modules/dnn/src/layers/eltwise_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_cuda.hpp"
  46 #include "../op_halide.hpp"
  47 #include "../op_inf_engine.hpp"
  48
  49 #ifdef HAVE_OPENCL
  50 #include "opencl_kernels_dnn.hpp"
  51 #endif
  52
  53 #ifdef HAVE_CUDA
  54 #include "../cuda4dnn/primitives/eltwise.hpp"
  55 using namespace cv::dnn::cuda4dnn;
  56 #endif
  57
  58 namespace cv
  59 {
  60 namespace dnn
  61 {
  62
  63 class EltwiseLayerImpl CV_FINAL : public EltwiseLayer
  64 {
  65 public:
  66     enum EltwiseOp
  67     {
  68         PROD = 0,
  69         SUM = 1,
  70         MAX = 2,
  71     } op;
  72     std::vector<float> coeffs;
  73
  74     EltwiseLayerImpl(const LayerParams& params)
  75     {
  76         setParamsFrom(params);
  77         op = SUM;
  78         if (params.has("operation"))
  79         {
  80             String operation = toLowerCase(params.get<String>("operation"));
  81             if (operation == "prod")
  82                 op = PROD;
  83             else if (operation == "sum")
  84                 op = SUM;
  85             else if (operation == "max")
  86                 op = MAX;
  87             else
  88                 CV_Error(cv::Error::StsBadArg, "Unknown operation type \"" + operation + "\"");
  89         }
  90
  91         if (params.has("coeff"))
  92         {
  93             DictValue paramCoeff = params.get("coeff");
  94             int i, n = paramCoeff.size();
  95             coeffs.resize(n);
  96             for (i = 0; i < n; i++)
  97             {
  98                 coeffs[i] = paramCoeff.get<float>(i);
  99             }
 100         }
 101     }
 102
 103     virtual bool supportBackend(int backendId) CV_OVERRIDE
 104     {
 105         return backendId == DNN_BACKEND_OPENCV ||
 106                backendId == DNN_BACKEND_CUDA ||
 107                backendId == DNN_BACKEND_HALIDE ||
 108                (backendId == DNN_BACKEND_INFERENCE_ENGINE &&
 109                 (preferableTarget != DNN_TARGET_OPENCL || coeffs.empty()));
 110     }
 111
 112     bool getMemoryShapes(const std::vector<MatShape> &inputs,
 113                          const int requiredOutputs,
 114                          std::vector<MatShape> &outputs,
 115                          std::vector<MatShape> &internals) const CV_OVERRIDE
 116     {
 117         CV_Assert(inputs.size() >= 2);
 118         CV_Assert(coeffs.size() == 0 || coeffs.size() == inputs.size());
 119         CV_Assert(op == SUM || coeffs.size() == 0);
 120
 121         for (int i = 1; i < inputs.size(); i++)
 122         {
 123             CV_Assert(inputs[0] == inputs[i]);
 124         }
 125
 126         outputs.assign(1, inputs[0]);
 127
 128         return false;
 129     }
 130
 131     class EltwiseInvoker : public ParallelLoopBody
 132     {
 133     public:
 134         const Mat* srcs;
 135         int nsrcs;
 136         Mat* dst;
 137         const std::vector<float>* coeffs;
 138         EltwiseOp op;
 139         int nstripes;
 140         const ActivationLayer* activ;
 141         int channels;
 142         size_t planeSize;
 143
 144         EltwiseInvoker() : srcs(0), nsrcs(0), dst(0), coeffs(0), op(PROD), nstripes(0), activ(0), channels(0), planeSize(0)  {}
 145
 146         static void run(const Mat* srcs, int nsrcs, Mat& dst,
 147                         const std::vector<float>& coeffs, EltwiseOp op,
 148                         const ActivationLayer* activ, int nstripes)
 149         {
 150             CV_Check(dst.dims, 1 < dst.dims && dst.dims <= 5, ""); CV_CheckTypeEQ(dst.type(), CV_32FC1, ""); CV_Assert(dst.isContinuous());
 151             CV_Assert(coeffs.empty() || coeffs.size() == (size_t)nsrcs);
 152
 153             for( int i = 0; i < nsrcs; i++ )
 154             {
 155                 CV_Assert(srcs[i].size == dst.size &&
 156                           srcs[i].type() == dst.type() &&
 157                           srcs[i].isContinuous());
 158             }
 159
 160             EltwiseInvoker p;
 161             p.srcs = srcs;
 162             p.nsrcs = nsrcs;
 163             p.dst = &dst;
 164             p.op = op;
 165             p.nstripes = nstripes;
 166             p.channels = (dst.dims >= 4 ? dst.size[1] : 1);
 167
 168             p.planeSize = dst.total(dst.dims >= 4 ? 2 : 1);
 169             CV_Assert(dst.total() == dst.size[0] * p.channels * p.planeSize);
 170
 171             bool simpleCoeffs = true;
 172             if( op == SUM && !coeffs.empty() )
 173             {
 174                 CV_Assert( coeffs.size() == (size_t)nsrcs );
 175
 176                 for( size_t i = 0; i < coeffs.size(); i++ )
 177                     if( coeffs[i] != 1 )
 178                     {
 179                         simpleCoeffs = false;
