modules/dnn/src/layers/pooling_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 "opencv2/core/hal/intrin.hpp"
  46 #include "op_halide.hpp"
  47 #include <float.h>
  48 #include <algorithm>
  49 using std::max;
  50 using std::min;
  51
  52 namespace cv
  53 {
  54 namespace dnn
  55 {
  56
  57 //TODO: add ceil_mode param
  58 class PoolingLayerImpl : public PoolingLayer
  59 {
  60 public:
  61     PoolingLayerImpl(const LayerParams& params)
  62     {
  63         type = PoolingLayer::MAX;
  64         computeMaxIdx = true;
  65
  66         if (params.has("pool"))
  67         {
  68             String pool = params.get<String>("pool").toLowerCase();
  69             if (pool == "max")
  70                 type = PoolingLayer::MAX;
  71             else if (pool == "ave")
  72                 type = PoolingLayer::AVE;
  73             else if (pool == "stochastic")
  74                 type = PoolingLayer::STOCHASTIC;
  75             else
  76                 CV_Error(Error::StsBadArg, "Unknown pooling type \"" + pool + "\"");
  77         }
  78
  79         getPoolingKernelParams(params, kernel.height, kernel.width, globalPooling,
  80                                pad.height, pad.width, stride.height, stride.width, padMode);
  81         setParamsFrom(params);
  82     }
  83
  84     void finalize(const std::vector<Mat*> &inputs, std::vector<Mat> &outputs)
  85     {
  86         CV_Assert(inputs.size() == 1);
  87
  88         cv::Size inp(inputs[0]->size[3], inputs[0]->size[2]),
  89                 out(outputs[0].size[3], outputs[0].size[2]);
  90
  91         if(globalPooling)
  92         {
  93             kernel = inp;
  94         }
  95
  96         getConvPoolPaddings(inp, out, kernel, stride, padMode, pad);
  97     }
  98
  99     virtual bool supportBackend(int backendId)
 100     {
 101         return backendId == DNN_BACKEND_DEFAULT ||
 102                backendId == DNN_BACKEND_HALIDE && haveHalide() &&
 103                (type == PoolingLayer::MAX ||
 104                 type == PoolingLayer::AVE && !pad.width && !pad.height);
 105     }
 106
 107     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
 108     {
 109         CV_TRACE_FUNCTION();
 110         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 111
 112         for (size_t ii = 0; ii < inputs.size(); ii++)
 113         {
 114             switch (type)
 115             {
 116                 case MAX:
 117                     maxPooling(*inputs[ii], outputs[2 * ii], outputs[2 * ii + 1]);
 118                     break;
 119                 case AVE:
 120                     avePooling(*inputs[ii], outputs[ii]);
 121                     break;
 122                 default:
 123                     CV_Error(Error::StsNotImplemented, "Not implemented");
 124                     break;
 125             }
 126         }
 127     }
 128
 129     virtual Ptr<BackendNode> initHalide(const std::vector<Ptr<BackendWrapper> > &inputs)
 130     {
 131         if (type == PoolingLayer::MAX)
 132             return initMaxPoolingHalide(inputs);
 133         else if (type == PoolingLayer::AVE)
 134             return initAvePoolingHalide(inputs);
 135         else
 136             return Ptr<BackendNode>();
 137     }
 138
 139     class PoolingInvoker : public ParallelLoopBody
 140     {
 141     public:
 142         const Mat* src;
 143         Mat *dst, *mask;
 144         Size kernel, stride, pad;
 145         int nstripes;
 146         bool computeMaxIdx;
 147         std::vector<int> ofsbuf;
 148         int poolingType;
 149
 150         PoolingInvoker() {}
 151
 152         static void run(const Mat& src, Mat& dst, Mat& mask, Size kernel,
 153                         Size stride, Size pad, int poolingType,
 154                         bool computeMaxIdx, int nstripes)
 155         {
 156             CV_Assert(src.isContinuous() && dst.isContinuous() &&
 157                       src.type() == CV_32F && src.type() == dst.type() &&
 158                       src.dims == 4 && dst.dims == 4 &&
 159                       src.size[0] == dst.size[0] && src.size[1] == dst.size[1] &&
 160                       (mask.empty() || (mask.type() == src.type() && mask.size == dst.size)));
 161
 162             PoolingInvoker p;
 163
 164             p.src = &src;
 165             p.dst = &dst;
 166             p.mask = &mask;
 167             p.kernel = kernel;
 168             p.stride = stride;
 169             p.pad = pad;
 170             p.nstripes = nstripes;
 171             p.computeMaxIdx = computeMaxIdx;
 172             p.poolingType = poolingType;
 173
 174             if( !computeMaxIdx )
 175             {
 176                 p.ofsbuf.resize(kernel.width*kernel.height);
 177                 for( int i = 0; i < kernel.height; i++ )
 178                     for( int j = 0; j < kernel.width; j++ )
 179                         p.ofsbuf[i*kernel.width + j] = src.size[3]*i + j;
 180             }
 181
 182             parallel_for_(Range(0, nstripes), p, nstripes);
 183         }
 184
 185         void operator()(const Range& r) const
 186         {
 187             int channels = dst->size[1], width = dst->size[3], height = dst->size[2];
