src/layers/mvn_layer.cpp

/*M///////////////////////////////////////////////////////////////////////////////////////
//
//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
//  By downloading, copying, installing or using the software you agree to this license.
//  If you do not agree to this license, do not download, install,
//  copy or use the software.
//
//
//                           License Agreement
//                For Open Source Computer Vision Library
//
// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
// Copyright (C) 2017, Intel Corporation, all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
//   * Redistribution's of source code must retain the above copyright notice,
//     this list of conditions and the following disclaimer.
//
//   * Redistribution's in binary form must reproduce the above copyright notice,
//     this list of conditions and the following disclaimer in the documentation
//     and/or other materials provided with the distribution.
//
//   * The name of the copyright holders may not be used to endorse or promote products
//     derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/

#include "../precomp.hpp"
#include "layers_common.hpp"
#include "../op_inf_engine.hpp"
#include "../ie_ngraph.hpp"
#include "../op_cuda.hpp"

#include <opencv2/dnn/shape_utils.hpp>

#ifdef HAVE_OPENCL
#include "../ocl4dnn/include/math_functions.hpp"
#include "opencl_kernels_dnn.hpp"
#endif

#ifdef HAVE_CUDA
#include "../cuda4dnn/primitives/mvn.hpp"
using namespace cv::dnn::cuda4dnn;
#endif

namespace cv
{
namespace dnn
{

class MVNLayerImpl CV_FINAL : public MVNLayer
{
public:
    MVNLayerImpl(const LayerParams& params)
    {
        setParamsFrom(params);
        normVariance = params.get<bool>("normalize_variance", true);
        acrossChannels = params.get<bool>("across_channels", false);
        eps = params.get<double>("eps", 1e-9);
        fuse_batch_norm = false;
        fuse_relu = false;
        relu_slope = 0.f;
        zeroDev = false;
    }

    Mat scale, shift;
#ifdef HAVE_OPENCL
    UMat umat_scale, umat_shift;
#endif
    bool fuse_batch_norm;

    Ptr<ReLULayer> activ_relu;
    float relu_slope;
    bool fuse_relu;
    bool zeroDev;  // TODO: Doesn't considered in Intel's Inference Engine backend.
    bool setActivation(const Ptr<ActivationLayer>& layer) CV_OVERRIDE
    {
        if (!layer.empty() && !fuse_relu && !fuse_batch_norm)
        {
            layer->getScaleShift(scale, shift);
            fuse_batch_norm = !scale.empty() || !shift.empty();
            return fuse_batch_norm;
        }

        if (!layer.empty() && preferableTarget == DNN_TARGET_OPENCL)
        {
            activ_relu = layer.dynamicCast<ReLULayer>();
            if( !activ_relu.empty() )
                relu_slope = activ_relu->negativeSlope;
        }
        fuse_relu = !activ_relu.empty();
        return fuse_relu;
    }

    void finalize(InputArrayOfArrays inputs_arr, OutputArrayOfArrays) CV_OVERRIDE
    {
        std::vector<Mat> inputs;
        inputs_arr.getMatVector(inputs);
        int splitDim = (acrossChannels) ? 1 : 2;
        int i, newRows = 1;
        for( i = 0; i < splitDim; i++ )
            newRows *= inputs[0].size[i];
        zeroDev = inputs[0].total() == newRows;
#ifdef HAVE_OPENCL
        umat_scale.release();
        umat_shift.release();
#endif
    }

    virtual bool supportBackend(int backendId) CV_OVERRIDE
    {
#ifdef HAVE_DNN_IE_NN_BUILDER_2019
        if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019)
        {
            bool isMyriad = preferableTarget == DNN_TARGET_MYRIAD || preferableTarget == DNN_TARGET_HDDL;
            return !zeroDev && (!isMyriad || eps <= 1e-7f);
        }
#endif
#ifdef HAVE_DNN_NGRAPH
        if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)
            return true;
#endif
        {
            return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_CUDA;
        }
    }

