modules/dnn/src/layers/convolution_layer.cpp

   1 /*M///////////////////////////////////////////////////////////////////////////////////////
   2 //
   3 //  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
   4 //
   5 //  By downloading, copying, installing or using the software you agree to this license.
   6 //  If you do not agree to this license, do not download, install,
   7 //  copy or use the software.
   8 //
   9 //
  10 //                           License Agreement
  11 //                For Open Source Computer Vision Library
  12 //
  13 // Copyright (C) 2013, OpenCV Foundation, all rights reserved.
  14 // Copyright (C) 2017, Intel Corporation, all rights reserved.
  15 // Third party copyrights are property of their respective owners.
  16 //
  17 // Redistribution and use in source and binary forms, with or without modification,
  18 // are permitted provided that the following conditions are met:
  19 //
  20 //   * Redistribution's of source code must retain the above copyright notice,
  21 //     this list of conditions and the following disclaimer.
  22 //
  23 //   * Redistribution's in binary form must reproduce the above copyright notice,
  24 //     this list of conditions and the following disclaimer in the documentation
  25 //     and/or other materials provided with the distribution.
  26 //
  27 //   * The name of the copyright holders may not be used to endorse or promote products
  28 //     derived from this software without specific prior written permission.
  29 //
  30 // This software is provided by the copyright holders and contributors "as is" and
  31 // any express or implied warranties, including, but not limited to, the implied
  32 // warranties of merchantability and fitness for a particular purpose are disclaimed.
  33 // In no event shall the Intel Corporation or contributors be liable for any direct,
  34 // indirect, incidental, special, exemplary, or consequential damages
  35 // (including, but not limited to, procurement of substitute goods or services;
  36 // loss of use, data, or profits; or business interruption) however caused
  37 // and on any theory of liability, whether in contract, strict liability,
  38 // or tort (including negligence or otherwise) arising in any way out of
  39 // the use of this software, even if advised of the possibility of such damage.
  40 //
  41 //M*/
  42
  43 #include "../precomp.hpp"
  44 #include "layers_common.hpp"
  45 #include "../op_halide.hpp"
  46 #include "../op_inf_engine.hpp"
  47 #include "opencv2/core/hal/hal.hpp"
  48 #include "opencv2/core/hal/intrin.hpp"
  49 #include <iostream>
  50 #include <numeric>
  51
  52 #ifdef HAVE_OPENCL
  53 #include "opencl_kernels_dnn.hpp"
  54 using namespace cv::dnn::ocl4dnn;
  55 #endif
  56
  57 namespace cv
  58 {
  59 namespace dnn
  60 {
  61
  62 class BaseConvolutionLayerImpl : public ConvolutionLayer
  63 {
  64 public:
  65     bool fusedWeights, fusedBias;
  66     std::vector<double> weightsMultipliers;
  67     BaseConvolutionLayerImpl(const LayerParams &params)
  68     {
  69         setParamsFrom(params);
  70         getConvolutionKernelParams(params, kernel_size, pads_begin, pads_end, strides, dilations, padMode, adjust_pads);
  71
  72         numOutput = params.get<int>("num_output");
  73         int ngroups = params.get<int>("group", 1);
  74         CV_Assert(numOutput % ngroups == 0);
  75
  76         if (kernel_size.size() == 2) {
  77             kernel = Size(kernel_size[1], kernel_size[0]);
  78             stride = Size(strides[1], strides[0]);
  79             for (int i = 0; i < pads_begin.size(); i++) {
  80                 if (pads_begin[i] != pads_end[i])
  81                     CV_Error(Error::StsNotImplemented, "Unsupported asymmetric padding in convolution layer");
  82             }
  83             pad = Size(pads_begin[1], pads_begin[0]);
  84             dilation = Size(dilations[1], dilations[0]);
  85
  86             adjustPad.height = adjust_pads[0];
  87             adjustPad.width = adjust_pads[1];
  88         }
  89
  90         for (int i = 0; i < adjust_pads.size(); i++) {
  91             CV_Assert(adjust_pads[i] < strides[i]);
  92         }
  93
  94         fusedWeights = false;
  95         fusedBias = false;
  96     }
  97
  98     virtual void finalize(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr) CV_OVERRIDE
  99     {
 100         std::vector<Mat> inputs, outputs;
 101         inputs_arr.getMatVector(inputs);
 102         outputs_arr.getMatVector(outputs);
 103
 104         CV_Assert(inputs.size() > 0);
 105
 106         CV_Assert(blobs.size() == 1 || blobs.size() == 2);
 107         CV_Assert(inputs[0].dims == outputs[0].dims);
 108         CV_Assert(blobs[0].dims == kernel_size.size() + 2);
 109         for (int i = 0; i < kernel_size.size(); i++) {
 110             CV_Assert(blobs[0].size[i + 2] == kernel_size[i]);
 111         }
 112
 113         const Mat &input = inputs[0];
 114         CV_Assert((input.dims == 4 || input.dims == 5) && (input.type() == CV_32F || input.type() == CV_16S));
 115         for (size_t i = 0; i < inputs.size(); i++)
 116         {
 117             CV_Assert(inputs[i].type() == input.type());
 118             CV_Assert((inputs[i].dims == 4 || inputs[i].dims == 5) && inputs[i].size[1] == input.size[1]);
 119             for (int j = 0; j < inputs[i].dims; j++) {
 120                 CV_Assert(inputs[i].size[j] == input.size[j]);
 121             }
 122         }
 123
 124         std::vector<int> inpShape;
 125         std::vector<int> outShape;
 126         for (int i = 2; i < inputs[0].dims; i++) {
 127             inpShape.push_back(inputs[0].size[i]);
 128             outShape.push_back(outputs[0].size[i]);
 129         }
 130         getConvPoolPaddings(inpShape, kernel_size, strides, padMode, pads_begin, pads_end);
 131         if (pads_begin.size() == 2) {
 132             for (int i = 0; i < pads_begin.size(); i++) {
 133                 if (pads_begin[i] != pads_end[i])
 134                     CV_Error(Error::StsNotImplemented, "Unsupported asymmetric padding in convolution layer");
 135             }
 136             pad = Size(pads_begin[1], pads_begin[0]);
 137         }
 138         fusedWeights = false;
 139         fusedBias = false;
 140     }
 141
 142     bool hasBias() const
 143     {
 144         return blobs.size() >= 2;
 145     }
 146
 147     virtual MatShape computeColRowShape(const MatShape &inpShape, const MatShape &outShape) const = 0;
 148     bool is1x1() const
 149     {
 150         return (kernel.height == 1 && kernel.width == 1) &&
 151                (stride.height == 1 && stride.width == 1) &&
 152                (dilation.height == 1 && dilation.width == 1);
 153     }
 154
 155     virtual bool tryFuse(Ptr<Layer>& top) CV_OVERRIDE
 156     {
 157         Mat w, b;
 158         top->getScaleShift(w, b);
 159         if (!w.empty() || !b.empty())
 160         {
 161             fuseWeights(w, b);
 162             fusedWeights = fusedWeights || !w.empty();
 163             fusedBias = fusedBias || (hasBias() && !w.empty()) || !b.empty();
 164             return true;
 165         }
 166         return false;
 167     }
 168
 169     virtual void fuseWeights(const Mat& w_, const Mat& b_) = 0;
 170
 171     virtual void applyHalideScheduler(Ptr<BackendNode>& node,
 172                                       const std::vector<Mat*> &inputs,
 173                                       const std::vector<Mat> &outputs,
 174                                       int targetId) const CV_OVERRIDE
 175     {
 176 #ifdef HAVE_HALIDE
 177         if (targetId != DNN_TARGET_CPU)
 178         {
 179             Layer::applyHalideScheduler(node, inputs, outputs, targetId);
 180             return;
 181         }
 182         Halide::Var x("x"), y("y"), c("c"), n("n"), tile("tile"), yi("yi"), yo("yo"), co("co"), ci("ci");
 183         Halide::Func& top = node.dynamicCast<HalideBackendNode>()->funcs[1];
 184         Halide::Func& padded_input = node.dynamicCast<HalideBackendNode>()->funcs[0];
 185
 186         int outW, outH, outC, outN;
 187         getCanonicalSize(outputs[0].size, &outW, &outH, &outC, &outN);
 188
 189         if (outW == 1 || outH <= 2)
 190             return;
 191
 192         if (is1x1() || outC <= 16)
 193             top.reorder(x, c, y)
 194                .split(y, yo, yi, 2)
 195                .fuse(yo, n, tile)
 196                .parallel(tile)
 197                .unroll(yi)
 198                .vectorize(x, outW >= 16 ? 16 : outW);
 199         else
 200             top.reorder(x, c, y)
 201                .split(y, yo, yi, 2)
 202                .split(c, co, ci, 16)
 203                .fuse(yo, co, tile).fuse(n, tile, tile)
 204                .parallel(tile)
 205                .unroll(yi)
 206                .vectorize(x, outW >= 16 ? 16 : outW);
 207         padded_input.compute_at(top, yi);
 208 #endif  // HAVE_HALIDE
 209     }
 210 };
 211
 212
 213 #define IS_POWER_LAYER(layer) \
 214             (!layer.empty() && !layer->type.compare("Power"))
 215 //TODO: simultaneously convolution and bias addition for cache optimization
 216 class ConvolutionLayerImpl CV_FINAL : public BaseConvolutionLayerImpl
 217 {
 218 public:
 219     enum { VEC_ALIGN = 8, DFT_TYPE = CV_32F };
 220     Mat weightsMat;
 221     std::vector<float> biasvec;
 222     std::vector<float> reluslope;
 223     Ptr<ActivationLayer> activ;
 224
 225 #ifdef HAVE_OPENCL
 226     Ptr<OCL4DNNConvSpatial<float> > convolutionOp;
 227     std::vector<UMat> umat_blobs;
 228     bool newActiv;
 229     ocl4dnnFusedActiv_t activType;
 230     float power;
 231 #endif
 232     ConvolutionLayerImpl(const LayerParams &params) : BaseConvolutionLayerImpl(params)
 233     {
 234 #ifdef HAVE_OPENCL
 235         newActiv = false;
 236         activType = OCL4DNN_CONV_FUSED_ACTIV_NONE;
 237         power = 0.f;
 238 #endif
 239     }
 240
 241     MatShape computeColRowShape(const MatShape &inpShape, const MatShape &outShape) const CV_OVERRIDE
 242     {
 243         Size out(outShape[3], outShape[2]);
 244         int inpGroupCn = blobs[0].size[1];
 245         int ksize = inpGroupCn * kernel.height * kernel.width;
 246         return shape(out.area(), ksize);
 247     }
 248
 249     virtual bool supportBackend(int backendId) CV_OVERRIDE
 250     {
 251 #ifdef HAVE_INF_ENGINE
 252         if (backendId == DNN_BACKEND_INFERENCE_ENGINE)
 253         {
 254             if (kernel_size.size() == 3)
