COMPMID-2257: Implement NEGenerateProposals.

Change-Id: I8d751f8b09f842a214c305a4530a71d82f16db8f
Signed-off-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1943
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

4 files changed, 533 insertions, 20 deletions

diff --git a/src/core/NEON/kernels/NECopyKernel.cpp b/src/core/NEON/kernels/NECopyKernel.cpp
index 4722c05507..83f3dded4f 100644
--- a/src/core/NEON/kernels/NECopyKernel.cpp
+++ b/src/core/NEON/kernels/NECopyKernel.cpp
@@ -29,28 +29,88 @@
 #include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
 
-using namespace arm_compute;
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding = PaddingList())
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON(padding.size() > 4);
+
+    // Validate output if initialized
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(misc::shape_calculator::compute_padded_shape(input->tensor_shape(), padding), output->tensor_shape());
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
+{
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output, *input);
+    return std::make_pair(Status{}, calculate_max_window(*output));
+}
+
+std::pair<Status, Window> validate_and_configure_window_with_padding(ITensorInfo *input, ITensorInfo *output, const PaddingList &padding)
+{
+    const TensorShape input_shape  = input->tensor_shape();
+    const TensorShape padded_shape = misc::shape_calculator::compute_padded_shape(input_shape, padding);
+    auto_init_if_empty(*output, input->clone()->set_tensor_shape(padded_shape));
+    // Configure window
+    const Window win = calculate_max_window(*output, output->dimension(0));
+    return std::make_pair(Status{}, win);
+}
+
+} // namespace
 
 NECopyKernel::NECopyKernel()
-    : _input(nullptr), _output(nullptr)
+    : _input(nullptr), _output(nullptr), _padding()
 {
 }
 
-void NECopyKernel::configure(const ITensor *input, ITensor *output)
+void NECopyKernel::configure(const ITensor *input, ITensor *output, const PaddingList &padding)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), padding));
+
+    _input   = input;
+    _output  = output;
+    _padding = padding;
 
-    _input  = input;
-    _output = output;
+    std::pair<Status, Window> win_config;
 
-    INEKernel::configure(calculate_max_window(*output->info()));
+    if(padding.empty())
+    {
+        win_config = validate_and_configure_window(input->info(), output->info());
+    }
+    else
+    {
+        win_config = validate_and_configure_window_with_padding(input->info(), output->info(), padding);
+    }
+
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    INEKernel::configure(win_config.second);
 }
 
-Status NECopyKernel::validate(const arm_compute::ITensorInfo *input, const arm_compute::ITensorInfo *output)
+Status NECopyKernel::validate(const arm_compute::ITensorInfo *input, const arm_compute::ITensorInfo *output, const PaddingList &padding)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, padding));
+
+    if(padding.empty())
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
+    }
+    else
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_with_padding(input->clone().get(), output->clone().get(), padding).first);
+    }
+
     return Status{};
 }
 
@@ -60,22 +120,38 @@ void NECopyKernel::run(const Window &window, const ThreadInfo &info)
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
 
-    Window output_window{ window };
-    output_window.set(Window::DimX, Window::Dimension(output_window.x().start(), output_window.x().end(), _input->info()->dimension(0)));
-
-    Window out_slice = output_window.first_slice_window_1D();
+    if(_padding.empty())
+    {
+        Window output_window{ window };
+        output_window.set(Window::DimX, Window::Dimension(output_window.x().start(), output_window.x().end(), _input->info()->dimension(0)));
+        Window out_slice = output_window.first_slice_window_1D();
+        do
+        {
+            Iterator input_it(_input, out_slice);
+            Iterator output_it(_output, out_slice);
 