 188             int inp_width = src->size[3], inp_height = src->size[2];
 189             size_t total = dst->total();
 190             size_t stripeSize = (total + nstripes - 1)/nstripes;
 191             size_t stripeStart = r.start*stripeSize;
 192             size_t stripeEnd = std::min(r.end*stripeSize, total);
 193             int kernel_w = kernel.width, kernel_h = kernel.height;
 194             int pad_w = pad.width, pad_h = pad.height;
 195             int stride_w = stride.width, stride_h = stride.height;
 196             bool compMaxIdx = computeMaxIdx;
 197
 198 #if CV_SIMD128
 199             const int* ofsptr = &ofsbuf[0];
 200             v_float32x4 idx00(0.f, (float)stride_w, (float)(stride_w*2), (float)(stride_w*3));
 201             v_float32x4 ones = v_setall_f32(1.f);
 202             v_float32x4 idx_delta = v_setall_f32((float)(inp_width - kernel_w));
 203 #endif
 204
 205             for( size_t ofs0 = stripeStart; ofs0 < stripeEnd; )
 206             {
 207                 size_t ofs = ofs0;
 208                 int x0 = (int)(ofs % width);
 209                 ofs /= width;
 210                 int y0 = (int)(ofs % height);
 211                 ofs /= height;
 212                 int c = (int)(ofs % channels);
 213                 int n = (int)(ofs / channels);
 214                 int ystart = y0 * stride_h - pad_h;
 215                 int yend = min(ystart + kernel_h, inp_height + pad_h);
 216                 int ydelta = yend - ystart;
 217                 ystart = max(ystart, 0);
 218                 yend = min(yend, inp_height);
 219                 const float *srcData = src->ptr<float>(n, c);
 220                 float *dstData = dst->ptr<float>(n, c, y0);
 221                 float *dstMaskData = mask->data ? mask->ptr<float>(n, c, y0) : 0;
 222
 223                 int delta = std::min((int)(stripeEnd - ofs0), width - x0);
 224                 ofs0 += delta;
 225                 int x1 = x0 + delta;
 226
 227                 if( poolingType == PoolingLayer::MAX )
 228                     for( ; x0 < x1; x0++ )
 229                     {
 230                         int xstart = x0 * stride_w - pad_w;
 231                         int xend = min(xstart + kernel_w, inp_width);
 232                         xstart = max(xstart, 0);
 233
 234 #if CV_SIMD128
 235                         if( xstart > 0 && x0 + 7 < x1 && (x0 + 7) * stride_w - pad_w + kernel_w < inp_width )
 236                         {
 237                             if( compMaxIdx )
 238                             {
 239                                 v_float32x4 max_val0 = v_setall_f32(-FLT_MAX);
 240                                 v_float32x4 max_val1 = max_val0;
 241                                 v_float32x4 max_idx0 = v_setall_f32(-1.f);
 242                                 v_float32x4 max_idx1 = max_idx0;
 243                                 int index0 = ystart * inp_width + xstart;
 244                                 v_float32x4 idx0 = idx00 + v_setall_f32((float)index0);
 245                                 v_float32x4 idx1 = idx0 + v_setall_f32((float)(stride_w*4));
 246
 247                                 for (int y = ystart; y < yend; ++y)
 248                                 {
 249                                     for (int x = xstart; x < xend; ++x, idx0 += ones, idx1 += ones)
 250                                     {
 251                                         const int index = y * inp_width + x;
 252                                         v_float32x4 v0(srcData[index], srcData[index + stride_w],
 253                                                        srcData[index + stride_w*2], srcData[index + stride_w*3]);
 254                                         v_float32x4 v1(srcData[index + stride_w*4], srcData[index + stride_w*5],
 255                                                        srcData[index + stride_w*6], srcData[index + stride_w*7]);
 256                                         max_idx0 = v_select(v0 > max_val0, idx0, max_idx0);
 257                                         max_idx1 = v_select(v1 > max_val1, idx1, max_idx1);
 258                                         max_val0 = v_max(max_val0, v0);
 259                                         max_val1 = v_max(max_val1, v1);
 260                                     }
 261                                     idx0 += idx_delta;
 262                                     idx1 += idx_delta;
 263                                 }
 264                                 v_store(dstData + x0, max_val0);
 265                                 v_store(dstData + x0 + 4, max_val1);
 266                                 v_store(dstMaskData + x0, max_idx0);
 267                                 v_store(dstMaskData + x0 + 4, max_idx1);
 268                                 x0 += 7;
 269                             }
 270                             else
 271                             {
 272                                 v_float32x4 max_val0 = v_setall_f32(-FLT_MAX);
 273                                 v_float32x4 max_val1 = max_val0;
 274
 275                                 if( yend - ystart == kernel_h )