 255                 return preferableTarget == DNN_TARGET_CPU;
 256             return (preferableTarget != DNN_TARGET_MYRIAD || dilation.width == dilation.height);
 257         }
 258         else
 259 #endif
 260             return (kernel_size.size() == 3 && preferableTarget == DNN_TARGET_CPU && backendId == DNN_BACKEND_OPENCV) ||
 261                    (kernel_size.size() == 2 && (backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE));
 262     }
 263
 264     bool getMemoryShapes(const std::vector<MatShape> &inputs,
 265                          const int requiredOutputs,
 266                          std::vector<MatShape> &outputs,
 267                          std::vector<MatShape> &internals) const CV_OVERRIDE
 268     {
 269         CV_Assert(blobs.size() != 0);
 270         CV_Assert(!hasBias() || blobs[1].total() == (size_t)blobs[0].size[0]);
 271         CV_Assert(inputs.size() == (size_t)1);
 272
 273         internals.clear();
 274
 275         CV_Assert(inputs.size() != 0);
 276         std::vector<int> inpShape(inputs[0].begin() + 2, inputs[0].end());
 277
 278         int outCn = blobs[0].size[0];
 279         std::vector<int> outShape;
 280         outShape.push_back(inputs[0][0]);
 281         outShape.push_back(outCn);
 282
 283         int inpCn = inputs[0][1];
 284         if (padMode.empty())
 285         {
 286             for (int i = 0; i < inpShape.size(); i++)
 287                 outShape.push_back((inpShape[i] + pads_begin[i] + pads_end[i] - dilations[i] * (kernel_size[i] - 1) - 1) / strides[i] + 1);
 288         }
 289         else
 290         {
 291             getConvPoolOutParams(inpShape, kernel_size, strides, padMode, dilations, outShape);
 292         }
 293
 294         int ngroups = inpCn / blobs[0].size[1];
 295         if (ngroups == 0 || ngroups * blobs[0].size[1] != inpCn)
 296             CV_Error(Error::StsError, format("Number of input channels should "
 297                      "be multiple of %d but got %d", blobs[0].size[1], inpCn));
 298         CV_Assert(ngroups > 0 && inpCn % ngroups == 0 && outCn % ngroups == 0);
 299
 300         outputs.resize(1, outShape);
 301
 302         return false;
 303     }
 304
 305     virtual void finalize(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr) CV_OVERRIDE
 306     {
 307         BaseConvolutionLayerImpl::finalize(inputs_arr, outputs_arr);
 308
 309         CV_Assert(!blobs.empty());
 310         const int outCn = blobs[0].size[0];
 311         // prepare weightsMat where each row is aligned and has enough zero padding on the right to
 312         // use vectorized (i.e. with intrinsics) loops without tail processing
 313         Mat wm = blobs[0].reshape(1, outCn);
 314         if( wm.step1() % VEC_ALIGN != 0 )
 315         {
 316             int newcols = (int)alignSize(wm.step1(), VEC_ALIGN);
 317             Mat wm_buffer = Mat(outCn, newcols, wm.type());
 318             Mat wm_padding = wm_buffer.colRange(wm.cols, newcols);
 319             wm_padding.setTo(Scalar::all(0.));
 320             Mat wm_aligned = wm_buffer.colRange(0, wm.cols);
 321             wm.copyTo(wm_aligned);
 322             wm = wm_aligned;
 323         }
 324         weightsMat = wm;
 325         weightsMultipliers.assign(outCn, 1.0);
 326
 327         Mat biasMat = hasBias() ? blobs[1].reshape(1, outCn) : Mat();
 328         biasvec.resize(outCn+2);
 329         if( biasMat.empty() )
 330         {
 331             for(int i = 0; i < outCn; i++ )
 332                 biasvec[i] = 0.f;
 333         }
 334         else
 335         {
 336             for(int i = 0; i < outCn; i++ )
 337                 biasvec[i] = biasMat.at<float>(i);
 338         }
 339 #ifdef HAVE_OPENCL
 340         convolutionOp.release();
 341 #endif
 342     }
 343
 344     bool setActivation(const Ptr<ActivationLayer>& layer) CV_OVERRIDE
 345     {
 346         if (!activ.empty() && !layer.empty())
 347             return false;
 348
 349         activ = layer;
 350         if (activ.empty())
 351             reluslope.clear();
 352 #ifdef HAVE_OPENCL
 353         newActiv = true;
 354         activType = OCL4DNN_CONV_FUSED_ACTIV_NONE;
 355
 356         if (IS_DNN_OPENCL_TARGET(preferableTarget))
 357         {
 358             Ptr<PowerLayer> activ_power = activ.dynamicCast<PowerLayer>();
 359             if (!activ_power.empty())
 360             {
 361                 if (activ_power->scale != 1.f || activ_power->shift != 0.f)
 362                 {
 363                     const int outCh = blobs[0].size[0];
 364                     fuseWeights(Mat(1, outCh, CV_32F, Scalar(activ_power->scale)),
 365                                 Mat(1, outCh, CV_32F, Scalar(activ_power->shift)));
 366                 }
 367
 368                 power = activ_power->power;
 369                 activType = OCL4DNN_CONV_FUSED_ACTIV_POWER;
 370             }
 371             Ptr<TanHLayer> activ_tanh = activ.dynamicCast<TanHLayer>();
 372             if (!activ_tanh.empty())
 373             {
 374                 activType = OCL4DNN_CONV_FUSED_ACTIV_TANH;
 375             }
 376         }
 377 #endif
 378         return !activ.empty();
 379     }
 380
 381     void fuseWeights(const Mat& w_, const Mat& b_) CV_OVERRIDE
 382     {
 383         // Convolution weights have OIHW data layout. Parameters fusion in case of
 384         // (conv(I) + b1 ) * w + b2
 385         // means to replace convolution's weights to [w*conv(I)] and bias to [b1 * w + b2]
 386         const int outCn = weightsMat.size[0];
 387         Mat w = w_.total() == 1 ? Mat(1, outCn, CV_32F, Scalar(w_.at<float>(0))) : w_;
 388         Mat b = b_.total() == 1 ? Mat(1, outCn, CV_32F, Scalar(b_.at<float>(0))) : b_;
 389         CV_Assert_N(!weightsMat.empty(), biasvec.size() == outCn + 2,
 390                     w.empty() || outCn == w.total(), b.empty() || outCn == b.total());
 391
 392         if (!w.empty())
 393         {
 394             // Keep origin weights unchanged.
 395             if (weightsMat.data == blobs[0].data)
 396                 weightsMat = weightsMat.clone();
 397
 398             Mat originWeights = blobs[0].reshape(1, outCn);
 399             for (int i = 0; i < outCn; ++i)
 400             {
 401                 double wi = w.at<float>(i);
 402                 weightsMultipliers[i] *= wi;
 403                 cv::multiply(originWeights.row(i), weightsMultipliers[i], weightsMat.row(i));
 404                 biasvec[i] *= wi;
 405             }
 406         }
 407
 408         if (!b.empty())
 409         {
 410             for (int i = 0; i < outCn; ++i)
 411                 biasvec[i] += b.at<float>(i);
 412         }
 413         biasvec[outCn] = biasvec[outCn+1] = biasvec[outCn-1];
 414     }
 415
 416     virtual Ptr<BackendNode> initHalide(const std::vector<Ptr<BackendWrapper> > &inputs) CV_OVERRIDE
 417     {
 418 #ifdef HAVE_HALIDE
 419         Halide::Buffer<float> inputBuffer = halideBuffer(inputs[0]);
 420
 421         const int inpCn = inputBuffer.channels();
 422         const int outCn = blobs[0].size[0];
 423         const int inpGroupCn = blobs[0].size[1];
 424         const int group = inpCn / inpGroupCn;
 425         const int outGroupCn = outCn / group;
 426
 427         Halide::Buffer<float> weights = wrapToHalideBuffer(blobs[0]);
 428
 429         Halide::Var x("x"), y("y"), c("c"), n("n");
 430         Halide::Func top = (name.empty() ? Halide::Func() : Halide::Func(name));
 431         Halide::Func padded_input(name + "_constant_exterior");
 432         if (pad.width || pad.height)
 433         {
 434             Halide::Func bounded =
 435                 Halide::BoundaryConditions::constant_exterior(inputBuffer, 0);
 436             padded_input(x, y, c, n) = bounded(x, y, c, n);
 437         }
 438         else
 439         {
 440             padded_input(x, y, c, n) = inputBuffer(x, y, c, n);
 441         }
 442
 443         Halide::RDom r(0, kernel.width, 0, kernel.height, 0, inpGroupCn);
 444         Halide::Expr kx = x * stride.width - pad.width + r.x * dilation.width;
 445         Halide::Expr ky = y * stride.height - pad.height + r.y * dilation.height;
 446         Halide::Expr kc = r.z;
 447         for (int i = 1; i < group; ++i)
 448         {
 449             kc = select(c < outGroupCn * i, kc, inpGroupCn * i + r.z);
 450         }
 451         Halide::Expr topExpr = sum(padded_input(kx, ky, kc, n) *
 452                                    weights(r.x, r.y, r.z, c));
 453         if (hasBias())
 454         {
 455             Halide::Buffer<float> bias = wrapToHalideBuffer(blobs[1], {outCn});
 456             topExpr += bias(c);
 457         }
 458         top(x, y, c, n) = topExpr;
 459         return Ptr<BackendNode>(new HalideBackendNode({ padded_input, top }));
 460 #endif  // HAVE_HALIDE
 461         return Ptr<BackendNode>();
 462     }
 463
 464 #ifdef HAVE_INF_ENGINE
 465     virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> > &inputs) CV_OVERRIDE
 466     {
 467         InferenceEngine::DataPtr input = infEngineDataNode(inputs[0]);
 468         std::vector<size_t> dims = input->getDims();
 469         CV_Assert(dims.size() == 4 || dims.size() == 5);
 470         const int inpCn = dims[1];
 471         const int outCn = blobs[0].size[0];
 472         const int inpGroupCn = blobs[0].size[1];
 473         const int group = inpCn / inpGroupCn;
 474         InferenceEngine::Layout layout = (dims.size() == 4) ? InferenceEngine::Layout::OIHW :
 475                                                               InferenceEngine::Layout::NCDHW;
 476
 477         auto ieWeights = wrapToInfEngineBlob(blobs[0], layout);
 478         if (fusedWeights)
 479         {
 480             if (weightsMat.isContinuous())
 481             {
 482                 Mat cvWeights = weightsMat.reshape(1, blobs[0].dims, blobs[0].size);
 483                 ieWeights = wrapToInfEngineBlob(cvWeights, layout);
 484             }
 485             else
 486             {
 487                 ieWeights = InferenceEngine::make_shared_blob<float>({
 488                                 InferenceEngine::Precision::FP32,
 489                                 ieWeights->getTensorDesc().getDims(), layout
 490                             });
 491                 ieWeights->allocate();
 492
 493                 Mat newWeights = infEngineBlobToMat(ieWeights).reshape(1, outCn);
 494                 Mat cvWeights = weightsMat.colRange(0, newWeights.cols);
 495                 cvWeights.copyTo(newWeights);
