src/runtime/CL/functions/CLGenerateProposalsLayer.cpp

/*
 * Copyright (c) 2019-2020 Arm Limited.
 *
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to
 * deal in the Software without restriction, including without limitation the
 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
 * sell copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in all
 * copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */
#include "arm_compute/runtime/CL/functions/CLGenerateProposalsLayer.h"

#include "arm_compute/core/CL/ICLTensor.h"
#include "arm_compute/core/Types.h"
#include "src/core/CL/kernels/CLBoundingBoxTransformKernel.h"
#include "src/core/CL/kernels/CLDequantizationLayerKernel.h"
#include "src/core/CL/kernels/CLGenerateProposalsLayerKernel.h"
#include "src/core/CL/kernels/CLPadLayerKernel.h"
#include "src/core/CL/kernels/CLPermuteKernel.h"
#include "src/core/CL/kernels/CLQuantizationLayerKernel.h"
#include "src/core/helpers/AutoConfiguration.h"

namespace arm_compute
{
CLGenerateProposalsLayer::CLGenerateProposalsLayer(std::shared_ptr<IMemoryManager> memory_manager)
    : _memory_group(memory_manager),
      _permute_deltas_kernel(std::make_unique<CLPermuteKernel>()),
      _flatten_deltas(),
      _permute_scores_kernel(std::make_unique<CLPermuteKernel>()),
      _flatten_scores(),
      _compute_anchors_kernel(std::make_unique<CLComputeAllAnchorsKernel>()),
      _bounding_box_kernel(std::make_unique<CLBoundingBoxTransformKernel>()),
      _pad_kernel(std::make_unique<CLPadLayerKernel>()),
      _dequantize_anchors(std::make_unique<CLDequantizationLayerKernel>()),
      _dequantize_deltas(std::make_unique<CLDequantizationLayerKernel>()),
      _quantize_all_proposals(std::make_unique<CLQuantizationLayerKernel>()),
      _cpp_nms(memory_manager),
      _is_nhwc(false),
      _is_qasymm8(false),
      _deltas_permuted(),
      _deltas_flattened(),
      _deltas_flattened_f32(),
      _scores_permuted(),
      _scores_flattened(),
      _all_anchors(),
      _all_anchors_f32(),
      _all_proposals(),
      _all_proposals_quantized(),
      _keeps_nms_unused(),
      _classes_nms_unused(),
      _proposals_4_roi_values(),
      _all_proposals_to_use(nullptr),
      _num_valid_proposals(nullptr),
      _scores_out(nullptr)
{
}

CLGenerateProposalsLayer::~CLGenerateProposalsLayer() = default;

void CLGenerateProposalsLayer::configure(const ICLTensor *scores, const ICLTensor *deltas, const ICLTensor *anchors, ICLTensor *proposals, ICLTensor *scores_out, ICLTensor *num_valid_proposals,
                                         const GenerateProposalsInfo &info)
{
    configure(CLKernelLibrary::get().get_compile_context(), scores, deltas, anchors, proposals, scores_out, num_valid_proposals, info);
}

void CLGenerateProposalsLayer::configure(const CLCompileContext &compile_context, const ICLTensor *scores, const ICLTensor *deltas, const ICLTensor *anchors, ICLTensor *proposals,
                                         ICLTensor *scores_out,
                                         ICLTensor *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(CLGenerateProposalsLayer::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 scores_data_type = scores->info()->data_type();
    _is_qasymm8                     = scores_data_type == DataType::QASYMM8;
    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();

    const QuantizationInfo scores_qinfo   = scores->info()->quantization_info();
    const DataType         rois_data_type = (_is_qasymm8) ? DataType::QASYMM16 : scores_data_type;
    const QuantizationInfo rois_qinfo     = (_is_qasymm8) ? QuantizationInfo(0.125f, 0) : scores->info()->quantization_info();

    // Compute all the anchors
    _memory_group.manage(&_all_anchors);
    _compute_anchors_kernel->configure(compile_context, 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, scores_data_type, deltas->info()->quantization_info()));

    // Permute and reshape deltas
    _memory_group.manage(&_deltas_flattened);
    if(!_is_nhwc)
    {
        _memory_group.manage(&_deltas_permuted);
        _permute_deltas_kernel->configure(compile_context, deltas, &_deltas_permuted, PermutationVector{ 2, 0, 1 });
        _flatten_deltas.configure(compile_context, &_deltas_permuted, &_deltas_flattened);
        _deltas_permuted.allocator()->allocate();
    }
    else
    {
        _flatten_deltas.configure(compile_context, deltas, &_deltas_flattened);
    }

    const TensorShape flatten_shape_scores(1, total_num_anchors);
    _scores_flattened.allocator()->init(TensorInfo(flatten_shape_scores, 1, scores_data_type, scores_qinfo));