 276                                 {
 277                                     const float* srcData1 = srcData + ystart*inp_width + xstart;
 278                                     if( stride_w == 1 )
 279                                         for (int k = 0; k < kernel_w*kernel_h; k++)
 280                                         {
 281                                             int index = ofsptr[k];
 282                                             v_float32x4 v0 = v_load(srcData1 + index);
 283                                             v_float32x4 v1 = v_load(srcData1 + index + 4);
 284                                             max_val0 = v_max(max_val0, v0);
 285                                             max_val1 = v_max(max_val1, v1);
 286                                         }
 287 #if CV_SSE2
 288                                     else if( stride_w == 2 )
 289                                         for (int k = 0; k < kernel_w*kernel_h; k++)
 290                                         {
 291                                             int index = ofsptr[k];
 292                                             v_float32x4 v00 = v_load(srcData1 + index), v01 = v_load(srcData1 + index + 4);
 293                                             v_float32x4 v0(_mm_shuffle_ps(v00.val, v01.val, _MM_SHUFFLE(2, 0, 2, 0)));
 294                                             v_float32x4 v10 = v_load(srcData1 + index + 8), v11 = v_load(srcData1 + index + 12);
 295                                             v_float32x4 v1(_mm_shuffle_ps(v10.val, v11.val, _MM_SHUFFLE(2, 0, 2, 0)));
 296                                             max_val0 = v_max(max_val0, v0);
 297                                             max_val1 = v_max(max_val1, v1);
 298                                         }
 299 #endif
 300                                     else
 301                                         for (int k = 0; k < kernel_w*kernel_h; k++)
 302                                         {
 303                                             int index = ofsptr[k];
 304                                             v_float32x4 v0(srcData1[index], srcData1[index + stride_w],
 305                                                            srcData1[index + stride_w*2], srcData1[index + stride_w*3]);
 306                                             v_float32x4 v1(srcData1[index + stride_w*4], srcData1[index + stride_w*5],
 307                                                            srcData1[index + stride_w*6], srcData1[index + stride_w*7]);
 308                                             max_val0 = v_max(max_val0, v0);
 309                                             max_val1 = v_max(max_val1, v1);
 310                                         }
 311                                 }
 312                                 else
 313                                 {
 314                                     for (int y = ystart; y < yend; ++y)
 315                                     {
 316                                         for (int x = xstart; x < xend; ++x)
 317                                         {
 318                                             const int index = y * inp_width + x;
 319                                             v_float32x4 v0(srcData[index], srcData[index + stride_w],
 320                                                            srcData[index + stride_w*2], srcData[index + stride_w*3]);
 321                                             v_float32x4 v1(srcData[index + stride_w*4], srcData[index + stride_w*5],
 322                                                            srcData[index + stride_w*6], srcData[index + stride_w*7]);
 323                                             max_val0 = v_max(max_val0, v0);
 324                                             max_val1 = v_max(max_val1, v1);
 325                                         }
 326                                     }
 327                                 }
 328                                 v_store(dstData + x0, max_val0);
 329                                 v_store(dstData + x0 + 4, max_val1);
 330                                 x0 += 7;
 331                             }
 332                         }
 333                         else
 334 #endif
 335                         {
 336                             float max_val = -FLT_MAX;
 337                             if( compMaxIdx )
 338                             {
 339                                 int max_index = -1;
 340                                 for (int y = ystart; y < yend; ++y)
 341                                     for (int x = xstart; x < xend; ++x)
 342                                     {
 343                                         const int index = y * inp_width + x;
 344                                         float val = srcData[index];
 345                                         if (val > max_val)
 346                                         {
 347                                             max_val = val;
 348                                             max_index = index;
 349                                         }
 350                                     }