 496             }
 497         }
 498         InferenceEngine::Blob::Ptr ieBiases;
 499         if (hasBias() || fusedBias)
 500         {
 501             Mat biasesMat({outCn}, CV_32F, &biasvec[0]);
 502             ieBiases = wrapToInfEngineBlob(biasesMat, {(size_t)outCn}, InferenceEngine::Layout::C);
 503         }
 504
 505         InferenceEngine::Builder::ConvolutionLayer ieLayer(name);
 506
 507         ieLayer.setKernel(kernel_size);
 508         ieLayer.setStrides(strides);
 509         ieLayer.setDilation(dilations);
 510         ieLayer.setPaddingsBegin(pads_begin);
 511         ieLayer.setPaddingsEnd(pads_end);
 512         ieLayer.setGroup((size_t)group);
 513         ieLayer.setOutDepth((size_t)outCn);
 514
 515         InferenceEngine::Builder::Layer l = ieLayer;
 516         addConstantData("weights", ieWeights, l);
 517         if (ieBiases)
 518             addConstantData("biases", ieBiases, l);
 519
 520         if (!padMode.empty())
 521             l.getParameters()["auto_pad"] = padMode == "VALID" ? std::string("valid") : std::string("same_upper");
 522
 523         return Ptr<BackendNode>(new InfEngineBackendNode(l));
 524     }
 525 #endif  // HAVE_INF_ENGINE
 526
 527     class ParallelConv : public cv::ParallelLoopBody
 528     {
 529     public:
 530         enum { BLK_SIZE = 32, BLK_SIZE_CN = 64 };
 531
 532         const Mat* input_;
 533         const Mat* weights_;
 534         Mat* output_;
 535         int outShape[4]; // used only for conv2d
 536         std::vector<size_t> kernel_size, pads_begin, pads_end, strides, dilations;
 537         int ngroups_, nstripes_;
 538         std::vector<int> ofstab_;
 539         const std::vector<float>* biasvec_;
 540         const std::vector<float>* reluslope_;
 541         const ActivationLayer* activ_;
 542         bool is1x1_;
 543         bool useAVX;
 544         bool useAVX2;
 545         bool useAVX512;
 546
 547         ParallelConv()
 548             : input_(0), weights_(0), output_(0), ngroups_(0), nstripes_(0),
 549               biasvec_(0), reluslope_(0), activ_(0), is1x1_(false), useAVX(false), useAVX2(false), useAVX512(false)
 550         {}
 551
 552         static void run( const Mat& input, Mat& output, const Mat& weights,
 553                          const std::vector<float>& biasvec,
 554                          const std::vector<float>& reluslope,
 555                          const std::vector<size_t>& kernel_size, const std::vector<size_t>& strides,
 556                          const std::vector<size_t>& pads_begin, const std::vector<size_t>& pads_end,
 557                          const std::vector<size_t>& dilations,
 558                          const ActivationLayer* activ, int ngroups, int nstripes )
 559         {
 560             size_t karea = std::accumulate(kernel_size.begin(), kernel_size.end(),
 561                                            1, std::multiplies<size_t>());
 562             CV_Assert_N(
 563                        (input.dims == 4 || input.dims == 5) && (input.dims == output.dims),
 564                        input.size[0] == output.size[0],
 565                        weights.rows == output.size[1],
 566                        weights.cols == (input.size[1]/ngroups)*karea,
 567                        input.type() == output.type(),
 568                        input.type() == weights.type(),
 569                        input.type() == CV_32FC1,
 570                        input.isContinuous(),
 571                        output.isContinuous(),
 572                        biasvec.size() == (size_t)output.size[1]+2);
 573             ParallelConv p;
 574
 575             p.input_ = &input;
 576             p.weights_ = &weights;
 577             p.output_ = &output;
 578             for( int i = 0; i < 4; i++ ) p.outShape[i] = output.size[i];
 579             p.outShape[1] /= ngroups;
 580
 581             p.kernel_size = kernel_size; p.strides = strides; p.dilations = dilations;
 582             p.pads_begin = pads_begin; p.pads_end = pads_end;
 583
 584             p.ngroups_ = ngroups;
 585             p.nstripes_ = nstripes;
 586
 587             int inpCnAll = input.size[1];
 588             int depth = (input.dims == 5) ? input.size[2] : 1;
 589             int width = input.size[input.dims - 1];
 590             int height = input.size[input.dims - 2];
 591             int inpCn = inpCnAll / ngroups;
 592
 593             bool isConv2D = kernel_size.size() == 2;
 594
 595             p.is1x1_ = isConv2D && kernel_size[0] == 1 && kernel_size[1] == 1 &&
 596                        pads_begin[0] == 0  && pads_begin[1] == 0;
 597
 598             p.useAVX    = checkHardwareSupport(CPU_AVX)  && isConv2D;
 599             p.useAVX2   = checkHardwareSupport(CPU_AVX2) && isConv2D;
 600             p.useAVX512 = CV_CPU_HAS_SUPPORT_AVX512_SKX  && isConv2D;
 601
 602             int ncn = std::min(inpCn, (int)BLK_SIZE_CN);
 603
 604             int kernel_d = !isConv2D? kernel_size[0] : 1;
 605             int kernel_h = kernel_size[kernel_size.size() - 2];
 606             int kernel_w = kernel_size.back();
 607
 608             int dil_d = !isConv2D? dilations[0] : 1;
 609             int dil_h = dilations[dilations.size() - 2];
 610             int dil_w = dilations.back();
 611
 612             p.ofstab_.resize(karea * ncn);
 613             int* ofstab = &p.ofstab_[0];
 614
 615             if (isConv2D)
 616             {
 617                 for( int k = 0; k < ncn; k++ )
 618                     for( int k_r = 0; k_r < kernel_h; k_r++ )
 619                         for( int k_c = 0; k_c < kernel_w; k_c++ )
 620                             ofstab[(k*kernel_h + k_r)*kernel_w + k_c] =
 621                                    (k*height + k_r*dil_h)*width + k_c*dil_w;
 622             }
 623             else
 624             {
 625                 for( int k = 0; k < ncn; k++ )
 626                     for (int k_d = 0; k_d < kernel_d; k_d++)
 627                         for( int k_r = 0; k_r < kernel_h; k_r++ )
 628                             for( int k_c = 0; k_c < kernel_w; k_c++ )
 629                                 ofstab[(k*kernel_d*kernel_h + k_d*kernel_h + k_r)*kernel_w + k_c] =
 630                                        (k*depth*height + k_d*dil_d*height + k_r*dil_h)*width + k_c*dil_w;
 631             }
 632
 633             p.biasvec_ = &biasvec;
 634             p.reluslope_ = &reluslope;
 635             p.activ_ = p.reluslope_->empty() ? activ : 0;
 636
 637             parallel_for_(Range(0, nstripes), p, nstripes);
 638         }
 639
 640         virtual void operator ()(const Range &r0) const CV_OVERRIDE
 641         {
 642             const int valign = ConvolutionLayerImpl::VEC_ALIGN;
 643             int ngroups = ngroups_, batchSize = input_->size[0]*ngroups;
 644             bool isConv2D = input_->dims == 4;
 645
 646             int outW = output_->size[output_->dims - 1];
 647             int outH = output_->size[output_->dims - 2];
 648             int outCn = output_->size[1]/ngroups;
 649
 650             int depth = !isConv2D? input_->size[2] : 1;
 651             int height = input_->size[input_->dims - 2];
 652             int width = input_->size[input_->dims - 1];
 653             int inpCn = input_->size[1]/ngroups;
 654
 655             const int nstripes = nstripes_;
 656
 657             int kernel_d = !isConv2D? kernel_size[0] : 1;
 658             int kernel_h = kernel_size[kernel_size.size() - 2];
 659             int kernel_w = kernel_size.back();
 660             int karea = kernel_w*kernel_h*kernel_d;
 661
 662             int pad_d = !isConv2D? pads_begin[0] : 0;
 663             int pad_t = pads_begin[pads_begin.size() - 2];
 664             int pad_l = pads_begin.back();
 665
 666             int stride_d = !isConv2D? strides[0] : 0;
 667             int stride_h = strides[strides.size() - 2];
 668             int stride_w = strides.back();
 669
 670             int dilation_d = !isConv2D? dilations[0] : 1;
 671             int dilation_h = dilations[dilations.size() - 2];
 672             int dilation_w = dilations.back();
 673
 674             int i, j, k, d;
 675             size_t inpPlaneSize = input_->total(2);
 676             size_t outPlaneSize = output_->total(2);
 677             bool is1x1 = is1x1_;
 678
 679             int stripesPerSample;
 680             size_t stripeSize;
 681             Range r = r0;
 682
 683             if( nstripes >= batchSize*2 )
 684             {
 685                 stripesPerSample = nstripes/batchSize;
 686                 stripeSize = alignSize((outPlaneSize + stripesPerSample - 1)/stripesPerSample, valign);
 687                 stripeSize = std::min(stripeSize, outPlaneSize);
 688             }
 689             else
 690             {
 691                 stripesPerSample = 1;
 692                 int samplesPerStripe = std::max((batchSize + nstripes - 1)/nstripes, 1);
 693                 r.start *= samplesPerStripe;
 694                 r.end *= samplesPerStripe;
 695                 stripeSize = outPlaneSize;
 696             }
 697
 698             const float* data_inp0_ = input_->ptr<float>();
 699             const int* ofstab = &ofstab_[0];
 700             const float* wptr_orig_ = weights_->ptr<float>();
 701             size_t wstep = weights_->step1();
 702             const float* biasptr_ = &biasvec_->at(0);
 703             const float* reluptr_ = reluslope_->empty() ? 0 : &reluslope_->at(0);
 704             float* data_out0_ = output_->ptr<float>();
 705             size_t rowbufsz = (size_t)karea*BLK_SIZE_CN*BLK_SIZE;
 706             AutoBuffer<float> rowbuf0_(rowbufsz + valign);
 707             float* rowbuf0 = alignPtr(rowbuf0_.data(), (int)(valign*sizeof(float)));
 708
 709             // we clear the buffer once; ultimately, it lets us to avoid
 710             // tail processing after running the unrolled/vectorized loop.
 711             // the main idea is to make sure that the tail (a.k.a. padding) of each row
 712             // (i.e. the elements with indices between vsz=karea*ncn and vsz_a)
 713             // does not contain NaNs or Infs. Because the padding in the weights
 714             // matrix is explicitly initialized with 0's, we handle all other
 715             // cases nicely, i.e. we can skip expliciting re-initialization
 716             // of the padding - we just retain elements from the previous iteration
 717             // of the loop over channels (cn0).
