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 static inline int scaleAndRoundRoi(float f, float scale)
  61 {
  62     return (int)(f * scale + (f >= 0.f ? 0.5f : -0.5f));
  63 }
  64
  65 class PoolingLayerImpl : public PoolingLayer
  66 {
  67 public:
  68     PoolingLayerImpl(const LayerParams& params)
  69     {
  70         type = MAX;
  71         computeMaxIdx = true;
  72         globalPooling = false;
  73
  74         if (params.has("pool"))
  75         {
  76             String pool = params.get<String>("pool").toLowerCase();
  77             if (pool == "max")
  78                 type = MAX;
  79             else if (pool == "ave")
  80                 type = AVE;
  81             else if (pool == "stochastic")
  82                 type = STOCHASTIC;
  83             else
  84                 CV_Error(Error::StsBadArg, "Unknown pooling type \"" + pool + "\"");
  85             getPoolingKernelParams(params, kernel.height, kernel.width, globalPooling,
  86                                    pad.height, pad.width, stride.height, stride.width, padMode);
  87         }
  88         else if (params.has("pooled_w") || params.has("pooled_h") || params.has("spatial_scale"))
  89         {
  90             type = ROI;
  91         }
  92         setParamsFrom(params);
  93         ceilMode = params.get<bool>("ceil_mode", true);
  94         pooledSize.width = params.get<uint32_t>("pooled_w", 1);
  95         pooledSize.height = params.get<uint32_t>("pooled_h", 1);
  96         spatialScale = params.get<float>("spatial_scale", 1);
  97     }
  98
  99 #ifdef HAVE_OPENCL
 100     Ptr<OCL4DNNPool<float> > poolOp;
 101 #endif
 102
 103     void finalize(const std::vector<Mat*> &inputs, std::vector<Mat> &outputs)
 104     {
 105         CV_Assert(!inputs.empty());
 106
 107         cv::Size inp(inputs[0]->size[3], inputs[0]->size[2]),
 108                 out(outputs[0].size[3], outputs[0].size[2]);
 109
 110         if(globalPooling)
 111         {
 112             kernel = inp;
 113         }
 114
 115         getConvPoolPaddings(inp, out, kernel, stride, padMode, Size(1, 1), pad);
 116     }
 117
 118     virtual bool supportBackend(int backendId)
 119     {
 120         return backendId == DNN_BACKEND_DEFAULT ||
 121                backendId == DNN_BACKEND_HALIDE && haveHalide() &&
 122                (type == MAX || type == AVE && !pad.width && !pad.height);
 123     }
 124
 125 #ifdef HAVE_OPENCL
 126     bool forward_ocl(InputArrayOfArrays inps, OutputArrayOfArrays outs, InputArrayOfArrays internals)
 127     {
 128         std::vector<UMat> inputs;
 129         std::vector<UMat> outputs;
 130
 131         inps.getUMatVector(inputs);
 132         outs.getUMatVector(outputs);
 133
 134         if (poolOp.empty())
 135         {
 136             OCL4DNNPoolConfig config;
 137
 138             config.in_shape = shape(inputs[0]);
 139             config.out_shape = shape(outputs[0]);
 140             config.kernel = kernel;
 141             config.pad = pad;
 142             config.stride = stride;
 143             config.channels = inputs[0].size[1];
 144             config.pool_method = type == MAX ? LIBDNN_POOLING_METHOD_MAX :
 145                                 (type == AVE ? LIBDNN_POOLING_METHOD_AVE :
 146                                                LIBDNN_POOLING_METHOD_STO);
 147             poolOp = Ptr<OCL4DNNPool<float> >(new OCL4DNNPool<float>(config));
 148         }
 149
 150         for (size_t ii = 0; ii < inputs.size(); ii++)
 151         {
 152             UMat& inpMat = inputs[ii];
 153             int out_index = (type == MAX) ? 2 : 1;
 154             UMat& outMat = outputs[out_index * ii];
 155             UMat maskMat = (type == MAX) ? outputs[2 * ii + 1] : UMat();
 156
 157             CV_Assert(inpMat.offset == 0 && outMat.offset == 0);
 158
 159             if (!poolOp->Forward(inpMat, outMat, maskMat))
 160                 return false;
 161         }
 162
 163         return true;
 164     }
 165 #endif
 166
 167     void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
 168     {
 169         CV_TRACE_FUNCTION();
 170         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 171
 172         CV_OCL_RUN((preferableTarget == DNN_TARGET_OPENCL) &&
 173                    OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
 174                    forward_ocl(inputs_arr, outputs_arr, internals_arr))
 175
 176         Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
 177     }
 178
 179     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
 180     {
 181         CV_TRACE_FUNCTION();
 182         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 183
 184         switch (type)
 185         {
 186             case MAX:
 187                 CV_Assert(inputs.size() == 1, outputs.size() == 2);
