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 "../op_vkcom.hpp"
  48 #include "opencv2/core/hal/hal.hpp"
  49 #include "opencv2/core/hal/intrin.hpp"
  50 #include <iostream>
  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);
  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             adjust_pads.push_back(params.get<int>("adj_h", 0));
  87             adjust_pads.push_back(params.get<int>("adj_w", 0));
  88
  89             adjustPad.height = adjust_pads[0];
  90             adjustPad.width = adjust_pads[1];
  91             CV_Assert(adjustPad.width < stride.width &&
  92                       adjustPad.height < stride.height);
  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         {
 261             if (kernel_size.size() != 2)
 262                 return false;
 263             return backendId == DNN_BACKEND_OPENCV ||
 264                    backendId == DNN_BACKEND_HALIDE ||
 265                    (backendId == DNN_BACKEND_VKCOM && haveVulkan());
 266         }
 267     }
 268
 269     bool getMemoryShapes(const std::vector<MatShape> &inputs,
 270                          const int requiredOutputs,
 271                          std::vector<MatShape> &outputs,
 272                          std::vector<MatShape> &internals) const CV_OVERRIDE
 273     {
 274         CV_Assert(blobs.size() != 0);
 275         CV_Assert(!hasBias() || blobs[1].total() == (size_t)blobs[0].size[0]);
 276         CV_Assert(inputs.size() == (size_t)1);
 277
 278         internals.clear();
 279
 280         CV_Assert(inputs.size() != 0);
 281         std::vector<int> inpShape(inputs[0].begin() + 2, inputs[0].end());
 282
 283         int outCn = blobs[0].size[0];
 284         std::vector<int> outShape;
 285         outShape.push_back(inputs[0][0]);
 286         outShape.push_back(outCn);
 287
 288         int inpCn = inputs[0][1];
 289         if (padMode.empty())
 290         {
 291             for (int i = 0; i < inpShape.size(); i++)
 292                 outShape.push_back((inpShape[i] + pads_begin[i] + pads_end[i] - dilations[i] * (kernel_size[i] - 1) - 1) / strides[i] + 1);
 293         }
 294         else
 295         {
 296             getConvPoolOutParams(inpShape, kernel_size, strides, padMode, dilations, outShape);
 297         }
 298
 299         int ngroups = inpCn / blobs[0].size[1];
 300         if (ngroups == 0 || ngroups * blobs[0].size[1] != inpCn)
 301             CV_Error(Error::StsError, format("Number of input channels should "
 302                      "be multiple of %d but got %d", blobs[0].size[1], inpCn));
 303         CV_Assert(ngroups > 0 && inpCn % ngroups == 0 && outCn % ngroups == 0);
 304
 305         outputs.resize(1, outShape);
 306
 307         return false;
 308     }
 309
 310     virtual void finalize(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr) CV_OVERRIDE
 311     {
 312         BaseConvolutionLayerImpl::finalize(inputs_arr, outputs_arr);
 313
 314         CV_Assert(!blobs.empty());
 315         const int outCn = blobs[0].size[0];
 316         // prepare weightsMat where each row is aligned and has enough zero padding on the right to
 317         // use vectorized (i.e. with intrinsics) loops without tail processing
 318         Mat wm = blobs[0].reshape(1, outCn);
 319         if( wm.step1() % VEC_ALIGN != 0 )
 320         {
 321             int newcols = (int)alignSize(wm.step1(), VEC_ALIGN);
 322             Mat wm_buffer = Mat(outCn, newcols, wm.type());
 323             Mat wm_padding = wm_buffer.colRange(wm.cols, newcols);
 324             wm_padding.setTo(Scalar::all(0.));
 325             Mat wm_aligned = wm_buffer.colRange(0, wm.cols);
 326             wm.copyTo(wm_aligned);
 327             wm = wm_aligned;
 328         }
 329         weightsMat = wm;
 330         weightsMultipliers.assign(outCn, 1.0);
 331
 332         Mat biasMat = hasBias() ? blobs[1].reshape(1, outCn) : Mat();
 333         biasvec.resize(outCn+2);
 334         if( biasMat.empty() )
 335         {
 336             for(int i = 0; i < outCn; i++ )
 337                 biasvec[i] = 0.f;
 338         }
 339         else
 340         {
 341             for(int i = 0; i < outCn; i++ )
 342                 biasvec[i] = biasMat.at<float>(i);
 343         }
 344 #ifdef HAVE_OPENCL
 345         convolutionOp.release();
 346 #endif
 347     }
 348
 349     bool setActivation(const Ptr<ActivationLayer>& layer) CV_OVERRIDE
 350     {
 351         if (!activ.empty() && !layer.empty())
 352             return false;
 353
 354         activ = layer;
 355         if (activ.empty())
 356             reluslope.clear();
 357 #ifdef HAVE_OPENCL
 358         newActiv = true;
 359         activType = OCL4DNN_CONV_FUSED_ACTIV_NONE;
 360
 361         if (IS_DNN_OPENCL_TARGET(preferableTarget))
 362         {
 363             Ptr<PowerLayer> activ_power = activ.dynamicCast<PowerLayer>();
 364             if (!activ_power.empty())
 365             {
 366                 if (activ_power->scale != 1.f || activ_power->shift != 0.f)
 367                 {
 368                     const int outCh = blobs[0].size[0];
 369                     fuseWeights(Mat(1, outCh, CV_32F, Scalar(activ_power->scale)),
 370                                 Mat(1, outCh, CV_32F, Scalar(activ_power->shift)));
 371                 }
 372
 373                 power = activ_power->power;
 374                 activType = OCL4DNN_CONV_FUSED_ACTIV_POWER;
 375             }
 376             Ptr<TanHLayer> activ_tanh = activ.dynamicCast<TanHLayer>();
 377             if (!activ_tanh.empty())
 378             {
 379                 activType = OCL4DNN_CONV_FUSED_ACTIV_TANH;
 380             }
 381         }
 382 #endif
 383         return !activ.empty();
 384     }
 385
 386     void fuseWeights(const Mat& w_, const Mat& b_) CV_OVERRIDE
 387     {
 388         // Convolution weights have OIHW data layout. Parameters fusion in case of
 389         // (conv(I) + b1 ) * w + b2
 390         // means to replace convolution's weights to [w*conv(I)] and bias to [b1 * w + b2]
 391         const int outCn = weightsMat.size[0];
 392         Mat w = w_.total() == 1 ? Mat(1, outCn, CV_32F, Scalar(w_.at<float>(0))) : w_;
 393         Mat b = b_.total() == 1 ? Mat(1, outCn, CV_32F, Scalar(b_.at<float>(0))) : b_;
 394         CV_Assert_N(!weightsMat.empty(), biasvec.size() == outCn + 2,
 395                     w.empty() || outCn == w.total(), b.empty() || outCn == b.total());
 396
 397         if (!w.empty())
 398         {
 399             // Keep origin weights unchanged.