-    do
+            execute_window_loop(out_slice, [&](const Coordinates &)
+            {
+                memcpy(output_it.ptr(), input_it.ptr(), _output->info()->dimension(0) * _output->info()->element_size());
+            },
+            input_it, output_it);
+        }
+        while(output_window.slide_window_slice_1D(out_slice));
+    }
+    else
     {
-        Iterator input_it(_input, out_slice);
-        Iterator output_it(_output, out_slice);
+        Window input_window{ window };
+        input_window.set(Window::DimX, Window::Dimension(0, window.x().end() - _padding[0].first, _input->info()->dimension(0)));
 
-        execute_window_loop(out_slice, [&](const Coordinates &)
+        Iterator     input_it(_input, input_window);
+        Iterator     output_it(_output, window);
+        const size_t row_size_in_bytes = _input->info()->dimension(0) * _input->info()->element_size();
+        execute_window_loop(window, [&](const Coordinates &)
         {
-            memcpy(output_it.ptr(), input_it.ptr(), _output->info()->dimension(0) * _output->info()->element_size());
+            auto dst_ptr = output_it.ptr() + _padding[0].first * _output->info()->element_size();
+            std::memcpy(dst_ptr, input_it.ptr(), row_size_in_bytes);
         },
         input_it, output_it);
-
     }
-    while(output_window.slide_window_slice_1D(out_slice));
 }
+} // namespace arm_compute
diff --git a/src/core/NEON/kernels/NEGenerateProposalsLayerKernel.cpp b/src/core/NEON/kernels/NEGenerateProposalsLayerKernel.cpp
new file mode 100644
index 0000000000..4a585b70fd
--- /dev/null
+++ b/src/core/NEON/kernels/NEGenerateProposalsLayerKernel.cpp
@@ -0,0 +1,131 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/NEON/kernels/NEGenerateProposalsLayerKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CPP/Validate.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Window.h"
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *anchors, const ITensorInfo *all_anchors, const ComputeAnchorsInfo &info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(anchors, all_anchors);
+    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(anchors);
+    ARM_COMPUTE_RETURN_ERROR_ON(anchors->dimension(0) != info.values_per_roi());
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(anchors, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON(anchors->num_dimensions() > 2);
+    if(all_anchors->total_size() > 0)
+    {
+        const size_t feature_height = info.feat_height();
+        const size_t feature_width  = info.feat_width();
+        const size_t num_anchors    = anchors->dimension(1);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(all_anchors, anchors);
+        ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->num_dimensions() > 2);
+        ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->dimension(0) != info.values_per_roi());
+        ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->dimension(1) != feature_height * feature_width * num_anchors);
+    }
+    return Status{};
+}
+
+} // namespace
+
+NEComputeAllAnchorsKernel::NEComputeAllAnchorsKernel()
+    : _anchors(nullptr), _all_anchors(nullptr), _anchors_info(0.f, 0.f, 0.f)
+{
+}
+
+void NEComputeAllAnchorsKernel::configure(const ITensor *anchors, ITensor *all_anchors, const ComputeAnchorsInfo &info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(anchors, all_anchors);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(anchors->info(), all_anchors->info(), info));
+
+    // Metadata
+    const size_t   num_anchors = anchors->info()->dimension(1);
+    const DataType data_type   = anchors->info()->data_type();
+    const float    width       = info.feat_width();
+    const float    height      = info.feat_height();
+
+    // Initialize the output if empty
+    const TensorShape output_shape(info.values_per_roi(), width * height * num_anchors);