    // Permute and reshape scores
    _memory_group.manage(&_scores_flattened);
    if(!_is_nhwc)
    {
        _memory_group.manage(&_scores_permuted);
        _permute_scores_kernel->configure(compile_context, scores, &_scores_permuted, PermutationVector{ 2, 0, 1 });
        _flatten_scores.configure(compile_context, &_scores_permuted, &_scores_flattened);
        _scores_permuted.allocator()->allocate();
    }
    else
    {
        _flatten_scores.configure(compile_context, scores, &_scores_flattened);
    }

    CLTensor *anchors_to_use = &_all_anchors;
    CLTensor *deltas_to_use  = &_deltas_flattened;
    if(_is_qasymm8)
    {
        _all_anchors_f32.allocator()->init(TensorInfo(_all_anchors.info()->tensor_shape(), 1, DataType::F32));
        _deltas_flattened_f32.allocator()->init(TensorInfo(_deltas_flattened.info()->tensor_shape(), 1, DataType::F32));
        _memory_group.manage(&_all_anchors_f32);
        _memory_group.manage(&_deltas_flattened_f32);
        // Dequantize anchors to float
        _dequantize_anchors->configure(compile_context, &_all_anchors, &_all_anchors_f32);
        _all_anchors.allocator()->allocate();
        anchors_to_use = &_all_anchors_f32;
        // Dequantize deltas to float
        _dequantize_deltas->configure(compile_context, &_deltas_flattened, &_deltas_flattened_f32);
        _deltas_flattened.allocator()->allocate();
        deltas_to_use = &_deltas_flattened_f32;
    }
    // Bounding box transform
    _memory_group.manage(&_all_proposals);
    BoundingBoxTransformInfo bbox_info(info.im_width(), info.im_height(), 1.f);
    _bounding_box_kernel->configure(compile_context, anchors_to_use, &_all_proposals, deltas_to_use, bbox_info);
    deltas_to_use->allocator()->allocate();
    anchors_to_use->allocator()->allocate();

    _all_proposals_to_use = &_all_proposals;
    if(_is_qasymm8)
    {
        _memory_group.manage(&_all_proposals_quantized);
        // Requantize all_proposals to QASYMM16 with 0.125 scale and 0 offset
        _all_proposals_quantized.allocator()->init(TensorInfo(_all_proposals.info()->tensor_shape(), 1, DataType::QASYMM16, QuantizationInfo(0.125f, 0)));
        _quantize_all_proposals->configure(compile_context, &_all_proposals, &_all_proposals_quantized);
        _all_proposals.allocator()->allocate();
        _all_proposals_to_use = &_all_proposals_quantized;
    }
    // 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, scores_data_type, scores_qinfo);
    auto_init_if_empty(*_proposals_4_roi_values.info(), TensorShape(values_per_roi, scores_nms_size), 1, rois_data_type, rois_qinfo);
    auto_init_if_empty(*num_valid_proposals->info(), TensorShape(1), 1, DataType::U32);

    // Initialize temporaries (unused) outputs
    _classes_nms_unused.allocator()->init(TensorInfo(TensorShape(scores_nms_size), 1, scores_data_type, scores_qinfo));
    _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);
    _cpp_nms.configure(&_scores_flattened, _all_proposals_to_use, nullptr, scores_out, &_proposals_4_roi_values, &_classes_nms_unused, nullptr, &_keeps_nms_unused, num_valid_proposals,
                       BoxNMSLimitInfo(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()));
    _keeps_nms_unused.allocator()->allocate();
    _classes_nms_unused.allocator()->allocate();
    _all_proposals_to_use->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
    _pad_kernel->configure(compile_context, &_proposals_4_roi_values, proposals, PaddingList{ { 1, 0 } });
    _proposals_4_roi_values.allocator()->allocate();
}

Status CLGenerateProposalsLayer::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_TYPE_CHANNEL_NOT_IN(scores, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
    ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(scores, DataLayout::NCHW, DataLayout::NHWC);
    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(scores, deltas);
    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(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();

    const bool is_qasymm8 = scores->data_type() == DataType::QASYMM8;

    ARM_COMPUTE_RETURN_ERROR_ON(num_images > 1);

    if(is_qasymm8)
    {
        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(anchors, 1, DataType::QSYMM16);
        const UniformQuantizationInfo anchors_qinfo = anchors->quantization_info().uniform();
        ARM_COMPUTE_RETURN_ERROR_ON(anchors_qinfo.scale != 0.125f);
    }