 400             if (weightsMat.data == blobs[0].data)
 401                 weightsMat = weightsMat.clone();
 402
 403             Mat originWeights = blobs[0].reshape(1, outCn);
 404             for (int i = 0; i < outCn; ++i)
 405             {
 406                 double wi = w.at<float>(i);
 407                 weightsMultipliers[i] *= wi;
 408                 cv::multiply(originWeights.row(i), weightsMultipliers[i], weightsMat.row(i));
 409                 biasvec[i] *= wi;
 410             }
 411         }
 412
 413         if (!b.empty())
 414         {
 415             for (int i = 0; i < outCn; ++i)
 416                 biasvec[i] += b.at<float>(i);
 417         }
 418         biasvec[outCn] = biasvec[outCn+1] = biasvec[outCn-1];
 419     }
 420
 421     virtual Ptr<BackendNode> initVkCom(const std::vector<Ptr<BackendWrapper> > &inputs) CV_OVERRIDE
 422     {
 423 #ifdef HAVE_VULKAN
 424         int out_channel = blobs[0].size[0];
 425         bool has_bias = hasBias() || fusedBias;
 426         int filter_size[2] = {kernel.height, kernel.width};
 427         int pad_size[2] = {pad.height, pad.width};
 428         int stride_size[2] = {stride.height, stride.width};
 429         int dilation_size[2] = {dilation.height, dilation.width};
 430         int activation = 0;
 431         vkcom::Tensor input_tensor = VkComTensor(inputs[0]);
 432         int in_channel = input_tensor.dimSize(1);
 433         int group = in_channel / blobs[0].size[1];
 434
 435         // TODO: support group > 1
 436         if (group != 1)
 437             return Ptr<BackendNode>();
 438
 439         int padding_mode;
 440         if (padMode.empty())
 441         {
 442             padding_mode = vkcom::kPaddingModeCaffe;
 443         }
 444         else if (padMode == "VALID")
 445         {
 446             padding_mode = vkcom::kPaddingModeValid;
 447         }
 448         else if (padMode == "SAME")
 449         {
 450             padding_mode = vkcom::kPaddingModeSame;
 451         }
 452         else
 453             CV_Error(Error::StsError, "Unsupported padding mode " + padMode);
 454
 455         std::shared_ptr<vkcom::OpBase> op(new vkcom::OpConv(out_channel, has_bias,
 456                     filter_size, pad_size,
 457                     stride_size, dilation_size,
 458                     activation, group,
 459                     padding_mode));
 460
 461         std::vector<Ptr<BackendWrapper> > blobsWrapper;
 462
 463         if (fusedWeights)
 464         {
 465             Mat wm;
 466             weightsMat.copyTo(wm); // to handle the case of isContinuous() == false
 467             wm = wm.reshape(1, blobs[0].dims, blobs[0].size);
 468             blobsWrapper.push_back(Ptr<BackendWrapper>(new VkComBackendWrapper(wm)));
 469         }
 470         else
 471         {
 472             blobsWrapper.push_back(Ptr<BackendWrapper>(new VkComBackendWrapper(blobs[0])));
 473         }
 474
 475         if (has_bias)
 476         {
 477             Mat biasesMat({out_channel}, CV_32F, &biasvec[0]);
 478             blobsWrapper.push_back(Ptr<BackendWrapper>(new VkComBackendWrapper(biasesMat)));
 479         }
 480
 481         return Ptr<BackendNode>(new VkComBackendNode(inputs, op, blobsWrapper));
 482 #endif  // HAVE_VULKAN
 483         return Ptr<BackendNode>();
 484     }
 485
 486
 487
 488     virtual Ptr<BackendNode> initHalide(const std::vector<Ptr<BackendWrapper> > &inputs) CV_OVERRIDE
 489     {
 490 #ifdef HAVE_HALIDE
 491         Halide::Buffer<float> inputBuffer = halideBuffer(inputs[0]);
 492
 493         const int inpCn = inputBuffer.channels();
 494         const int outCn = blobs[0].size[0];
 495         const int inpGroupCn = blobs[0].size[1];
 496         const int group = inpCn / inpGroupCn;
 497         const int outGroupCn = outCn / group;
 498
 499         Halide::Buffer<float> weights = wrapToHalideBuffer(blobs[0]);
 500
 501         Halide::Var x("x"), y("y"), c("c"), n("n");
 502         Halide::Func top = (name.empty() ? Halide::Func() : Halide::Func(name));
 503         Halide::Func padded_input(name + "_constant_exterior");
 504         if (pad.width || pad.height)
 505         {
 506             Halide::Func bounded =
 507                 Halide::BoundaryConditions::constant_exterior(inputBuffer, 0);
 508             padded_input(x, y, c, n) = bounded(x, y, c, n);
 509         }
 510         else
 511         {
 512             padded_input(x, y, c, n) = inputBuffer(x, y, c, n);
 513         }
 514
 515         Halide::RDom r(0, kernel.width, 0, kernel.height, 0, inpGroupCn);
 516         Halide::Expr kx = x * stride.width - pad.width + r.x * dilation.width;
 517         Halide::Expr ky = y * stride.height - pad.height + r.y * dilation.height;
 518         Halide::Expr kc = r.z;
 519         for (int i = 1; i < group; ++i)
 520         {
 521             kc = select(c < outGroupCn * i, kc, inpGroupCn * i + r.z);
 522         }
 523         Halide::Expr topExpr = sum(padded_input(kx, ky, kc, n) *
 524                                    weights(r.x, r.y, r.z, c));
 525         if (hasBias())
 526         {
 527             Halide::Buffer<float> bias = wrapToHalideBuffer(blobs[1], {outCn});
 528             topExpr += bias(c);
 529         }
 530         top(x, y, c, n) = topExpr;
 531         return Ptr<BackendNode>(new HalideBackendNode({ padded_input, top }));
 532 #endif  // HAVE_HALIDE
 533         return Ptr<BackendNode>();
 534     }
 535
 536 #ifdef HAVE_INF_ENGINE
 537     virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> > &inputs) CV_OVERRIDE
 538     {
 539         InferenceEngine::DataPtr input = infEngineDataNode(inputs[0]);
 540         CV_Assert(input->dims.size() == 4 || input->dims.size() == 5);
 541
 542         const int inpCn = input->dims[input->dims.size() - 2];  // NOTE: input->dims are reversed (WHIO or WHDIO)
 543         const int outCn = blobs[0].size[0];
 544         const int inpGroupCn = blobs[0].size[1];
 545         const int group = inpCn / inpGroupCn;
 546
 547         InferenceEngine::Layout layout = (input->dims.size() == 4) ? InferenceEngine::Layout::OIHW :
 548                                                                      InferenceEngine::Layout::NCDHW;
 549
 550         auto ieWeights = wrapToInfEngineBlob(blobs[0], layout);
 551         if (fusedWeights)
 552         {
 553             if (weightsMat.isContinuous())
 554             {
 555                 Mat cvWeights = weightsMat.reshape(1, blobs[0].dims, blobs[0].size);
 556                 ieWeights = wrapToInfEngineBlob(cvWeights, layout);
 557             }
 558             else
 559             {
 560                 ieWeights = InferenceEngine::make_shared_blob<float>(
 561                                     InferenceEngine::Precision::FP32, layout,
 562                                     ieWeights->dims());
 563                 ieWeights->allocate();
 564
 565                 Mat newWeights = infEngineBlobToMat(ieWeights).reshape(1, outCn);
 566                 Mat cvWeights = weightsMat.colRange(0, newWeights.cols);
 567                 cvWeights.copyTo(newWeights);
 568             }
 569         }
 570         InferenceEngine::Blob::Ptr ieBiases;
 571         if (hasBias() || fusedBias)
 572         {
 573             Mat biasesMat({outCn}, CV_32F, &biasvec[0]);
 574             ieBiases = wrapToInfEngineBlob(biasesMat, {(size_t)outCn}, InferenceEngine::Layout::C);
 575         }
 576
 577         InferenceEngine::Builder::ConvolutionLayer ieLayer(name);
 578
 579         ieLayer.setKernel(kernel_size);
 580         ieLayer.setStrides(strides);
 581         ieLayer.setDilation(dilations);
 582         ieLayer.setPaddingsBegin(pads_begin);
 583         ieLayer.setPaddingsEnd(pads_end);
 584         ieLayer.setGroup((size_t)group);
 585         ieLayer.setOutDepth((size_t)outCn);
 586
 587         InferenceEngine::Builder::Layer l = ieLayer;
 588         addConstantData("weights", ieWeights, l);
 589         if (ieBiases)
 590             addConstantData("biases", ieBiases, l);
 591
 592         if (!padMode.empty())
 593             l.getParameters()["auto_pad"] = padMode == "VALID" ? std::string("valid") : std::string("same_upper");
 594
 595         return Ptr<BackendNode>(new InfEngineBackendNode(l));
 596     }
 597 #endif  // HAVE_INF_ENGINE
 598
 599     class ParallelConv : public cv::ParallelLoopBody
 600     {
 601     public:
 602         enum { BLK_SIZE = 32, BLK_SIZE_CN = 64 };
 603
 604         const Mat* input_;
 605         const Mat* weights_;
 606         Mat* output_;
 607         int outShape[4];
 608         Size kernel_, pad_, stride_, dilation_;
 609         int ngroups_, nstripes_;
 610         std::vector<int> ofstab_;
 611         const std::vector<float>* biasvec_;
 612         const std::vector<float>* reluslope_;
 613         const ActivationLayer* activ_;
 614         bool is1x1_;
 615         bool useAVX;
 616         bool useAVX2;
 617         bool useAVX512;
 618
 619         ParallelConv()
 620             : input_(0), weights_(0), output_(0), ngroups_(0), nstripes_(0),
 621               biasvec_(0), reluslope_(0), activ_(0), is1x1_(false), useAVX(false), useAVX2(false), useAVX512(false)
 622         {}
 623
 624         static void run( const Mat& input, Mat& output, const Mat& weights,
 625                          const std::vector<float>& biasvec,
 626                          const std::vector<float>& reluslope,
 627                          Size kernel, Size pad, Size stride, Size dilation,
 628                          const ActivationLayer* activ, int ngroups, int nstripes )
 629         {
 630             CV_Assert_N(
 631                        input.dims == 4 && output.dims == 4,
 632                        input.size[0] == output.size[0],
 633                        weights.rows == output.size[1],
 634                        weights.cols == (input.size[1]/ngroups)*kernel.width*kernel.height,
 635                        input.type() == output.type(),
 636                        input.type() == weights.type(),
 637                        input.type() == CV_32FC1,
 638                        input.isContinuous(),
 639                        output.isContinuous(),
 640                        biasvec.size() == (size_t)output.size[1]+2);