+    auto_init_if_empty(*all_anchors->info(), output_shape, 1, data_type);
+
+    // Set instance variables
+    _anchors      = anchors;
+    _all_anchors  = all_anchors;
+    _anchors_info = info;
+
+    Window win = calculate_max_window(*all_anchors->info(), Steps(info.values_per_roi()));
+
+    INEKernel::configure(win);
+}
+
+Status NEComputeAllAnchorsKernel::validate(const ITensorInfo *anchors, const ITensorInfo *all_anchors, const ComputeAnchorsInfo &info)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(anchors, all_anchors, info));
+    return Status{};
+}
+
+void NEComputeAllAnchorsKernel::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);
+
+    Iterator all_anchors_it(_all_anchors, window);
+    Iterator anchors_it(_all_anchors, window);
+
+    const size_t num_anchors = _anchors->info()->dimension(1);
+    const float  stride      = 1.f / _anchors_info.spatial_scale();
+    const size_t feat_width  = _anchors_info.feat_width();
+
+    execute_window_loop(window, [&](const Coordinates & id)
+    {
+        const size_t anchor_offset = id.y() % num_anchors;
+
+        const auto out_anchor_ptr = reinterpret_cast<float *>(all_anchors_it.ptr());
+        const auto anchor_ptr     = reinterpret_cast<float *>(_anchors->ptr_to_element(Coordinates(0, anchor_offset)));
+
+        *out_anchor_ptr       = *anchor_ptr;
+        *(1 + out_anchor_ptr) = *(1 + anchor_ptr);
+        *(2 + out_anchor_ptr) = *(2 + anchor_ptr);
+        *(3 + out_anchor_ptr) = *(3 + anchor_ptr);
+
+        const size_t shift_idy = id.y() / num_anchors;
+        const float  shiftx    = (shift_idy % feat_width) * stride;
+        const float  shifty    = (shift_idy / feat_width) * stride;
+
+        *out_anchor_ptr += shiftx;
+        *(out_anchor_ptr + 1) += shifty;
+        *(out_anchor_ptr + 2) += shiftx;
+        *(out_anchor_ptr + 3) += shifty;
+    },
+    all_anchors_it);
+}
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEComputeAllAnchors.cpp b/src/runtime/NEON/functions/NEComputeAllAnchors.cpp
new file mode 100644
index 0000000000..7702fb026d
--- /dev/null
+++ b/src/runtime/NEON/functions/NEComputeAllAnchors.cpp
@@ -0,0 +1,42 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/runtime/NEON/functions/NEComputeAllAnchors.h"
+
+#include "support/ToolchainSupport.h"
+
+namespace arm_compute
+{
+void NEComputeAllAnchors::configure(const ITensor *anchors, ITensor *all_anchors, const ComputeAnchorsInfo &info)
+{
+    // Configure ComputeAllAnchors kernel
+    auto k = arm_compute::support::cpp14::make_unique<NEComputeAllAnchorsKernel>();
+    k->configure(anchors, all_anchors, info);
+    _kernel = std::move(k);
+}
+
+Status NEComputeAllAnchors::validate(const ITensorInfo *anchors, const ITensorInfo *all_anchors, const ComputeAnchorsInfo &info)
+{
+    return NEComputeAllAnchorsKernel::validate(anchors, all_anchors, info);
+}
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEGenerateProposalsLayer.cpp b/src/runtime/NEON/functions/NEGenerateProposalsLayer.cpp
new file mode 100644
index 0000000000..6e5da43a94
--- /dev/null
+++ b/src/runtime/NEON/functions/NEGenerateProposalsLayer.cpp
@@ -0,0 +1,264 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/runtime/NEON/functions/NEGenerateProposalsLayer.h"
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/NEScheduler.h"
+#include "support/ToolchainSupport.h"
+
+namespace arm_compute
+{
+NEGenerateProposalsLayer::NEGenerateProposalsLayer(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)),
+      _permute_deltas_kernel(),
+      _flatten_deltas_kernel(),
+      _permute_scores_kernel(),
+      _flatten_scores_kernel(),
+      _compute_anchors_kernel(),
+      _bounding_box_kernel(),
+      _memset_kernel(),
+      _padded_copy_kernel(),
+      _cpp_nms_kernel(),
+      _is_nhwc(false),
+      _deltas_permuted(),
+      _deltas_flattened(),
