• Dmitry Matveev's avatar
    Merge pull request #15090 from dmatveev:dm/ng-0001-g-api-inference-api · 0757a51e
    Dmitry Matveev authored
    * G-API-NG/API: Introduced inference API and IE-based backend
    
    - Very quick-n-dirty implementation
    - OpenCV's own DNN module is not used
    - No tests so far
    
    * G-API-NG/IE: Refined IE backend, added more tests
    
    * G-API-NG/IE: Fixed various CI warnings & build issues + tests
    
    - Added tests on multi-dimensional own::Mat
    - Added tests on GMatDesc with dimensions
    - Documentation on infer.hpp
    - Fixed more warnings + added a ROI list test
    - Fix descr_of clash for vector<Mat> & standalone mode
    - Fix build issue with gcc-4.8x
    - Addressed review comments
    
    * G-API-NG/IE: Addressed review comments
    
    - Pass `false` to findDataFile()
    - Add deprecation warning suppression macros for IE
    0757a51e
gexecutor.cpp 8.6 KB
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.
//
// Copyright (C) 2018 Intel Corporation


#include "precomp.hpp"

#include <iostream>

#include <ade/util/zip_range.hpp>

#include <opencv2/gapi/opencv_includes.hpp>
#include "executor/gexecutor.hpp"
#include "compiler/passes/passes.hpp"

cv::gimpl::GExecutor::GExecutor(std::unique_ptr<ade::Graph> &&g_model)
    : m_orig_graph(std::move(g_model))
    , m_island_graph(GModel::Graph(*m_orig_graph).metadata()
                     .get<IslandModel>().model)
    , m_gm(*m_orig_graph)
    , m_gim(*m_island_graph)
{
    // NB: Right now GIslandModel is acyclic, so for a naive execution,
    // simple unrolling to a list of triggers is enough

    // Naive execution model is similar to current CPU (OpenCV) plugin
    // execution model:
    // 1. Allocate all internal resources first (NB - CPU plugin doesn't do it)
    // 2. Put input/output GComputation arguments to the storage
    // 3. For every Island, prepare vectors of input/output parameter descs
    // 4. Iterate over a list of operations (sorted in the topological order)
    // 5. For every operation, form a list of input/output data objects
    // 6. Run GIslandExecutable
    // 7. writeBack

    auto sorted = m_gim.metadata().get<ade::passes::TopologicalSortData>();
    for (auto nh : sorted.nodes())
    {
        switch (m_gim.metadata(nh).get<NodeKind>().k)
        {
        case NodeKind::ISLAND:
            {
                std::vector<RcDesc> input_rcs;
                std::vector<RcDesc> output_rcs;
                input_rcs.reserve(nh->inNodes().size());
                output_rcs.reserve(nh->outNodes().size());

                auto xtract = [&](ade::NodeHandle slot_nh, std::vector<RcDesc> &vec) {
                    const auto orig_data_nh
                        = m_gim.metadata(slot_nh).get<DataSlot>().original_data_node;
                    const auto &orig_data_info
                        = m_gm.metadata(orig_data_nh).get<Data>();
                    vec.emplace_back(RcDesc{ orig_data_info.rc
                                           , orig_data_info.shape
                                           , orig_data_info.ctor});
                };
                // (3)
                for (auto in_slot_nh  : nh->inNodes())  xtract(in_slot_nh,  input_rcs);
                for (auto out_slot_nh : nh->outNodes()) xtract(out_slot_nh, output_rcs);

                m_ops.emplace_back(OpDesc{ std::move(input_rcs)
                                         , std::move(output_rcs)
                                         , m_gim.metadata(nh).get<IslandExec>().object});
            }
            break;

        case NodeKind::SLOT:
            {
                const auto orig_data_nh
                    = m_gim.metadata(nh).get<DataSlot>().original_data_node;
                // (1)
                initResource(orig_data_nh);
                m_slots.emplace_back(DataDesc{nh, orig_data_nh});
            }
            break;

        default:
            GAPI_Assert(false);
            break;
        } // switch(kind)
    } // for(gim nodes)
}

void cv::gimpl::GExecutor::initResource(const ade::NodeHandle &orig_nh)
{
    const Data &d = m_gm.metadata(orig_nh).get<Data>();

    if (   d.storage != Data::Storage::INTERNAL
        && d.storage != Data::Storage::CONST_VAL)
        return;

    // INTERNALS+CONST only! no need to allocate/reset output objects
    // to as it is bound externally (e.g. already in the m_res)

    switch (d.shape)
    {
    case GShape::GMAT:
        {
            const auto desc = util::get<cv::GMatDesc>(d.meta);
            auto& mat = m_res.slot<cv::gapi::own::Mat>()[d.rc];
            createMat(desc, mat);
        }
        break;

    case GShape::GSCALAR:
        if (d.storage == Data::Storage::CONST_VAL)
        {
            auto rc = RcDesc{d.rc, d.shape, d.ctor};
            magazine::bindInArg(m_res, rc, m_gm.metadata(orig_nh).get<ConstValue>().arg);
        }
        break;

    case GShape::GARRAY:
        // Constructed on Reset, do nothing here
        break;

    default:
        GAPI_Assert(false);
    }
}

void cv::gimpl::GExecutor::run(cv::gimpl::GRuntimeArgs &&args)
{
    // (2)
    const auto proto = m_gm.metadata().get<Protocol>();

