modules/ocl/src/haar.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) 2010-2012, Institute Of Software Chinese Academy Of Science, all rights reserved.
  14 // Copyright (C) 2010-2012, Advanced Micro Devices, Inc., all rights reserved.
  15 // Third party copyrights are property of their respective owners.
  16 //
  17 // @Authors
  18 //    Niko Li, newlife20080214@gmail.com
  19 //    Wang Weiyan, wangweiyanster@gmail.com
  20 //    Jia Haipeng, jiahaipeng95@gmail.com
  21 //    Wu Xinglong, wxl370@126.com
  22 //    Wang Yao, bitwangyaoyao@gmail.com
  23 //    Sen Liu, swjtuls1987@126.com
  24 //
  25 // Redistribution and use in source and binary forms, with or without modification,
  26 // are permitted provided that the following conditions are met:
  27 //
  28 //   * Redistribution's of source code must retain the above copyright notice,
  29 //     this list of conditions and the following disclaimer.
  30 //
  31 //   * Redistribution's in binary form must reproduce the above copyright notice,
  32 //     this list of conditions and the following disclaimer in the documentation
  33 //     and/or other oclMaterials provided with the distribution.
  34 //
  35 //   * The name of the copyright holders may not be used to endorse or promote products
  36 //     derived from this software without specific prior written permission.
  37 //
  38 // This software is provided by the copyright holders and contributors "as is" and
  39 // any express or implied warranties, including, but not limited to, the implied
  40 // warranties of merchantability and fitness for a particular purpose are disclaimed.
  41 // In no event shall the Intel Corporation or contributors be liable for any direct,
  42 // indirect, incidental, special, exemplary, or consequential damages
  43 // (including, but not limited to, procurement of substitute goods or services;
  44 // loss of use, data, or profits; or business interruption) however caused
  45 // and on any theory of liability, whether in contract, strict liability,
  46 // or tort (including negligence or otherwise) arising in any way out of
  47 // the use of this software, even if advised of the possibility of such damage.
  48 //
  49 //M*/
  50
  51 #include "precomp.hpp"
  52 #include "opencl_kernels.hpp"
  53
  54 using namespace cv;
  55 using namespace cv::ocl;
  56
  57 /* these settings affect the quality of detection: change with care */
  58 #define CV_ADJUST_FEATURES 1
  59 #define CV_ADJUST_WEIGHTS  0
  60
  61 typedef int sumtype;
  62 typedef double sqsumtype;
  63
  64 typedef struct CvHidHaarFeature
  65 {
  66     struct
  67     {
  68         sumtype *p0, *p1, *p2, *p3;
  69         float weight;
  70     }
  71     rect[CV_HAAR_FEATURE_MAX];
  72 }
  73 CvHidHaarFeature;
  74
  75
  76 typedef struct CvHidHaarTreeNode
  77 {
  78     CvHidHaarFeature feature;
  79     float threshold;
  80     int left;
  81     int right;
  82 }
  83 CvHidHaarTreeNode;
  84
  85
  86 typedef struct CvHidHaarClassifier
  87 {
  88     int count;
  89     //CvHaarFeature* orig_feature;
  90     CvHidHaarTreeNode *node;
  91     float *alpha;
  92 }
  93 CvHidHaarClassifier;
  94
  95
  96 typedef struct CvHidHaarStageClassifier
  97 {
  98     int  count;
  99     float threshold;
 100     CvHidHaarClassifier *classifier;
 101     int two_rects;
 102
 103     struct CvHidHaarStageClassifier *next;
 104     struct CvHidHaarStageClassifier *child;
 105     struct CvHidHaarStageClassifier *parent;
 106 }
 107 CvHidHaarStageClassifier;
 108
 109
