opencv/modules/dnn/src/layers/softmax_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_cuda.hpp"
#include "../op_halide.hpp"
#include "../op_inf_engine.hpp"
#include "../ie_ngraph.hpp"
#include "../op_vkcom.hpp"
#include "../op_webnn.hpp"

#include <algorithm>
#include <stdlib.h>
#include <opencv2/core/utils/logger.hpp>
using std::max;

#ifdef HAVE_OPENCL
#include "opencl_kernels_dnn.hpp"
using namespace cv::dnn::ocl4dnn;
#endif

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

namespace cv
{
namespace dnn
{

class SoftMaxLayerImpl CV_FINAL : public SoftmaxLayer
{
public:

    SoftMaxLayerImpl(const LayerParams& params)
    {
        axisRaw = params.get<int>("axis", 1);
        logSoftMax = params.get<bool>("log_softmax", false);
        setParamsFrom(params);
    }

#ifdef HAVE_OPENCL
    Ptr<OCL4DNNSoftmax<float> > softmaxOp;
#endif

    bool getMemoryShapes(const std::vector<MatShape> &inputs,
                         const int requiredOutputs,
                         std::vector<MatShape> &outputs,
                         std::vector<MatShape> &internals) const CV_OVERRIDE
    {
        bool inplace = Layer::getMemoryShapes(inputs, requiredOutputs, outputs, internals);
        MatShape shape = inputs[0];
        int cAxis = normalize_axis(axisRaw, shape.size());
        shape[cAxis] = 1;
        internals.assign(1, shape);
        return inplace;
    }

    virtual bool supportBackend(int backendId) CV_OVERRIDE
    {
#ifdef HAVE_INF_ENGINE
        if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)
            return true;
#endif
#ifdef HAVE_WEBNN
        if (backendId == DNN_BACKEND_WEBNN) {
            // TODO: support logSoftMax
            if (logSoftMax)
            {
                CV_LOG_WARNING(NULL, "logSoftMax is not supported by WebNN backend.")
            }
            return !logSoftMax;
        }
#endif
        return backendId == DNN_BACKEND_OPENCV ||
               backendId == DNN_BACKEND_CUDA ||
               (backendId == DNN_BACKEND_HALIDE && haveHalide() && axisRaw == 1) ||
               (backendId == DNN_BACKEND_VKCOM && haveVulkan());
    }

#ifdef HAVE_OPENCL
    virtual void finalize(const std::vector<Mat*> &inputs, std::vector<Mat> &outputs) CV_OVERRIDE
    {
        softmaxOp.release();
    }

    bool forward_ocl(InputArrayOfArrays inputs_, OutputArrayOfArrays outputs_, OutputArrayOfArrays internals_)
    {
        std::vector<UMat> inputs;
        std::vector<UMat> outputs;
        std::vector<UMat> internals;

        bool use_half = (inputs_.depth() == CV_16S);
        inputs_.getUMatVector(inputs);
        outputs_.getUMatVector(outputs);
        internals_.getUMatVector(internals);

        UMat& src = inputs[0];
        UMat& dstMat = outputs[0];
        int axis = normalize_axis(axisRaw, src.dims);

        if (softmaxOp.empty())
        {
            OCL4DNNSoftmaxConfig config;
            config.in_shape = shape(inputs[0]);
            config.axis = axis;
            config.channels = inputs[0].size[axis];
            config.logsoftmax = logSoftMax;
            config.use_half = use_half;

            softmaxOp = Ptr<OCL4DNNSoftmax<float> >(new OCL4DNNSoftmax<float>(config));
        }

        if (softmaxOp->Forward(src, dstMat))
            return true;

        UMat& bufMat = internals[0];
        MatShape s = shape(src);
        size_t outerSize = total(s, 0, axis);
        size_t channels = src.size[axis];
        size_t innerSize = total(s, axis + 1);