 641             ParallelConv p;
 642
 643             p.input_ = &input;
 644             p.weights_ = &weights;
 645             p.output_ = &output;
 646             for( int i = 0; i < 4; i++ ) p.outShape[i] = output.size[i];
 647             p.outShape[1] /= ngroups;
 648             p.kernel_ = kernel; p.pad_ = pad; p.stride_ = stride; p.dilation_ = dilation;
 649             p.ngroups_ = ngroups;
 650             p.nstripes_ = nstripes;
 651
 652             int inpCnAll = input.size[1], width = input.size[3], height = input.size[2];
 653             int inpCn = inpCnAll / ngroups;
 654             p.is1x1_ = kernel == Size(1,1) && pad == Size(0, 0);
 655             p.useAVX = checkHardwareSupport(CPU_AVX);
 656             p.useAVX2 = checkHardwareSupport(CPU_AVX2);
 657             p.useAVX512 = CV_CPU_HAS_SUPPORT_AVX512_SKX;
 658
 659             int ncn = std::min(inpCn, (int)BLK_SIZE_CN);
 660             p.ofstab_.resize(kernel.width*kernel.height*ncn);
 661             int* ofstab = &p.ofstab_[0];
 662
 663             for( int k = 0; k < ncn; k++ )
 664                 for( int k_r = 0; k_r < kernel.height; k_r++ )
 665                     for( int k_c = 0; k_c < kernel.width; k_c++ )
 666                         ofstab[(k*kernel.height + k_r)*kernel.width + k_c] =
 667                         (k*height + k_r*dilation.height)*width + k_c*dilation.width;
 668
 669             p.biasvec_ = &biasvec;
 670             p.reluslope_ = &reluslope;
 671             p.activ_ = p.reluslope_->empty() ? activ : 0;
 672
 673             parallel_for_(Range(0, nstripes), p, nstripes);
 674         }
 675
 676         virtual void operator ()(const Range &r0) const CV_OVERRIDE
 677         {
 678             const int valign = ConvolutionLayerImpl::VEC_ALIGN;
 679             int ngroups = ngroups_, batchSize = input_->size[0]*ngroups;
 680             int outW = output_->size[3], outH = output_->size[2], outCn = output_->size[1]/ngroups;
 681             int width = input_->size[3], height = input_->size[2], inpCn = input_->size[1]/ngroups;
 682             const int nstripes = nstripes_;
 683             int kernel_w = kernel_.width, kernel_h = kernel_.height;
 684             int pad_w = pad_.width, pad_h = pad_.height;
 685             int stride_w = stride_.width, stride_h = stride_.height;
 686             int dilation_w = dilation_.width, dilation_h = dilation_.height;
 687             int karea = kernel_w*kernel_h;
 688             int i, j, k;
 689             size_t inpPlaneSize = width*height;
 690             size_t outPlaneSize = outW*outH;
 691             bool is1x1 = is1x1_;
 692
 693             int stripesPerSample;
 694             size_t stripeSize;
 695             Range r = r0;
 696
 697             if( nstripes >= batchSize*2 )
 698             {
 699                 stripesPerSample = nstripes/batchSize;
 700                 stripeSize = alignSize((outPlaneSize + stripesPerSample - 1)/stripesPerSample, valign);
 701                 stripeSize = std::min(stripeSize, outPlaneSize);
 702             }
 703             else
 704             {
 705                 stripesPerSample = 1;
 706                 int samplesPerStripe = std::max((batchSize + nstripes - 1)/nstripes, 1);
 707                 r.start *= samplesPerStripe;
 708                 r.end *= samplesPerStripe;
 709                 stripeSize = outPlaneSize;
 710             }
 711
 712             const float* data_inp0_ = input_->ptr<float>();
 713             const int* ofstab = &ofstab_[0];
 714             const float* wptr_orig_ = weights_->ptr<float>();
 715             size_t wstep = weights_->step1();
 716             const float* biasptr_ = &biasvec_->at(0);
 717             const float* reluptr_ = reluslope_->empty() ? 0 : &reluslope_->at(0);
 718             float* data_out0_ = output_->ptr<float>();
 719             size_t rowbufsz = (size_t)karea*BLK_SIZE_CN*BLK_SIZE;
 720             AutoBuffer<float> rowbuf0_(rowbufsz + valign);
 721             float* rowbuf0 = alignPtr(rowbuf0_.data(), (int)(valign*sizeof(float)));
 722
 723             // we clear the buffer once; ultimately, it lets us to avoid
 724             // tail processing after running the unrolled/vectorized loop.
 725             // the main idea is to make sure that the tail (a.k.a. padding) of each row
 726             // (i.e. the elements with indices between vsz=karea*ncn and vsz_a)
 727             // does not contain NaNs or Infs. Because the padding in the weights
 728             // matrix is explicitly initialized with 0's, we handle all other
 729             // cases nicely, i.e. we can skip expliciting re-initialization
 730             // of the padding - we just retain elements from the previous iteration
 731             // of the loop over channels (cn0).
 732             memset(rowbuf0, 0, rowbufsz*sizeof(rowbuf0[0]) );
 733
 734             for( int stripe = r.start; stripe < r.end; stripe++ )
 735             {
 736                 int subsampleIdx = stripe/stripesPerSample;
 737                 if( subsampleIdx >= batchSize )
 738                     break;
 739                 int stripeStart = (int)((stripe - subsampleIdx*stripesPerSample)*stripeSize);
 740                 int stripeEnd = (int)std::min(stripeStart + stripeSize, outPlaneSize);
 741                 const float* data_inp0 = data_inp0_ + subsampleIdx*inpPlaneSize*inpCn;
 742                 float* data_out0 = data_out0_ + subsampleIdx*outPlaneSize*outCn;
 743                 int startOutCn = (subsampleIdx % ngroups)*outCn;
 744                 const float* wptr_orig = wptr_orig_ + wstep*startOutCn;
 745                 const float* biasptr = biasptr_ + startOutCn;
 746
 747                 for( int cn0 = 0; cn0 < inpCn; cn0 += BLK_SIZE_CN )
 748                 {
 749                     int cn1 = std::min(cn0 + BLK_SIZE_CN, inpCn);
 750                     int ncn = cn1 - cn0, vsz = karea*ncn;
 751                     int vsz_a = (int)alignSize(vsz, valign);
 752                     const float* wptr = wptr_orig + cn0*karea;
 753                     // we apply [Channels][P]ReLU (if any) during the final pass only.
 754                     const float* relu = cn1 == inpCn && reluptr_ ? reluptr_ + startOutCn : 0;
 755
 756                     for( int ofs0 = stripeStart; ofs0 < stripeEnd; ofs0 += BLK_SIZE )
 757                     {
 758                         int ofs, ofs1 = std::min(ofs0 + BLK_SIZE, stripeEnd);
 759                         int out_i = ofs0 / outW;
 760                         int out_j = ofs0 - out_i * outW;
 761
 762                         // do im2row for a part of input tensor
 763                         float* rowbuf = rowbuf0;
 764                         for( ofs = ofs0; ofs < ofs1; out_j = 0, ++out_i )
 765                         {
 766                             int delta = std::min(ofs1 - ofs, outW - out_j);
 767                             int out_j1 = out_j + delta;
 768                             int in_i = out_i * stride_h - pad_h;
 769                             int in_j = out_j * stride_w - pad_w;
 770                             const float* imgptr = data_inp0 + (cn0*height + in_i)*width + in_j;
 771                             ofs += delta;
 772
 773                             // do im2row for a part of input tensor
 774                             if( is1x1 )
 775                             {
 776                                 for( ; out_j < out_j1; out_j++, rowbuf += vsz_a, imgptr += stride_w )
 777                                 {
 778                                     for( k = 0; k < vsz; k++ )
 779                                         rowbuf[k] = imgptr[k*inpPlaneSize];
 780                                 }
 781                             }
 782                             else
 783                             {
 784                                 bool ok_i = 0 <= in_i && in_i < height - (kernel_h-1)*dilation_h;
 785                                 int i0 = std::max(0, (-in_i + dilation_h-1)/dilation_h);
 786                                 int i1 = std::min(kernel_h, (height - in_i + dilation_h-1)/dilation_h);
 787
 788                                 for( ; out_j < out_j1; out_j++, rowbuf += vsz_a, imgptr += stride_w, in_j += stride_w )
 789                                 {
 790                                     // this condition should be true for most of the tensor elements, i.e.
 791                                     // most of the time the kernel aperture is inside the tensor X-Y plane.
 792                                     if( ok_i && out_j + 2 <= out_j1 && 0 <= in_j && in_j + stride_w*2 <= width - (kernel_w-1)*dilation_w )
 793                                     {
 794                                         for( k = 0; k < vsz; k++ )
 795                                         {
 796                                             int k1 = ofstab[k];
 797                                             float v0 = imgptr[k1];
 798                                             float v1 = imgptr[k1 + stride_w];
 799                                             rowbuf[k] = v0;
 800                                             rowbuf[k+vsz_a] = v1;
 801                                         }
 802                                         out_j++;
 803                                         rowbuf += vsz_a;
 804                                         imgptr += stride_w;
 805                                         in_j += stride_w;
 806                                     }
 807                                     else
 808                                     {
 809                                         int j0 = std::max(0, (-in_j + dilation_w-1)/dilation_w);
 810                                         int j1 = std::min(kernel_w, (width - in_j + dilation_w-1)/dilation_w);
 811
 812                                         // here some non-continuous sub-row of the row will not be
 813                                         // filled from the tensor; we need to make sure that the uncovered
 814                                         // elements are explicitly set to 0's. the easiest way is to
 815                                         // set all the elements to 0's before the loop.