 110 struct CvHidHaarClassifierCascade
 111 {
 112     int  count;
 113     int  is_stump_based;
 114     int  has_tilted_features;
 115     int  is_tree;
 116     double inv_window_area;
 117     CvMat sum, sqsum, tilted;
 118     CvHidHaarStageClassifier *stage_classifier;
 119     sqsumtype *pq0, *pq1, *pq2, *pq3;
 120     sumtype *p0, *p1, *p2, *p3;
 121
 122     void **ipp_stages;
 123 };
 124 typedef struct
 125 {
 126     int width_height;
 127     int grpnumperline_totalgrp;
 128     int imgoff;
 129     float factor;
 130 } detect_piramid_info;
 131 #ifdef _MSC_VER
 132 #define _ALIGNED_ON(_ALIGNMENT) __declspec(align(_ALIGNMENT))
 133
 134 typedef _ALIGNED_ON(128) struct  GpuHidHaarTreeNode
 135 {
 136     _ALIGNED_ON(64) int p[CV_HAAR_FEATURE_MAX][4];
 137     float weight[CV_HAAR_FEATURE_MAX] ;
 138     float threshold ;
 139     _ALIGNED_ON(16) float alpha[3] ;
 140     _ALIGNED_ON(4) int left ;
 141     _ALIGNED_ON(4) int right ;
 142 }
 143 GpuHidHaarTreeNode;
 144
 145
 146 typedef  _ALIGNED_ON(32) struct  GpuHidHaarClassifier
 147 {
 148     _ALIGNED_ON(4) int count;
 149     _ALIGNED_ON(8) GpuHidHaarTreeNode *node ;
 150     _ALIGNED_ON(8) float *alpha ;
 151 }
 152 GpuHidHaarClassifier;
 153
 154
 155 typedef _ALIGNED_ON(64) struct   GpuHidHaarStageClassifier
 156 {
 157     _ALIGNED_ON(4) int  count ;
 158     _ALIGNED_ON(4) float threshold ;
 159     _ALIGNED_ON(4) int two_rects ;
 160     _ALIGNED_ON(8) GpuHidHaarClassifier *classifier ;
 161     _ALIGNED_ON(8) struct GpuHidHaarStageClassifier *next;
 162     _ALIGNED_ON(8) struct GpuHidHaarStageClassifier *child ;
 163     _ALIGNED_ON(8) struct GpuHidHaarStageClassifier *parent ;
 164 }
 165 GpuHidHaarStageClassifier;
 166
 167
 168 typedef _ALIGNED_ON(64) struct  GpuHidHaarClassifierCascade
 169 {
 170     _ALIGNED_ON(4) int  count ;
 171     _ALIGNED_ON(4) int  is_stump_based ;
 172     _ALIGNED_ON(4) int  has_tilted_features ;
 173     _ALIGNED_ON(4) int  is_tree ;
 174     _ALIGNED_ON(4) int pq0 ;
 175     _ALIGNED_ON(4) int pq1 ;
 176     _ALIGNED_ON(4) int pq2 ;
 177     _ALIGNED_ON(4) int pq3 ;
 178     _ALIGNED_ON(4) int p0 ;
 179     _ALIGNED_ON(4) int p1 ;
 180     _ALIGNED_ON(4) int p2 ;
 181     _ALIGNED_ON(4) int p3 ;
 182     _ALIGNED_ON(4) float inv_window_area ;
 183 } GpuHidHaarClassifierCascade;
 184 #else
 185 #define _ALIGNED_ON(_ALIGNMENT) __attribute__((aligned(_ALIGNMENT) ))
 186
 187 typedef struct _ALIGNED_ON(128) GpuHidHaarTreeNode
 188 {
 189     int p[CV_HAAR_FEATURE_MAX][4] _ALIGNED_ON(64);
 190     float weight[CV_HAAR_FEATURE_MAX];// _ALIGNED_ON(16);
 191     float threshold;// _ALIGNED_ON(4);
 192     float alpha[3] _ALIGNED_ON(16);
 193     int left _ALIGNED_ON(4);
 194     int right _ALIGNED_ON(4);
 195 }
 196 GpuHidHaarTreeNode;
 197
 198 typedef struct _ALIGNED_ON(32) GpuHidHaarClassifier
 199 {
 200     int count _ALIGNED_ON(4);
 201     GpuHidHaarTreeNode *node _ALIGNED_ON(8);
 202     float *alpha _ALIGNED_ON(8);
 203 }
 204 GpuHidHaarClassifier;
 205
 206
 207 typedef struct _ALIGNED_ON(64) GpuHidHaarStageClassifier
 208 {
 209     int  count _ALIGNED_ON(4);
 210     float threshold _ALIGNED_ON(4);
 211     int two_rects _ALIGNED_ON(4);
 212     GpuHidHaarClassifier *classifier _ALIGNED_ON(8);
 213     struct GpuHidHaarStageClassifier *next _ALIGNED_ON(8);
 214     struct GpuHidHaarStageClassifier *child _ALIGNED_ON(8);
 215     struct GpuHidHaarStageClassifier *parent _ALIGNED_ON(8);
 216 }
 217 GpuHidHaarStageClassifier;
 218
 219
 220 typedef struct _ALIGNED_ON(64) GpuHidHaarClassifierCascade
 221 {
