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