 816                                         memset(rowbuf, 0, vsz*sizeof(rowbuf[0]));
 817                                         for( k = 0; k < ncn; k++ )
 818                                         {
 819                                             for( i = i0; i < i1; i++ )
 820                                             {
 821                                                 for( j = j0; j < j1; j++ )
 822                                                 {
 823                                                     int imgofs = k*(width*height) + i*(dilation_h*width) + j*dilation_w;
 824                                                     rowbuf[(k*kernel_h + i)*kernel_w + j] = imgptr[imgofs];
 825                                                 }
 826                                             }
 827                                         }
 828                                     }
 829                                 }
 830                             }
 831                         }
 832
 833                         // now compute dot product of the weights
 834                         // and im2row-transformed part of the tensor
 835                         int bsz = ofs1 - ofs0;
 836                     #if CV_TRY_AVX512_SKX
 837                         /* AVX512 convolution requires an alignment of 16, and ROI is only there for larger vector sizes */
 838                         if(useAVX512)
 839                             opt_AVX512_SKX::fastConv(wptr, wstep, biasptr, rowbuf0, data_out0 + ofs0,
 840                                           outShape, bsz, vsz, vsz_a, relu, cn0 == 0);
 841                         else
 842                     #endif
 843                     #if CV_TRY_AVX2
 844                         if(useAVX2)
 845                             opt_AVX2::fastConv(wptr, wstep, biasptr, rowbuf0, data_out0 + ofs0,
 846                                           outShape, bsz, vsz, vsz_a, relu, cn0 == 0);
 847                         else
 848                     #endif
 849                     #if CV_TRY_AVX
 850                         if(useAVX)
 851                             opt_AVX::fastConv(wptr, wstep, biasptr, rowbuf0, data_out0 + ofs0,
 852                                          outShape, bsz, vsz, vsz_a, relu, cn0 == 0);
 853                         else
 854                     #endif
 855                         for( int i = 0; i < outCn; i += 2 )
 856                         {
 857                             const float* wptr0 = wptr + i*wstep;
 858                             const float* wptr1 = wptr0 + wstep;
 859                             float* outptr0 = data_out0 + ofs0 + i*outPlaneSize;
 860                             float* outptr1 = outptr0 + outPlaneSize;
 861                             float bias0 = biasptr[i], bias1 = biasptr[i+1];
 862                             float r0 = 1.f, r1 = 1.f;
 863
 864                             if( i+1 >= outCn )
 865                             {
 866                                 wptr1 = wptr0;
 867                                 outptr1 = outptr0;
 868                                 bias1 = bias0;
 869                             }
 870
 871                             if( relu )
 872                             {
 873                                 r0 = relu[i]; r1 = relu[i+1];
 874                                 if( i+1 >= outCn )
 875                                     r1 = r0;
 876                             }
 877
 878                             int j = 0;
 879                         #if CV_SIMD128
 880                             v_float32x4 vr0 = v_setall_f32(r0), vr1 = v_setall_f32(r1), z = v_setzero_f32();
 881
 882                             for( ; j <= bsz - 4; j += 4 )
 883                             {
 884                                 const float* rptr = rowbuf0 + j*vsz_a;
 885                                 v_float32x4 s0, s1;
 886
 887                                 if( cn0 == 0 )
 888                                 {
 889                                     s0 = v_setall_f32(bias0);
 890                                     s1 = v_setall_f32(bias1);
 891                                 }
 892                                 else
 893                                 {
 894                                     s0 = v_load(outptr0 + j);
 895                                     s1 = v_load(outptr1 + j);
 896                                 }
 897
 898                                 v_float32x4 vs00 = v_setzero_f32(), vs01 = v_setzero_f32(),
 899                                             vs02 = v_setzero_f32(), vs03 = v_setzero_f32(),
 900                                             vs10 = v_setzero_f32(), vs11 = v_setzero_f32(),
 901                                             vs12 = v_setzero_f32(), vs13 = v_setzero_f32();
 902                                 for( k = 0; k < vsz; k += 4, rptr += 4 )
 903                                 {
 904                                     v_float32x4 w0 = v_load_aligned(wptr0 + k), w1 = v_load_aligned(wptr1 + k);
 905                                     v_float32x4 r0 = v_load_aligned(rptr), r1 = v_load_aligned(rptr + vsz_a),
 906                                                 r2 = v_load_aligned(rptr + vsz_a*2), r3 = v_load_aligned(rptr + vsz_a*3);
 907
 908                                     vs00 += w0*r0;
 909                                     vs01 += w0*r1;
 910                                     vs02 += w0*r2;
 911                                     vs03 += w0*r3;
 912
 913                                     vs10 += w1*r0;
 914                                     vs11 += w1*r1;
 915                                     vs12 += w1*r2;
 916                                     vs13 += w1*r3;
 917                                 }
 918                                 s0 += v_reduce_sum4(vs00, vs01, vs02, vs03);
 919                                 s1 += v_reduce_sum4(vs10, vs11, vs12, vs13);
 920                                 if( relu )
 921                                 {
 922                                     s0 = v_select(s0 > z, s0, s0*vr0);
 923                                     s1 = v_select(s1 > z, s1, s1*vr1);
 924                                 }
 925
 926                                 v_store(outptr0 + j, s0);
 927                                 v_store(outptr1 + j, s1);
 928                             }
 929                         #endif
 930                             for( ; j < bsz; j++ )
 931                             {
 932                                 const float* rptr = rowbuf0 + j*vsz_a;
 933                                 float s00, s10;
 934
 935                                 if( cn0 == 0 )
 936                                 {
 937                                     s00 = bias0;
 938                                     s10 = bias1;
 939                                 }
 940                                 else
 941                                 {
 942                                     s00 = outptr0[j];
 943                                     s10 = outptr1[j];
 944                                 }
 945
 946                                 for( k = 0; k < vsz; k++ )
 947                                 {
 948                                     float r0 = rptr[k];
 949                                     s00 += wptr0[k]*r0;
 950                                     s10 += wptr1[k]*r0;
 951                                 }
 952                                 if( relu )
 953                                 {
 954                                     s00 = s00 > 0.f ? s00 : s00*r0;
 955                                     s10 = s10 > 0.f ? s10 : s10*r1;
 956                                 }
 957
 958                                 outptr0[j] = s00;
 959                                 outptr1[j] = s10;
 960                             }
 961                         }
 962                     }
 963                 }
 964
 965                 if( activ_ )
 966                     activ_->forwardSlice(data_out0 + stripeStart, data_out0 + stripeStart,
 967                                          (int)(stripeEnd - stripeStart),
 968                                          outPlaneSize, startOutCn, startOutCn + outCn);
 969             }
 970         }
 971     };
 972
 973 #ifdef HAVE_OPENCL
 974     bool forward_ocl(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
 975     {
 976         std::vector<UMat> inputs;
 977         std::vector<UMat> outputs;
 978
 979         bool use_half = (inps.depth() == CV_16S);
 980         inps.getUMatVector(inputs);
 981         outs.getUMatVector(outputs);
 982
 983         CV_Assert(outputs.size() == 1);
 984         for (int i = 0; i < inputs.size(); ++i)
 985             CV_Assert(inputs[i].u != outputs[0].u);
 986
 987         if (umat_blobs.empty())
 988         {
 989             size_t n = blobs.size();
 990             umat_blobs.resize(n);
 991             for (size_t i = 0; i < n; i++)
 992             {
 993                 blobs[i].copyTo(umat_blobs[i]);
 994             }
 995         }
 996
 997         if (convolutionOp.empty())
 998         {
 999             OCL4DNNConvConfig config;
1000             config.in_shape = shape(inputs[0]);
1001             config.out_shape = shape(outputs[0]);
1002             config.kernel = kernel;
1003             config.pad = pad;
1004             config.stride = stride;
1005             config.dilation = dilation;
1006             config.group = inputs[0].size[1] / umat_blobs[0].size[1];
1007             config.bias_term = (hasBias()) ? true : false;
1008             config.use_half = use_half;
1009