 351
 352                                 dstData[x0] = max_val;
 353                                 dstMaskData[x0] = max_index;
 354                             }
 355                             else
 356                             {
 357                                 for (int y = ystart; y < yend; ++y)
 358                                     for (int x = xstart; x < xend; ++x)
 359                                     {
 360                                         const int index = y * inp_width + x;
 361                                         float val = srcData[index];
 362                                         max_val = std::max(max_val, val);
 363                                     }
 364
 365                                 dstData[x0] = max_val;
 366                             }
 367                         }
 368                     }
 369                 else
 370                 {
 371                     for( ; x0 < x1; x0++ )
 372                     {
 373                         int xstart = x0 * stride_w - pad_w;
 374                         int xend = min(xstart + kernel_w, inp_width + pad_w);
 375                         int xdelta = xend - xstart;
 376                         xstart = max(xstart, 0);
 377                         xend = min(xend, inp_width);
 378                         float inv_kernel_area = 1.f/(ydelta*xdelta);
 379
 380 #if CV_SIMD128
 381                         if( xstart > 0 && x0 + 7 < x1 && (x0 + 7) * stride_w - pad_w + kernel_w < inp_width )
 382                         {
 383                             v_float32x4 sum_val0 = v_setzero_f32(), sum_val1 = v_setzero_f32();
 384                             v_float32x4 ikarea = v_setall_f32(inv_kernel_area);
 385
 386                             for (int y = ystart; y < yend; ++y)
 387                             {
 388                                 for (int x = xstart; x < xend; ++x)
 389                                 {
 390                                     const int index = y * inp_width + x;
 391                                     v_float32x4 v0(srcData[index], srcData[index + stride_w],
 392                                                    srcData[index + stride_w*2], srcData[index + stride_w*3]);
 393                                     v_float32x4 v1(srcData[index + stride_w*4], srcData[index + stride_w*5],
 394                                                    srcData[index + stride_w*6], srcData[index + stride_w*7]);
 395                                     sum_val0 += v0;
 396                                     sum_val1 += v1;
 397                                 }
 398                             }
 399                             v_store(dstData + x0, sum_val0*ikarea);
 400                             v_store(dstData + x0 + 4, sum_val1*ikarea);
 401                             x0 += 7;
 402                         }
 403                         else
 404 #endif
 405                         {
 406                             float sum_val = 0.f;
 407                             for (int y = ystart; y < yend; ++y)
 408                                 for (int x = xstart; x < xend; ++x)
 409                                 {
 410                                     const int index = y * inp_width + x;
 411                                     float val = srcData[index];
 412                                     sum_val += val;
 413                                 }
 414
 415                             dstData[x0] = sum_val*inv_kernel_area;
 416                         }
 417                     }
 418                 }
 419             }
 420         }
 421     };
 422
 423     void maxPooling(Mat &src, Mat &dst, Mat &mask)
 424     {
 425         const int nstripes = getNumThreads();
 426         PoolingInvoker::run(src, dst, mask, kernel, stride, pad, type, computeMaxIdx, nstripes);
 427     }
 428
 429     void avePooling(Mat &src, Mat &dst)
 430     {
 431         const int nstripes = getNumThreads();
 432         Mat mask;
 433         PoolingInvoker::run(src, dst, mask, kernel, stride, pad, type, computeMaxIdx, nstripes);
 434     }
 435
 436     virtual Ptr<BackendNode> initMaxPoolingHalide(const std::vector<Ptr<BackendWrapper> > &inputs)
 437     {
 438 #ifdef HAVE_HALIDE
 439         Halide::Buffer<float> inputBuffer = halideBuffer(inputs[0]);
 440         const int inWidth = inputBuffer.width();
 441         const int inHeight = inputBuffer.height();
 442
 443         Halide::Var x("x"), y("y"), c("c"), n("n");
 444         Halide::Func top = (name.empty() ? Halide::Func() : Halide::Func(name));
 445         Halide::RDom r(0, kernel.width, 0, kernel.height);
 446         Halide::Expr kx, ky;
 447         if (pad.width || pad.height)
 448         {
 449             kx = clamp(x * stride.width + r.x - pad.width, 0, inWidth - 1);
 450             ky = clamp(y * stride.height + r.y - pad.height, 0, inHeight - 1);
 451         }
 452         else
 453         {
 454             kx = min(x * stride.width + r.x, inWidth - 1);
 455             ky = min(y * stride.height + r.y, inHeight - 1);
 456         }
 457
 458         // Halide::argmax returns tuple (r.x, r.y, max).