 180                         break;
 181                     }
 182             }
 183             p.coeffs = simpleCoeffs ? 0 : &coeffs;
 184             p.activ = activ;
 185
 186             parallel_for_(Range(0, nstripes), p, nstripes);
 187         }
 188
 189         void operator()(const Range& r) const CV_OVERRIDE
 190         {
 191             size_t total = dst->size[0]*planeSize;
 192             size_t stripeSize = (total + nstripes - 1)/nstripes;
 193             size_t stripeStart = r.start*stripeSize;
 194             size_t stripeEnd = std::min(r.end*stripeSize, total);
 195             int c, j, k, n = nsrcs;
 196             const float* coeffsptr = coeffs && !coeffs->empty() ? &coeffs->at(0) : 0;
 197             float* dstptr0 = dst->ptr<float>();
 198             int blockSize0 = 1 << 12, blockSize;
 199
 200             for( size_t ofs = stripeStart; ofs < stripeEnd; ofs += blockSize )
 201             {
 202                 int sampleIdx = (int)(ofs / planeSize);
 203                 int delta = (int)ofs - sampleIdx * planeSize;
 204                 blockSize = std::min(blockSize0, std::min((int)(stripeEnd - ofs), (int)planeSize - delta));
 205                 if( blockSize <= 0 )
 206                     break;
 207
 208                 for( c = 0; c < channels; c++ )
 209                 {
 210                     size_t globalDelta = delta + (sampleIdx*channels + c)*planeSize;
 211                     const float* srcptr0 = srcs[0].ptr<float>() + globalDelta;
 212                     float* dstptr = dstptr0 + globalDelta;
 213
 214                     if( op == PROD )
 215                     {
 216                         for( k = 1; k < n; k++ )
 217                         {
 218                             const float* srcptr1 = srcs[k].ptr<float>() + globalDelta;
 219                             for( j = 0; j < blockSize; j++ )
 220                             {
 221                                 dstptr[j] = srcptr0[j]*srcptr1[j];
 222                             }
 223                             srcptr0 = (const float*)dstptr;
 224                         }
 225                     }
 226                     else if( op == MAX )
 227                     {
 228                         for( k = 1; k < n; k++ )
 229                         {
 230                             const float* srcptr1 = srcs[k].ptr<float>() + globalDelta;
 231                             for( j = 0; j < blockSize; j++ )
 232                             {
 233                                 dstptr[j] = std::max(srcptr0[j], srcptr1[j]);
 234                             }
 235                             srcptr0 = (const float*)dstptr;
 236                         }
 237                     }
 238                     else if( !coeffsptr )
 239                     {
 240                         for( k = 1; k < n; k++ )
 241                         {
 242                             const float* srcptr1 = srcs[k].ptr<float>() + globalDelta;
 243                             for( j = 0; j < blockSize; j++ )
 244                             {
 245                                 dstptr[j] = srcptr0[j] + srcptr1[j];
 246                             }
 247                             srcptr0 = (const float*)dstptr;
 248                         }
 249                     }
 250                     else
 251                     {
 252                         float c0 = coeffsptr[0];
 253                         for( k = 1; k < n; k++ )
 254                         {
 255                             const float* srcptr1 = srcs[k].ptr<float>() + globalDelta;
 256                             float c1 = coeffsptr[k];
 257                             for( j = 0; j < blockSize; j++ )
 258                             {
 259                                 dstptr[j] = c0*srcptr0[j] + c1*srcptr1[j];
 260                             }
 261                             srcptr0 = (const float*)dstptr;
 262                             c0 = 1;
 263                         }
 264                     }
 265                 }
 266
 267                 if( activ )
 268                 {
 269                     float* ptr = dstptr0 + delta + sampleIdx*channels*planeSize;
 270                     activ->forwardSlice(ptr, ptr, blockSize, planeSize, 0, channels);
 271                 }
 272             }
 273         }
 274     };
 275
 276 #ifdef HAVE_OPENCL
 277     bool forward_ocl(InputArrayOfArrays inputs_, OutputArrayOfArrays outputs_, OutputArrayOfArrays internals_)
 278     {
 279         std::vector<UMat> inputs;
 280         std::vector<UMat> outputs;
 281
 282         if (inputs_.depth() == CV_16S && op != SUM)
 283             return false;
 284
 285         inputs_.getUMatVector(inputs);
 286         outputs_.getUMatVector(outputs);