1010             convolutionOp = Ptr<OCL4DNNConvSpatial<float> >(new OCL4DNNConvSpatial<float>(config));
1011         }
1012
1013         int outCn = umat_blobs[0].size[0];
1014
1015         reluslope.clear();
1016         if( activ )
1017         {
1018             Ptr<ReLULayer> activ_relu = activ.dynamicCast<ReLULayer>();
1019             if( !activ_relu.empty() )
1020             {
1021                 reluslope.assign(outCn+2, activ_relu->negativeSlope);
1022                 activType = OCL4DNN_CONV_FUSED_ACTIV_RELU;
1023             }
1024
1025             Ptr<ReLU6Layer> activ_relu6 = activ.dynamicCast<ReLU6Layer>();
1026             if( !activ_relu6.empty() )
1027             {
1028                 reluslope.resize(2);
1029                 reluslope[0] = activ_relu6->minValue;
1030                 reluslope[1] = activ_relu6->maxValue;
1031                 activType = OCL4DNN_CONV_FUSED_ACTIV_RELU6;
1032             }
1033
1034             Ptr<ChannelsPReLULayer> activ_chprelu = activ.dynamicCast<ChannelsPReLULayer>();
1035             if( !activ_chprelu.empty() )
1036             {
1037                 const Mat& m = activ_chprelu->blobs[0];
1038                 CV_Assert(m.isContinuous() && m.type() == CV_32F && (int)m.total() == outCn);
1039                 const float* mdata = m.ptr<float>();
1040                 reluslope.resize(outCn+2);
1041                 std::copy(mdata, mdata + outCn, reluslope.begin());
1042                 reluslope[outCn] = reluslope[outCn+1] = reluslope[outCn-1];
1043                 activType = OCL4DNN_CONV_FUSED_ACTIV_PRELU;
1044             }
1045         }
1046
1047         if (fusedWeights)
1048         {
1049             weightsMat.copyTo(umat_blobs[0]);
1050             fusedWeights = false;
1051         }
1052         if (fusedBias)
1053         {
1054             if ( umat_blobs.size() < 2 )
1055                 umat_blobs.resize(2);
1056             umat_blobs[1] = UMat(biasvec, true);
1057             convolutionOp->setBias(true);
1058             fusedBias = false;
1059         }
1060
1061         if ( newActiv )
1062         {
1063             if ( activType == OCL4DNN_CONV_FUSED_ACTIV_RELU )
1064             {
1065                 CV_Assert(!reluslope.empty());
1066                 convolutionOp->setActivReLU(true, reluslope[0]);
1067             }
1068             else if ( activType == OCL4DNN_CONV_FUSED_ACTIV_PRELU)
1069             {
1070                 CV_Assert(!reluslope.empty());
1071                 convolutionOp->setActivPReLU(true, reluslope);
1072             }
1073             else if ( activType == OCL4DNN_CONV_FUSED_ACTIV_POWER)
1074             {
1075                 convolutionOp->setActivPower(true, power);
1076             }
1077             else if ( activType == OCL4DNN_CONV_FUSED_ACTIV_TANH)
1078             {
1079                 convolutionOp->setActivTanh(true);
1080             }
1081             else if ( activType == OCL4DNN_CONV_FUSED_ACTIV_RELU6)
1082             {
1083                 convolutionOp->setActivReLU6(true, reluslope[0], reluslope[1]);
1084             }
1085             else
1086             {
1087                 convolutionOp->setActivReLU(false, 0);
1088                 convolutionOp->setActivPReLU(false, reluslope);
1089                 convolutionOp->setActivPower(false, 1.f);
1090                 convolutionOp->setActivTanh(false);
1091                 convolutionOp->setActivReLU6(false, 0, 0);
1092             }
1093             newActiv = false;
1094         }
1095
1096         UMat& inpMat = inputs[0];
1097         UMat& outMat = outputs[0];
1098         int batch_size = inpMat.size[0];
1099
1100         return convolutionOp->Forward(inpMat,
1101                                       inputs.size() == 2 ? inputs[1] : UMat(),
1102                                       umat_blobs[0],
1103                                       umat_blobs.size() > 1 ? umat_blobs[1] : UMat(),
1104                                       outMat,
1105                                       batch_size);
1106     }
1107 #endif
1108
1109     void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr) CV_OVERRIDE
1110     {
1111         CV_TRACE_FUNCTION();
1112         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
1113
1114         CV_OCL_RUN(IS_DNN_OPENCL_TARGET(preferableTarget),
1115                    forward_ocl(inputs_arr, outputs_arr, internals_arr))
1116
1117         if (inputs_arr.depth() == CV_16S)
1118         {
1119             forward_fallback(inputs_arr, outputs_arr, internals_arr);
1120             return;
1121         }
1122
1123         std::vector<Mat> inputs, outputs;
1124         inputs_arr.getMatVector(inputs);
1125         outputs_arr.getMatVector(outputs);
1126
1127         /*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",
1128                name.c_str(), inputs[0].size[0], inputs[0].size[1], inputs[0].size[2], inputs[0].size[3],
1129                kernel.width, kernel.height, pad.width, pad.height,
1130                stride.width, stride.height, dilation.width, dilation.height);*/
1131         CV_Assert_N(inputs.size() == (size_t)1, inputs[0].size[1] % blobs[0].size[1] == 0,
1132                     outputs.size() == 1, inputs[0].data != outputs[0].data);
1133
1134         if (inputs[0].dims == 5) {
1135             CV_Error(Error::StsNotImplemented, "Convolution3D layer is not supported on OCV backend");
1136         }
1137
1138         int ngroups = inputs[0].size[1]/blobs[0].size[1];
1139         CV_Assert(outputs[0].size[1] % ngroups == 0);
1140         int outCn = blobs[0].size[0];
1141
1142         reluslope.clear();
1143         if( activ )
1144         {
1145             Ptr<ReLULayer> activ_relu = activ.dynamicCast<ReLULayer>();
1146             if( !activ_relu.empty() )
1147             {
1148                 reluslope.assign(outCn+2, activ_relu->negativeSlope);
1149             }
1150
1151             Ptr<ChannelsPReLULayer> activ_chprelu = activ.dynamicCast<ChannelsPReLULayer>();
1152             if( !activ_chprelu.empty() )
1153             {
1154                 const Mat& m = activ_chprelu->blobs[0];
1155                 CV_Assert(m.isContinuous() && m.type() == CV_32F && (int)m.total() == outCn);
1156                 const float* mdata = m.ptr<float>();
1157                 reluslope.resize(outCn+2);
1158                 std::copy(mdata, mdata + outCn, reluslope.begin());
1159                 reluslope[outCn] = reluslope[outCn+1] = reluslope[outCn-1];
1160             }
1161         }
1162
1163         int nstripes = std::max(getNumThreads(), 1);
1164
1165         ParallelConv::run(inputs[0], outputs[0], weightsMat, biasvec, reluslope,
1166                           kernel, pad, stride, dilation, activ.get(), ngroups, nstripes);
1167     }
1168
1169     virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
1170                            const std::vector<MatShape> &outputs) const CV_OVERRIDE
1171     {
1172         CV_Assert(inputs.size() == outputs.size());
1173
1174         int64 flops = 0;
1175         for (int i = 0; i < inputs.size(); i++)
1176         {
1177             flops += total(outputs[i])*(CV_BIG_INT(2)*kernel.area()*inputs[i][1] + 1);
1178         }
1179
1180         return flops;
1181     }
1182 };
1183
1184 class DeConvolutionLayerImpl CV_FINAL : public BaseConvolutionLayerImpl
1185 {
1186 public:
1187     Mat weightsMat, biasesMat;
1188     UMat umat_weights;
1189     UMat umat_biases;
1190
1191     DeConvolutionLayerImpl(const LayerParams& params) : BaseConvolutionLayerImpl(params) {}
1192
1193     MatShape computeColRowShape(const MatShape &inpShape, const MatShape &outShape) const CV_OVERRIDE
1194     {
1195         int inpCn = inpShape[1];
1196         int inpH = inpShape[2];
1197         int inpW = inpShape[3];
1198         int outCn = outShape[1];
1199         int ngroups = inpCn / blobs[0].size[0];
1200         int outGroupCn = outCn / ngroups;
1201         int ksize = outGroupCn * kernel.height * kernel.width;
1202         return shape(ksize, inpH * inpW);
1203     }
1204
1205     virtual bool supportBackend(int backendId) CV_OVERRIDE
1206     {
1207 #ifdef HAVE_INF_ENGINE
1208         const int outGroupCn = blobs[0].size[1];  // Weights are in IOHW layout
1209         const int group = numOutput / outGroupCn;
1210
1211         if (backendId == DNN_BACKEND_INFERENCE_ENGINE)
1212         {
1213             if (kernel_size.size() == 3)
1214                 CV_Error(Error::StsNotImplemented, "Unsupported deconvolution3D layer");
1215
1216             if (adjustPad.height || adjustPad.width)
1217             {
1218                 if (padMode.empty())
1219                 {
1220                     if (preferableTarget != DNN_TARGET_CPU && group != 1)
1221                     {
1222                         if ((adjustPad.height && pad.height) || (adjustPad.width && pad.width))
1223                             return false;
1224                     }
1225                     return pad.width >= adjustPad.width && pad.height >= adjustPad.height;
1226                 }
1227                 else if (padMode == "SAME")
1228                 {
1229                     return kernel.width >= pad.width + 1 + adjustPad.width &&
1230                            kernel.height >= pad.height + 1 + adjustPad.height;
1231                 }
1232                 else if (padMode == "VALID")
1233                     return false;
1234             }
1235
1236             if (group != 1)
1237             {
1238                 return preferableTarget == DNN_TARGET_CPU;
1239             }
1240             if (preferableTarget == DNN_TARGET_OPENCL || preferableTarget == DNN_TARGET_OPENCL_FP16)