 459         Halide::Tuple res = argmax(inputBuffer(kx, ky, c, n));
 460
 461         // Compute offset from argmax in range [0, kernel_size).
 462         Halide::Expr max_index;
 463         if (pad.width || pad.height)
 464         {
 465             max_index = clamp(y * stride.height + res[1] - pad.height,
 466                               0, inHeight - 1) * inWidth +
 467                         clamp(x * stride.width + res[0] - pad.width,
 468                               0, inWidth - 1);
 469         }
 470         else
 471         {
 472             max_index = min(y * stride.height + res[1], inHeight - 1) * inWidth +
 473                         min(x * stride.width + res[0], inWidth - 1);
 474         }
 475         top(x, y, c, n) = { res[2], Halide::cast<float>(max_index) };
 476         return Ptr<BackendNode>(new HalideBackendNode(top));
 477 #endif  // HAVE_HALIDE
 478         return Ptr<BackendNode>();
 479     }
 480
 481     virtual Ptr<BackendNode> initAvePoolingHalide(const std::vector<Ptr<BackendWrapper> > &inputs)
 482     {
 483 #ifdef HAVE_HALIDE
 484         Halide::Buffer<float> inputBuffer = halideBuffer(inputs[0]);
 485
 486         const int inW = inputBuffer.width(), inH = inputBuffer.height();
 487         if ((inW - kernel.width) % stride.width || (inH - kernel.height) % stride.height)
 488         {
 489             CV_Error(cv::Error::StsNotImplemented,
 490                      "Halide backend for average pooling with partial "
 491                      "kernels is not implemented");
 492         }
 493
 494         const float norm = 1.0f / (kernel.width * kernel.height);
 495
 496         Halide::Var x("x"), y("y"), c("c"), n("n");
 497         Halide::Func top = (name.empty() ? Halide::Func() : Halide::Func(name));
 498         Halide::RDom r(0, kernel.width, 0, kernel.height);
 499         top(x, y, c, n) = sum(
 500             inputBuffer(x * stride.width + r.x,
 501                         y * stride.height + r.y, c, n)) * norm;
 502         return Ptr<BackendNode>(new HalideBackendNode(top));
 503 #endif  // HAVE_HALIDE
 504         return Ptr<BackendNode>();
 505     }
 506
 507     virtual void applyHalideScheduler(Ptr<BackendNode>& node,
 508                                       const std::vector<Mat*> &inputs,
 509                                       const std::vector<Mat> &outputs,
 510                                       int targetId) const
 511     {
 512 #ifdef  HAVE_HALIDE
 513         if (targetId != DNN_TARGET_CPU)
 514         {
 515             Layer::applyHalideScheduler(node, inputs, outputs, targetId);
 516             return;
 517         }
 518         Halide::Var x("x"), y("y"), c("c"), n("n"), tile("tile"),
 519                     xi("xi"), yi("yi"), ci("ci"), xo("xo"), yo("yo"), co("co");
 520         Halide::Func& top = node.dynamicCast<HalideBackendNode>()->funcs.back();
 521
 522         int outW, outH, outC, outN;
 523         getCanonicalSize(outputs[0].size, &outW, &outH, &outC, &outN);
 524
 525         if (outW < 8 || outH < 8)
 526         {
 527             if (outC > 8)
 528                 top.split(c, co, ci, 8)
 529                    .fuse(x, y, tile).fuse(co, tile, tile).fuse(n, tile, tile)
 530                    .parallel(tile)
 531                    .vectorize(ci);
 532             else
 533             {
 534                 top.fuse(y, c, tile).fuse(n, tile, tile)
 535                    .parallel(tile);
 536                 if (outW > 1)
 537                     top.vectorize(x);
 538             }
 539         }
 540         else
 541         {
 542             if (outC > 8)
 543                 top.split(x, xo, xi, 8).split(y, yo, yi, 8).split(c, co, ci, 8)