 222     int  count _ALIGNED_ON(4);
 223     int  is_stump_based _ALIGNED_ON(4);
 224     int  has_tilted_features _ALIGNED_ON(4);
 225     int  is_tree _ALIGNED_ON(4);
 226     int pq0 _ALIGNED_ON(4);
 227     int pq1 _ALIGNED_ON(4);
 228     int pq2 _ALIGNED_ON(4);
 229     int pq3 _ALIGNED_ON(4);
 230     int p0 _ALIGNED_ON(4);
 231     int p1 _ALIGNED_ON(4);
 232     int p2 _ALIGNED_ON(4);
 233     int p3 _ALIGNED_ON(4);
 234     float inv_window_area _ALIGNED_ON(4);
 235 } GpuHidHaarClassifierCascade;
 236 #endif
 237
 238 const int icv_object_win_border = 1;
 239 const float icv_stage_threshold_bias = 0.0001f;
 240 double globaltime = 0;
 241
 242 /* create more efficient internal representation of haar classifier cascade */
 243 static GpuHidHaarClassifierCascade * gpuCreateHidHaarClassifierCascade( CvHaarClassifierCascade *cascade, int *size, int *totalclassifier)
 244 {
 245     GpuHidHaarClassifierCascade *out = 0;
 246
 247     int i, j, k, l;
 248     int datasize;
 249     int total_classifiers = 0;
 250     int total_nodes = 0;
 251     char errorstr[256];
 252
 253     GpuHidHaarStageClassifier *stage_classifier_ptr;
 254     GpuHidHaarClassifier *haar_classifier_ptr;
 255     GpuHidHaarTreeNode *haar_node_ptr;
 256
 257     CvSize orig_window_size;
 258     int has_tilted_features = 0;
 259
 260     if( !CV_IS_HAAR_CLASSIFIER(cascade) )
 261         CV_Error( !cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier pointer" );
 262
 263     if( cascade->hid_cascade )
 264         CV_Error( CV_StsError, "hid_cascade has been already created" );
 265
 266     if( !cascade->stage_classifier )
 267         CV_Error( CV_StsNullPtr, "" );
 268
 269     if( cascade->count <= 0 )
 270         CV_Error( CV_StsOutOfRange, "Negative number of cascade stages" );
 271
 272     orig_window_size = cascade->orig_window_size;
 273
 274     /* check input structure correctness and calculate total memory size needed for
 275     internal representation of the classifier cascade */
 276     for( i = 0; i < cascade->count; i++ )
 277     {
 278         CvHaarStageClassifier *stage_classifier = cascade->stage_classifier + i;
 279
 280         if( !stage_classifier->classifier ||
 281                 stage_classifier->count <= 0 )
 282         {
 283             sprintf( errorstr, "header of the stage classifier #%d is invalid "
 284                      "(has null pointers or non-positive classfier count)", i );
 285             CV_Error( CV_StsError, errorstr );
 286         }
 287
 288         total_classifiers += stage_classifier->count;
 289
 290         for( j = 0; j < stage_classifier->count; j++ )
 291         {
 292             CvHaarClassifier *classifier = stage_classifier->classifier + j;
 293
 294             total_nodes += classifier->count;
 295             for( l = 0; l < classifier->count; l++ )
 296             {
 297                 for( k = 0; k < CV_HAAR_FEATURE_MAX; k++ )
 298                 {
 299                     if( classifier->haar_feature[l].rect[k].r.width )
 300                     {
 301                         CvRect r = classifier->haar_feature[l].rect[k].r;
 302                         int tilted = classifier->haar_feature[l].tilted;
 303                         has_tilted_features |= tilted != 0;
 304                         if( r.width < 0 || r.height < 0 || r.y < 0 ||
 305                                 r.x + r.width > orig_window_size.width
 306                                 ||
 307                                 (!tilted &&
 308                                  (r.x < 0 || r.y + r.height > orig_window_size.height))