1241                 return dilation.width == 1 && dilation.height == 1;
1242             return true;
1243         }
1244         else
1245 #endif  // HAVE_INF_ENGINE
1246             return kernel_size.size() == 2 && (backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE);
1247     }
1248
1249     bool getMemoryShapes(const std::vector<MatShape> &inputs,
1250                          const int requiredOutputs,
1251                          std::vector<MatShape> &outputs,
1252                          std::vector<MatShape> &internals) const CV_OVERRIDE
1253     {
1254         CV_Assert(!hasBias() || blobs[1].total() == (size_t)numOutput);
1255         CV_Assert(inputs.size() != 0);
1256
1257         int outCn = numOutput;
1258         std::vector<int> outShape;
1259         outShape.push_back(inputs[0][0]);  // batch
1260         outShape.push_back(outCn);
1261         if (padMode.empty())
1262         {
1263             for (int i = 0; i < kernel_size.size(); i++)
1264                 outShape.push_back(strides[i] * (inputs[0][2 + i] - 1) + kernel_size[i] - pads_begin[i] - pads_end[i] + adjust_pads[i]);
1265         }
1266         else if (padMode == "VALID")
1267         {
1268             for (int i = 0; i < kernel_size.size(); i++)
1269                 outShape.push_back(strides[i] * (inputs[0][2 + i] - 1) + kernel_size[i] + adjust_pads[i]);
1270         }
1271         else if (padMode == "SAME")
1272         {
1273             for (int i = 0; i < kernel_size.size(); i++)
1274                 outShape.push_back(strides[i] * (inputs[0][2 + i] - 1) + 1 + adjust_pads[i]);
1275         }
1276         else
1277             CV_Error(Error::StsError, "Unsupported padding mode " + padMode);
1278
1279         CV_Assert(outCn % blobs[0].size[1] == 0);
1280         int ngroups = outCn / blobs[0].size[1];
1281
1282         int inpCn = inputs[0][1];
1283         CV_Assert(inpCn % ngroups == 0 && outCn % ngroups == 0);
1284         CV_Assert(blobs[0].size[0] == inpCn);
1285
1286         outputs.resize(1, outShape);
1287
1288         if (!is1x1())
1289             internals.push_back(computeColRowShape(inputs[0], outputs[0]));
1290
1291         return false;
1292     }
1293
1294     void finalize(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr) CV_OVERRIDE
1295     {
1296         BaseConvolutionLayerImpl::finalize(inputs_arr, outputs_arr);
1297
1298         std::vector<Mat> inputs, outputs;
1299         inputs_arr.getMatVector(inputs);
1300         outputs_arr.getMatVector(outputs);
1301
1302         std::vector<int> inpShape;
1303         std::vector<int> outShape;
1304         for (int i = 2; i < inputs[0].dims; i++) {
1305             inpShape.push_back(inputs[0].size[i]);
1306             outShape.push_back(outputs[0].size[i]);
1307         }
1308         getConvPoolPaddings(outShape, kernel_size, strides, padMode, pads_begin, pads_end);
1309         if (pads_begin.size() == 2) {
1310             for (int i = 0; i < pads_begin.size(); i++) {
1311                 if (pads_begin[i] != pads_end[i])
1312                     CV_Error(Error::StsNotImplemented, "Unsupported asymmetric padding in deconvolution layer");
1313             }
1314             pad = Size(pads_begin[1], pads_begin[0]);
1315         }
1316
1317         weightsMultipliers.assign(numOutput, 1.0);
1318         if (weightsMat.empty())
1319         {
1320             transpose(blobs[0].reshape(1, blobs[0].size[0]), weightsMat);
1321             biasesMat = hasBias() ? blobs[1].reshape(1, numOutput)
1322                                   : Mat::zeros(numOutput, 1, CV_32F);
1323         }
1324     }
1325
1326     void fuseWeights(const Mat& w_, const Mat& b_) CV_OVERRIDE
1327     {
1328         Mat w = w_.total() == 1 ? Mat(1, numOutput, CV_32F, Scalar(w_.at<float>(0))) : w_;
1329         Mat b = b_.total() == 1 ? Mat(1, numOutput, CV_32F, Scalar(b_.at<float>(0))) : b_;
1330
1331         CV_Assert_N(!weightsMat.empty(),
1332                      w.empty() || numOutput == w.total(),
1333                      b.empty() || numOutput == b.total());
1334
1335         if (!w.empty())
1336         {
1337             transpose(blobs[0].reshape(1, blobs[0].size[0]), weightsMat);
1338             weightsMat = weightsMat.reshape(1, numOutput);
1339             for (int i = 0; i < numOutput; ++i)
1340             {
1341                 double wi = w.at<float>(i);
1342                 weightsMultipliers[i] *= wi;
1343                 cv::multiply(weightsMat.row(i), weightsMultipliers[i], weightsMat.row(i));
1344                 biasesMat.at<float>(i) *= wi;
1345             }
1346             weightsMat = weightsMat.reshape(1, weightsMat.total() / blobs[0].size[0]);
1347         }
1348
1349         if (!b.empty())
1350         {
1351             cv::add(biasesMat, b.reshape(1, numOutput), biasesMat);
1352         }
1353     }
1354
1355     class MatMulInvoker : public ParallelLoopBody
1356     {
1357     public:
1358         MatMulInvoker(const Mat& a, const Mat& b, Mat& c, int nstripes)
1359         {
1360             a_ = &a;
1361             b_ = &b;
1362             c_ = &c;
1363             nstripes_ = nstripes;
1364             useAVX = checkHardwareSupport(CPU_AVX);
1365             useAVX2 = checkHardwareSupport(CPU_AVX2);
1366             useAVX512 = CV_CPU_HAS_SUPPORT_AVX512_SKX;
1367         }
1368
1369         void operator()(const Range& range_) const CV_OVERRIDE
1370         {
1371             int stripeSize = (int)alignSize((b_->cols + nstripes_ - 1)/nstripes_, 16);
1372             Range range(range_.start*stripeSize, std::min(range_.end*stripeSize, b_->cols));
1373             int mmax = a_->rows;
1374             int nmax = range.end - range.start;
1375             int kmax = a_->cols;
1376             int m, n, k;
1377             const float* aptr = a_->ptr<float>();
1378             const float* bptr = b_->ptr<float>() + range.start;
1379             float* cptr = c_->ptr<float>() + range.start;
1380             size_t astep = a_->step1();
1381             size_t bstep = b_->step1();
1382             size_t cstep = c_->step1();
1383
1384         #if CV_TRY_AVX512_SKX
1385             if( useAVX512 )
1386                 opt_AVX512_SKX::fastGEMM( aptr, astep, bptr, bstep, cptr, cstep, mmax, kmax, nmax );
1387             else
1388         #endif
1389         #if CV_TRY_AVX2
1390             if( useAVX2 )
1391                 opt_AVX2::fastGEMM( aptr, astep, bptr, bstep, cptr, cstep, mmax, kmax, nmax );
1392             else
1393         #endif
1394         #if CV_TRY_AVX
1395             if( useAVX )
1396                 opt_AVX::fastGEMM( aptr, astep, bptr, bstep, cptr, cstep, mmax, kmax, nmax );
1397             else
1398         #endif
1399             for( m = 0; m < mmax; m += 2 )
1400             {
1401                 float* dst0 = cptr + cstep*m;
1402                 float* dst1 = cptr + cstep*std::min(m+1, mmax-1);
1403                 const float* aptr0 = aptr + astep*m;
1404                 const float* aptr1 = aptr + astep*std::min(m+1, mmax-1);
1405
1406                 for( n = 0; n < nmax; n++ )
1407                 {
1408                     dst0[n] = 0.f;
1409                     dst1[n] = 0.f;
1410                 }
1411
1412                 for( k = 0; k < kmax; k += 4 )
1413                 {
1414                     float alpha00 = aptr0[k];
1415                     float alpha01 = aptr1[k];
1416                     float alpha10 = 0.f, alpha11 = 0.f;
1417                     float alpha20 = 0.f, alpha21 = 0.f;
1418                     float alpha30 = 0.f, alpha31 = 0.f;
1419                     const float* bptr0 = bptr + k*bstep;
1420                     const float* bptr1 = bptr0;
1421                     const float* bptr2 = bptr0;
1422                     const float* bptr3 = bptr0;
1423
1424                     if( k+1 < kmax )
1425                     {
1426                         alpha10 = aptr0[k+1];
1427                         alpha11 = aptr1[k+1];
1428                         bptr1 = bptr0 + bstep;
1429                         if( k+2 < kmax )
1430                         {
1431                             alpha20 = aptr0[k+2];
1432                             alpha21 = aptr1[k+2];
1433                             bptr2 = bptr1 + bstep;
1434                             if( k+3 < kmax )
1435                             {
1436                                 alpha30 = aptr0[k+3];
1437                                 alpha31 = aptr1[k+3];
1438                                 bptr3 = bptr2 + bstep;
1439                             }
1440                         }
1441                     }
1442                     n = 0;
1443
1444                 #if CV_SIMD128
1445                     v_float32x4 a00 = v_setall_f32(alpha00);
1446                     v_float32x4 a01 = v_setall_f32(alpha01);
1447                     v_float32x4 a10 = v_setall_f32(alpha10);
1448                     v_float32x4 a11 = v_setall_f32(alpha11);
1449                     v_float32x4 a20 = v_setall_f32(alpha20);
1450                     v_float32x4 a21 = v_setall_f32(alpha21);
1451                     v_float32x4 a30 = v_setall_f32(alpha30);
1452                     v_float32x4 a31 = v_setall_f32(alpha31);
1453
1454                     for( ; n <= nmax - 4; n += 4 )
1455                     {
1456                         v_float32x4 b0 = v_load(bptr0 + n);
1457                         v_float32x4 b1 = v_load(bptr1 + n);
1458                         v_float32x4 b2 = v_load(bptr2 + n);
1459                         v_float32x4 b3 = v_load(bptr3 + n);
1460                         v_float32x4 d0 = v_load(dst0 + n);
1461                         v_float32x4 d1 = v_load(dst1 + n);
1462                         d0 += b0*a00;