 718             memset(rowbuf0, 0, rowbufsz*sizeof(rowbuf0[0]) );
 719
 720             for( int stripe = r.start; stripe < r.end; stripe++ )
 721             {
 722                 int subsampleIdx = stripe/stripesPerSample;
 723                 if( subsampleIdx >= batchSize )
 724                     break;
 725                 int stripeStart = (int)((stripe - subsampleIdx*stripesPerSample)*stripeSize);
 726                 int stripeEnd = (int)std::min(stripeStart + stripeSize, outPlaneSize);
 727                 const float* data_inp0 = data_inp0_ + subsampleIdx*inpPlaneSize*inpCn;
 728                 float* data_out0 = data_out0_ + subsampleIdx*outPlaneSize*outCn;
 729                 int startOutCn = (subsampleIdx % ngroups)*outCn;
 730                 const float* wptr_orig = wptr_orig_ + wstep*startOutCn;
 731                 const float* biasptr = biasptr_ + startOutCn;
 732
 733                 for( int cn0 = 0; cn0 < inpCn; cn0 += BLK_SIZE_CN )
 734                 {
 735                     int cn1 = std::min(cn0 + BLK_SIZE_CN, inpCn);
 736                     int ncn = cn1 - cn0, vsz = karea*ncn;
 737                     int vsz_a = (int)alignSize(vsz, valign);
 738                     const float* wptr = wptr_orig + cn0*karea;
 739                     // we apply [Channels][P]ReLU (if any) during the final pass only.
 740                     const float* relu = cn1 == inpCn && reluptr_ ? reluptr_ + startOutCn : 0;
 741
 742                     for( int ofs0 = stripeStart; ofs0 < stripeEnd; ofs0 += BLK_SIZE )
 743                     {
 744                         int ofs, ofs1 = std::min(ofs0 + BLK_SIZE, stripeEnd);
 745
 746                         int out_d = ofs0 / (outH * outW);
 747                         int out_i = (ofs0 - out_d * outH * outW) / outW;
 748                         int out_j = ofs0 % outW;
 749
 750                         // do im2row for a part of input tensor
 751                         float* rowbuf = rowbuf0;
 752
 753                         if (isConv2D)
 754                         {
 755                             for( ofs = ofs0; ofs < ofs1; out_j = 0, ++out_i )
 756                             {
 757                                 int delta = std::min(ofs1 - ofs, outW - out_j);
 758                                 int out_j1 = out_j + delta;
 759
 760                                 int in_i = out_i * stride_h - pad_t;
 761                                 int in_j = out_j * stride_w - pad_l;
 762                                 const float* imgptr = data_inp0 + (cn0*height + in_i)*width + in_j;
 763                                 ofs += delta;
 764
 765                                 // do im2row for a part of input tensor
 766                                 if( is1x1 )
 767                                 {
 768                                     for( ; out_j < out_j1; out_j++, rowbuf += vsz_a, imgptr += stride_w )
 769                                     {
 770                                         for( k = 0; k < vsz; k++ )
 771                                             rowbuf[k] = imgptr[k*inpPlaneSize];
 772                                     }
 773                                 }
 774                                 else
 775                                 {
 776                                     bool ok_i = 0 <= in_i && in_i < height - (kernel_h-1)*dilation_h;
 777                                     int i0 = std::max(0, (-in_i + dilation_h-1)/dilation_h);
 778                                     int i1 = std::min(kernel_h, (height - in_i + dilation_h-1)/dilation_h);
 779
 780                                     for( ; out_j < out_j1; out_j++, rowbuf += vsz_a, imgptr += stride_w, in_j += stride_w )
 781                                     {
 782                                         // this condition should be true for most of the tensor elements, i.e.
 783                                         // most of the time the kernel aperture is inside the tensor X-Y plane.
 784                                         if( ok_i && out_j + 2 <= out_j1 && 0 <= in_j && in_j + stride_w*2 <= width - (kernel_w-1)*dilation_w )
 785                                         {
 786                                             for( k = 0; k < vsz; k++ )
 787                                             {
 788                                                 int k1 = ofstab[k];
 789                                                 float v0 = imgptr[k1];
 790                                                 float v1 = imgptr[k1 + stride_w];
 791                                                 rowbuf[k] = v0;
 792                                                 rowbuf[k+vsz_a] = v1;
 793                                             }
 794                                             out_j++;
 795                                             rowbuf += vsz_a;
 796                                             imgptr += stride_w;
 797                                             in_j += stride_w;
 798                                         }
 799                                         else
 800                                         {
 801                                             int j0 = std::max(0, (-in_j + dilation_w-1)/dilation_w);
 802                                             int j1 = std::min(kernel_w, (width - in_j + dilation_w-1)/dilation_w);
 803
 804                                             // here some non-continuous sub-row of the row will not be
 805                                             // filled from the tensor; we need to make sure that the uncovered
 806                                             // elements are explicitly set to 0's. the easiest way is to
 807                                             // set all the elements to 0's before the loop.
 808                                             memset(rowbuf, 0, vsz*sizeof(rowbuf[0]));
 809                                             for( k = 0; k < ncn; k++ )
 810                                             {
 811                                                 for( i = i0; i < i1; i++ )
 812                                                 {
 813                                                     for( j = j0; j < j1; j++ )
 814                                                     {
 815                                                         int imgofs = k*(width*height) + i*(dilation_h*width) + j*dilation_w;
 816                                                         rowbuf[(k*kernel_h + i)*kernel_w + j] = imgptr[imgofs];
 817                                                     }
 818                                                 }
 819                                             }
 820                                         }
 821                                     }
 822                                 }
 823                             }
 824                         }
 825                         else
 826                         {
 827                             for( ofs = ofs0; ofs < ofs1; out_d += (out_i + 1) / outH, out_i = (out_i + 1) % outH, out_j = 0 )
 828                             {
 829                                 int delta = std::min(ofs1 - ofs, outW - out_j);
 830                                 int out_j1 = out_j + delta;
 831
 832                                 int in_d = out_d * stride_d - pad_d;
 833                                 int in_i = out_i * stride_h - pad_t;
 834                                 int in_j = out_j * stride_w - pad_l;
 835                                 const float* imgptr = data_inp0 + (cn0*depth*height + in_d*height + in_i)*width + in_j;
 836                                 ofs += delta;
 837
 838                                 int d0 = std::max(0, (-in_d + dilation_d - 1) / dilation_d);
 839                                 int d1 = std::min(kernel_d, (depth - in_d + dilation_d - 1) / dilation_d);
 840
 841                                 int i0 = std::max(0, (-in_i + dilation_h-1)/dilation_h);
 842                                 int i1 = std::min(kernel_h, (height - in_i + dilation_h-1)/dilation_h);
 843
 844                                 for( ; out_j < out_j1; out_j++, rowbuf += vsz_a, imgptr += stride_w, in_j += stride_w )
 845                                 {
 846                                     int j0 = std::max(0, (-in_j + dilation_w-1)/dilation_w);
 847                                     int j1 = std::min(kernel_w, (width - in_j + dilation_w-1)/dilation_w);
 848
 849                                     // here some non-continuous sub-row of the row will not be
 850                                     // filled from the tensor; we need to make sure that the uncovered
 851                                     // elements are explicitly set to 0's. the easiest way is to
 852                                     // set all the elements to 0's before the loop.
 853                                     memset(rowbuf, 0, vsz*sizeof(rowbuf[0]));
 854                                     for( k = 0; k < ncn; k++ )
 855                                     {
 856                                         for ( d = d0; d < d1; d++)
 857                                         {
 858                                             for( i = i0; i < i1; i++ )
 859                                             {
 860                                                 for( j = j0; j < j1; j++ )
 861                                                 {
 862                                                     int imgofs = k*(depth*width*height) + d*dilation_d*width*height + i*(dilation_h*width) + j*dilation_w;
 863                                                     rowbuf[(k*kernel_d*kernel_h + d*kernel_h + i)*kernel_w + j] = imgptr[imgofs];
 864                                                 }
 865                                             }
 866                                         }
 867                                     }
 868                                 }
 869                             }
 870                         }
 871
 872                         // now compute dot product of the weights
 873                         // and im2row-transformed part of the tensor
 874                         int bsz = ofs1 - ofs0;
 875                     #if CV_TRY_AVX512_SKX
 876                         /* AVX512 convolution requires an alignment of 16, and ROI is only there for larger vector sizes */
 877                         if(useAVX512)
 878                             opt_AVX512_SKX::fastConv(wptr, wstep, biasptr, rowbuf0, data_out0 + ofs0,
 879                                           outShape, bsz, vsz, vsz_a, relu, cn0 == 0);
 880                         else
 881                     #endif
 882                     #if CV_TRY_AVX2
 883                         if(useAVX2)
 884                             opt_AVX2::fastConv(wptr, wstep, biasptr, rowbuf0, data_out0 + ofs0,
 885                                           outShape, bsz, vsz, vsz_a, relu, cn0 == 0);
 886                         else
 887                     #endif
 888                     #if CV_TRY_AVX
 889                         if(useAVX)
 890                             opt_AVX::fastConv(wptr, wstep, biasptr, rowbuf0, data_out0 + ofs0,
 891                                          outShape, bsz, vsz, vsz_a, relu, cn0 == 0);
 892                         else
 893                     #endif
 894                         for( int i = 0; i < outCn; i += 2 )
 895                         {
 896                             const float* wptr0 = wptr + i*wstep;
 897                             const float* wptr1 = wptr0 + wstep;
 898                             float* outptr0 = data_out0 + ofs0 + i*outPlaneSize;