 544                    .fuse(xo, yo, tile).fuse(co, tile, tile).fuse(n, tile, tile)
 545                    .parallel(tile)
 546                    .vectorize(xi);
 547             else
 548                 top.split(x, xo, xi, 8).split(y, yo, yi, 8)
 549                    .fuse(xo, yo, tile).fuse(c, tile, tile).fuse(n, tile, tile)
 550                    .parallel(tile)
 551                    .vectorize(xi);
 552         }
 553 #endif  // HAVE_HALIDE
 554     }
 555
 556     bool getMemoryShapes(const std::vector<MatShape> &inputs,
 557                          const int requiredOutputs,
 558                          std::vector<MatShape> &outputs,
 559                          std::vector<MatShape> &internals) const
 560     {
 561         CV_Assert(inputs.size() != 0);
 562         Size in(inputs[0][3], inputs[0][2]), out;
 563
 564         if (globalPooling)
 565         {
 566             out.height = 1;
 567             out.width = 1;
 568         }
 569         else if (padMode.empty())
 570         {
 571             //Yeah, something strange Caffe scheme-)
 572             out.height = static_cast<int>(ceil(static_cast<float>(in.height + 2 * pad.height -
 573                                                                   kernel.height) / stride.height)) + 1;
 574             out.width = static_cast<int>(ceil(static_cast<float>(in.width + 2 * pad.width -
 575                                                                  kernel.width) / stride.width)) + 1;
 576
 577             if (pad.height || pad.width)
 578             {
 579                 // If we have padding, ensure that the last pooling starts strictly
 580                 // inside the image (instead of at the padding); otherwise clip the last.
 581                 if ((out.height - 1) * stride.height >= in.height + pad.height)
 582                     --out.height;
 583                 if ((out.width - 1) * stride.width >= in.width + pad.width)
 584                     --out.width;
 585                 CV_Assert((out.height - 1) * stride.height < in.height + pad.height);
 586                 CV_Assert((out.width - 1) * stride.width < in.width + pad.width);
 587             }
 588         }
 589         else
 590         {
 591             getConvPoolOutParams(in, kernel, stride,
 592                                  padMode, out);
 593         }
 594
 595         outputs.resize(type == MAX ? 2 * inputs.size() : inputs.size());
 596         for (size_t i = 0; i < inputs.size(); i++)
 597         {
 598             size_t index = type == MAX ? 2*i : i;
 599             int dims[] = {inputs[i][0], inputs[i][1], out.height, out.width};
 600             outputs[index] = shape(dims);
 601
 602             if (type == MAX)
 603                 outputs[index + 1] = shape(dims);
 604         }
 605
 606         return false;
 607     }
 608
 609     virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
 610                            const std::vector<MatShape> &outputs) const
 611     {
 612         (void)inputs; // suppress unused variable warning
 613         long flops = 0;
 614
 615         for(int i = 0; i < outputs.size(); i++)
 616         {
 617             if (type == MAX)
 618             {
 619                 if (i%2 == 0)
 620                     flops += total(outputs[i])*kernel.area();
 621             }
 622             else
 623             {
 624                 flops += total(outputs[i])*(kernel.area() + 1);
 625             }
 626         }
 627         return flops;
 628     }
 629 };
 630
 631 Ptr<PoolingLayer> PoolingLayer::create(const LayerParams& params)
 632 {
 633     return Ptr<PoolingLayer>(new PoolingLayerImpl(params));
 634 }
 635
 636 }
 637 }