clone().get(), output->clone().get(), info)
+                                .first);
+
+    return Status{};
+}
+void NEPriorBoxLayerKernel::run(const Window &window, const ThreadInfo &info)
+{
+    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);
+
+    // Run function
+    (this->*_func)(window);
+}
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/runtime/NEON/functions/NEPriorBoxLayer.cpp b/src/runtime/NEON/functions/NEPriorBoxLayer.cpp
new file mode 100644
index 0000000000..6e7d4ab667
--- /dev/null
+++ b/src/runtime/NEON/functions/NEPriorBoxLayer.cpp
@@ -0,0 +1,47 @@
+/*
+ * 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, 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/NEPriorBoxLayer.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/NEON/NEScheduler.h"
+
+namespace arm_compute
+{
+void NEPriorBoxLayer::configure(const ITensor *input1, const ITensor *input2, ITensor *output, const PriorBoxLayerInfo &info)
+{
+    auto k = arm_compute::support::cpp14::make_unique<NEPriorBoxLayerKernel>();
+    k->configure(input1, input2, output, info);
+    _kernel = std::move(k);
+}
+
+Status NEPriorBoxLayer::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const PriorBoxLayerInfo &info)
+{
+    return NEPriorBoxLayerKernel::validate(input1, input2, output, info);
+}
+} // namespace arm_compute
diff --git a/tests/datasets/PriorBoxLayerDataset.h b/tests/datasets/PriorBoxLayerDataset.h
index 22a0d97057..c63e941171 100644
--- a/tests/datasets/PriorBoxLayerDataset.h
+++ b/tests/datasets/PriorBoxLayerDataset.h
@@ -110,7 +110,6 @@ public:
         std::vector<float> max_val      = { 60.f };
         std::vector<float> aspect_ratio = { 2.f };
         std::array<float, 2> steps = { 8.f, 8.f };
-
         add_config(TensorShape(4U, 4U), PriorBoxLayerInfo(min_val, var, 0.5f, true, false, max_val, aspect_ratio, Coordinates2D{ 8, 8 }, steps));
     }
 };
diff --git a/tests/validation/NEON/PriorBoxLayer.cpp b/tests/validation/NEON/PriorBoxLayer.cpp
new file mode 100644
index 0000000000..ed11120a10
--- /dev/null
+++ b/tests/validation/NEON/PriorBoxLayer.cpp
@@ -0,0 +1,100 @@
+/*
+ * 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, 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/Types.h"
+#include "arm_compute/runtime/NEON/functions/NEPriorBoxLayer.h"
+#include "arm_compute/runtime/Tensor.h"
+#include "arm_compute/runtime/TensorAllocator.h"
+#include "tests/NEON/Accessor.h"
+#include "tests/datasets/PriorBoxLayerDataset.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/PriorBoxLayerFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+constexpr AbsoluteTolerance<float> tolerance_f32(0.00001f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
+} // namespace
+
+TEST_SUITE(NEON)
+TEST_SUITE(PriorBoxLayer)
+
+template <typename T>
+using NEPriorBoxLayerFixture = PriorBoxLayerValidationFixture<Tensor, Accessor, NEPriorBoxLayer, T>;
+
+// *INDENT-OFF*
+// clang-format off
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(
+               framework::dataset::make("Input1Info", { TensorInfo(TensorShape(10U, 10U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(10U, 10U, 2U), 1, DataType::F32),    // Window shrink
+                                                     }),
+               framework::dataset::make("Input2Info", { TensorInfo(TensorShape(10U, 10U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(10U, 10U, 2U), 1, DataType::F32),
+                                                     })),
+               framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(1200U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(1000U, 2U), 1, DataType::F32),
+                                                     })),
+               framework::dataset::make("PriorBoxInfo",{ PriorBoxLayerInfo(std::vector<float>(1), std::vector<float>(1), 0, true, true, std::vector<float>(1), std::vector<float>(1), Coordinates2D{8, 8}, std::array<float, 2>()),
+                                                         PriorBoxLayerInfo(std::vector<float>(1), std::vector<float>(1), 0, true, true, std::vector<float>(1), std::vector<float>(1), Coordinates2D{8, 8}, std::array<float, 2>()),
+                                                     })),
+               framework::dataset::make("Expected", { true, false})),
+               input1_info, input2_info, output_info, info, expected)
+{
+    bool has_error = bool(NEPriorBoxLayer::validate(&input1_info.clone()->set_is_resizable(false), &input2_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), info));
+    ARM_COMPUTE_EXPECT(has_error == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+
+TEST_SUITE(Float)
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEPriorBoxLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallPriorBoxLayerDataset(),
+                                                                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                                                           framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_f32, 0);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEPriorBoxLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargePriorBoxLayerDataset(),
+                                                                                                                 framework::dataset::make("DataType", DataType::F32)),
+                                                                                                         framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_f32, 0);
+}
+TEST_SUITE_END() // Float
+TEST_SUITE_END() // FP32
+
+TEST_SUITE_END() // PriorBoxLayer
+TEST_SUITE_END() // NEON
+} // namespace validation
+} // namespace test
+} // namespace arm_compute