 899                             float* outptr1 = outptr0 + outPlaneSize;
 900                             float bias0 = biasptr[i], bias1 = biasptr[i+1];
 901                             float r0 = 1.f, r1 = 1.f;
 902
 903                             if( i+1 >= outCn )
 904                             {
 905                                 wptr1 = wptr0;
 906                                 outptr1 = outptr0;
 907                                 bias1 = bias0;
 908                             }
 909
 910                             if( relu )
 911                             {
 912                                 r0 = relu[i]; r1 = relu[i+1];
 913                                 if( i+1 >= outCn )
 914                                     r1 = r0;
 915                             }
 916
 917                             int j = 0;
 918                         #if CV_SIMD128
 919                             v_float32x4 vr0 = v_setall_f32(r0), vr1 = v_setall_f32(r1), z = v_setzero_f32();
 920
 921                             for( ; j <= bsz - 4; j += 4 )
 922                             {
 923                                 const float* rptr = rowbuf0 + j*vsz_a;
 924                                 v_float32x4 s0, s1;
 925
 926                                 if( cn0 == 0 )
 927                                 {
 928                                     s0 = v_setall_f32(bias0);
 929                                     s1 = v_setall_f32(bias1);
 930                                 }
 931                                 else
 932                                 {
 933                                     s0 = v_load(outptr0 + j);
 934                                     s1 = v_load(outptr1 + j);
 935                                 }
 936
 937                                 v_float32x4 vs00 = v_setzero_f32(), vs01 = v_setzero_f32(),
 938                                             vs02 = v_setzero_f32(), vs03 = v_setzero_f32(),
 939                                             vs10 = v_setzero_f32(), vs11 = v_setzero_f32(),
 940                                             vs12 = v_setzero_f32(), vs13 = v_setzero_f32();
 941                                 for( k = 0; k < vsz; k += 4, rptr += 4 )
 942                                 {
 943                                     v_float32x4 w0 = v_load_aligned(wptr0 + k), w1 = v_load_aligned(wptr1 + k);
 944                                     v_float32x4 r0 = v_load_aligned(rptr), r1 = v_load_aligned(rptr + vsz_a),
 945                                                 r2 = v_load_aligned(rptr + vsz_a*2), r3 = v_load_aligned(rptr + vsz_a*3);
 946
 947                                     vs00 += w0*r0;
 948                                     vs01 += w0*r1;
 949                                     vs02 += w0*r2;
 950                                     vs03 += w0*r3;
 951
 952                                     vs10 += w1*r0;
 953                                     vs11 += w1*r1;
 954                                     vs12 += w1*r2;
 955                                     vs13 += w1*r3;
 956                                 }
 957                                 s0 += v_reduce_sum4(vs00, vs01, vs02, vs03);
 958                                 s1 += v_reduce_sum4(vs10, vs11, vs12, vs13);
 959                                 if( relu )
 960                                 {
 961                                     s0 = v_select(s0 > z, s0, s0*vr0);
 962                                     s1 = v_select(s1 > z, s1, s1*vr1);
 963                                 }
 964
 965                                 v_store(outptr0 + j, s0);
 966                                 v_store(outptr1 + j, s1);
 967                             }
 968                         #endif
 969                             for( ; j < bsz; j++ )
 970                             {
 971                                 const float* rptr = rowbuf0 + j*vsz_a;
 972                                 float s00, s10;
 973
 974                                 if( cn0 == 0 )
 975                                 {
 976                                     s00 = bias0;
 977                                     s10 = bias1;
 978                                 }
 979                                 else
 980                                 {
 981                                     s00 = outptr0[j];
 982                                     s10 = outptr1[j];
 983                                 }
 984
 985                                 for( k = 0; k < vsz; k++ )
 986                                 {
 987                                     float r0 = rptr[k];
 988                                     s00 += wptr0[k]*r0;
 989                                     s10 += wptr1[k]*r0;
 990                                 }
 991                                 if( relu )
 992                                 {
 993                                     s00 = s00 > 0.f ? s00 : s00*r0;
 994                                     s10 = s10 > 0.f ? s10 : s10*r1;
 995                                 }
 996
 997                                 outptr0[j] = s00;
 998                                 outptr1[j] = s10;
 999                             }
1000                         }
1001                     }
1002                 }
1003
1004                 if( activ_ )
1005                     activ_->forwardSlice(data_out0 + stripeStart, data_out0 + stripeStart,
1006                                          (int)(stripeEnd - stripeStart),
1007                                          outPlaneSize, startOutCn, startOutCn + outCn);
1008             }
1009         }
1010     };
1011
1012 #ifdef HAVE_OPENCL
1013     bool forward_ocl(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
1014     {
1015         std::vector<UMat> inputs;
1016         std::vector<UMat> outputs;
1017
1018         bool use_half = (inps.depth() == CV_16S);
1019         inps.getUMatVector(inputs);
1020         outs.getUMatVector(outputs);
1021
1022         CV_Assert(outputs.size() == 1);
1023         for (int i = 0; i < inputs.size(); ++i)
1024             CV_Assert(inputs[i].u != outputs[0].u);
1025
1026         if (umat_blobs.empty())
1027         {
1028             size_t n = blobs.size();
1029             umat_blobs.resize(n);
1030             for (size_t i = 0; i < n; i++)
1031             {
1032                 blobs[i].copyTo(umat_blobs[i]);
1033             }
1034         }
1035
1036         if (convolutionOp.empty())
1037         {
1038             OCL4DNNConvConfig config;
1039             config.in_shape = shape(inputs[0]);
1040             config.out_shape = shape(outputs[0]);
1041             config.kernel = kernel;
1042             config.pad = pad;
1043             config.stride = stride;
1044             config.dilation = dilation;
1045             config.group = inputs[0].size[1] / umat_blobs[0].size[1];
1046             config.bias_term = (hasBias()) ? true : false;
1047             config.use_half = use_half;
1048
1049             convolutionOp = Ptr<OCL4DNNConvSpatial<float> >(new OCL4DNNConvSpatial<float>(config));
1050         }
1051
1052         int outCn = umat_blobs[0].size[0];
1053
1054         reluslope.clear();
1055         if( activ )
1056         {
1057             Ptr<ReLULayer> activ_relu = activ.dynamicCast<ReLULayer>();
1058             if( !activ_relu.empty() )
1059             {
1060                 reluslope.assign(outCn+2, activ_relu->negativeSlope);
1061                 activType = OCL4DNN_CONV_FUSED_ACTIV_RELU;
1062             }
1063
1064             Ptr<ReLU6Layer> activ_relu6 = activ.dynamicCast<ReLU6Layer>();
1065             if( !activ_relu6.empty() )
1066             {
1067                 reluslope.resize(2);
1068                 reluslope[0] = activ_relu6->minValue;
1069                 reluslope[1] = activ_relu6->maxValue;
1070                 activType = OCL4DNN_CONV_FUSED_ACTIV_RELU6;
1071             }
1072
1073             Ptr<ChannelsPReLULayer> activ_chprelu = activ.dynamicCast<ChannelsPReLULayer>();
1074             if( !activ_chprelu.empty() )
1075             {
1076                 const Mat& m = activ_chprelu->blobs[0];
1077                 CV_Assert(m.isContinuous() && m.type() == CV_32F && (int)m.total() == outCn);
1078                 const float* mdata = m.ptr<float>();
1079                 reluslope.resize(outCn+2);
1080                 std::copy(mdata, mdata + outCn, reluslope.begin());
1081                 reluslope[outCn] = reluslope[outCn+1] = reluslope[outCn-1];
1082                 activType = OCL4DNN_CONV_FUSED_ACTIV_PRELU;
1083             }
1084         }
1085
1086         if (fusedWeights)
1087         {
1088             weightsMat.copyTo(umat_blobs[0]);
1089             fusedWeights = false;
1090         }
1091         if (fusedBias)
1092         {
1093             if ( umat_blobs.size() < 2 )
1094                 umat_blobs.resize(2);
1095             umat_blobs[1] = UMat(biasvec, true);
1096             convolutionOp->setBias(true);
1097             fusedBias = false;
1098         }
1099
1100         if ( newActiv )
1101         {
1102             if ( activType == OCL4DNN_CONV_FUSED_ACTIV_RELU )
1103             {
1104                 CV_Assert(!reluslope.empty());
1105                 convolutionOp->setActivReLU(true, reluslope[0]);
1106             }
1107             else if ( activType == OCL4DNN_CONV_FUSED_ACTIV_PRELU)
1108             {
1109                 CV_Assert(!reluslope.empty());
1110                 convolutionOp->setActivPReLU(true, reluslope);
1111             }
1112             else if ( activType == OCL4DNN_CONV_FUSED_ACTIV_POWER)
1113             {
1114                 convolutionOp->setActivPower(true, power);
1115             }
1116             else if ( activType == OCL4DNN_CONV_FUSED_ACTIV_TANH)
1117             {
1118                 convolutionOp->setActivTanh(true);
1119             }
1120             else if ( activType == OCL4DNN_CONV_FUSED_ACTIV_RELU6)
1121             {
1122                 convolutionOp->setActivReLU6(true, reluslope[0], reluslope[1]);
1123             }
1124             else
1125             {
1126                 convolutionOp->setActivReLU(false, 0);
1127                 convolutionOp->setActivPReLU(false, reluslope);
1128                 convolutionOp->setActivPower(false, 1.f);
1129                 convolutionOp->setActivTanh(false);
1130                 convolutionOp->setActivReLU6(false, 0, 0);
1131             }
1132             newActiv = false;
1133         }
1134
1135         UMat& inpMat = inputs[0];
1136         UMat& outMat = outputs[0];
1137         int batch_size = inpMat.size[0];
1138
1139         return convolutionOp->Forward(inpMat,
1140                                       inputs.size() == 2 ? inputs[1] : UMat(),
1141                                       umat_blobs[0],
1142                                       umat_blobs.size() > 1 ? umat_blobs[1] : UMat(),
1143                                       outMat,
1144                                       batch_size);
1145     }
1146 #endif
1147
1148     void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr) CV_OVERRIDE
1149     {
1150         CV_TRACE_FUNCTION();
1151         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
1152
1153         CV_OCL_RUN(IS_DNN_OPENCL_TARGET(preferableTarget),
1154                    forward_ocl(inputs_arr, outputs_arr, internals_arr))
1155
1156         if (inputs_arr.depth() == CV_16S)
1157         {
1158             forward_fallback(inputs_arr, outputs_arr, internals_arr);
1159             return;
1160         }
1161
1162         std::vector<Mat> inputs, outputs;
1163         inputs_arr.getMatVector(inputs);