#ifdef HAVE_OPENCL
    bool fast_forward_ocl(std::vector<UMat> &inputs, std::vector<UMat> &outputs)
    {
        if (umat_scale.empty() && !scale.empty())
            scale.copyTo(umat_scale);
        if (umat_shift.empty() && !shift.empty())
            shift.copyTo(umat_shift);
        UMat& bnorm_weight = umat_scale;
        UMat& bnorm_bias = umat_shift;

        const unsigned LOCAL_SIZE = 128;
        bool use_half = (inputs[0].depth() == CV_16S);
        String opts = format(" -DT=%s -DT4=%s -Dconvert_T=%s -DLOCAL_SIZE=%u", use_half ? "half" : "float",
                             use_half ? "half4" : "float4", use_half ? "convert_half4" : "convert_float4",
                             LOCAL_SIZE
        );

        int splitDim = (acrossChannels) ? 1 : 2;
        for (size_t inpIdx = 0; inpIdx < inputs.size(); inpIdx++)
        {
            UMat &inpMat = inputs[inpIdx];
            UMat &outMat = outputs[inpIdx];
            int newRows = total(shape(inpMat), 0, splitDim);
            CV_Assert(newRows != 0);

            MatShape s = shape(newRows, inpMat.total() / newRows);
            UMat meanMat = UMat(s[0], 1, (use_half) ? CV_16S : CV_32F);
            UMat tmpMat  = UMat(s[0], s[1], CV_32F);
            float alpha = 1.0f / s[1];

            String buildopt = "-DNUM=4" + opts;
            ocl::Kernel k("mean_fuse4", ocl::dnn::mvn_oclsrc, buildopt + " -DKERNEL_MEAN_FUSE");
            size_t localsize[] = { LOCAL_SIZE };
            size_t globalsize[] = { (size_t)s[0] / 4 * localsize[0] };

            int argId = 0;
            k.set(argId++, ocl::KernelArg::PtrReadOnly(inpMat));
            k.set(argId++, (int)s[1]);
            k.set(argId++, alpha);
            k.set(argId++, ocl::KernelArg::PtrWriteOnly(meanMat));
            k.set(argId++, ocl::KernelArg::PtrWriteOnly(tmpMat));
            bool ret = k.run(1, globalsize, localsize, false);
            if (!ret)
                return false;

            buildopt += format(" %s %s", (fuse_batch_norm) ? "-DFUSE_BATCH_NORM" : "",
                               (fuse_relu) ? "-DFUSE_RELU" : "");

            ocl::Kernel k1("mvn_fuse4", ocl::dnn::mvn_oclsrc, buildopt + " -DKERNEL_MVN_FUSE");
            argId = 0;
            k1.set(argId++, ocl::KernelArg::PtrReadOnly(tmpMat));
            k1.set(argId++, ocl::KernelArg::PtrReadOnly(inpMat));
            k1.set(argId++, ocl::KernelArg::PtrReadOnly(meanMat));
            k1.set(argId++, (int)s[1]);
            k1.set(argId++, (float)alpha);
            k1.set(argId++, (float)eps);
            k1.set(argId++, (float)relu_slope);
            k1.set(argId++, ocl::KernelArg::PtrReadOnly(bnorm_weight));
            k1.set(argId++, ocl::KernelArg::PtrReadOnly(bnorm_bias));
            k1.set(argId++, ocl::KernelArg::PtrWriteOnly(outMat));
            ret = k1.run(1, globalsize, localsize, false);
            if (!ret)
                return false;
        }
        return true;
    }

    bool forward_ocl(InputArrayOfArrays inputs_, OutputArrayOfArrays outputs_, OutputArrayOfArrays internals_)
    {
        if (umat_scale.empty() && !scale.empty())
            scale.copyTo(umat_scale);
        if (umat_shift.empty() && !shift.empty())
            shift.copyTo(umat_shift);
        UMat& bnorm_weight = umat_scale;
        UMat& bnorm_bias = umat_shift;

        std::vector<UMat> inputs;
        std::vector<UMat> outputs;

        inputs_.getUMatVector(inputs);
        outputs_.getUMatVector(outputs);

        int splitDim = (acrossChannels) ? 1 : 2;
        int row_size = total(shape(inputs[0]), 0, splitDim);
        int plane_size = total(shape(inputs[0]), splitDim);
        if (normVariance && (row_size % 4 == 0) && (plane_size % 4 == 0))
            return fast_forward_ocl(inputs, outputs);

        if (inputs[0].depth() == CV_16S)
            return false;

        String opts = format(" -DT=float -DT4=float4 -Dconvert_T=convert_float4");

        for (size_t inpIdx = 0; inpIdx < inputs.size(); inpIdx++)
        {
            UMat &inpMat = inputs[inpIdx];
            UMat &outMat = outputs[inpIdx];
            int newRows = total(shape(inpMat), 0, splitDim);
            CV_Assert(newRows != 0);