1463                         d1 += b0*a01;
1464                         d0 += b1*a10;
1465                         d1 += b1*a11;
1466                         d0 += b2*a20;
1467                         d1 += b2*a21;
1468                         d0 += b3*a30;
1469                         d1 += b3*a31;
1470                         v_store(dst0 + n, d0);
1471                         v_store(dst1 + n, d1);
1472                     }
1473                 #endif
1474
1475                     for( ; n < nmax; n++ )
1476                     {
1477                         float b0 = bptr0[n], b1 = bptr1[n];
1478                         float b2 = bptr2[n], b3 = bptr3[n];
1479                         float d0 = dst0[n] + alpha00*b0 + alpha10*b1 + alpha20*b2 + alpha30*b3;
1480                         float d1 = dst1[n] + alpha01*b0 + alpha11*b1 + alpha21*b2 + alpha31*b3;
1481                         dst0[n] = d0;
1482                         dst1[n] = d1;
1483                     }
1484                 }
1485             }
1486         }
1487
1488         const Mat *a_, *b_;
1489         Mat* c_;
1490         int nstripes_;
1491         bool useAVX;
1492         bool useAVX2;
1493         bool useAVX512;
1494     };
1495
1496     class Col2ImInvoker : public cv::ParallelLoopBody
1497     {
1498     public:
1499         const float* data_col;
1500         const float* biasvec;
1501         int channels, height, width;
1502         int kernel_h, kernel_w;
1503         int pad_h, pad_w;
1504         int stride_h, stride_w;
1505         float* data_im;
1506         int height_col, width_col;
1507         int nstripes;
1508         bool is1x1;
1509
1510         Col2ImInvoker()
1511             : data_col(0), biasvec(0), channels(0), height(0), width(0),
1512               kernel_h(0), kernel_w(0), pad_h(0), pad_w(0), stride_h(0), stride_w(0), data_im(0),
1513               height_col(0), width_col(0), nstripes(0), is1x1(0)
1514         {}
1515
1516         static void run(const float* data_col,
1517                         int channels, int height, int width,
1518                         int kernel_h, int kernel_w,
1519                         int pad_h, int pad_w,
1520                         int stride_h, int stride_w,
1521                         int height_col, int width_col,
1522                         float* data_im,
1523                         const float* biasvec,
1524                         bool is1x1)
1525         {
1526             const int nstripes = getNumThreads();
1527
1528             Col2ImInvoker t;
1529             t.data_col = data_col;
1530             t.data_im = data_im;
1531             t.channels = channels; t.height = height; t.width = width;
1532             t.kernel_h = kernel_h; t.kernel_w = kernel_w;
1533             t.pad_h = pad_h; t.pad_w = pad_w;
1534             t.stride_h = stride_h; t.stride_w = stride_w;
1535             t.height_col = height_col;
1536             t.width_col = width_col;
1537             t.nstripes = nstripes;
1538             t.is1x1 = is1x1;
1539             t.biasvec = biasvec;
1540
1541             parallel_for_(Range(0, nstripes), t, nstripes);
1542         }
1543
1544         virtual void operator ()(const Range &r) const CV_OVERRIDE
1545         {
1546             const float* data_col_ = data_col;
1547             float* data_im_ = data_im;
1548             int coeff_h = (1 - stride_h * kernel_w * height_col) * width_col;
1549             int coeff_w = (1 - stride_w * height_col * width_col);
1550             size_t total = (size_t)channels * height * width;
1551             size_t stripeSize = (total + nstripes - 1)/nstripes;
1552             size_t startIndex = r.start*stripeSize;
1553             size_t endIndex = std::min(r.end*stripeSize, total);
1554             int w = (int)(startIndex % width + pad_w);
1555             int h = (int)((startIndex / width) % height + pad_h);
1556             int c = (int)(startIndex / (width * height));
1557             int h_col_start = (h < kernel_h) ? 0 : (h - kernel_h) / stride_h + 1;
1558             int h_col_end = std::min(h / stride_h + 1, height_col);
1559             int plane_size_col = height_col * width_col;
1560             int offset = (c * kernel_h * kernel_w + h * kernel_w + w) * plane_size_col;
1561             bool is1x1_ = is1x1;
1562             const float* biasvec_ = biasvec;
1563
1564             for (size_t index = startIndex; index < endIndex; index++)
1565             {
1566                 // compute the start and end of the output
1567                 int w_col_start = (w < kernel_w) ? 0 : (w - kernel_w) / stride_w + 1;
1568                 int w_col_end = std::min(w / stride_w + 1, width_col);
1569                 float val;
1570
1571                 if( is1x1_ )
1572                     val = data_im_[index];
1573                 else
1574                 {
1575                     val = 0.f;
1576                     for (int h_col = h_col_start; h_col < h_col_end; ++h_col) {
1577                         for (int w_col = w_col_start; w_col < w_col_end; ++w_col) {
1578                             val += data_col_[offset + h_col * coeff_h + w_col * coeff_w];
1579                         }
1580                     }
1581                 }
1582                 data_im_[index] = val + biasvec_[c];
1583
1584                 offset += plane_size_col;
1585                 if( ++w >= width + pad_w )
1586                 {
1587                     w = (int)((index + 1)% width + pad_w);
1588                     h = (int)(((index + 1) / width) % height + pad_h);
1589                     c = (int)((index + 1) / (width * height));
1590                     h_col_start = (h < kernel_h) ? 0 : (h - kernel_h) / stride_h + 1;
1591                     h_col_end = std::min(h / stride_h + 1, height_col);
1592                     offset = (c * kernel_h * kernel_w + h * kernel_w + w) * plane_size_col;
1593                 }
1594             }
1595         }
1596     };
1597
1598 #ifdef HAVE_OPENCL
1599     bool forward_ocl(InputArrayOfArrays inputs_, OutputArrayOfArrays outputs_, OutputArrayOfArrays internals_)
1600     {
1601         std::vector<UMat> inputs;
1602         std::vector<UMat> outputs;
1603         std::vector<UMat> internals;
1604
1605         if (inputs_.depth() == CV_16S)
1606             return false;
1607
1608         inputs_.getUMatVector(inputs);
1609         outputs_.getUMatVector(outputs);
1610         internals_.getUMatVector(internals);
1611
1612         int outCn = numOutput;
1613         int inpCn = inputs[0].size[1];
1614
1615         if (is1x1())
1616             return false;
1617
1618         if (umat_weights.empty())
1619         {
1620             if (fusedWeights)
1621                 weightsMat.copyTo(umat_weights);
1622             else
1623                 transpose(blobs[0].reshape(1, inpCn), umat_weights);
1624
1625             if (fusedBias)
1626                 biasesMat.copyTo(umat_biases);
1627             else
1628             {
1629                 if (hasBias())
1630                     blobs[1].reshape(1, outCn).copyTo(umat_biases);
1631                 else
1632                     umat_biases = UMat::zeros(outCn, 1, CV_32F);
1633             }
1634         }
1635
1636         String buildopt = format("-DT=%s ", ocl::typeToStr(inputs[0].type()));
1637         buildopt += format("-DPAD_H=%d -DPAD_W=%d -DKERNEL_H=%d -DKERNEL_W=%d -DSTRIDE_H=%d -DSTRIDE_W=%d ",
1638                            pad.height, pad.width, kernel.height, kernel.width, stride.height, stride.width);
1639
1640         for (size_t ii = 0; ii < outputs.size(); ii++)
1641         {
1642             int ngroups = outCn / blobs[0].size[1];
1643             int inpGroupCn = inpCn / ngroups;
1644             int outGroupCn = blobs[0].size[1];
1645             const UMat& inp = inputs[ii];
1646             UMat& out = outputs[ii];
1647             int numImg = inp.size[0];
1648             int inpH = inp.size[2], inpW = inp.size[3];
1649             int outH = out.size[2], outW = out.size[3];
1650
1651             MatShape inpshape = shape(numImg*inpCn, inpH*inpW);
1652             MatShape outshape = shape(numImg*outCn, outH*outW);
1653             UMat convBlob = inputs[ii].reshape(1, inpshape.size(), &inpshape[0]);
1654             UMat decnBlob = out.reshape(1, outshape.size(), &outshape[0]);
1655             int rows = internals[0].rows / ngroups;
1656
1657             for (int n = 0; n < numImg; n++)
1658             {
1659                 for (int g = 0; g < ngroups; g++)
1660                 {
1661                     UMat colMat = internals[0].rowRange(_Range(g * rows, rows));
1662                     UMat convMat = convBlob.rowRange(_Range((g + n * ngroups) * inpGroupCn, inpGroupCn));
1663                     UMat wghtMat = umat_weights.colRange(_Range(g * inpGroupCn, inpGroupCn));
1664                     gemm(wghtMat, convMat, 1, noArray(), 0, colMat, 0);
1665                 }
1666
1667                 for (int g = 0; g < ngroups; g++)
1668                 {
1669                     int total = outGroupCn * decnBlob.cols;
1670                     int index = 0;
1671                     int height_col = inpH;
1672                     int width_col = inpW;
1673                     int coeff_h = (1 - stride.height * kernel.width * height_col) * width_col;
1674                     int coeff_w = (1 - stride.width * height_col * width_col);
1675
1676                     ocl::Kernel k("col2im", ocl::dnn::col2im_oclsrc, buildopt);
1677                     k.set(index++, total);
1678                     k.set(index++, ocl::KernelArg::PtrReadOnly(internals[0]));
1679                     k.set(index++, (int)(g * rows * internals[0].cols));
