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 "opencl_kernels_dnn.hpp"
  48 #include <float.h>
  49 #include <algorithm>
  50 using std::max;
  51 using std::min;
  52 #ifdef HAVE_OPENCL
  53 using namespace cv::dnn::ocl4dnn;
  54 #endif
  55
  56 namespace cv
  57 {
  58 namespace dnn
  59 {
  60
  61 class PoolingLayerImpl : public PoolingLayer
  62 {
  63 public:
  64     PoolingLayerImpl(const LayerParams& params)
  65     {
  66         type = PoolingLayer::MAX;
  67         computeMaxIdx = true;
  68         globalPooling = false;
  69
  70         if (params.has("pool"))
  71         {
  72             String pool = params.get<String>("pool").toLowerCase();
  73             if (pool == "max")
  74                 type = PoolingLayer::MAX;
  75             else if (pool == "ave")
  76                 type = PoolingLayer::AVE;
  77             else if (pool == "stochastic")
  78                 type = PoolingLayer::STOCHASTIC;
  79             else
  80                 CV_Error(Error::StsBadArg, "Unknown pooling type \"" + pool + "\"");
  81             getPoolingKernelParams(params, kernel.height, kernel.width, globalPooling,
  82                                    pad.height, pad.width, stride.height, stride.width, padMode);
  83         }
  84         else if (params.has("pooled_w") || params.has("pooled_h") || params.has("spatial_scale"))
  85         {
  86             type = PoolingLayer::ROI;
  87         }
  88         setParamsFrom(params);
  89         ceilMode = params.get<bool>("ceil_mode", true);
  90         pooledSize.width = params.get<uint32_t>("pooled_w", 1);
  91         pooledSize.height = params.get<uint32_t>("pooled_h", 1);
  92         spatialScale = params.get<float>("spatial_scale", 1);
  93     }
  94
  95 #ifdef HAVE_OPENCL
  96     Ptr<OCL4DNNPool<float> > poolOp;
  97 #endif
  98
  99     void finalize(const std::vector<Mat*> &inputs, std::vector<Mat> &outputs)
 100     {
 101         CV_Assert(!inputs.empty());
 102
 103         cv::Size inp(inputs[0]->size[3], inputs[0]->size[2]),
 104                 out(outputs[0].size[3], outputs[0].size[2]);
 105
 106         if(globalPooling)
 107         {
 108             kernel = inp;
 109         }
 110
 111         getConvPoolPaddings(inp, out, kernel, stride, padMode, Size(1, 1), pad);
 112     }
 113
 114     virtual bool supportBackend(int backendId)
 115     {
 116         return backendId == DNN_BACKEND_DEFAULT ||
 117                backendId == DNN_BACKEND_HALIDE && haveHalide() &&
 118                (type == PoolingLayer::MAX ||
 119                 type == PoolingLayer::AVE && !pad.width && !pad.height);
 120     }
 121
 122 #ifdef HAVE_OPENCL
 123     bool forward_ocl(InputArrayOfArrays inps, OutputArrayOfArrays outs, InputArrayOfArrays internals)
 124     {
 125         std::vector<UMat> inputs;
 126         std::vector<UMat> outputs;
 127
 128         inps.getUMatVector(inputs);
 129         outs.getUMatVector(outputs);
 130
 131         if (poolOp.empty())
 132         {
 133             OCL4DNNPoolConfig config;
 134
 135             config.in_shape = shape(inputs[0]);
 136             config.out_shape = shape(outputs[0]);
 137             config.kernel = kernel;
 138             config.pad = pad;
 139             config.stride = stride;
 140             config.channels = inputs[0].size[1];
 141             config.pool_method = type == MAX ? LIBDNN_POOLING_METHOD_MAX :
 142                                 (type == AVE ? LIBDNN_POOLING_METHOD_AVE :
 143                                                LIBDNN_POOLING_METHOD_STO);
 144             poolOp = Ptr<OCL4DNNPool<float> >(new OCL4DNNPool<float>(config));
 145         }
 146
 147         for (size_t ii = 0; ii < inputs.size(); ii++)
 148         {
 149             UMat& inpMat = inputs[ii];
 150             int out_index = (type == MAX) ? 2 : 1;
 151             UMat& outMat = outputs[out_index * ii];
 152             UMat maskMat = (type == MAX) ? outputs[2 * ii + 1] : UMat();
 153
 154             CV_Assert(inpMat.offset == 0 && outMat.offset == 0);
 155
 156             if (!poolOp->Forward(inpMat, outMat, maskMat))
 157                 return false;
 158         }
 159
 160         return true;
 161     }
 162 #endif
 163
 164     void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
 165     {
 166         CV_TRACE_FUNCTION();
 167         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 168