+      _scores_permuted(),
+      _scores_flattened(),
+      _all_anchors(),
+      _all_proposals(),
+      _keeps_nms_unused(),
+      _classes_nms_unused(),
+      _proposals_4_roi_values(),
+      _num_valid_proposals(nullptr),
+      _scores_out(nullptr)
+{
+}
+
+void NEGenerateProposalsLayer::configure(const ITensor *scores, const ITensor *deltas, const ITensor *anchors, ITensor *proposals, ITensor *scores_out, ITensor *num_valid_proposals,
+                                         const GenerateProposalsInfo &info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(scores, deltas, anchors, proposals, scores_out, num_valid_proposals);
+    ARM_COMPUTE_ERROR_THROW_ON(NEGenerateProposalsLayer::validate(scores->info(), deltas->info(), anchors->info(), proposals->info(), scores_out->info(), num_valid_proposals->info(), info));
+
+    _is_nhwc                         = scores->info()->data_layout() == DataLayout::NHWC;
+    const DataType data_type         = deltas->info()->data_type();
+    const int      num_anchors       = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::CHANNEL));
+    const int      feat_width        = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::WIDTH));
+    const int      feat_height       = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::HEIGHT));
+    const int      total_num_anchors = num_anchors * feat_width * feat_height;
+    const int      pre_nms_topN      = info.pre_nms_topN();
+    const int      post_nms_topN     = info.post_nms_topN();
+    const size_t   values_per_roi    = info.values_per_roi();
+
+    // Compute all the anchors
+    _memory_group.manage(&_all_anchors);
+    _compute_anchors_kernel.configure(anchors, &_all_anchors, ComputeAnchorsInfo(feat_width, feat_height, info.spatial_scale()));
+
+    const TensorShape flatten_shape_deltas(values_per_roi, total_num_anchors);
+    _deltas_flattened.allocator()->init(TensorInfo(flatten_shape_deltas, 1, data_type));
+    _memory_group.manage(&_deltas_flattened);
+
+    // Permute and reshape deltas
+    if(!_is_nhwc)
+    {
+        _memory_group.manage(&_deltas_permuted);
+        _permute_deltas_kernel.configure(deltas, &_deltas_permuted, PermutationVector{ 2, 0, 1 });
+        _flatten_deltas_kernel.configure(&_deltas_permuted, &_deltas_flattened);
+        _deltas_permuted.allocator()->allocate();
+    }
+    else
+    {
+        _flatten_deltas_kernel.configure(deltas, &_deltas_flattened);
+    }
+
+    const TensorShape flatten_shape_scores(1, total_num_anchors);
+    _scores_flattened.allocator()->init(TensorInfo(flatten_shape_scores, 1, data_type));
+    _memory_group.manage(&_scores_flattened);
+    // Permute and reshape scores
+    if(!_is_nhwc)
+    {
+        _memory_group.manage(&_scores_permuted);
+        _permute_scores_kernel.configure(scores, &_scores_permuted, PermutationVector{ 2, 0, 1 });
+        _flatten_scores_kernel.configure(&_scores_permuted, &_scores_flattened);
+        _scores_permuted.allocator()->allocate();
+    }
+    else
+    {
+        _flatten_scores_kernel.configure(scores, &_scores_flattened);
+    }
+
+    // Bounding box transform
+    _memory_group.manage(&_all_proposals);
+    BoundingBoxTransformInfo bbox_info(info.im_width(), info.im_height(), 1.f);
+    _bounding_box_kernel.configure(&_all_anchors, &_all_proposals, &_deltas_flattened, bbox_info);
+    _deltas_flattened.allocator()->allocate();
+    _all_anchors.allocator()->allocate();
+
+    // The original layer implementation first selects the best pre_nms_topN anchors (thus having a lightweight sort)
+    // that are then transformed by bbox_transform. The boxes generated are then fed into a non-sorting NMS operation.
+    // Since we are reusing the NMS layer and we don't implement any CL/sort, we let NMS do the sorting (of all the input)
+    // and the filtering