 287
 288         switch (op)
 289         {
 290             case SUM:
 291                 {
 292                     int channels = total(shape(outputs[0]), 0, 2);
 293                     int plane_size = total(shape(outputs[0]), 2);
 294                     if (channels % 4 == 0 && plane_size % 4 == 0)
 295                     {
 296                         size_t localsize[] = { 128 };
 297                         size_t globalsize[] = { (size_t)channels / 4 * localsize[0] };
 298                         String opts;
 299                         if (inputs_.depth() == CV_16S)
 300                             opts = " -DDtype=half -DDtype4=half4 -DDtype8=half8";
 301                         else
 302                             opts = " -DDtype=float -DDtype4=float4 -DDtype8=float8";
 303
 304                         for (int i = 0; i < (inputs.size() - 1); ++i)
 305                         {
 306                             String buildopt = format("-DLOOP=%d", i) + opts;
 307                             ocl::Kernel kernel("op_sum4", ocl::dnn::eltwise_oclsrc, buildopt);
 308                             int idx = 0;
 309                             UMat inpMat = (i == 0) ? inputs[0] : UMat();
 310                             float coeff1 = (coeffs.empty() || i > 0) ? 1.0f : coeffs[i];
 311                             float coeff2 = coeffs.empty() ? 1.0f : coeffs[i + 1];
 312                             kernel.set(idx++, ocl::KernelArg::PtrReadOnly(inputs[0]));
 313                             kernel.set(idx++, ocl::KernelArg::PtrReadOnly(inputs[1]));
 314                             kernel.set(idx++, (int)plane_size);
 315                             kernel.set(idx++, (float)coeff1);
 316                             kernel.set(idx++, (float)coeff2);
 317                             kernel.set(idx++, ocl::KernelArg::PtrReadWrite(outputs[0]));
 318                             bool ret = kernel.run(1, globalsize, localsize, false);
 319                             if (!ret)
 320                                 return false;
 321                         }
 322                     }
 323                     else
 324                     {
 325                         if (inputs_.depth() == CV_16S)
 326                             return false;
 327
 328                         float coeff1 = coeffs.empty() ? 1.f : coeffs[0];
 329                         float coeff2 = coeffs.empty() ? 1.f : coeffs[1];
 330                         UMat mul0, mul1;
 331                         multiply(coeff1, inputs[0], mul0);
 332                         multiply(coeff2, inputs[1], mul1);
 333                         add(mul0, mul1, outputs[0]);
 334                         for (int i = 2; i < inputs.size(); ++i)
 335                         {
 336                             float coeff = coeffs.empty() ? 1.f : coeffs[i];
 337                             multiply(coeff, inputs[i], mul0);
 338                             add(mul0, outputs[0], outputs[0]);
 339                         }
 340                     }
 341                 }
 342                 break;
 343             case PROD:
 344                 multiply(inputs[0], inputs[1], outputs[0]);
 345                 for (int i = 2; i < inputs.size(); ++i)
 346                     multiply(inputs[i], outputs[0], outputs[0]);
 347                 break;
 348             case MAX:
 349                 max(inputs[0], inputs[1], outputs[0]);
 350                 for (int i = 2; i < inputs.size(); ++i)
 351                     max(inputs[i], outputs[0], outputs[0]);
 352                 break;
 353             default:
 354                 return false;
 355         }
 356         return true;
 357     }
 358 #endif
 359
 360     void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr) CV_OVERRIDE
 361     {
 362         CV_TRACE_FUNCTION();
 363         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 364
 365         CV_OCL_RUN(IS_DNN_OPENCL_TARGET(preferableTarget),
 366                    forward_ocl(inputs_arr, outputs_arr, internals_arr))
 367
 368         if (inputs_arr.depth() == CV_16S)
 369         {
 370             forward_fallback(inputs_arr, outputs_arr, internals_arr);
 371             return;
 372         }
 373
 374         std::vector<Mat> inputs, outputs;
 375         inputs_arr.getMatVector(inputs);
 376         outputs_arr.getMatVector(outputs);
 377
 378         CV_Assert(outputs.size() == 1);
 379         const int nstripes = getNumThreads();
 380         EltwiseInvoker::run(&inputs[0], (int)inputs.size(), outputs[0],
 381                             coeffs, op, activ.get(), nstripes);
 382     }
 383
 384 #ifdef HAVE_CUDA
 385     Ptr<BackendNode> initCUDA(
 386         void *context_,
 387         const std::vector<Ptr<BackendWrapper>>& inputs,