    // Basic check if input/output arguments are correct
    // FIXME: Move to GCompiled (do once for all GExecutors)
    if (proto.inputs.size() != args.inObjs.size()) // TODO: Also check types
    {
        util::throw_error(std::logic_error
                          ("Computation's input protocol doesn\'t "
                           "match actual arguments!"));
    }
    if (proto.outputs.size() != args.outObjs.size()) // TODO: Also check types
    {
        util::throw_error(std::logic_error
                          ("Computation's output protocol doesn\'t "
                           "match actual arguments!"));
    }

    namespace util = ade::util;

    //ensure that output Mat parameters are correctly allocated
    for (auto index : util::iota(proto.out_nhs.size()) )     //FIXME: avoid copy of NodeHandle and GRunRsltComp ?
    {
        auto& nh = proto.out_nhs.at(index);
        const Data &d = m_gm.metadata(nh).get<Data>();
        if (d.shape == GShape::GMAT)
        {
            using cv::util::get;
            const auto desc = get<cv::GMatDesc>(d.meta);

            auto check_own_mat = [&desc, &args, &index]()
            {
                auto& out_mat = *get<cv::gapi::own::Mat*>(args.outObjs.at(index));
                GAPI_Assert(out_mat.data != nullptr &&
                        desc.canDescribe(out_mat));
            };

#if !defined(GAPI_STANDALONE)
            // Building as part of OpenCV - follow OpenCV behavior
            // In the case of cv::Mat if output buffer is not enough to hold the result, reallocate it
            if (cv::util::holds_alternative<cv::Mat*>(args.outObjs.at(index)))
            {
                auto& out_mat = *get<cv::Mat*>(args.outObjs.at(index));
                createMat(desc, out_mat);
            }
            // In the case of own::Mat never reallocated, checked to perfectly fit required meta
            else
            {
                check_own_mat();
            }
#else
            // Building standalone - output buffer should always exist,
            // and _exact_ match our inferred metadata
            check_own_mat();
#endif // !defined(GAPI_STANDALONE)
        }
    }
    // Update storage with user-passed objects
    for (auto it : ade::util::zip(ade::util::toRange(proto.inputs),
                                  ade::util::toRange(args.inObjs)))
    {
        magazine::bindInArg(m_res, std::get<0>(it), std::get<1>(it));
    }
    for (auto it : ade::util::zip(ade::util::toRange(proto.outputs),
                                  ade::util::toRange(args.outObjs)))
    {
        magazine::bindOutArg(m_res, std::get<0>(it), std::get<1>(it));
    }

    // Reset internal data
    for (auto &sd : m_slots)
    {
        const auto& data = m_gm.metadata(sd.data_nh).get<Data>();
        magazine::resetInternalData(m_res, data);
    }

    // Run the script
    for (auto &op : m_ops)
    {
        // (5)
        using InObj  = GIslandExecutable::InObj;
        using OutObj = GIslandExecutable::OutObj;
        std::vector<InObj>  in_objs;
        std::vector<OutObj> out_objs;
        in_objs.reserve (op.in_objects.size());
        out_objs.reserve(op.out_objects.size());

        for (const auto &rc : op.in_objects)
        {
            in_objs.emplace_back(InObj{rc, magazine::getArg(m_res, rc)});
        }
        for (const auto &rc : op.out_objects)
        {
            out_objs.emplace_back(OutObj{rc, magazine::getObjPtr(m_res, rc)});
        }

        // (6)
        op.isl_exec->run(std::move(in_objs), std::move(out_objs));
    }

    // (7)
    for (auto it : ade::util::zip(ade::util::toRange(proto.outputs),
                                  ade::util::toRange(args.outObjs)))
    {
        magazine::writeBack(m_res, std::get<0>(it), std::get<1>(it));
    }
}

const cv::gimpl::GModel::Graph& cv::gimpl::GExecutor::model() const
{
    return m_gm;
}

bool cv::gimpl::GExecutor::canReshape() const
{
    // FIXME: Introduce proper reshaping support on GExecutor level
    // for all cases!
    return (m_ops.size() == 1) && m_ops[0].isl_exec->canReshape();
}

void cv::gimpl::GExecutor::reshape(const GMetaArgs& inMetas, const GCompileArgs& args)
{
    GAPI_Assert(canReshape());
    auto& g = *m_orig_graph.get();
    ade::passes::PassContext ctx{g};
    passes::initMeta(ctx, inMetas);
    passes::inferMeta(ctx, true);
    m_ops[0].isl_exec->reshape(g, args);
}