1164         outputs_arr.getMatVector(outputs);
1165
1166         /*printf("conv %s: input (%d x %d x %d x %d), kernel (%d x %d), pad (%d x %d), stride (%d x %d), dilation (%d x %d)\n",
1167                name.c_str(), inputs[0].size[0], inputs[0].size[1], inputs[0].size[2], inputs[0].size[3],
1168                kernel.width, kernel.height, pad.width, pad.height,
1169                stride.width, stride.height, dilation.width, dilation.height);*/
1170         CV_Assert_N(inputs.size() == (size_t)1, inputs[0].size[1] % blobs[0].size[1] == 0,
1171                     outputs.size() == 1, inputs[0].data != outputs[0].data);
1172
1173         int ngroups = inputs[0].size[1]/blobs[0].size[1];
1174         CV_Assert(outputs[0].size[1] % ngroups == 0);
1175         int outCn = blobs[0].size[0];
1176
1177         reluslope.clear();
1178         if( activ )
1179         {
1180             Ptr<ReLULayer> activ_relu = activ.dynamicCast<ReLULayer>();
1181             if( !activ_relu.empty() )
1182             {
1183                 reluslope.assign(outCn+2, activ_relu->negativeSlope);
1184             }
1185
1186             Ptr<ChannelsPReLULayer> activ_chprelu = activ.dynamicCast<ChannelsPReLULayer>();
1187             if( !activ_chprelu.empty() )
1188             {
1189                 const Mat& m = activ_chprelu->blobs[0];
1190                 CV_Assert(m.isContinuous() && m.type() == CV_32F && (int)m.total() == outCn);
1191                 const float* mdata = m.ptr<float>();
1192                 reluslope.resize(outCn+2);
1193                 std::copy(mdata, mdata + outCn, reluslope.begin());
1194                 reluslope[outCn] = reluslope[outCn+1] = reluslope[outCn-1];
1195             }
1196         }
1197
1198         int nstripes = std::max(getNumThreads(), 1);
1199
1200         ParallelConv::run(inputs[0], outputs[0], weightsMat, biasvec, reluslope,
1201                           kernel_size, strides, pads_begin, pads_end, dilations, activ.get(), ngroups, nstripes);
1202     }
1203
1204     virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
1205                            const std::vector<MatShape> &outputs) const CV_OVERRIDE
1206     {
1207         CV_Assert(inputs.size() == outputs.size());
1208
1209         int64 flops = 0;
1210         int karea = std::accumulate(kernel_size.begin(), kernel_size.end(), 1, std::multiplies<size_t>());
1211         for (int i = 0; i < inputs.size(); i++)
1212         {
1213             flops += total(outputs[i])*(CV_BIG_INT(2)*karea*inputs[i][1] + 1);
1214         }
1215
1216         return flops;
1217     }
1218 };
1219
1220 class DeConvolutionLayerImpl CV_FINAL : public BaseConvolutionLayerImpl
1221 {
1222 public:
1223     Mat weightsMat, biasesMat;
1224     UMat umat_weights;
1225     UMat umat_biases;
1226
1227     DeConvolutionLayerImpl(const LayerParams& params) : BaseConvolutionLayerImpl(params) {}
1228
1229     MatShape computeColRowShape(const MatShape &inpShape, const MatShape &outShape) const CV_OVERRIDE
1230     {
1231         int inpCn = inpShape[1];
1232         int inpH = inpShape[2];
1233         int inpW = inpShape[3];
1234         int outCn = outShape[1];
1235         int ngroups = inpCn / blobs[0].size[0];
1236         int outGroupCn = outCn / ngroups;
1237         int ksize = outGroupCn * kernel.height * kernel.width;
1238         return shape(ksize, inpH * inpW);
1239     }
1240
1241     virtual bool supportBackend(int backendId) CV_OVERRIDE
1242     {
1243 #ifdef HAVE_INF_ENGINE
1244         const int outGroupCn = blobs[0].size[1];  // Weights are in IOHW or IODHW layout
1245         const int group = numOutput / outGroupCn;
1246
1247         if (backendId == DNN_BACKEND_INFERENCE_ENGINE)
1248         {
1249             if (kernel_size.size() == 3 && preferableTarget != DNN_TARGET_CPU) {
1250                 return false;
1251             }
1252
1253             if (std::accumulate(adjust_pads.begin(), adjust_pads.end(), 0, std::plus<size_t>()) > 0)
1254             {
1255                 if (padMode.empty())
1256                 {
1257                     if (preferableTarget != DNN_TARGET_CPU && group != 1)
1258                     {
1259                         for (int i = 0; i < adjust_pads.size(); i++) {
1260                             if (adjust_pads[i] && pads_begin[i])
1261                                 return false;
1262                         }
1263                     }
1264                     for (int i = 0; i < adjust_pads.size(); i++) {
1265                         if (pads_end[i] < adjust_pads[i])
1266                             return false;
1267                     }
1268                     return true;
1269                 }
1270                 else if (padMode == "SAME")
1271                 {
1272                     for (int i = 0; i < adjust_pads.size(); i++) {
1273                         if (kernel_size[i] < pads_begin[i] + 1 + adjust_pads[i])
1274                             return false;
1275                     }
1276                     return true;
1277                 }
1278                 else if (padMode == "VALID")
1279                     return false;
1280             }
1281
1282             if (group != 1)
1283             {
1284                 return preferableTarget == DNN_TARGET_CPU;
1285             }
1286             if (preferableTarget == DNN_TARGET_OPENCL || preferableTarget == DNN_TARGET_OPENCL_FP16)
1287                 return std::accumulate(dilations.begin(), dilations.end(), 1, std::multiplies<size_t>()) == 1;
1288             return true;
1289         }
1290         else
1291 #endif  // HAVE_INF_ENGINE
1292             return kernel_size.size() == 2 && (backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE);
1293     }
1294
1295     bool getMemoryShapes(const std::vector<MatShape> &inputs,
1296                          const int requiredOutputs,
1297                          std::vector<MatShape> &outputs,
1298                          std::vector<MatShape> &internals) const CV_OVERRIDE
1299     {
1300         CV_Assert(!hasBias() || blobs[1].total() == (size_t)numOutput);
1301         CV_Assert(inputs.size() != 0);
1302
1303         int outCn = numOutput;
1304         std::vector<int> outShape;
1305         outShape.push_back(inputs[0][0]);  // batch
1306         outShape.push_back(outCn);
1307         if (padMode.empty())
1308         {
1309             for (int i = 0; i < kernel_size.size(); i++)
1310                 outShape.push_back(strides[i] * (inputs[0][2 + i] - 1) + kernel_size[i] - pads_begin[i] - pads_end[i] + adjust_pads[i]);
1311         }
1312         else if (padMode == "VALID")
1313         {
1314             for (int i = 0; i < kernel_size.size(); i++)
1315                 outShape.push_back(strides[i] * (inputs[0][2 + i] - 1) + kernel_size[i] + adjust_pads[i]);
1316         }
1317         else if (padMode == "SAME")
1318         {
1319             for (int i = 0; i < kernel_size.size(); i++)
1320                 outShape.push_back(strides[i] * (inputs[0][2 + i] - 1) + 1 + adjust_pads[i]);
1321         }
1322         else
1323             CV_Error(Error::StsError, "Unsupported padding mode " + padMode);
1324
1325         CV_Assert(outCn % blobs[0].size[1] == 0);
1326         int ngroups = outCn / blobs[0].size[1];
1327
1328         int inpCn = inputs[0][1];
1329         CV_Assert(inpCn % ngroups == 0 && outCn % ngroups == 0);
1330         CV_Assert(blobs[0].size[0] == inpCn);
1331
1332         outputs.resize(1, outShape);
1333
1334         if (!is1x1())
1335             internals.push_back(computeColRowShape(inputs[0], outputs[0]));
1336
1337         return false;
1338     }
1339
1340     void finalize(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr) CV_OVERRIDE
1341     {
1342         BaseConvolutionLayerImpl::finalize(inputs_arr, outputs_arr);
1343
1344         std::vector<Mat> inputs, outputs;
1345         inputs_arr.getMatVector(inputs);
1346         outputs_arr.getMatVector(outputs);
1347
1348         std::vector<int> inpShape;
1349         std::vector<int> outShape;
1350         for (int i = 2; i < inputs[0].dims; i++) {
1351             inpShape.push_back(inputs[0].size[i]);
1352             outShape.push_back(outputs[0].size[i]);
1353         }
1354         getConvPoolPaddings(outShape, kernel_size, strides, padMode, pads_begin, pads_end);
1355         if (pads_begin.size() == 2) {
1356             for (int i = 0; i < pads_begin.size(); i++) {
1357                 if (pads_begin[i] != pads_end[i])
1358                     CV_Error(Error::StsNotImplemented, "Unsupported asymmetric padding in deconvolution layer");
1359             }
1360             pad = Size(pads_begin[1], pads_begin[0]);
1361         }
1362
1363         weightsMultipliers.assign(numOutput, 1.0);
1364         if (weightsMat.empty())
1365         {
1366             transpose(blobs[0].reshape(1, blobs[0].size[0]), weightsMat);
1367             biasesMat = hasBias() ? blobs[1].reshape(1, numOutput)
1368                                   : Mat::zeros(numOutput, 1, CV_32F);
1369         }
1370     }
1371
1372     void fuseWeights(const Mat& w_, const Mat& b_) CV_OVERRIDE
1373     {
1374         Mat w = w_.total() == 1 ? Mat(1, numOutput, CV_32F, Scalar(w_.at<float>(0))) : w_;
1375         Mat b = b_.total() == 1 ? Mat(1, numOutput, CV_32F, Scalar(b_.at<float>(0))) : b_;
1376
1377         CV_Assert_N(!weightsMat.empty(),
1378                      w.empty() || numOutput == w.total(),
1379                      b.empty() || numOutput == b.total());
1380
1381         if (!w.empty())
1382         {
1383             transpose(blobs[0].reshape(1, blobs[0].size[0]), weightsMat);
1384             weightsMat = weightsMat.reshape(1, numOutput);
1385             for (int i = 0; i < numOutput; ++i)
1386             {
1387                 double wi = w.at<float>(i);
1388                 weightsMultipliers[i] *= wi;
1389                 cv::multiply(weightsMat.row(i), weightsMultipliers[i], weightsMat.row(i));
1390                 biasesMat.at<float>(i) *= wi;
1391             }
1392             weightsMat = weightsMat.reshape(1, weightsMat.total() / blobs[0].size[0]);
1393         }
1394
1395         if (!b.empty())
1396         {
1397             cv::add(biasesMat, b.reshape(1, numOutput), biasesMat);
1398         }
1399     }
1400
1401     class MatMulInvoker : public ParallelLoopBody
1402     {
1403     public:
1404         MatMulInvoker(const Mat& a, const Mat& b, Mat& c, int nstripes)
1405         {
1406             a_ = &a;
1407             b_ = &b;
1408             c_ = &c;
1409             nstripes_ = nstripes;
1410             useAVX = checkHardwareSupport(CPU_AVX);
1411             useAVX2 = checkHardwareSupport(CPU_AVX2);
1412             useAVX512 = CV_CPU_HAS_SUPPORT_AVX512_SKX;
1413         }
1414
1415         void operator()(const Range& range_) const CV_OVERRIDE
1416         {
1417             int stripeSize = (int)alignSize((b_->cols + nstripes_ - 1)/nstripes_, 16);
1418             Range range(range_.start*stripeSize, std::min(range_.end*stripeSize, b_->cols));
1419             int mmax = a_->rows;
1420             int nmax = range.end - range.start;
1421             int kmax = a_->cols;
1422             int m, n, k;
1423             const float* aptr = a_->ptr<float>();
1424             const float* bptr = b_->ptr<float>() + range.start;
1425             float* cptr = c_->ptr<float>() + range.start;
1426             size_t astep = a_->step1();
1427             size_t bstep = b_->step1();
1428             size_t cstep = c_->step1();
1429
1430         #if CV_TRY_AVX512_SKX
1431             if( useAVX512 )
1432                 opt_AVX512_SKX::fastGEMM( aptr, astep, bptr, bstep, cptr, cstep, mmax, kmax, nmax );