 188                 maxPooling(*inputs[0], outputs[0], outputs[1]);
 189                 break;
 190             case AVE:
 191                 CV_Assert(inputs.size() == 1, outputs.size() == 1);
 192                 avePooling(*inputs[0], outputs[0]);
 193                 break;
 194             case ROI:
 195                 CV_Assert(inputs.size() == 2, outputs.size() == 1);
 196                 roiPooling(*inputs[0], *inputs[1], outputs[0]);
 197                 break;
 198             default:
 199                 CV_Error(Error::StsNotImplemented, "Not implemented");
 200                 break;
 201         }
 202     }
 203
 204     virtual Ptr<BackendNode> initHalide(const std::vector<Ptr<BackendWrapper> > &inputs)
 205     {
 206         if (type == MAX)
 207             return initMaxPoolingHalide(inputs);
 208         else if (type == AVE)
 209             return initAvePoolingHalide(inputs);
 210         else
 211             return Ptr<BackendNode>();
 212     }
 213
 214     class PoolingInvoker : public ParallelLoopBody
 215     {
 216     public:
 217         const Mat* src, *rois;
 218         Mat *dst, *mask;
 219         Size kernel, stride, pad;
 220         int nstripes;
 221         bool computeMaxIdx;
 222         std::vector<int> ofsbuf;
 223         int poolingType;
 224         float spatialScale;
 225
 226         PoolingInvoker() : src(0), rois(0), dst(0), mask(0), nstripes(0),
 227                            computeMaxIdx(0), poolingType(MAX), spatialScale(0) {}
 228
 229         static void run(const Mat& src, const Mat& rois, Mat& dst, Mat& mask, Size kernel,
 230                         Size stride, Size pad, int poolingType, float spatialScale,
 231                         bool computeMaxIdx, int nstripes)
 232         {
 233             CV_Assert(src.isContinuous() && dst.isContinuous() &&
 234                       src.type() == CV_32F && src.type() == dst.type() &&
 235                       src.dims == 4 && dst.dims == 4 &&
 236                       (poolingType == ROI && dst.size[0] == rois.size[0] ||
 237                        src.size[0] == dst.size[0]) && src.size[1] == dst.size[1] &&
 238                       (mask.empty() || (mask.type() == src.type() && mask.size == dst.size)));
 239
 240             PoolingInvoker p;
 241
 242             p.src = &src;
 243             p.rois = &rois;
 244             p.dst = &dst;
 245             p.mask = &mask;
 246             p.kernel = kernel;
 247             p.stride = stride;
 248             p.pad = pad;
 249             p.nstripes = nstripes;
 250             p.computeMaxIdx = computeMaxIdx;
 251             p.poolingType = poolingType;
 252             p.spatialScale = spatialScale;
 253
 254             if( !computeMaxIdx )
 255             {
 256                 p.ofsbuf.resize(kernel.width*kernel.height);
 257                 for( int i = 0; i < kernel.height; i++ )
 258                     for( int j = 0; j < kernel.width; j++ )
 259                         p.ofsbuf[i*kernel.width + j] = src.size[3]*i + j;
 260             }
 261
 262             parallel_for_(Range(0, nstripes), p, nstripes);
 263         }
 264
 265         void operator()(const Range& r) const
 266         {
 267             int channels = dst->size[1], width = dst->size[3], height = dst->size[2];
 268             int inp_width = src->size[3], inp_height = src->size[2];
 269             size_t total = dst->total();
 270             size_t stripeSize = (total + nstripes - 1)/nstripes;
 271             size_t stripeStart = r.start*stripeSize;
 272             size_t stripeEnd = std::min(r.end*stripeSize, total);
 273             int kernel_w = kernel.width, kernel_h = kernel.height;
 274             int pad_w = pad.width, pad_h = pad.height;
 275             int stride_w = stride.width, stride_h = stride.height;
 276             bool compMaxIdx = computeMaxIdx;
 277
 278 #if CV_SIMD128
 279             const int* ofsptr = &ofsbuf[0];
 280             v_float32x4 idx00(0.f, (float)stride_w, (float)(stride_w*2), (float)(stride_w*3));
 281             v_float32x4 ones = v_setall_f32(1.f);
 282             v_float32x4 idx_delta = v_setall_f32((float)(inp_width - kernel_w));
 283 #endif
 284
 285             for( size_t ofs0 = stripeStart; ofs0 < stripeEnd; )
 286             {
 287                 size_t ofs = ofs0;
 288                 int x0 = (int)(ofs % width);
 289                 ofs /= width;
 290                 int y0 = (int)(ofs % height);
 291                 ofs /= height;
 292                 int c = (int)(ofs % channels);
 293                 int n = (int)(ofs / channels);
 294                 int ystart, yend;
 295
 296                 const float *srcData;
 297                 int xstartROI = 0;
 298                 float roiRatio = 0;
 299                 if (poolingType == ROI)