 388         const std::vector<Ptr<BackendWrapper>>& outputs
 389     ) override
 390     {
 391         auto context = reinterpret_cast<csl::CSLContext*>(context_);
 392
 393         auto op_ = [this] {
 394             switch (op) {
 395             case MAX: return cuda4dnn::EltwiseOpType::MAX;
 396             case SUM: return cuda4dnn::EltwiseOpType::SUM;
 397             case PROD: return cuda4dnn::EltwiseOpType::PRODUCT;
 398             }
 399             return cuda4dnn::EltwiseOpType::SUM;
 400         }();
 401
 402         return make_cuda_node<cuda4dnn::EltwiseOp>(preferableTarget, std::move(context->stream), op_, coeffs);
 403     }
 404 #endif
 405
 406     virtual Ptr<BackendNode> initHalide(const std::vector<Ptr<BackendWrapper> > &input) CV_OVERRIDE
 407     {
 408 #ifdef HAVE_HALIDE
 409         Halide::Var x("x"), y("y"), c("c"), n("n");
 410         Halide::Func top = (name.empty() ? Halide::Func() : Halide::Func(name));
 411         Halide::Expr topExpr;
 412         std::vector<Halide::Buffer<> > inputBuffers = halideBuffers(input);
 413         switch (op)
 414         {
 415             case SUM:
 416                 if (coeffs.empty())
 417                 {
 418                     topExpr = inputBuffers[0](x, y, c, n) +
 419                               inputBuffers[1](x, y, c, n);
 420                     for (int i = 2; i < inputBuffers.size(); ++i)
 421                         topExpr += inputBuffers[i](x, y, c, n);
 422                 }
 423                 else
 424                 {
 425                   topExpr = coeffs[0] * inputBuffers[0](x, y, c, n) +
 426                             coeffs[1] * inputBuffers[1](x, y, c, n);
 427                   for (int i = 2; i < inputBuffers.size(); ++i)
 428                       topExpr += coeffs[i] * inputBuffers[i](x, y, c, n);
 429                 }
 430                 break;
 431             case PROD:
 432                 topExpr = inputBuffers[0](x, y, c, n) *
 433                           inputBuffers[1](x, y, c, n);
 434                 for (int i = 2; i < inputBuffers.size(); ++i)
 435                     topExpr *= inputBuffers[i](x, y, c, n);
 436                 break;
 437             case MAX:
 438                 topExpr = max(inputBuffers[0](x, y, c, n),
 439                               inputBuffers[1](x, y, c, n));
 440                 for (int i = 2; i < inputBuffers.size(); ++i)
 441                     topExpr = max(topExpr, inputBuffers[i](x, y, c, n));
 442                 break;
 443             default:
 444                 return Ptr<BackendNode>();
 445         }
 446         top(x, y, c, n) = topExpr;
 447         return Ptr<BackendNode>(new HalideBackendNode(top));
 448 #endif  // HAVE_HALIDE
 449         return Ptr<BackendNode>();
 450     }
 451
 452 #ifdef HAVE_INF_ENGINE
 453     virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >& inputs) CV_OVERRIDE
 454     {
 455         InferenceEngine::Builder::EltwiseLayer ieLayer(name);
 456
 457         ieLayer.setInputPorts(std::vector<InferenceEngine::Port>(inputs.size()));
 458
 459         if (op == SUM)
 460             ieLayer.setEltwiseType(InferenceEngine::Builder::EltwiseLayer::EltwiseType::SUM);
 461         else if (op == PROD)
 462             ieLayer.setEltwiseType(InferenceEngine::Builder::EltwiseLayer::EltwiseType::MUL);
 463         else if (op == MAX)
 464             ieLayer.setEltwiseType(InferenceEngine::Builder::EltwiseLayer::EltwiseType::MAX);
 465         else
 466             CV_Error(Error::StsNotImplemented, "Unsupported eltwise operation");
 467
 468         InferenceEngine::Builder::Layer l = ieLayer;
 469         if (!coeffs.empty())
 470             l.getParameters()["coeff"] = coeffs;
 471
 472         return Ptr<BackendNode>(new InfEngineBackendNode(l));
 473     }
 474 #endif  // HAVE_INF_ENGINE
 475
 476     virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
 477                            const std::vector<MatShape> &outputs) const CV_OVERRIDE
 478     {
 479         CV_UNUSED(outputs); // suppress unused variable warning
 480         CV_Assert(inputs.size());
 481
 482         long flops = inputs.size() * total(inputs[0]);
 483
 484         return flops;
 485     }
 486
 487     bool setActivation(const Ptr<ActivationLayer>& layer) CV_OVERRIDE
 488     {
 489         if (activ.empty() || layer.empty())
 490         {
 491             activ = layer;
 492             return !activ.empty();
 493         }
 494         else
 495             return false;
 496     }
 497
 498     Ptr<ActivationLayer> activ;
 499 };
 500
 501 Ptr<EltwiseLayer> EltwiseLayer::create(const LayerParams& params)
 502 {
 503     return Ptr<EltwiseLayer>(new EltwiseLayerImpl(params));
 504 }
 505
 506 }
 507 }