 169         CV_OCL_RUN((preferableTarget == DNN_TARGET_OPENCL) &&
 170                    OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
 171                    forward_ocl(inputs_arr, outputs_arr, internals_arr))
 172
 173         Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
 174     }
 175
 176     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
 177     {
 178         CV_TRACE_FUNCTION();
 179         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 180
 181         switch (type)
 182         {
 183             case MAX:
 184                 CV_Assert(inputs.size() == 1, outputs.size() == 2);
 185                 maxPooling(*inputs[0], outputs[0], outputs[1]);
 186                 break;
 187             case AVE:
 188                 CV_Assert(inputs.size() == 1, outputs.size() == 1);
 189                 avePooling(*inputs[0], outputs[0]);
 190                 break;
 191             case ROI:
 192                 CV_Assert(inputs.size() == 2, outputs.size() == 1);
 193                 roiPooling(*inputs[0], *inputs[1], outputs[0]);
 194                 break;
 195             default:
 196                 CV_Error(Error::StsNotImplemented, "Not implemented");
 197                 break;
 198         }
 199     }
 200
 201     virtual Ptr<BackendNode> initHalide(const std::vector<Ptr<BackendWrapper> > &inputs)
 202     {
 203         if (type == PoolingLayer::MAX)
 204             return initMaxPoolingHalide(inputs);
 205         else if (type == PoolingLayer::AVE)
 206             return initAvePoolingHalide(inputs);
 207         else
 208             return Ptr<BackendNode>();
 209     }
 210
 211     class PoolingInvoker : public ParallelLoopBody
 212     {
 213     public:
 214         const Mat* src, *rois;
 215         Mat *dst, *mask;
 216         Size kernel, stride, pad;
 217         int nstripes;
 218         bool computeMaxIdx;
 219         std::vector<int> ofsbuf;
 220         int poolingType;
 221         float spatialScale;
 222
 223         PoolingInvoker() : src(0), rois(0), dst(0), mask(0), nstripes(0),
 224                            computeMaxIdx(0), poolingType(PoolingLayer::MAX), spatialScale(0) {}
 225
 226         static void run(const Mat& src, const Mat& rois, Mat& dst, Mat& mask, Size kernel,
 227                         Size stride, Size pad, int poolingType, float spatialScale,
 228                         bool computeMaxIdx, int nstripes)
 229         {
 230             CV_Assert(src.isContinuous() && dst.isContinuous() &&
 231                       src.type() == CV_32F && src.type() == dst.type() &&
 232                       src.dims == 4 && dst.dims == 4 &&
 233                       (poolingType == ROI && dst.size[0] == rois.size[0] ||
 234                        src.size[0] == dst.size[0]) && src.size[1] == dst.size[1] &&
 235                       (mask.empty() || (mask.type() == src.type() && mask.size == dst.size)));
 236
 237             PoolingInvoker p;
 238
 239             p.src = &src;
 240             p.rois = &rois;
 241             p.dst = &dst;
 242             p.mask = &mask;
 243             p.kernel = kernel;
 244             p.stride = stride;
 245             p.pad = pad;
 246             p.nstripes = nstripes;
 247             p.computeMaxIdx = computeMaxIdx;
 248             p.poolingType = poolingType;
 249             p.spatialScale = spatialScale;
 250
 251             if( !computeMaxIdx )
 252             {
 253                 p.ofsbuf.resize(kernel.width*kernel.height);
 254                 for( int i = 0; i < kernel.height; i++ )
 255                     for( int j = 0; j < kernel.width; j++ )
 256                         p.ofsbuf[i*kernel.width + j] = src.size[3]*i + j;
 257             }
 258
 259             parallel_for_(Range(0, nstripes), p, nstripes);
 260         }
 261
 262         void operator()(const Range& r) const
 263         {
 264             int channels = dst->size[1], width = dst->size[3], height = dst->size[2];
 265             int inp_width = src->size[3], inp_height = src->size[2];
 266             size_t total = dst->total();
 267             size_t stripeSize = (total + nstripes - 1)/nstripes;
 268             size_t stripeStart = r.start*stripeSize;
 269             size_t stripeEnd = std::min(r.end*stripeSize, total);
 270             int kernel_w = kernel.width, kernel_h = kernel.height;
 271             int pad_w = pad.width, pad_h = pad.height;