1680                     k.set(index++, outGroupCn);
1681                     k.set(index++, outH);
1682                     k.set(index++, outW);
1683                     k.set(index++, height_col);
1684                     k.set(index++, width_col);
1685                     k.set(index++, coeff_h);
1686                     k.set(index++, coeff_w);
1687                     k.set(index++, ocl::KernelArg::PtrReadOnly(umat_biases));
1688                     k.set(index++, (int)(g * outGroupCn * umat_biases.cols));
1689                     k.set(index++, ocl::KernelArg::PtrWriteOnly(decnBlob));
1690                     k.set(index++, (int)((g + n * ngroups) * outGroupCn * decnBlob.cols));
1691
1692                     size_t global[] = { (size_t)total };
1693                     bool ret = k.run(1, global, NULL, false);
1694                     if (!ret)
1695                         return false;
1696                 }
1697             }
1698         }
1699
1700         return true;
1701     }
1702 #endif
1703
1704     void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr) CV_OVERRIDE
1705     {
1706         CV_TRACE_FUNCTION();
1707         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
1708
1709         CV_OCL_RUN(IS_DNN_OPENCL_TARGET(preferableTarget),
1710                    forward_ocl(inputs_arr, outputs_arr, internals_arr));
1711
1712         if (inputs_arr.depth() == CV_16S)
1713         {
1714             forward_fallback(inputs_arr, outputs_arr, internals_arr);
1715             return;
1716         }
1717
1718         std::vector<Mat> inputs, outputs, internals;
1719         inputs_arr.getMatVector(inputs);
1720         outputs_arr.getMatVector(outputs);
1721         internals_arr.getMatVector(internals);
1722
1723         int outCn = numOutput;
1724         int inpCn = inputs[0].size[1];
1725         bool is1x1flag = is1x1();
1726         int nstripes = getNumThreads();
1727
1728         if( weightsMat.empty() )
1729         {
1730             transpose(blobs[0].reshape(1, inpCn), weightsMat);
1731             biasesMat = hasBias() ? blobs[1].reshape(1, outCn) : Mat::zeros(outCn, 1, CV_32F);
1732         }
1733
1734         for (size_t ii = 0; ii < outputs.size(); ii++)
1735         {
1736             int ngroups = outCn / blobs[0].size[1];
1737             int inpGroupCn = inpCn / ngroups;
1738             int outGroupCn = blobs[0].size[1];
1739             const Mat& inp = inputs[ii];
1740             Mat& out = outputs[ii];
1741             int numImg = inp.size[0];
1742             int inpH = inp.size[2], inpW = inp.size[3];
1743             int outH = out.size[2], outW = out.size[3];
1744
1745             Mat convBlob = inputs[ii].reshape(1, numImg*inpCn);
1746             Mat decnBlob = out.reshape(1, numImg*outCn);
1747
1748             for (int n = 0; n < numImg; n++)
1749             {
1750                 for (int g = 0; g < ngroups; g++)
1751                 {
1752                     Mat dstMat = decnBlob.rowRange(_Range((g + n * ngroups) * outGroupCn, outGroupCn));
1753                     Mat &colMat = is1x1flag ? dstMat : internals[0];
1754
1755                     Mat convMat = convBlob.rowRange(_Range((g + n * ngroups) * inpGroupCn, inpGroupCn));
1756                     Mat wghtMat = weightsMat.colRange(_Range(g * inpGroupCn, inpGroupCn));
1757                     Mat curBiasMat = biasesMat.rowRange(_Range(g * outGroupCn, outGroupCn));
1758
1759                     //gemm(wghtMat, convMat, 1, colMat, 0, colMat, 0);
1760                     MatMulInvoker mminvoker(wghtMat, convMat, colMat, nstripes);
1761                     parallel_for_(Range(0, nstripes), mminvoker, nstripes);
1762
1763                     Col2ImInvoker::run(colMat.ptr<float>(), outGroupCn, outH, outW,
1764                                        kernel.height, kernel.width, pad.height, pad.width,
1765                                        stride.height, stride.width, inpH, inpW, dstMat.ptr<float>(),
1766                                        curBiasMat.ptr<float>(), is1x1flag);
1767                 }
1768             }
1769         }
1770     }
1771
1772     virtual Ptr<BackendNode> initHalide(const std::vector<Ptr<BackendWrapper> > &inputs) CV_OVERRIDE
1773     {
1774 #ifdef HAVE_HALIDE
1775         Halide::Buffer<float> inputBuffer = halideBuffer(inputs[0]);
1776
1777         int inW, inH, inC, inN;
1778         getCanonicalSize(inputBuffer, &inW, &inH, &inC, &inN);
1779         const int outGroupCn = blobs[0].size[1];
1780         const int group = numOutput / outGroupCn;
1781         const int inpGroupCn = blobs[0].size[0] / group;
1782
1783         Halide::Var x("x"), y("y"), c("c"), n("n");
1784         Halide::Func top = (name.empty() ? Halide::Func() : Halide::Func(name));
1785         Halide::Func padded_input(name + "_constant_exterior");
1786         auto weights = wrapToHalideBuffer(blobs[0]);
1787
1788         Halide::Func dilated_input("dilated_input");
1789         dilated_input(x, y, c, n) = 0.0f;
1790         Halide::RDom r1(0, inW, 0, inH);
1791         dilated_input(r1.x * stride.width, r1.y * stride.height, c, n) =
1792               inputBuffer(r1.x, r1.y, c, n);
1793         dilated_input.compute_root();
1794
1795         Halide::Func bounded =
1796             Halide::BoundaryConditions::constant_exterior(dilated_input, 0,
1797                                                           0, (inW - 1) * stride.width + 1,
1798                                                           0, (inH - 1) * stride.height + 1,
1799                                                           0, inC, 0, inN);
1800         padded_input(x, y, c, n) = bounded(x, y, c, n);
1801
1802         Halide::RDom r(0, kernel.width, 0, kernel.height, 0, inpGroupCn);
1803         Halide::Expr kx = x + pad.width - r.x;
1804         Halide::Expr ky = y + pad.height - r.y;
1805         Halide::Expr kInC = r.z;
1806         Halide::Expr kOutC = c;
1807         for (int i = 1; i < group; ++i)
1808         {
1809             kInC = select(c < outGroupCn * i, kInC, inpGroupCn * i + r.z);
1810             kOutC = select(c < outGroupCn * i, kOutC, c - outGroupCn * i);
1811         }
1812         Halide::Expr topExpr = sum(padded_input(kx, ky, kInC, n) *
1813                                    weights(r.x, r.y, kOutC, kInC));
1814         if (hasBias())
1815         {
1816             auto bias = wrapToHalideBuffer(blobs[1], {numOutput});
1817             topExpr += bias(c);
1818         }
1819         top(x, y, c, n) = topExpr;
1820         return Ptr<BackendNode>(new HalideBackendNode({ padded_input, top }));
1821 #endif  // HAVE_HALIDE
1822         return Ptr<BackendNode>();
1823     }
1824
1825 #ifdef HAVE_INF_ENGINE
1826     virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> > &) CV_OVERRIDE
1827     {
1828         auto ieWeights = wrapToInfEngineBlob(blobs[0], InferenceEngine::Layout::OIHW);
1829         if (fusedWeights)
1830         {
1831             ieWeights = InferenceEngine::make_shared_blob<float>(
1832                                 InferenceEngine::Precision::FP32, InferenceEngine::Layout::OIHW,
1833                                 ieWeights->dims());
1834             ieWeights->allocate();
1835
1836             int inpCn = blobs[0].size[0];
1837             Mat newWeights = infEngineBlobToMat(ieWeights).reshape(1, inpCn);
1838             transpose(weightsMat, newWeights);
1839         }
1840
1841         const int outGroupCn = blobs[0].size[1];  // Weights are in IOHW layout
1842         const int group = numOutput / outGroupCn;
1843
1844         InferenceEngine::Builder::DeconvolutionLayer ieLayer(name);
1845
1846         ieLayer.setKernel(kernel_size);
1847         ieLayer.setStrides(strides);
1848         ieLayer.setDilation(dilations);
1849         ieLayer.setPaddingsBegin(pads_begin);
1850
1851         if (padMode.empty())
1852         {
1853             ieLayer.setPaddingsEnd({pads_end[0] - adjust_pads[0], pads_end[1] - adjust_pads[1]});
1854         }
1855         else if (padMode == "SAME")
1856         {
1857             ieLayer.setPaddingsEnd({kernel_size[0] - pads_begin[0] - 1 - adjust_pads[0],
1858                                     kernel_size[1] - pads_begin[1] - 1 - adjust_pads[1]});
1859         }
1860         ieLayer.setGroup((size_t)group);
1861         ieLayer.setOutDepth((size_t)numOutput);
1862
1863         InferenceEngine::Builder::Layer l = ieLayer;
1864         addConstantData("weights", ieWeights, l);
1865         if (hasBias())
1866             addConstantData("biases", wrapToInfEngineBlob(biasesMat, {(size_t)numOutput}, InferenceEngine::Layout::C), l);
1867         return Ptr<BackendNode>(new InfEngineBackendNode(l));
1868     }
1869 #endif  // HAVE_INF_ENGINE
1870
1871     virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
1872                            const std::vector<MatShape> &outputs) const CV_OVERRIDE
1873     {
1874         CV_Assert(inputs.size() == outputs.size());
1875
1876         float flops = 0;
1877         int outChannels = blobs[0].size[0];
1878
1879         for (int i = 0; i < inputs.size(); i++)
1880         {
1881             flops += CV_BIG_INT(2)*outChannels*kernel.area()*total(inputs[i]);
1882         }
1883
1884         return flops;
1885     }
1886 };
1887
1888 Ptr<BaseConvolutionLayer> ConvolutionLayer::create(const LayerParams &params)
1889 {
1890     Ptr<ConvolutionLayerImpl> l(new ConvolutionLayerImpl(params));
1891     return l;
1892 }
1893
1894 Ptr<BaseConvolutionLayer> DeconvolutionLayer::create(const LayerParams &params)
1895 {
1896     return Ptr<BaseConvolutionLayer>(new DeConvolutionLayerImpl(params));
1897 }
1898
1899 }
1900 }