+    const int   scores_nms_size = std::min<int>(std::min<int>(post_nms_topN, pre_nms_topN), total_num_anchors);
+    const float min_size_scaled = info.min_size() * info.im_scale();
+    _memory_group.manage(&_classes_nms_unused);
+    _memory_group.manage(&_keeps_nms_unused);
+
+    // Note that NMS needs outputs preinitialized.
+    auto_init_if_empty(*scores_out->info(), TensorShape(scores_nms_size), 1, data_type);
+    auto_init_if_empty(*_proposals_4_roi_values.info(), TensorShape(values_per_roi, scores_nms_size), 1, data_type);
+    auto_init_if_empty(*num_valid_proposals->info(), TensorShape(scores_nms_size), 1, DataType::U32);
+
+    // Initialize temporaries (unused) outputs
+    _classes_nms_unused.allocator()->init(TensorInfo(TensorShape(8, 1), 1, data_type));
+    _keeps_nms_unused.allocator()->init(*scores_out->info());
+
+    // Save the output (to map and unmap them at run)
+    _scores_out          = scores_out;
+    _num_valid_proposals = num_valid_proposals;
+
+    _memory_group.manage(&_proposals_4_roi_values);
+
+    const BoxNMSLimitInfo box_nms_info(0.0f, info.nms_thres(), scores_nms_size, false, NMSType::LINEAR, 0.5f, 0.001f, true, min_size_scaled, info.im_width(), info.im_height());
+    _cpp_nms_kernel.configure(&_scores_flattened /*scores_in*/,
+                              &_all_proposals /*boxes_in,*/,
+                              nullptr /* batch_splits_in*/,
+                              scores_out /* scores_out*/,
+                              &_proposals_4_roi_values /*boxes_out*/,
+                              &_classes_nms_unused /*classes*/,
+                              nullptr /*batch_splits_out*/,
+                              &_keeps_nms_unused /*keeps*/,
+                              num_valid_proposals /* keeps_size*/,
+                              box_nms_info);
+
+    _keeps_nms_unused.allocator()->allocate();
+    _classes_nms_unused.allocator()->allocate();
+    _all_proposals.allocator()->allocate();
+    _scores_flattened.allocator()->allocate();
+
+    // Add the first column that represents the batch id. This will be all zeros, as we don't support multiple images
+    _padded_copy_kernel.configure(&_proposals_4_roi_values, proposals, PaddingList{ { 1, 0 } });
+    _proposals_4_roi_values.allocator()->allocate();
+
+    _memset_kernel.configure(proposals, PixelValue());
+}
+
+Status NEGenerateProposalsLayer::validate(const ITensorInfo *scores, const ITensorInfo *deltas, const ITensorInfo *anchors, const ITensorInfo *proposals, const ITensorInfo *scores_out,
+                                          const ITensorInfo *num_valid_proposals, const GenerateProposalsInfo &info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(scores, deltas, anchors, proposals, scores_out, num_valid_proposals);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(scores, DataLayout::NCHW, DataLayout::NHWC);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(scores, deltas);
+
+    const int num_anchors       = scores->dimension(get_data_layout_dimension_index(scores->data_layout(), DataLayoutDimension::CHANNEL));
+    const int feat_width        = scores->dimension(get_data_layout_dimension_index(scores->data_layout(), DataLayoutDimension::WIDTH));
+    const int feat_height       = scores->dimension(get_data_layout_dimension_index(scores->data_layout(), DataLayoutDimension::HEIGHT));
+    const int num_images        = scores->dimension(3);
+    const int total_num_anchors = num_anchors * feat_width * feat_height;
+    const int values_per_roi    = info.values_per_roi();
+
+    ARM_COMPUTE_RETURN_ERROR_ON(num_images > 1);
+
+    TensorInfo all_anchors_info(anchors->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
+    ARM_COMPUTE_RETURN_ON_ERROR(NEComputeAllAnchorsKernel::validate(anchors, &all_anchors_info, ComputeAnchorsInfo(feat_width, feat_height, info.spatial_scale())));