 272             int stride_w = stride.width, stride_h = stride.height;
 273             bool compMaxIdx = computeMaxIdx;
 274
 275 #if CV_SIMD128
 276             const int* ofsptr = &ofsbuf[0];
 277             v_float32x4 idx00(0.f, (float)stride_w, (float)(stride_w*2), (float)(stride_w*3));
 278             v_float32x4 ones = v_setall_f32(1.f);
 279             v_float32x4 idx_delta = v_setall_f32((float)(inp_width - kernel_w));
 280 #endif
 281
 282             for( size_t ofs0 = stripeStart; ofs0 < stripeEnd; )
 283             {
 284                 size_t ofs = ofs0;
 285                 int x0 = (int)(ofs % width);
 286                 ofs /= width;
 287                 int y0 = (int)(ofs % height);
 288                 ofs /= height;
 289                 int c = (int)(ofs % channels);
 290                 int n = (int)(ofs / channels);
 291                 int ystart, yend;
 292
 293                 const float *srcData;
 294                 int xstartROI = 0;
 295                 float roiRatio = 0;
 296                 if (poolingType == ROI)
 297                 {
 298                     const float *roisData = rois->ptr<float>(n);
 299                     int ystartROI = round(roisData[2] * spatialScale);
 300                     int yendROI = round(roisData[4] * spatialScale);
 301                     int roiHeight = std::max(yendROI - ystartROI + 1, 1);
 302                     roiRatio = (float)roiHeight / height;
 303
 304                     ystart = ystartROI + y0 * roiRatio;
 305                     yend = ystartROI + std::ceil((y0 + 1) * roiRatio);
 306
 307                     xstartROI = round(roisData[1] * spatialScale);
 308                     int xendROI = round(roisData[3] * spatialScale);
 309                     int roiWidth = std::max(xendROI - xstartROI + 1, 1);
 310                     roiRatio = (float)roiWidth / width;
 311
 312                     CV_Assert(roisData[0] < src->size[0]);
 313                     srcData = src->ptr<float>(roisData[0], c);
 314                 }
 315                 else
 316                 {
 317                     ystart = y0 * stride_h - pad_h;
 318                     yend = min(ystart + kernel_h, inp_height + pad_h);
 319                     srcData = src->ptr<float>(n, c);
 320                 }
 321                 int ydelta = yend - ystart;
 322                 ystart = max(ystart, 0);
 323                 yend = min(yend, inp_height);
 324                 float *dstData = dst->ptr<float>(n, c, y0);
 325                 float *dstMaskData = mask->data ? mask->ptr<float>(n, c, y0) : 0;
 326
 327                 int delta = std::min((int)(stripeEnd - ofs0), width - x0);
 328                 ofs0 += delta;
 329                 int x1 = x0 + delta;
 330
 331                 if( poolingType == MAX || poolingType == ROI)
 332                     for( ; x0 < x1; x0++ )
 333                     {
 334                         int xstart, xend;
 335                         if (poolingType == ROI)
 336                         {
 337                             xstart = xstartROI + x0 * roiRatio;
 338                             xend = xstartROI + std::ceil((x0 + 1) * roiRatio);
 339                         }
 340                         else
 341                         {
 342                             xstart = x0 * stride_w - pad_w;
 343                             xend = xstart + kernel_w;
 344                         }
 345                         xstart = max(xstart, 0);
 346                         xend = min(xend, inp_width);
 347                         if (xstart >= xend || ystart >= yend)
 348                         {
 349                             dstData[x0] = 0;
 350                             if (compMaxIdx && dstMaskData)
 351                                 dstMaskData[x0] = -1;
 352                             continue;
 353                         }
 354 #if CV_SIMD128
 355                         if( xstart > 0 && x0 + 7 < x1 && (x0 + 7) * stride_w - pad_w + kernel_w < inp_width )
 356                         {
 357                             if( compMaxIdx )
 358                             {
 359                                 v_float32x4 max_val0 = v_setall_f32(-FLT_MAX);
 360                                 v_float32x4 max_val1 = max_val0;
 361                                 v_float32x4 max_idx0 = v_setall_f32(-1.f);