        String buildOpts = format("-DT=%s", use_half ? "half" : "float");
        ocl::Kernel kmax, ksub, ksum, kdiv;

        if (!kmax.create("kernel_channel_max", ocl::dnn::softmax_oclsrc, buildOpts))
            return false;

        if (!ksub.create("kernel_channel_subtract", ocl::dnn::softmax_oclsrc, buildOpts))
            return false;

        if (!ksum.create("kernel_channel_sum", ocl::dnn::softmax_oclsrc, buildOpts))
            return false;

        if (logSoftMax) buildOpts += " -DLOG_SOFTMAX ";
        if (!kdiv.create("kernel_channel_div", ocl::dnn::softmax_oclsrc, buildOpts))
            return false;

        size_t bufSize = internals[0].total();
        size_t totalSize = src.total();

        size_t internal_globalSize[1] = { bufSize };
        size_t total_globalSize[1] = { totalSize };

        kmax.args((int)outerSize, (int)channels, (int)innerSize,
                  ocl::KernelArg::PtrReadOnly(src), ocl::KernelArg::PtrReadWrite(bufMat));
        if (!kmax.run(1, internal_globalSize, NULL, false))
            return false;

        ksub.args((int)totalSize, (int)outerSize, (int)channels, (int)innerSize,
                  ocl::KernelArg::PtrReadOnly(bufMat),
                  ocl::KernelArg::PtrReadOnly(src), ocl::KernelArg::PtrWriteOnly(dstMat));
        if (!ksub.run(1, total_globalSize, NULL, false))
            return false;

        ksum.args((int)outerSize, (int)channels, (int)innerSize,
                  ocl::KernelArg::PtrReadOnly(dstMat), ocl::KernelArg::PtrReadWrite(bufMat));
        if (!ksum.run(1, internal_globalSize, NULL, false))
            return false;

        kdiv.args((int)totalSize, (int)outerSize, (int)channels, (int)innerSize,
                  ocl::KernelArg::PtrReadOnly(bufMat), ocl::KernelArg::PtrReadWrite(dstMat));
        if (!kdiv.run(1, total_globalSize, NULL, false))
            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);

        const Mat &src = inputs[0];
        Mat &dst = outputs[0];

        int axis = normalize_axis(axisRaw, src.dims);
        size_t outerSize = src.total(0, axis), channels = src.size[axis],
                innerSize = src.total(axis + 1);

        CV_Assert(src.type() == CV_32F);
        CV_Assert(src.isContinuous() && dst.isContinuous());

        const float *srcPtr = src.ptr<float>();
        float *dstPtr = dst.ptr<float>();
        float *bufPtr = internals[0].ptr<float>();

        size_t outerStep = src.total(axis);
        size_t cnStep = src.total(axis + 1);

        //compute max along axis
        for (size_t outerDim = 0; outerDim < outerSize; outerDim++)
        {
            size_t srcOffset = outerDim * outerStep;
            size_t bufOffset = outerDim * cnStep;

            memcpy(bufPtr + bufOffset, srcPtr + srcOffset, innerSize * sizeof(float));

            for (size_t cnDim = 1; cnDim < channels; cnDim++)
            {
                for (size_t i = 0; i < innerSize; i++)
                    bufPtr[bufOffset + i] = std::max(bufPtr[bufOffset + i], srcPtr[srcOffset + cnDim * cnStep + i]);
            }
        }

        //subtract max
        for (size_t outerDim = 0; outerDim < outerSize; outerDim++)
        {
            size_t srcOffset = outerDim * outerStep;
            size_t bufOffset = outerDim * cnStep;

            for (size_t cnDim = 0; cnDim < channels; cnDim++)
            {
                const int offset = srcOffset + cnDim * cnStep;
                for (size_t i = 0; i < innerSize; i++)
                    dstPtr[offset + i] = srcPtr[offset + i] - bufPtr[bufOffset + i];
            }
        }

        cv::exp(dst, dst);

        for (size_t outerDim = 0; outerDim < outerSize; outerDim++)
        {
            size_t srcOffset = outerDim * outerStep;
            size_t bufOffset = outerDim * cnStep;

            //sum exp along axis
            for (size_t i = 0; i < innerSize; i++)
                bufPtr[bufOffset + i] = 0.f;

            for (size_t cnDim = 0; cnDim < channels; cnDim++)
            {
                const int offset = srcOffset + cnDim * cnStep;
                for (size_t i = 0; i < innerSize; i++)
                    bufPtr[bufOffset + i] += dstPtr[offset + i];
            }

            //divide by computed sum
            for (size_t cnDim = 0; cnDim < channels; cnDim++)
            {
                const int offset = srcOffset + cnDim * cnStep;
                for (size_t i = 0; i < innerSize; i++)
                    dstPtr[offset + i] /= bufPtr[bufOffset + i];
            }
            if (logSoftMax)
            {
                for (size_t cnDim = 0; cnDim < channels; cnDim++)
                {
                    const int offset = srcOffset + cnDim * cnStep;
                    for (size_t i = 0; i < innerSize; i++)
                        dstPtr[offset + i] = log(dstPtr[offset + i]);
                }
            }
        }
    }