            MatShape s = shape(newRows, inpMat.total() / newRows);
            UMat oneMat = UMat::ones(s[1], 1, CV_32F);
            UMat meanMat = UMat(s[0], 1, CV_32F);
            UMat devMat  = UMat(s[0], 1, CV_32F);
            UMat tmpMat  = UMat(s[0], s[1], CV_32F);
            float alpha = 1.0f / s[1];

            bool ret = ocl4dnn::ocl4dnnGEMV<float>(ocl4dnn::CblasNoTrans, s[0], s[1], alpha,
                                                   inpMat, 0, oneMat, 0, 0.0f, meanMat, 0);
            if (!ret)
                return false;

            int number = (s[1] % 8 == 0) ? 8 : ((s[1] % 4 == 0) ? 4 : 1);
            size_t global[] = { (size_t)s[0], (size_t)(s[1] / number) };
            String buildopt = format("-DNUM=%d", number) + opts;
            if (normVariance)
            {
                String kname = format("calc_mean%d", number);
                ocl::Kernel kernel(kname.c_str(), ocl::dnn::mvn_oclsrc, buildopt + " -DKERNEL_MEAN");
                if (kernel.empty())
                    return false;

                kernel.set(0, ocl::KernelArg::PtrReadOnly(inpMat));
                kernel.set(1, (int)s[0]);
                kernel.set(2, (int)s[1]);
                kernel.set(3, ocl::KernelArg::PtrReadOnly(meanMat));
                kernel.set(4, ocl::KernelArg::PtrWriteOnly(tmpMat));
                ret = kernel.run(2, global, NULL, false);
                if (!ret)
                    return false;

                ret = ocl4dnn::ocl4dnnGEMV<float>(ocl4dnn::CblasNoTrans, s[0], s[1], alpha,
                                                  tmpMat, 0, oneMat, 0, 0.0f, devMat, 0);
                if (!ret)
                    return false;
            }

            String kname = format("mvn%d", number);
            buildopt += format("%s%s%s -DKERNEL_MVN", (normVariance) ? " -DNORM_VARIANCE" : "",
                               (fuse_batch_norm) ? " -DFUSE_BATCH_NORM" : "",
                               (fuse_relu) ? " -DFUSE_RELU" : "");
            ocl::Kernel kernel1(kname.c_str(), ocl::dnn::mvn_oclsrc, buildopt);
            if (kernel1.empty())
                return false;
            kernel1.set(0, ocl::KernelArg::PtrReadOnly(inpMat));
            kernel1.set(1, (int)s[0]);
            kernel1.set(2, (int)s[1]);
            kernel1.set(3, (float)eps);
            kernel1.set(4, ocl::KernelArg::PtrReadOnly(meanMat));
            kernel1.set(5, ocl::KernelArg::PtrReadOnly(devMat));
            kernel1.set(6, ocl::KernelArg::PtrReadOnly(bnorm_weight));
            kernel1.set(7, ocl::KernelArg::PtrReadOnly(bnorm_bias));
            kernel1.set(8, (int)inpMat.size[1]);
            kernel1.set(9, (float)relu_slope);
            kernel1.set(10, ocl::KernelArg::PtrWriteOnly(outMat));
            ret = kernel1.run(2, global, NULL, false);
            if (!ret)
                return false;
        }
        return true;
    }
#endif

    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr) CV_OVERRIDE
    {
        CV_TRACE_FUNCTION();
        CV_TRACE_ARG_VALUE(name, "name", name.c_str());

        CV_OCL_RUN(IS_DNN_OPENCL_TARGET(preferableTarget),
                   forward_ocl(inputs_arr, outputs_arr, internals_arr))

        if (inputs_arr.depth() == CV_16S)
        {
            forward_fallback(inputs_arr, outputs_arr, internals_arr);
            return;
        }

        std::vector<Mat> inputs, outputs, internals;
        inputs_arr.getMatVector(inputs);
        outputs_arr.getMatVector(outputs);
        internals_arr.getMatVector(internals);

        for (size_t inpIdx = 0; inpIdx < inputs.size(); inpIdx++)
        {
            Mat &inpBlob = inputs[inpIdx];
            Mat &outBlob = outputs[inpIdx];

            int splitDim = (acrossChannels) ? 1 : 2;
            int i, newRows = 1;
            for( i = 0; i < splitDim; i++ )
                newRows *= inpBlob.size[i];