 300                 {
 301                     const float *roisData = rois->ptr<float>(n);
 302                     int ystartROI = scaleAndRoundRoi(roisData[2], spatialScale);
 303                     int yendROI = scaleAndRoundRoi(roisData[4], spatialScale);
 304                     int roiHeight = std::max(yendROI - ystartROI + 1, 1);
 305                     roiRatio = (float)roiHeight / height;
 306
 307                     ystart = ystartROI + y0 * roiRatio;
 308                     yend = ystartROI + std::ceil((y0 + 1) * roiRatio);
 309
 310                     xstartROI = scaleAndRoundRoi(roisData[1], spatialScale);
 311                     int xendROI = scaleAndRoundRoi(roisData[3], spatialScale);
 312                     int roiWidth = std::max(xendROI - xstartROI + 1, 1);
 313                     roiRatio = (float)roiWidth / width;
 314
 315                     CV_Assert(roisData[0] < src->size[0]);
 316                     srcData = src->ptr<float>(roisData[0], c);
 317                 }
 318                 else
 319                 {
 320                     ystart = y0 * stride_h - pad_h;
 321                     yend = min(ystart + kernel_h, inp_height + pad_h);
 322                     srcData = src->ptr<float>(n, c);
 323                 }
 324                 int ydelta = yend - ystart;
 325                 ystart = max(ystart, 0);
 326                 yend = min(yend, inp_height);
 327                 float *dstData = dst->ptr<float>(n, c, y0);
 328                 float *dstMaskData = mask->data ? mask->ptr<float>(n, c, y0) : 0;
 329
 330                 int delta = std::min((int)(stripeEnd - ofs0), width - x0);
 331                 ofs0 += delta;
 332                 int x1 = x0 + delta;
 333
 334                 if( poolingType == MAX || poolingType == ROI)
 335                     for( ; x0 < x1; x0++ )
 336                     {
 337                         int xstart, xend;
 338                         if (poolingType == ROI)
 339                         {
 340                             xstart = xstartROI + x0 * roiRatio;
 341                             xend = xstartROI + std::ceil((x0 + 1) * roiRatio);
 342                         }
 343                         else
 344                         {
 345                             xstart = x0 * stride_w - pad_w;
 346                             xend = xstart + kernel_w;
 347                         }
 348                         xstart = max(xstart, 0);
 349                         xend = min(xend, inp_width);
 350                         if (xstart >= xend || ystart >= yend)
 351                         {
 352                             dstData[x0] = 0;
 353                             if (compMaxIdx && dstMaskData)
 354                                 dstMaskData[x0] = -1;
 355                             continue;
 356                         }
 357 #if CV_SIMD128
 358                         if( xstart > 0 && x0 + 7 < x1 && (x0 + 7) * stride_w - pad_w + kernel_w < inp_width )
 359                         {
 360                             if( compMaxIdx )
 361                             {
 362                                 v_float32x4 max_val0 = v_setall_f32(-FLT_MAX);
 363                                 v_float32x4 max_val1 = max_val0;
 364                                 v_float32x4 max_idx0 = v_setall_f32(-1.f);
 365                                 v_float32x4 max_idx1 = max_idx0;
 366                                 int index0 = ystart * inp_width + xstart;
 367                                 v_float32x4 idx0 = idx00 + v_setall_f32((float)index0);
 368                                 v_float32x4 idx1 = idx0 + v_setall_f32((float)(stride_w*4));
 369
 370                                 for (int y = ystart; y < yend; ++y)
 371                                 {
 372                                     for (int x = xstart; x < xend; ++x, idx0 += ones, idx1 += ones)
 373                                     {
 374                                         const int index = y * inp_width + x;
 375                                         v_float32x4 v0(srcData[index], srcData[index + stride_w],
 376                                                        srcData[index + stride_w*2], srcData[index + stride_w*3]);
 377                                         v_float32x4 v1(srcData[index + stride_w*4], srcData[index + stride_w*5],
 378                                                        srcData[index + stride_w*6], srcData[index + stride_w*7]);
 379                                         max_idx0 = v_select(v0 > max_val0, idx0, max_idx0);
 380                                         max_idx1 = v_select(v1 > max_val1, idx1, max_idx1);
 381                                         max_val0 = v_max(max_val0, v0);