1433             else
1434         #endif
1435         #if CV_TRY_AVX2
1436             if( useAVX2 )
1437                 opt_AVX2::fastGEMM( aptr, astep, bptr, bstep, cptr, cstep, mmax, kmax, nmax );
1438             else
1439         #endif
1440         #if CV_TRY_AVX
1441             if( useAVX )
1442                 opt_AVX::fastGEMM( aptr, astep, bptr, bstep, cptr, cstep, mmax, kmax, nmax );
1443             else
1444         #endif
1445             for( m = 0; m < mmax; m += 2 )
1446             {
1447                 float* dst0 = cptr + cstep*m;
1448                 float* dst1 = cptr + cstep*std::min(m+1, mmax-1);
1449                 const float* aptr0 = aptr + astep*m;
1450                 const float* aptr1 = aptr + astep*std::min(m+1, mmax-1);
1451
1452                 for( n = 0; n < nmax; n++ )
1453                 {
1454                     dst0[n] = 0.f;
1455                     dst1[n] = 0.f;
1456                 }
1457
1458                 for( k = 0; k < kmax; k += 4 )
1459                 {
1460                     float alpha00 = aptr0[k];
1461                     float alpha01 = aptr1[k];
1462                     float alpha10 = 0.f, alpha11 = 0.f;
1463                     float alpha20 = 0.f, alpha21 = 0.f;
1464                     float alpha30 = 0.f, alpha31 = 0.f;
1465                     const float* bptr0 = bptr + k*bstep;
1466                     const float* bptr1 = bptr0;
1467                     const float* bptr2 = bptr0;
1468                     const float* bptr3 = bptr0;
1469
1470                     if( k+1 < kmax )
1471                     {
1472                         alpha10 = aptr0[k+1];
1473                         alpha11 = aptr1[k+1];
1474                         bptr1 = bptr0 + bstep;
1475                         if( k+2 < kmax )
1476                         {
1477                             alpha20 = aptr0[k+2];
1478                             alpha21 = aptr1[k+2];
1479                             bptr2 = bptr1 + bstep;
1480                             if( k+3 < kmax )
1481                             {
1482                                 alpha30 = aptr0[k+3];
1483                                 alpha31 = aptr1[k+3];
1484                                 bptr3 = bptr2 + bstep;
1485                             }
1486                         }
1487                     }
1488                     n = 0;
1489
1490                 #if CV_SIMD128
1491                     v_float32x4 a00 = v_setall_f32(alpha00);
1492                     v_float32x4 a01 = v_setall_f32(alpha01);
1493                     v_float32x4 a10 = v_setall_f32(alpha10);
1494                     v_float32x4 a11 = v_setall_f32(alpha11);
1495                     v_float32x4 a20 = v_setall_f32(alpha20);
1496                     v_float32x4 a21 = v_setall_f32(alpha21);
1497                     v_float32x4 a30 = v_setall_f32(alpha30);
1498                     v_float32x4 a31 = v_setall_f32(alpha31);
1499
1500                     for( ; n <= nmax - 4; n += 4 )
1501                     {
1502                         v_float32x4 b0 = v_load(bptr0 + n);
1503                         v_float32x4 b1 = v_load(bptr1 + n);
1504                         v_float32x4 b2 = v_load(bptr2 + n);
1505                         v_float32x4 b3 = v_load(bptr3 + n);
1506                         v_float32x4 d0 = v_load(dst0 + n);
1507                         v_float32x4 d1 = v_load(dst1 + n);
1508                         d0 += b0*a00;
1509                         d1 += b0*a01;
1510                         d0 += b1*a10;
1511                         d1 += b1*a11;
1512                         d0 += b2*a20;
1513                         d1 += b2*a21;
1514                         d0 += b3*a30;
1515                         d1 += b3*a31;
1516                         v_store(dst0 + n, d0);
1517                         v_store(dst1 + n, d1);
1518                     }
1519                 #endif
1520
1521                     for( ; n < nmax; n++ )
1522                     {
1523                         float b0 = bptr0[n], b1 = bptr1[n];
1524                         float b2 = bptr2[n], b3 = bptr3[n];
1525                         float d0 = dst0[n] + alpha00*b0 + alpha10*b1 + alpha20*b2 + alpha30*b3;
1526                         float d1 = dst1[n] + alpha01*b0 + alpha11*b1 + alpha21*b2 + alpha31*b3;
1527                         dst0[n] = d0;
1528                         dst1[n] = d1;
1529                     }
1530                 }
1531             }
1532         }
1533
1534         const Mat *a_, *b_;
1535         Mat* c_;
1536         int nstripes_;
1537         bool useAVX;
1538         bool useAVX2;
1539         bool useAVX512;
1540     };
1541
1542     class Col2ImInvoker : public cv::ParallelLoopBody
1543     {
1544     public:
1545         const float* data_col;
1546         const float* biasvec;
1547         int channels, height, width;
1548         int kernel_h, kernel_w;
1549         int pad_h, pad_w;
1550         int stride_h, stride_w;
1551         float* data_im;
1552         int height_col, width_col;
1553         int nstripes;
1554         bool is1x1;
1555
1556         Col2ImInvoker()
1557             : data_col(0), biasvec(0), channels(0), height(0), width(0),
1558               kernel_h(0), kernel_w(0), pad_h(0), pad_w(0), stride_h(0), stride_w(0), data_im(0),
1559               height_col(0), width_col(0), nstripes(0), is1x1(0)
1560         {}
1561
1562         static void run(const float* data_col,
1563                         int channels, int height, int width,
1564                         int kernel_h, int kernel_w,
1565                         int pad_h, int pad_w,
1566                         int stride_h, int stride_w,
1567                         int height_col, int width_col,
1568                         float* data_im,
1569                         const float* biasvec,
1570                         bool is1x1)
1571         {
1572             const int nstripes = getNumThreads();
1573
1574             Col2ImInvoker t;
1575             t.data_col = data_col;
1576             t.data_im = data_im;
1577             t.channels = channels; t.height = height; t.width = width;
1578             t.kernel_h = kernel_h; t.kernel_w = kernel_w;
1579             t.pad_h = pad_h; t.pad_w = pad_w;
1580             t.stride_h = stride_h; t.stride_w = stride_w;
1581             t.height_col = height_col;
1582             t.width_col = width_col;
1583             t.nstripes = nstripes;
1584             t.is1x1 = is1x1;
1585             t.biasvec = biasvec;
1586
1587             parallel_for_(Range(0, nstripes), t, nstripes);
1588         }
1589
1590         virtual void operator ()(const Range &r) const CV_OVERRIDE
1591         {
1592             const float* data_col_ = data_col;
1593             float* data_im_ = data_im;
1594             int coeff_h = (1 - stride_h * kernel_w * height_col) * width_col;
1595             int coeff_w = (1 - stride_w * height_col * width_col);
1596             size_t total = (size_t)channels * height * width;
1597             size_t stripeSize = (total + nstripes - 1)/nstripes;
1598             size_t startIndex = r.start*stripeSize;
1599             size_t endIndex = std::min(r.end*stripeSize, total);
1600             int w = (int)(startIndex % width + pad_w);
1601             int h = (int)((startIndex / width) % height + pad_h);
1602             int c = (int)(startIndex / (width * height));
1603             int h_col_start = (h < kernel_h) ? 0 : (h - kernel_h) / stride_h + 1;
1604             int h_col_end = std::min(h / stride_h + 1, height_col);
1605             int plane_size_col = height_col * width_col;
1606             int offset = (c * kernel_h * kernel_w + h * kernel_w + w) * plane_size_col;
1607             bool is1x1_ = is1x1;
1608             const float* biasvec_ = biasvec;
1609
1610             for (size_t index = startIndex; index < endIndex; index++)
1611             {
1612                 // compute the start and end of the output
1613                 int w_col_start = (w < kernel_w) ? 0 : (w - kernel_w) / stride_w + 1;
1614                 int w_col_end = std::min(w / stride_w + 1, width_col);
1615                 float val;
1616
1617                 if( is1x1_ )
1618                     val = data_im_[index];
1619                 else
1620                 {
1621                     val = 0.f;
1622                     for (int h_col = h_col_start; h_col < h_col_end; ++h_col) {
1623                         for (int w_col = w_col_start; w_col < w_col_end; ++w_col) {
1624                             val += data_col_[offset + h_col * coeff_h + w_col * coeff_w];
1625                         }
1626                     }
1627                 }
1628                 data_im_[index] = val + biasvec_[c];
1629
1630                 offset += plane_size_col;
1631                 if( ++w >= width + pad_w )
1632                 {
1633                     w = (int)((index + 1)% width + pad_w);
1634                     h = (int)(((index + 1) / width) % height + pad_h);
1635                     c = (int)((index + 1) / (width * height));
1636                     h_col_start = (h < kernel_h) ? 0 : (h - kernel_h) / stride_h + 1;
1637                     h_col_end = std::min(h / stride_h + 1, height_col);
1638                     offset = (c * kernel_h * kernel_w + h * kernel_w + w) * plane_size_col;
1639                 }
1640             }
1641         }
1642     };
1643
1644 #ifdef HAVE_OPENCL
1645     bool forward_ocl(InputArrayOfArrays inputs_, OutputArrayOfArrays outputs_, OutputArrayOfArrays internals_)
1646     {
1647         std::vector<UMat> inputs;
1648         std::vector<UMat> outputs;
1649         std::vector<UMat> internals;
1650
1651         if (inputs_.depth() == CV_16S)
1652             return false;
1653
1654         inputs_.getUMatVector(inputs);
1655         outputs_.getUMatVector(outputs);
1656         internals_.getUMatVector(internals);
1657
1658         int outCn = numOutput;
1659         int inpCn = inputs[0].size[1];
1660
1661         if (is1x1())
1662             return false;
1663
1664         if (umat_weights.empty())
1665         {
1666             if (fusedWeights)
1667                 weightsMat.copyTo(umat_weights);
1668             else
1669                 transpose(blobs[0].reshape(1, inpCn), umat_weights);
1670
1671             if (fusedBias)
1672                 biasesMat.copyTo(umat_biases);
1673             else
1674             {
1675                 if (hasBias())
1676                     blobs[1].reshape(1, outCn).copyTo(umat_biases);
1677                 else
1678                     umat_biases = UMat::zeros(outCn, 1, CV_32F);
1679             }
1680         }
1681
1682         String buildopt = format("-DT=%s ", ocl::typeToStr(inputs[0].type()));
1683         buildopt += format("-DPAD_H=%d -DPAD_W=%d -DKERNEL_H=%d -DKERNEL_W=%d -DSTRIDE_H=%d -DSTRIDE_W=%d ",
1684                            pad.height, pad.width, kernel.height, kernel.width, stride.height, stride.width);
1685
1686         for (size_t ii = 0; ii < outputs.size(); ii++)
1687         {
1688             int ngroups = outCn / blobs[0].size[1];
1689             int inpGroupCn = inpCn / ngroups;
1690             int outGroupCn = blobs[0].size[1];
1691             const UMat& inp = inputs[ii];
1692             UMat& out = outputs[ii];
1693             int numImg = inp.size[0];
1694             int inpH = inp.size[2], inpW = inp.size[3];
1695             int outH = out.size[2], outW = out.size[3];
1696
1697             MatShape inpshape = shape(numImg*inpCn, inpH*inpW);
1698             MatShape outshape = shape(numImg*outCn, outH*outW);
1699             UMat convBlob = inputs[ii].reshape(1, inpshape.size(), &inpshape[0]);