            Mat inpMat = inpBlob.reshape(1, newRows);
            Mat outMat = outBlob.reshape(1, newRows);

            if ( inpBlob.total() == newRows )
            {
                // MVN is applied to single values at an every row.
                if (shift.empty())
                {
                    outBlob.setTo(0);
                }
                else
                {
                    for ( i = 0; i < newRows; i++ )
                    {
                        outMat.row(i).setTo(((float*)shift.data)[i]);
                    }
                }
                return;
            }

            Scalar mean, dev;
            for ( i = 0; i < newRows; i++)
            {
                Mat inpRow = inpMat.row(i);
                Mat outRow = outMat.row(i);
                float weight = 1.f;
                float bias = 0.f;
                if (fuse_batch_norm)
                {
                    weight = i < scale.cols ? ((float*)scale.data)[i] : weight;
                    bias = i < shift.cols ? ((float*)shift.data)[i] : bias;
                }
                cv::meanStdDev(inpRow, mean, (normVariance) ? dev : noArray());
                double alpha = 1;
                if (normVariance)
                {
                    alpha = 1 / std::sqrt(eps + dev[0]*dev[0]);
                }
                double normalizationScale = 1.0;
                double normalizationShift = 0.0;
                if (fuse_batch_norm)
                {
                    normalizationScale = alpha * weight;
                    normalizationShift = -mean[0] * normalizationScale + bias;
                }
                else
                {
                    normalizationScale = alpha;
                    normalizationShift = -mean[0] * alpha;
                }
                inpRow.convertTo(outRow, outRow.type(), normalizationScale, normalizationShift);
            }
        }
    }

#ifdef HAVE_DNN_IE_NN_BUILDER_2019
    virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
    {
        InferenceEngine::Builder::MVNLayer ieLayer(name);
        ieLayer.setAcrossChannels(acrossChannels);
        ieLayer.setNormalize(normVariance);
        ieLayer.setEpsilon(eps);
        return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
    }
#endif  // HAVE_DNN_IE_NN_BUILDER_2019

#ifdef HAVE_DNN_NGRAPH
    virtual Ptr<BackendNode> initNgraph(const std::vector<Ptr<BackendWrapper> >& inputs,
                                        const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE
    {
        auto& ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
#if INF_ENGINE_VER_MAJOR_LE(INF_ENGINE_RELEASE_2021_2)
        auto mvn = std::make_shared<ngraph::op::MVN>(ieInpNode, acrossChannels, normVariance, eps);
#else
        int64_t start_axis = acrossChannels ? 1 : 2;
        std::vector<int64_t> axes_v(ieInpNode->get_shape().size() - start_axis);
        std::iota(axes_v.begin(), axes_v.end(), start_axis);
        auto axes = std::make_shared<ngraph::op::Constant>(ngraph::element::i64, ngraph::Shape{axes_v.size()}, axes_v.data());
        auto mvn = std::make_shared<ngraph::op::v6::MVN>(ieInpNode, axes, normVariance, eps, ngraph::op::MVNEpsMode::INSIDE_SQRT);
#endif
        return Ptr<BackendNode>(new InfEngineNgraphNode(mvn));
    }
#endif  // HAVE_DNN_NGRAPH

#ifdef HAVE_CUDA
    Ptr<BackendNode> initCUDA(
        void *context_,
        const std::vector<Ptr<BackendWrapper>>& inputs,
        const std::vector<Ptr<BackendWrapper>>& outputs
    ) override
    {
        auto context = reinterpret_cast<csl::CSLContext*>(context_);

        cuda4dnn::MVNConfiguration config;
        config.split_axis = acrossChannels ? 1 : 2;
        config.normalize_variance = normVariance;
        config.epsilon = eps;
        config.input_shapes.resize(inputs.size());
        for (int i = 0; i < inputs.size(); i++)
        {
            auto wrapper = inputs[i].dynamicCast<CUDABackendWrapper>();
            auto shape = wrapper->getShape();
            config.input_shapes[i].assign(std::begin(shape), std::end(shape));
        }

        return make_cuda_node<cuda4dnn::MVNOp>(preferableTarget, std::move(context->stream), config);
    }
#endif

    virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
                           const std::vector<MatShape> &outputs) const CV_OVERRIDE
    {
        CV_UNUSED(outputs); // suppress unused variable warning
        long flops = 0;
        for(int i = 0; i < inputs.size(); i++)
        {
            flops += 6*total(inputs[i]) + 3*total(inputs[i], 0, normVariance ? 2 : 1);
        }
        return flops;
    }
};

Ptr<MVNLayer> MVNLayer::create(const LayerParams& params)
{
    return Ptr<MVNLayer>(new MVNLayerImpl(params));
}

}
}