 362                                 v_float32x4 max_idx1 = max_idx0;
 363                                 int index0 = ystart * inp_width + xstart;
 364                                 v_float32x4 idx0 = idx00 + v_setall_f32((float)index0);
 365                                 v_float32x4 idx1 = idx0 + v_setall_f32((float)(stride_w*4));
 366
 367                                 for (int y = ystart; y < yend; ++y)
 368                                 {
 369                                     for (int x = xstart; x < xend; ++x, idx0 += ones, idx1 += ones)
 370                                     {
 371                                         const int index = y * inp_width + x;
 372                                         v_float32x4 v0(srcData[index], srcData[index + stride_w],
 373                                                        srcData[index + stride_w*2], srcData[index + stride_w*3]);
 374                                         v_float32x4 v1(srcData[index + stride_w*4], srcData[index + stride_w*5],
 375                                                        srcData[index + stride_w*6], srcData[index + stride_w*7]);
 376                                         max_idx0 = v_select(v0 > max_val0, idx0, max_idx0);
 377                                         max_idx1 = v_select(v1 > max_val1, idx1, max_idx1);
 378                                         max_val0 = v_max(max_val0, v0);
 379                                         max_val1 = v_max(max_val1, v1);
 380                                     }
 381                                     idx0 += idx_delta;
 382                                     idx1 += idx_delta;
 383                                 }
 384                                 v_store(dstData + x0, max_val0);
 385                                 v_store(dstData + x0 + 4, max_val1);
 386                                 if (dstMaskData)
 387                                 {
 388                                     v_store(dstMaskData + x0, max_idx0);
 389                                     v_store(dstMaskData + x0 + 4, max_idx1);
 390                                 }
 391                                 x0 += 7;
 392                             }
 393                             else
 394                             {
 395                                 v_float32x4 max_val0 = v_setall_f32(-FLT_MAX);
 396                                 v_float32x4 max_val1 = max_val0;
 397
 398                                 if( yend - ystart == kernel_h )
 399                                 {
 400                                     const float* srcData1 = srcData + ystart*inp_width + xstart;
 401                                     if( stride_w == 1 )
 402                                         for (int k = 0; k < kernel_w*kernel_h; k++)
 403                                         {
 404                                             int index = ofsptr[k];
 405                                             v_float32x4 v0 = v_load(srcData1 + index);
 406                                             v_float32x4 v1 = v_load(srcData1 + index + 4);
 407                                             max_val0 = v_max(max_val0, v0);
 408                                             max_val1 = v_max(max_val1, v1);
 409                                         }
 410 #if CV_SSE2
 411                                     else if( stride_w == 2 )
 412                                         for (int k = 0; k < kernel_w*kernel_h; k++)
 413                                         {
 414                                             int index = ofsptr[k];
 415                                             v_float32x4 v00 = v_load(srcData1 + index), v01 = v_load(srcData1 + index + 4);
 416                                             v_float32x4 v0(_mm_shuffle_ps(v00.val, v01.val, _MM_SHUFFLE(2, 0, 2, 0)));
 417                                             v_float32x4 v10 = v_load(srcData1 + index + 8), v11 = v_load(srcData1 + index + 12);
 418                                             v_float32x4 v1(_mm_shuffle_ps(v10.val, v11.val, _MM_SHUFFLE(2, 0, 2, 0)));
 419                                             max_val0 = v_max(max_val0, v0);
 420                                             max_val1 = v_max(max_val1, v1);
 421                                         }
 422 #endif
 423                                     else
 424                                         for (int k = 0; k < kernel_w*kernel_h; k++)