 382                                         max_val1 = v_max(max_val1, v1);
 383                                     }
 384                                     idx0 += idx_delta;
 385                                     idx1 += idx_delta;
 386                                 }
 387                                 v_store(dstData + x0, max_val0);
 388                                 v_store(dstData + x0 + 4, max_val1);
 389                                 if (dstMaskData)
 390                                 {
 391                                     v_store(dstMaskData + x0, max_idx0);
 392                                     v_store(dstMaskData + x0 + 4, max_idx1);
 393                                 }
 394                                 x0 += 7;
 395                             }
 396                             else
 397                             {
 398                                 v_float32x4 max_val0 = v_setall_f32(-FLT_MAX);
 399                                 v_float32x4 max_val1 = max_val0;
 400
 401                                 if( yend - ystart == kernel_h )
 402                                 {
 403                                     const float* srcData1 = srcData + ystart*inp_width + xstart;
 404                                     if( stride_w == 1 )
 405                                         for (int k = 0; k < kernel_w*kernel_h; k++)
 406                                         {
 407                                             int index = ofsptr[k];
 408                                             v_float32x4 v0 = v_load(srcData1 + index);
 409                                             v_float32x4 v1 = v_load(srcData1 + index + 4);
 410                                             max_val0 = v_max(max_val0, v0);
 411                                             max_val1 = v_max(max_val1, v1);
 412                                         }
 413 #if CV_SSE2
 414                                     else if( stride_w == 2 )
 415                                         for (int k = 0; k < kernel_w*kernel_h; k++)
 416                                         {
 417                                             int index = ofsptr[k];
 418                                             v_float32x4 v00 = v_load(srcData1 + index), v01 = v_load(srcData1 + index + 4);
 419                                             v_float32x4 v0(_mm_shuffle_ps(v00.val, v01.val, _MM_SHUFFLE(2, 0, 2, 0)));
 420                                             v_float32x4 v10 = v_load(srcData1 + index + 8), v11 = v_load(srcData1 + index + 12);
 421                                             v_float32x4 v1(_mm_shuffle_ps(v10.val, v11.val, _MM_SHUFFLE(2, 0, 2, 0)));
 422                                             max_val0 = v_max(max_val0, v0);
 423                                             max_val1 = v_max(max_val1, v1);
 424                                         }
 425 #endif
 426                                     else
 427                                         for (int k = 0; k < kernel_w*kernel_h; k++)
 428                                         {
 429                                             int index = ofsptr[k];
 430                                             v_float32x4 v0(srcData1[index], srcData1[index + stride_w],
 431                                                            srcData1[index + stride_w*2], srcData1[index + stride_w*3]);
 432                                             v_float32x4 v1(srcData1[index + stride_w*4], srcData1[index + stride_w*5],
 433                                                            srcData1[index + stride_w*6], srcData1[index + stride_w*7]);
 434                                             max_val0 = v_max(max_val0, v0);
 435                                             max_val1 = v_max(max_val1, v1);
 436                                         }
 437                                 }
 438                                 else
 439                                 {
 440                                     for (int y = ystart; y < yend; ++y)
 441                                     {
 442                                         for (int x = xstart; x < xend; ++x)
 443                                         {
 444                                             const int index = y * inp_width + x;
 445                                             v_float32x4 v0(srcData[index], srcData[index + stride_w],
 446                                                            srcData[index + stride_w*2], srcData[index + stride_w*3]);
 447                                             v_float32x4 v1(srcData[index + stride_w*4], srcData[index + stride_w*5],
 448                                                            srcData[index + stride_w*6], srcData[index + stride_w*7]);
 449                                             max_val0 = v_max(max_val0, v0);