    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(CLComputeAllAnchorsKernel::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(CLPermuteKernel::validate(deltas, &deltas_permuted_info, PermutationVector{ 2, 0, 1 }));
        ARM_COMPUTE_RETURN_ON_ERROR(CLPermuteKernel::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(CLReshapeLayer::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(CLReshapeLayer::validate(&scores_permuted_info, &scores_flattened_info));

    TensorInfo *proposals_4_roi_values_to_use = &proposals_4_roi_values;
    TensorInfo  proposals_4_roi_values_quantized(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
    proposals_4_roi_values_quantized.set_data_type(DataType::QASYMM16).set_quantization_info(QuantizationInfo(0.125f, 0));
    if(is_qasymm8)
    {
        TensorInfo all_anchors_f32_info(anchors->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true).set_data_type(DataType::F32));
        ARM_COMPUTE_RETURN_ON_ERROR(CLDequantizationLayerKernel::validate(&all_anchors_info, &all_anchors_f32_info));

        TensorInfo deltas_flattened_f32_info(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true).set_data_type(DataType::F32));
        ARM_COMPUTE_RETURN_ON_ERROR(CLDequantizationLayerKernel::validate(&deltas_flattened_info, &deltas_flattened_f32_info));

        TensorInfo proposals_4_roi_values_f32(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true).set_data_type(DataType::F32));
        ARM_COMPUTE_RETURN_ON_ERROR(CLBoundingBoxTransformKernel::validate(&all_anchors_f32_info, &proposals_4_roi_values_f32, &deltas_flattened_f32_info,
                                                                           BoundingBoxTransformInfo(info.im_width(), info.im_height(), 1.f)));

        ARM_COMPUTE_RETURN_ON_ERROR(CLQuantizationLayerKernel::validate(&proposals_4_roi_values_f32, &proposals_4_roi_values_quantized));
        proposals_4_roi_values_to_use = &proposals_4_roi_values_quantized;
    }
    else
    {
        ARM_COMPUTE_RETURN_ON_ERROR(CLBoundingBoxTransformKernel::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(CLPadLayerKernel::validate(proposals_4_roi_values_to_use, proposals, PaddingList{ { 1, 0 } }));

    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));
        if(is_qasymm8)
        {
            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(proposals, 1, DataType::QASYMM16);
            const UniformQuantizationInfo proposals_qinfo = proposals->quantization_info().uniform();
            ARM_COMPUTE_RETURN_ERROR_ON(proposals_qinfo.scale != 0.125f);
            ARM_COMPUTE_RETURN_ERROR_ON(proposals_qinfo.offset != 0);
        }
        else
        {
            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(proposals, scores);
        }
    }

    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 CLGenerateProposalsLayer::run_cpp_nms_kernel()
{
    // Map inputs
    _scores_flattened.map(true);
    _all_proposals_to_use->map(true);

    // Map outputs
    _scores_out->map(CLScheduler::get().queue(), true);
    _proposals_4_roi_values.map(CLScheduler::get().queue(), true);
    _num_valid_proposals->map(CLScheduler::get().queue(), true);
    _keeps_nms_unused.map(true);
    _classes_nms_unused.map(true);

    // Run nms
    _cpp_nms.run();

    // Unmap outputs
    _keeps_nms_unused.unmap();
    _classes_nms_unused.unmap();
    _scores_out->unmap(CLScheduler::get().queue());
    _proposals_4_roi_values.unmap(CLScheduler::get().queue());
    _num_valid_proposals->unmap(CLScheduler::get().queue());

    // Unmap inputs
    _scores_flattened.unmap();
    _all_proposals_to_use->unmap();
}

void CLGenerateProposalsLayer::run()
{
    // Acquire all the temporaries
    MemoryGroupResourceScope scope_mg(_memory_group);

    // Compute all the anchors
    CLScheduler::get().enqueue(*_compute_anchors_kernel, false);

    // Transpose and reshape the inputs
    if(!_is_nhwc)
    {
        CLScheduler::get().enqueue(*_permute_deltas_kernel, false);
        CLScheduler::get().enqueue(*_permute_scores_kernel, false);
    }
    _flatten_deltas.run();
    _flatten_scores.run();

    if(_is_qasymm8)
    {
        CLScheduler::get().enqueue(*_dequantize_anchors, false);
        CLScheduler::get().enqueue(*_dequantize_deltas, false);
    }

    // Build the boxes
    CLScheduler::get().enqueue(*_bounding_box_kernel, false);

    if(_is_qasymm8)
    {
        CLScheduler::get().enqueue(*_quantize_all_proposals, false);
    }

    // Non maxima suppression
    run_cpp_nms_kernel();
    // Add dummy batch indexes
    CLScheduler::get().enqueue(*_pad_kernel, true);
}
} // namespace arm_compute