 425                                         {
 426                                             int index = ofsptr[k];
 427                                             v_float32x4 v0(srcData1[index], srcData1[index + stride_w],
 428                                                            srcData1[index + stride_w*2], srcData1[index + stride_w*3]);
 429                                             v_float32x4 v1(srcData1[index + stride_w*4], srcData1[index + stride_w*5],
 430                                                            srcData1[index + stride_w*6], srcData1[index + stride_w*7]);
 431                                             max_val0 = v_max(max_val0, v0);
 432                                             max_val1 = v_max(max_val1, v1);
 433                                         }
 434                                 }
 435                                 else
 436                                 {
 437                                     for (int y = ystart; y < yend; ++y)
 438                                     {
 439                                         for (int x = xstart; x < xend; ++x)
 440                                         {
 441                                             const int index = y * inp_width + x;
 442                                             v_float32x4 v0(srcData[index], srcData[index + stride_w],
 443                                                            srcData[index + stride_w*2], srcData[index + stride_w*3]);
 444                                             v_float32x4 v1(srcData[index + stride_w*4], srcData[index + stride_w*5],
 445                                                            srcData[index + stride_w*6], srcData[index + stride_w*7]);
 446                                             max_val0 = v_max(max_val0, v0);
 447                                             max_val1 = v_max(max_val1, v1);
 448                                         }
 449                                     }
 450                                 }
 451                                 v_store(dstData + x0, max_val0);
 452                                 v_store(dstData + x0 + 4, max_val1);
 453                                 x0 += 7;
 454                             }
 455                         }
 456                         else
 457 #endif
 458                         {
 459                             float max_val = -FLT_MAX;
 460                             if( compMaxIdx )
 461                             {
 462                                 int max_index = -1;
 463                                 for (int y = ystart; y < yend; ++y)
 464                                     for (int x = xstart; x < xend; ++x)
 465                                     {
 466                                         const int index = y * inp_width + x;
 467                                         float val = srcData[index];
 468                                         if (val > max_val)
 469                                         {
 470                                             max_val = val;
 471                                             max_index = index;
 472                                         }
 473                                     }
 474
 475                                 dstData[x0] = max_val;
 476                                 if (dstMaskData)
 477                                     dstMaskData[x0] = max_index;
 478                             }
 479                             else
 480                             {
 481                                 for (int y = ystart; y < yend; ++y)
 482                                     for (int x = xstart; x < xend; ++x)
 483                                     {
 484                                         const int index = y * inp_width + x;
 485                                         float val = srcData[index];
 486                                         max_val = std::max(max_val, val);
 487                                     }
 488
 489                                 dstData[x0] = max_val;
 490                             }
 491                         }
 492                     }
 493                 else
 494                 {
 495                     for( ; x0 < x1; x0++ )
 496                     {
 497                         int xstart = x0 * stride_w - pad_w;
 498                         int xend = min(xstart + kernel_w, inp_width + pad_w);