 450                                             max_val1 = v_max(max_val1, v1);
 451                                         }
 452                                     }
 453                                 }
 454                                 v_store(dstData + x0, max_val0);
 455                                 v_store(dstData + x0 + 4, max_val1);
 456                                 x0 += 7;
 457                             }
 458                         }
 459                         else
 460 #endif
 461                         {
 462                             float max_val = -FLT_MAX;
 463                             if( compMaxIdx )
 464                             {
 465                                 int max_index = -1;
 466                                 for (int y = ystart; y < yend; ++y)
 467                                     for (int x = xstart; x < xend; ++x)
 468                                     {
 469                                         const int index = y * inp_width + x;
 470                                         float val = srcData[index];
 471                                         if (val > max_val)
 472                                         {
 473                                             max_val = val;
 474                                             max_index = index;
 475                                         }
 476                                     }
 477
 478                                 dstData[x0] = max_val;
 479                                 if (dstMaskData)
 480                                     dstMaskData[x0] = max_index;
 481                             }
 482                             else
 483                             {
 484                                 for (int y = ystart; y < yend; ++y)
 485                                     for (int x = xstart; x < xend; ++x)
 486                                     {
 487                                         const int index = y * inp_width + x;
 488                                         float val = srcData[index];
 489                                         max_val = std::max(max_val, val);
 490                                     }
 491
 492                                 dstData[x0] = max_val;
 493                             }
 494                         }
 495                     }
 496                 else
 497                 {
 498                     for( ; x0 < x1; x0++ )
 499                     {
 500                         int xstart = x0 * stride_w - pad_w;
 501                         int xend = min(xstart + kernel_w, inp_width + pad_w);
 502                         int xdelta = xend - xstart;
 503                         xstart = max(xstart, 0);
 504                         xend = min(xend, inp_width);
 505                         float inv_kernel_area = 1.f/(ydelta*xdelta);
 506
 507 #if CV_SIMD128
 508                         if( xstart > 0 && x0 + 7 < x1 && (x0 + 7) * stride_w - pad_w + kernel_w < inp_width )
 509                         {
 510                             v_float32x4 sum_val0 = v_setzero_f32(), sum_val1 = v_setzero_f32();
 511                             v_float32x4 ikarea = v_setall_f32(inv_kernel_area);
 512
 513                             for (int y = ystart; y < yend; ++y)
 514                             {
 515                                 for (int x = xstart; x < xend; ++x)
 516                                 {
 517                                     const int index = y * inp_width + x;
 518                                     v_float32x4 v0(srcData[index], srcData[index + stride_w],
 519                                                    srcData[index + stride_w*2], srcData[index + stride_w*3]);
 520                                     v_float32x4 v1(srcData[index + stride_w*4], srcData[index + stride_w*5],
 521                                                    srcData[index + stride_w*6], srcData[index + stride_w*7]);
 522                                     sum_val0 += v0;
 523                                     sum_val1 += v1;
 524                                 }
 525                             }
 526                             v_store(dstData + x0, sum_val0*ikarea);
 527                             v_store(dstData + x0 + 4, sum_val1*ikarea);
 528                             x0 += 7;
 529                         }
 530                         else
 531 #endif
 532                         {
 533                             float sum_val = 0.f;
 534                             for (int y = ystart; y < yend; ++y)
 535                                 for (int x = xstart; x < xend; ++x)
 536                                 {
 537                                     const int index = y * inp_width + x;
 538                                     float val = srcData[index];