+
+    TensorInfo deltas_permuted_info = deltas->clone()->set_tensor_shape(TensorShape(values_per_roi * num_anchors, feat_width, feat_height)).set_is_resizable(true);
+    TensorInfo scores_permuted_info = scores->clone()->set_tensor_shape(TensorShape(num_anchors, feat_width, feat_height)).set_is_resizable(true);
+    if(scores->data_layout() == DataLayout::NHWC)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(deltas, &deltas_permuted_info);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(scores, &scores_permuted_info);
+    }
+    else
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(NEPermuteKernel::validate(deltas, &deltas_permuted_info, PermutationVector{ 2, 0, 1 }));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEPermuteKernel::validate(scores, &scores_permuted_info, PermutationVector{ 2, 0, 1 }));
+    }
+
+    TensorInfo deltas_flattened_info(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
+    ARM_COMPUTE_RETURN_ON_ERROR(NEReshapeLayerKernel::validate(&deltas_permuted_info, &deltas_flattened_info));
+
+    TensorInfo scores_flattened_info(scores->clone()->set_tensor_shape(TensorShape(1, total_num_anchors)).set_is_resizable(true));
+    TensorInfo proposals_4_roi_values(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
+
+    ARM_COMPUTE_RETURN_ON_ERROR(NEReshapeLayerKernel::validate(&scores_permuted_info, &scores_flattened_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(NEBoundingBoxTransformKernel::validate(&all_anchors_info, &proposals_4_roi_values, &deltas_flattened_info, BoundingBoxTransformInfo(info.im_width(), info.im_height(),
+                                                                       1.f)));
+
+    ARM_COMPUTE_RETURN_ON_ERROR(NECopyKernel::validate(&proposals_4_roi_values, proposals, PaddingList{ { 0, 1 } }));
+
+    if(num_valid_proposals->total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(num_valid_proposals->num_dimensions() > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(num_valid_proposals->dimension(0) > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(num_valid_proposals, 1, DataType::U32);
+    }
+
+    if(proposals->total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(proposals->num_dimensions() > 2);
+        ARM_COMPUTE_RETURN_ERROR_ON(proposals->dimension(0) != size_t(values_per_roi) + 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(proposals->dimension(1) != size_t(total_num_anchors));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(proposals, deltas);
+    }
+
+    if(scores_out->total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(scores_out->num_dimensions() > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(scores_out->dimension(0) != size_t(total_num_anchors));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(scores_out, scores);
+    }
+
+    return Status{};
+}
+
+void NEGenerateProposalsLayer::run()
+{
+    // Acquire all the temporaries
+    MemoryGroupResourceScope scope_mg(_memory_group);
+
+    // Compute all the anchors
+    NEScheduler::get().schedule(&_compute_anchors_kernel, Window::DimY);
+
+    // Transpose and reshape the inputs
+    if(!_is_nhwc)
+    {
+        NEScheduler::get().schedule(&_permute_deltas_kernel, Window::DimY);
+        NEScheduler::get().schedule(&_permute_scores_kernel, Window::DimY);
+    }
+
+    NEScheduler::get().schedule(&_flatten_deltas_kernel, Window::DimY);
+    NEScheduler::get().schedule(&_flatten_scores_kernel, Window::DimY);
+
+    // Build the boxes
+    NEScheduler::get().schedule(&_bounding_box_kernel, Window::DimY);
+
+    // Non maxima suppression
+    CPPScheduler::get().schedule(&_cpp_nms_kernel, Window::DimX);
+
+    // Add dummy batch indexes
+
+    NEScheduler::get().schedule(&_memset_kernel, Window::DimY);
+    NEScheduler::get().schedule(&_padded_copy_kernel, Window::DimY);
+}
+} // namespace arm_compute