 499                         int xdelta = xend - xstart;
 500                         xstart = max(xstart, 0);
 501                         xend = min(xend, inp_width);
 502                         float inv_kernel_area = 1.f/(ydelta*xdelta);
 503
 504 #if CV_SIMD128
 505                         if( xstart > 0 && x0 + 7 < x1 && (x0 + 7) * stride_w - pad_w + kernel_w < inp_width )
 506                         {
 507                             v_float32x4 sum_val0 = v_setzero_f32(), sum_val1 = v_setzero_f32();
 508                             v_float32x4 ikarea = v_setall_f32(inv_kernel_area);
 509
 510                             for (int y = ystart; y < yend; ++y)
 511                             {
 512                                 for (int x = xstart; x < xend; ++x)
 513                                 {
 514                                     const int index = y * inp_width + x;
 515                                     v_float32x4 v0(srcData[index], srcData[index + stride_w],
 516                                                    srcData[index + stride_w*2], srcData[index + stride_w*3]);
 517                                     v_float32x4 v1(srcData[index + stride_w*4], srcData[index + stride_w*5],
 518                                                    srcData[index + stride_w*6], srcData[index + stride_w*7]);
 519                                     sum_val0 += v0;
 520                                     sum_val1 += v1;
 521                                 }
 522                             }
 523                             v_store(dstData + x0, sum_val0*ikarea);
 524                             v_store(dstData + x0 + 4, sum_val1*ikarea);
 525                             x0 += 7;
 526                         }
 527                         else
 528 #endif
 529                         {
 530                             float sum_val = 0.f;
 531                             for (int y = ystart; y < yend; ++y)
 532                                 for (int x = xstart; x < xend; ++x)
 533                                 {
 534                                     const int index = y * inp_width + x;
 535                                     float val = srcData[index];
 536                                     sum_val += val;
 537                                 }
 538
 539                             dstData[x0] = sum_val*inv_kernel_area;
 540                         }
 541                     }
 542                 }
 543             }
 544         }
 545     };
 546
 547     void maxPooling(Mat &src, Mat &dst, Mat &mask)
 548     {
 549         const int nstripes = getNumThreads();
 550         Mat rois;
 551         PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad, type, spatialScale, computeMaxIdx, nstripes);
 552     }
 553
 554     void avePooling(Mat &src, Mat &dst)
 555     {
 556         const int nstripes = getNumThreads();
 557         Mat rois, mask;
 558         PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad, type, spatialScale, computeMaxIdx, nstripes);
 559     }
 560
 561     void roiPooling(const Mat &src, const Mat &rois, Mat &dst)
 562     {
 563         const int nstripes = getNumThreads();
 564         Mat mask;
 565         PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad, type, spatialScale, computeMaxIdx, nstripes);
 566     }
 567
 568     virtual Ptr<BackendNode> initMaxPoolingHalide(const std::vector<Ptr<BackendWrapper> > &inputs)
 569     {
 570 #ifdef HAVE_HALIDE
 571         Halide::Buffer<float> inputBuffer = halideBuffer(inputs[0]);
 572         const int inWidth = inputBuffer.width();
 573         const int inHeight = inputBuffer.height();
 574
 575         Halide::Var x("x"), y("y"), c("c"), n("n");
 576         Halide::Func top = (name.empty() ? Halide::Func() : Halide::Func(name));
 577         Halide::RDom r(0, kernel.width, 0, kernel.height);
 578         Halide::Expr kx, ky;
 579         if (pad.width || pad.height)
 580         {
 581             kx = clamp(x * stride.width + r.x - pad.width, 0, inWidth - 1);
 582             ky = clamp(y * stride.height + r.y - pad.height, 0, inHeight - 1);
 583         }
 584         else
 585         {
 586             kx = min(x * stride.width + r.x, inWidth - 1);
 587             ky = min(y * stride.height + r.y, inHeight - 1);
 588         }
 589
 590         // Halide::argmax returns tuple (r.x, r.y, max).