 539                                     sum_val += val;
 540                                 }
 541
 542                             dstData[x0] = sum_val*inv_kernel_area;
 543                         }
 544                     }
 545                 }
 546             }
 547         }
 548     };
 549
 550     void maxPooling(Mat &src, Mat &dst, Mat &mask)
 551     {
 552         const int nstripes = getNumThreads();
 553         Mat rois;
 554         PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad, type, spatialScale, computeMaxIdx, nstripes);
 555     }
 556
 557     void avePooling(Mat &src, Mat &dst)
 558     {
 559         const int nstripes = getNumThreads();
 560         Mat rois, mask;
 561         PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad, type, spatialScale, computeMaxIdx, nstripes);
 562     }
 563
 564     void roiPooling(const Mat &src, const Mat &rois, Mat &dst)
 565     {
 566         const int nstripes = getNumThreads();
 567         Mat mask;
 568         PoolingInvoker::run(src, rois, dst, mask, kernel, stride, pad, type, spatialScale, computeMaxIdx, nstripes);
 569     }
 570
 571     virtual Ptr<BackendNode> initMaxPoolingHalide(const std::vector<Ptr<BackendWrapper> > &inputs)
 572     {
 573 #ifdef HAVE_HALIDE
 574         Halide::Buffer<float> inputBuffer = halideBuffer(inputs[0]);
 575         const int inWidth = inputBuffer.width();
 576         const int inHeight = inputBuffer.height();
 577
 578         Halide::Var x("x"), y("y"), c("c"), n("n");
 579         Halide::Func top = (name.empty() ? Halide::Func() : Halide::Func(name));
 580         Halide::RDom r(0, kernel.width, 0, kernel.height);
 581         Halide::Expr kx, ky;
 582         if (pad.width || pad.height)
 583         {
 584             kx = clamp(x * stride.width + r.x - pad.width, 0, inWidth - 1);
 585             ky = clamp(y * stride.height + r.y - pad.height, 0, inHeight - 1);
 586         }
 587         else
 588         {
 589             kx = min(x * stride.width + r.x, inWidth - 1);
 590             ky = min(y * stride.height + r.y, inHeight - 1);
 591         }
 592
 593         // Halide::argmax returns tuple (r.x, r.y, max).
 594         Halide::Tuple res = argmax(inputBuffer(kx, ky, c, n));
 595
 596         // Compute offset from argmax in range [0, kernel_size).
 597         Halide::Expr max_index;
 598         if (pad.width || pad.height)
 599         {
 600             max_index = clamp(y * stride.height + res[1] - pad.height,
 601                               0, inHeight - 1) * inWidth +
 602                         clamp(x * stride.width + res[0] - pad.width,
 603                               0, inWidth - 1);
 604         }
 605         else
 606         {
 607             max_index = min(y * stride.height + res[1], inHeight - 1) * inWidth +
 608                         min(x * stride.width + res[0], inWidth - 1);
 609         }
 610         top(x, y, c, n) = { res[2], Halide::cast<float>(max_index) };
 611         return Ptr<BackendNode>(new HalideBackendNode(top));
 612 #endif  // HAVE_HALIDE
 613         return Ptr<BackendNode>();
 614     }
 615
 616     virtual Ptr<BackendNode> initAvePoolingHalide(const std::vector<Ptr<BackendWrapper> > &inputs)
 617     {
 618 #ifdef HAVE_HALIDE
 619         Halide::Buffer<float> inputBuffer = halideBuffer(inputs[0]);
 620
 621         const int inW = inputBuffer.width(), inH = inputBuffer.height();
 622         if ((inW - kernel.width) % stride.width || (inH - kernel.height) % stride.height)
 623         {
 624             CV_Error(cv::Error::StsNotImplemented,
 625                      "Halide backend for average pooling with partial "
 626                      "kernels is not implemented");
 627         }
 628
 629         const float norm = 1.0f / (kernel.width * kernel.height);
 630
 631         Halide::Var x("x"), y("y"), c("c"), n("n");
 632         Halide::Func top = (name.empty() ? Halide::Func() : Halide::Func(name));
 633         Halide::RDom r(0, kernel.width, 0, kernel.height);
 634         top(x, y, c, n) = sum(
 635             inputBuffer(x * stride.width + r.x,
 636                         y * stride.height + r.y, c, n)) * norm;
 637         return Ptr<BackendNode>(new HalideBackendNode(top));
 638 #endif  // HAVE_HALIDE
 639         return Ptr<BackendNode>();
 640     }
 641
 642     virtual void applyHalideScheduler(Ptr<BackendNode>& node,
 643                                       const std::vector<Mat*> &inputs,
 644                                       const std::vector<Mat> &outputs,
 645                                       int targetId) const
 646     {
 647 #ifdef  HAVE_HALIDE
 648         if (targetId != DNN_TARGET_CPU)
 649         {
 650             Layer::applyHalideScheduler(node, inputs, outputs, targetId);