1700             UMat decnBlob = out.reshape(1, outshape.size(), &outshape[0]);
1701             int rows = internals[0].rows / ngroups;
1702
1703             for (int n = 0; n < numImg; n++)
1704             {
1705                 for (int g = 0; g < ngroups; g++)
1706                 {
1707                     UMat colMat = internals[0].rowRange(_Range(g * rows, rows));
1708                     UMat convMat = convBlob.rowRange(_Range((g + n * ngroups) * inpGroupCn, inpGroupCn));
1709                     UMat wghtMat = umat_weights.colRange(_Range(g * inpGroupCn, inpGroupCn));
1710                     gemm(wghtMat, convMat, 1, noArray(), 0, colMat, 0);
1711                 }
1712
1713                 for (int g = 0; g < ngroups; g++)
1714                 {
1715                     int total = outGroupCn * decnBlob.cols;
1716                     int index = 0;
1717                     int height_col = inpH;
1718                     int width_col = inpW;
1719                     int coeff_h = (1 - stride.height * kernel.width * height_col) * width_col;
1720                     int coeff_w = (1 - stride.width * height_col * width_col);
1721
1722                     ocl::Kernel k("col2im", ocl::dnn::col2im_oclsrc, buildopt);
1723                     k.set(index++, total);
1724                     k.set(index++, ocl::KernelArg::PtrReadOnly(internals[0]));
1725                     k.set(index++, (int)(g * rows * internals[0].cols));
1726                     k.set(index++, outGroupCn);
1727                     k.set(index++, outH);
1728                     k.set(index++, outW);
1729                     k.set(index++, height_col);
1730                     k.set(index++, width_col);
1731                     k.set(index++, coeff_h);
1732                     k.set(index++, coeff_w);
1733                     k.set(index++, ocl::KernelArg::PtrReadOnly(umat_biases));
1734                     k.set(index++, (int)(g * outGroupCn * umat_biases.cols));
1735                     k.set(index++, ocl::KernelArg::PtrWriteOnly(decnBlob));
1736                     k.set(index++, (int)((g + n * ngroups) * outGroupCn * decnBlob.cols));
1737
1738                     size_t global[] = { (size_t)total };
1739                     bool ret = k.run(1, global, NULL, false);
1740                     if (!ret)
1741                         return false;
1742                 }
1743             }
1744         }
1745
1746         return true;
1747     }
1748 #endif
1749
1750     void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr) CV_OVERRIDE
1751     {
1752         CV_TRACE_FUNCTION();
1753         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
1754
1755         CV_OCL_RUN(IS_DNN_OPENCL_TARGET(preferableTarget),
1756                    forward_ocl(inputs_arr, outputs_arr, internals_arr));
1757
1758         if (inputs_arr.depth() == CV_16S)
1759         {
1760             forward_fallback(inputs_arr, outputs_arr, internals_arr);
1761             return;
1762         }
1763
1764         std::vector<Mat> inputs, outputs, internals;
1765         inputs_arr.getMatVector(inputs);
1766         outputs_arr.getMatVector(outputs);
1767         internals_arr.getMatVector(internals);
1768
1769         int outCn = numOutput;
1770         int inpCn = inputs[0].size[1];
1771         bool is1x1flag = is1x1();
1772         int nstripes = getNumThreads();
1773
1774         if( weightsMat.empty() )
1775         {
1776             transpose(blobs[0].reshape(1, inpCn), weightsMat);
1777             biasesMat = hasBias() ? blobs[1].reshape(1, outCn) : Mat::zeros(outCn, 1, CV_32F);
1778         }
1779
1780         for (size_t ii = 0; ii < outputs.size(); ii++)
1781         {
1782             int ngroups = outCn / blobs[0].size[1];
1783             int inpGroupCn = inpCn / ngroups;
1784             int outGroupCn = blobs[0].size[1];
1785             const Mat& inp = inputs[ii];
1786             Mat& out = outputs[ii];
1787             int numImg = inp.size[0];
1788             int inpH = inp.size[2], inpW = inp.size[3];
1789             int outH = out.size[2], outW = out.size[3];
1790
1791             Mat convBlob = inputs[ii].reshape(1, numImg*inpCn);
1792             Mat decnBlob = out.reshape(1, numImg*outCn);
1793
1794             for (int n = 0; n < numImg; n++)
1795             {
1796                 for (int g = 0; g < ngroups; g++)
1797                 {
1798                     Mat dstMat = decnBlob.rowRange(_Range((g + n * ngroups) * outGroupCn, outGroupCn));
1799                     Mat &colMat = is1x1flag ? dstMat : internals[0];
1800
1801                     Mat convMat = convBlob.rowRange(_Range((g + n * ngroups) * inpGroupCn, inpGroupCn));
1802                     Mat wghtMat = weightsMat.colRange(_Range(g * inpGroupCn, inpGroupCn));
1803                     Mat curBiasMat = biasesMat.rowRange(_Range(g * outGroupCn, outGroupCn));
1804
1805                     //gemm(wghtMat, convMat, 1, colMat, 0, colMat, 0);
1806                     MatMulInvoker mminvoker(wghtMat, convMat, colMat, nstripes);
1807                     parallel_for_(Range(0, nstripes), mminvoker, nstripes);
1808
1809                     Col2ImInvoker::run(colMat.ptr<float>(), outGroupCn, outH, outW,
1810                                        kernel.height, kernel.width, pad.height, pad.width,
1811                                        stride.height, stride.width, inpH, inpW, dstMat.ptr<float>(),
1812                                        curBiasMat.ptr<float>(), is1x1flag);
1813                 }
1814             }
1815         }
1816     }
1817
1818     virtual Ptr<BackendNode> initHalide(const std::vector<Ptr<BackendWrapper> > &inputs) CV_OVERRIDE
1819     {
1820 #ifdef HAVE_HALIDE
1821         Halide::Buffer<float> inputBuffer = halideBuffer(inputs[0]);
1822
1823         int inW, inH, inC, inN;
1824         getCanonicalSize(inputBuffer, &inW, &inH, &inC, &inN);
1825         const int outGroupCn = blobs[0].size[1];
1826         const int group = numOutput / outGroupCn;
1827         const int inpGroupCn = blobs[0].size[0] / group;
1828
1829         Halide::Var x("x"), y("y"), c("c"), n("n");
1830         Halide::Func top = (name.empty() ? Halide::Func() : Halide::Func(name));
1831         Halide::Func padded_input(name + "_constant_exterior");
1832         auto weights = wrapToHalideBuffer(blobs[0]);
1833
1834         Halide::Func dilated_input("dilated_input");
1835         dilated_input(x, y, c, n) = 0.0f;
1836         Halide::RDom r1(0, inW, 0, inH);
1837         dilated_input(r1.x * stride.width, r1.y * stride.height, c, n) =
1838               inputBuffer(r1.x, r1.y, c, n);
1839         dilated_input.compute_root();
1840
1841         Halide::Func bounded =
1842             Halide::BoundaryConditions::constant_exterior(dilated_input, 0,
1843                                                           0, (inW - 1) * stride.width + 1,
1844                                                           0, (inH - 1) * stride.height + 1,
1845                                                           0, inC, 0, inN);
1846         padded_input(x, y, c, n) = bounded(x, y, c, n);
1847
1848         Halide::RDom r(0, kernel.width, 0, kernel.height, 0, inpGroupCn);
1849         Halide::Expr kx = x + pad.width - r.x;
1850         Halide::Expr ky = y + pad.height - r.y;
1851         Halide::Expr kInC = r.z;
1852         Halide::Expr kOutC = c;
1853         for (int i = 1; i < group; ++i)
1854         {
1855             kInC = select(c < outGroupCn * i, kInC, inpGroupCn * i + r.z);
1856             kOutC = select(c < outGroupCn * i, kOutC, c - outGroupCn * i);
1857         }
1858         Halide::Expr topExpr = sum(padded_input(kx, ky, kInC, n) *
1859                                    weights(r.x, r.y, kOutC, kInC));
1860         if (hasBias())
1861         {
1862             auto bias = wrapToHalideBuffer(blobs[1], {numOutput});
1863             topExpr += bias(c);
1864         }
1865         top(x, y, c, n) = topExpr;
1866         return Ptr<BackendNode>(new HalideBackendNode({ padded_input, top }));
1867 #endif  // HAVE_HALIDE
1868         return Ptr<BackendNode>();
1869     }
1870
1871 #ifdef HAVE_INF_ENGINE
1872     virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> > &) CV_OVERRIDE
1873     {
1874         InferenceEngine::Layout layout = blobs[0].dims == 5? InferenceEngine::Layout::NCDHW :
1875                                                              InferenceEngine::Layout::OIHW;
1876
1877         auto ieWeights = wrapToInfEngineBlob(blobs[0], layout);
1878         if (fusedWeights)
1879         {
1880             ieWeights = InferenceEngine::make_shared_blob<float>({
1881                             InferenceEngine::Precision::FP32,
1882                             ieWeights->getTensorDesc().getDims(), layout
1883                         });
1884             ieWeights->allocate();
1885
1886             int inpCn = blobs[0].size[0];
1887             Mat newWeights = infEngineBlobToMat(ieWeights).reshape(1, inpCn);
1888             transpose(weightsMat, newWeights);
1889         }
1890
1891         const int outGroupCn = blobs[0].size[1];  // Weights are in IOHW or OIDHW layout
1892         const int group = numOutput / outGroupCn;
1893
1894         InferenceEngine::Builder::DeconvolutionLayer ieLayer(name);
1895
1896         ieLayer.setKernel(kernel_size);
1897         ieLayer.setStrides(strides);
1898         ieLayer.setDilation(dilations);
1899         ieLayer.setPaddingsBegin(pads_begin);
1900
1901         if (padMode.empty())
1902         {
1903             std::vector<size_t> paddings_end;
1904             for (int i = 0; i < pads_end.size(); i++) {
1905                 paddings_end.push_back(pads_end[i] - adjust_pads[i]);
1906             }
1907             ieLayer.setPaddingsEnd(paddings_end);
1908         }
1909         else if (padMode == "SAME")
1910         {
1911             std::vector<size_t> paddings_end;
1912             for (int i = 0; i < pads_begin.size(); i++) {
1913                 paddings_end.push_back(kernel_size[i] - pads_begin[i] - 1 - adjust_pads[i]);
1914             }
1915             ieLayer.setPaddingsEnd(paddings_end);
1916         }
1917         ieLayer.setGroup((size_t)group);
1918         ieLayer.setOutDepth((size_t)numOutput);
1919
1920         InferenceEngine::Builder::Layer l = ieLayer;
1921         addConstantData("weights", ieWeights, l);
1922         if (hasBias())
1923             addConstantData("biases", wrapToInfEngineBlob(biasesMat, {(size_t)numOutput}, InferenceEngine::Layout::C), l);
1924         return Ptr<BackendNode>(new InfEngineBackendNode(l));
1925     }
1926 #endif  // HAVE_INF_ENGINE
1927
1928     virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
1929                            const std::vector<MatShape> &outputs) const CV_OVERRIDE
1930     {
1931         CV_Assert(inputs.size() == outputs.size());
1932
1933         float flops = 0;
1934         int outChannels = blobs[0].size[0];
1935         size_t karea = std::accumulate(kernel_size.begin(), kernel_size.end(),
1936                                        1, std::multiplies<size_t>());
1937
1938         for (int i = 0; i < inputs.size(); i++)
1939         {
1940             flops += CV_BIG_INT(2)*outChannels*karea*total(inputs[i]);
1941         }
1942
1943         return flops;
1944     }
1945 };
1946
1947 Ptr<BaseConvolutionLayer> ConvolutionLayer::create(const LayerParams &params)
1948 {
1949     Ptr<ConvolutionLayerImpl> l(new ConvolutionLayerImpl(params));
1950     return l;
1951 }
1952
1953 Ptr<BaseConvolutionLayer> DeconvolutionLayer::create(const LayerParams &params)
1954 {
1955     return Ptr<BaseConvolutionLayer>(new DeConvolutionLayerImpl(params));
1956 }
1957
1958 }
1959 }