 591         Halide::Tuple res = argmax(inputBuffer(kx, ky, c, n));
 592
 593         // Compute offset from argmax in range [0, kernel_size).
 594         Halide::Expr max_index;
 595         if (pad.width || pad.height)
 596         {
 597             max_index = clamp(y * stride.height + res[1] - pad.height,
 598                               0, inHeight - 1) * inWidth +
 599                         clamp(x * stride.width + res[0] - pad.width,
 600                               0, inWidth - 1);
 601         }
 602         else
 603         {
 604             max_index = min(y * stride.height + res[1], inHeight - 1) * inWidth +
 605                         min(x * stride.width + res[0], inWidth - 1);
 606         }
 607         top(x, y, c, n) = { res[2], Halide::cast<float>(max_index) };
 608         return Ptr<BackendNode>(new HalideBackendNode(top));
 609 #endif  // HAVE_HALIDE
 610         return Ptr<BackendNode>();
 611     }
 612
 613     virtual Ptr<BackendNode> initAvePoolingHalide(const std::vector<Ptr<BackendWrapper> > &inputs)
 614     {
 615 #ifdef HAVE_HALIDE
 616         Halide::Buffer<float> inputBuffer = halideBuffer(inputs[0]);
 617
 618         const int inW = inputBuffer.width(), inH = inputBuffer.height();
 619         if ((inW - kernel.width) % stride.width || (inH - kernel.height) % stride.height)
 620         {
 621             CV_Error(cv::Error::StsNotImplemented,
 622                      "Halide backend for average pooling with partial "
 623                      "kernels is not implemented");
 624         }
 625
 626         const float norm = 1.0f / (kernel.width * kernel.height);
 627
 628         Halide::Var x("x"), y("y"), c("c"), n("n");
 629         Halide::Func top = (name.empty() ? Halide::Func() : Halide::Func(name));
 630         Halide::RDom r(0, kernel.width, 0, kernel.height);
 631         top(x, y, c, n) = sum(
 632             inputBuffer(x * stride.width + r.x,
 633                         y * stride.height + r.y, c, n)) * norm;
 634         return Ptr<BackendNode>(new HalideBackendNode(top));
 635 #endif  // HAVE_HALIDE
 636         return Ptr<BackendNode>();
 637     }
 638
 639     virtual void applyHalideScheduler(Ptr<BackendNode>& node,
 640                                       const std::vector<Mat*> &inputs,
 641                                       const std::vector<Mat> &outputs,
 642                                       int targetId) const
 643     {
 644 #ifdef  HAVE_HALIDE
 645         if (targetId != DNN_TARGET_CPU)
 646         {
 647             Layer::applyHalideScheduler(node, inputs, outputs, targetId);
 648             return;
 649         }
 650         Halide::Var x("x"), y("y"), c("c"), n("n"), tile("tile"),
 651                     xi("xi"), yi("yi"), ci("ci"), xo("xo"), yo("yo"), co("co");
 652         Halide::Func& top = node.dynamicCast<HalideBackendNode>()->funcs.back();
 653
 654         int outW, outH, outC, outN;
 655         getCanonicalSize(outputs[0].size, &outW, &outH, &outC, &outN);
 656
 657         if (outW < 8 || outH < 8)
 658         {
 659             if (outC > 8)
 660                 top.split(c, co, ci, 8)
 661                    .fuse(x, y, tile).fuse(co, tile, tile).fuse(n, tile, tile)
 662                    .parallel(tile)
 663                    .vectorize(ci);
 664             else
 665             {
 666                 top.fuse(y, c, tile).fuse(n, tile, tile)
 667                    .parallel(tile);
 668                 if (outW > 1)
 669                     top.vectorize(x);
 670             }
 671         }
 672         else
 673         {
 674             if (outC > 8)