 651             return;
 652         }
 653         Halide::Var x("x"), y("y"), c("c"), n("n"), tile("tile"),
 654                     xi("xi"), yi("yi"), ci("ci"), xo("xo"), yo("yo"), co("co");
 655         Halide::Func& top = node.dynamicCast<HalideBackendNode>()->funcs.back();
 656
 657         int outW, outH, outC, outN;
 658         getCanonicalSize(outputs[0].size, &outW, &outH, &outC, &outN);
 659
 660         if (outW < 8 || outH < 8)
 661         {
 662             if (outC > 8)
 663                 top.split(c, co, ci, 8)
 664                    .fuse(x, y, tile).fuse(co, tile, tile).fuse(n, tile, tile)
 665                    .parallel(tile)
 666                    .vectorize(ci);
 667             else
 668             {
 669                 top.fuse(y, c, tile).fuse(n, tile, tile)
 670                    .parallel(tile);
 671                 if (outW > 1)
 672                     top.vectorize(x);
 673             }
 674         }
 675         else
 676         {
 677             if (outC > 8)
 678                 top.split(x, xo, xi, 8).split(y, yo, yi, 8).split(c, co, ci, 8)
 679                    .fuse(xo, yo, tile).fuse(co, tile, tile).fuse(n, tile, tile)
 680                    .parallel(tile)
 681                    .vectorize(xi);
 682             else
 683                 top.split(x, xo, xi, 8).split(y, yo, yi, 8)
 684                    .fuse(xo, yo, tile).fuse(c, tile, tile).fuse(n, tile, tile)
 685                    .parallel(tile)
 686                    .vectorize(xi);
 687         }
 688 #endif  // HAVE_HALIDE
 689     }
 690
 691     bool getMemoryShapes(const std::vector<MatShape> &inputs,
 692                          const int requiredOutputs,
 693                          std::vector<MatShape> &outputs,
 694                          std::vector<MatShape> &internals) const
 695     {
 696         CV_Assert(inputs.size() != 0);
 697         Size in(inputs[0][3], inputs[0][2]), out;
 698
 699         if (globalPooling)
 700         {
 701             out.height = 1;
 702             out.width = 1;
 703         }
 704         else if (type == ROI)
 705         {
 706             out.height = pooledSize.height;
 707             out.width = pooledSize.width;
 708         }
 709         else if (padMode.empty())
 710         {
 711             float height = (float)(in.height + 2 * pad.height - kernel.height) / stride.height;
 712             float width = (float)(in.width + 2 * pad.width - kernel.width) / stride.width;
 713             out.height = 1 + (ceilMode ? ceil(height) : floor(height));
 714             out.width = 1 + (ceilMode ? ceil(width) : floor(width));
 715
 716             if (pad.height || pad.width)
 717             {
 718                 // If we have padding, ensure that the last pooling starts strictly
 719                 // inside the image (instead of at the padding); otherwise clip the last.
 720                 if ((out.height - 1) * stride.height >= in.height + pad.height)
 721                     --out.height;
 722                 if ((out.width - 1) * stride.width >= in.width + pad.width)
 723                     --out.width;
 724                 CV_Assert((out.height - 1) * stride.height < in.height + pad.height);
 725                 CV_Assert((out.width - 1) * stride.width < in.width + pad.width);
 726             }
 727         }
 728         else
 729         {
 730             getConvPoolOutParams(in, kernel, stride, padMode, Size(1, 1), out);
 731         }
 732
 733         int dims[] = {inputs[0][0], inputs[0][1], out.height, out.width};
 734         if (type == ROI)
 735         {
 736             CV_Assert(inputs.size() == 2);
 737             dims[0] = inputs[1][0];  // Number of proposals;
 738         }
 739         outputs.assign(type == MAX ? 2 : 1, shape(dims));
 740         return false;
 741     }
 742
 743     virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
 744                            const std::vector<MatShape> &outputs) const
 745     {
 746         (void)inputs; // suppress unused variable warning
 747         long flops = 0;
 748
 749         for(int i = 0; i < outputs.size(); i++)
 750         {
 751             if (type == MAX)
 752             {
 753                 if (i%2 == 0)
 754                     flops += total(outputs[i])*kernel.area();
 755             }
 756             else
 757             {
 758                 flops += total(outputs[i])*(kernel.area() + 1);
 759             }
 760         }
 761         return flops;
 762     }
 763 private:
 764     enum Type
 765     {
 766         MAX,
 767         AVE,
 768         STOCHASTIC,
 769         ROI
 770     };
 771 };
 772
 773 Ptr<PoolingLayer> PoolingLayer::create(const LayerParams& params)
 774 {
 775     return Ptr<PoolingLayer>(new PoolingLayerImpl(params));
 776 }
 777
 778 }
 779 }