#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_);

        auto input_wrapper = inputs[0].dynamicCast<CUDABackendWrapper>();
        auto channel_axis = normalize_axis(axisRaw, input_wrapper->getRank());
        return make_cuda_node<cuda4dnn::SoftmaxOp>(preferableTarget, std::move(context->cudnn_handle), channel_axis, logSoftMax);
    }
#endif

    virtual Ptr<BackendNode> initVkCom(const std::vector<Ptr<BackendWrapper> > &inputs) CV_OVERRIDE
    {
#ifdef HAVE_VULKAN
        vkcom::Tensor in = VkComTensor(inputs[0]);
        int cAxis = normalize_axis(axisRaw, in.dimNum());
        std::shared_ptr<vkcom::OpBase> op(new vkcom::OpSoftmax(cAxis, logSoftMax));
        return Ptr<BackendNode>(new VkComBackendNode(inputs, op));
#endif  // HAVE_VULKAN
        return Ptr<BackendNode>();
    }


    virtual Ptr<BackendNode> initHalide(const std::vector<Ptr<BackendWrapper> > &inputs) CV_OVERRIDE
    {
#ifdef HAVE_HALIDE
        Halide::Buffer<float> inputBuffer = halideBuffer(inputs[0]);
        int inW, inH, inC, inN;
        getCanonicalSize(inputBuffer, &inW, &inH, &inC, &inN);

        if (inW != 1 || inH != 1)
            CV_Error(cv::Error::StsNotImplemented,
                     "Halide backend for SoftMax with spatial size "
                     "more than 1x1 is not implemented");

        Halide::Var x("x"), y("y"), c("c"), n("n");
        Halide::Func top = (name.empty() ? Halide::Func() : Halide::Func(name));

        Halide::Func expInput("expInput");
        Halide::RDom r(0, inW, 0, inH, 0, inC);
        expInput(x, y, c, n) = exp(inputBuffer(x, y, c, n));
        Halide::Expr globalSum = sum(expInput(r.x, r.y, r.z, n));
        top(x, y, c, n) = expInput(x, y, c, n) / globalSum;
        return Ptr<BackendNode>(new HalideBackendNode(top));
#endif  // HAVE_HALIDE
        return Ptr<BackendNode>();
    }


#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;
        int axis = normalize_axis(axisRaw, ieInpNode->get_shape().size());
        auto softmax = std::make_shared<ngraph::op::v1::Softmax>(ieInpNode, axis);
        if (logSoftMax)
            return Ptr<BackendNode>(new InfEngineNgraphNode(std::make_shared<ngraph::op::v0::Log>(softmax)));

        return Ptr<BackendNode>(new InfEngineNgraphNode(softmax));
    }
#endif  // HAVE_DNN_NGRAPH

    virtual bool tryQuantize(const std::vector<std::vector<float> > &scales,
                             const std::vector<std::vector<int> > &zeropoints, LayerParams& params) CV_OVERRIDE
    {
        float inpScale = scales[0][0];
        Mat lookUpTable(1, 256, CV_32F);
        float* table = lookUpTable.ptr<float>();
        for (int i = -128; i < 128; i++)
        {
            float x = inpScale*(i - 127); // ensures exp(x) is always between (0, 1)
            table[i+128] = std::exp(x);
        }
        params.blobs.clear();
        params.blobs.push_back(lookUpTable);
        params.set("input_scale", inpScale);
        params.set("input_zeropoint", zeropoints[0][0]);
        return true;
    }

#ifdef HAVE_WEBNN
    virtual Ptr<BackendNode> initWebnn(const std::vector<Ptr<BackendWrapper> >& inputs, const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE
    {
        Ptr<WebnnBackendNode> node = nodes[0].dynamicCast<WebnnBackendNode>();
        auto& webnnInpOperand = node->operand;
        auto& webnnGraphBuilder = node->net->builder;
        auto operand = webnnGraphBuilder.Softmax(webnnInpOperand);
        return Ptr<BackendNode>(new WebnnBackendNode(operand));
    }

#endif

    int64 getFLOPS(const std::vector<MatShape> &inputs,
                  const std::vector<MatShape> &outputs) const CV_OVERRIDE
    {
        CV_UNUSED(outputs); // suppress unused variable warning
        int64 flops = 0;

        for (int i = 0; i < inputs.size(); i++)
        {
            flops += 4*total(inputs[i]);
        }

        return flops;
    }

    int axisRaw;
};

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

}
}