 675                 top.split(x, xo, xi, 8).split(y, yo, yi, 8).split(c, co, ci, 8)
 676                    .fuse(xo, yo, tile).fuse(co, tile, tile).fuse(n, tile, tile)
 677                    .parallel(tile)
 678                    .vectorize(xi);
 679             else
 680                 top.split(x, xo, xi, 8).split(y, yo, yi, 8)
 681                    .fuse(xo, yo, tile).fuse(c, tile, tile).fuse(n, tile, tile)
 682                    .parallel(tile)
 683                    .vectorize(xi);
 684         }
 685 #endif  // HAVE_HALIDE
 686     }
 687
 688     bool getMemoryShapes(const std::vector<MatShape> &inputs,
 689                          const int requiredOutputs,
 690                          std::vector<MatShape> &outputs,
 691                          std::vector<MatShape> &internals) const
 692     {
 693         CV_Assert(inputs.size() != 0);
 694         Size in(inputs[0][3], inputs[0][2]), out;
 695
 696         if (globalPooling)
 697         {
 698             out.height = 1;
 699             out.width = 1;
 700         }
 701         else if (type == PoolingLayer::ROI)
 702         {
 703             out.height = pooledSize.height;
 704             out.width = pooledSize.width;
 705         }
 706         else if (padMode.empty())
 707         {
 708             float height = (float)(in.height + 2 * pad.height - kernel.height) / stride.height;
 709             float width = (float)(in.width + 2 * pad.width - kernel.width) / stride.width;
 710             out.height = 1 + (ceilMode ? ceil(height) : floor(height));
 711             out.width = 1 + (ceilMode ? ceil(width) : floor(width));
 712
 713             if (pad.height || pad.width)
 714             {
 715                 // If we have padding, ensure that the last pooling starts strictly
 716                 // inside the image (instead of at the padding); otherwise clip the last.
 717                 if ((out.height - 1) * stride.height >= in.height + pad.height)
 718                     --out.height;
 719                 if ((out.width - 1) * stride.width >= in.width + pad.width)
 720                     --out.width;
 721                 CV_Assert((out.height - 1) * stride.height < in.height + pad.height);
 722                 CV_Assert((out.width - 1) * stride.width < in.width + pad.width);
 723             }
 724         }
 725         else
 726         {
 727             getConvPoolOutParams(in, kernel, stride, padMode, Size(1, 1), out);
 728         }
 729
 730         int dims[] = {inputs[0][0], inputs[0][1], out.height, out.width};
 731         if (type == ROI)
 732         {
 733             CV_Assert(inputs.size() == 2);
 734             dims[0] = inputs[1][0];  // Number of proposals;
 735         }
 736         outputs.assign(type == MAX ? 2 : 1, shape(dims));
 737         return false;
 738     }
 739
 740     virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
 741                            const std::vector<MatShape> &outputs) const
 742     {
 743         (void)inputs; // suppress unused variable warning
 744         long flops = 0;
 745
 746         for(int i = 0; i < outputs.size(); i++)
 747         {
 748             if (type == MAX)
 749             {
 750                 if (i%2 == 0)
 751                     flops += total(outputs[i])*kernel.area();
 752             }
 753             else
 754             {
 755                 flops += total(outputs[i])*(kernel.area() + 1);
 756             }
 757         }
 758         return flops;
 759     }
 760 };
 761
 762 Ptr<PoolingLayer> PoolingLayer::create(const LayerParams& params)
 763 {
 764     return Ptr<PoolingLayer>(new PoolingLayerImpl(params));
 765 }
 766
 767 }
 768 }