 309                                 ||
 310                                 (tilted && (r.x - r.height < 0 ||
 311                                             r.y + r.width + r.height > orig_window_size.height)))
 312                         {
 313                             sprintf( errorstr, "rectangle #%d of the classifier #%d of "
 314                                      "the stage classifier #%d is not inside "
 315                                      "the reference (original) cascade window", k, j, i );
 316                             CV_Error( CV_StsNullPtr, errorstr );
 317                         }
 318                     }
 319                 }
 320             }
 321         }
 322     }
 323
 324     // this is an upper boundary for the whole hidden cascade size
 325     datasize = sizeof(GpuHidHaarClassifierCascade)                   +
 326                sizeof(GpuHidHaarStageClassifier) * cascade->count    +
 327                sizeof(GpuHidHaarClassifier)      * total_classifiers +
 328                sizeof(GpuHidHaarTreeNode)        * total_nodes;
 329
 330     *totalclassifier = total_classifiers;
 331     *size = datasize;
 332     out = (GpuHidHaarClassifierCascade *)cvAlloc( datasize );
 333     memset( out, 0, sizeof(*out) );
 334
 335     /* init header */
 336     out->count = cascade->count;
 337     stage_classifier_ptr = (GpuHidHaarStageClassifier *)(out + 1);
 338     haar_classifier_ptr = (GpuHidHaarClassifier *)(stage_classifier_ptr + cascade->count);
 339     haar_node_ptr = (GpuHidHaarTreeNode *)(haar_classifier_ptr + total_classifiers);
 340
 341     out->is_stump_based = 1;
 342     out->has_tilted_features = has_tilted_features;
 343     out->is_tree = 0;
 344
 345     /* initialize internal representation */
 346     for( i = 0; i < cascade->count; i++ )
 347     {
 348         CvHaarStageClassifier *stage_classifier = cascade->stage_classifier + i;
 349         GpuHidHaarStageClassifier *hid_stage_classifier = stage_classifier_ptr + i;
 350
 351         hid_stage_classifier->count = stage_classifier->count;
 352         hid_stage_classifier->threshold = stage_classifier->threshold - icv_stage_threshold_bias;
 353         hid_stage_classifier->classifier = haar_classifier_ptr;
 354         hid_stage_classifier->two_rects = 1;
 355         haar_classifier_ptr += stage_classifier->count;
 356
 357         for( j = 0; j < stage_classifier->count; j++ )
 358         {
 359             CvHaarClassifier *classifier         = stage_classifier->classifier + j;
 360             GpuHidHaarClassifier *hid_classifier = hid_stage_classifier->classifier + j;
 361             int node_count = classifier->count;
 362
 363             float *alpha_ptr = &haar_node_ptr->alpha[0];
 364
 365             hid_classifier->count = node_count;
 366             hid_classifier->node = haar_node_ptr;
 367             hid_classifier->alpha = alpha_ptr;
 368
 369             for( l = 0; l < node_count; l++ )
 370             {
 371                 GpuHidHaarTreeNode *node     = hid_classifier->node + l;
 372                 CvHaarFeature      *feature = classifier->haar_feature + l;
 373
 374                 memset( node, -1, sizeof(*node) );
 375                 node->threshold = classifier->threshold[l];
 376                 node->left      = classifier->left[l];
 377                 node->right     = classifier->right[l];
 378
 379                 if( fabs(feature->rect[2].weight) < DBL_EPSILON ||
 380                         feature->rect[2].r.width == 0 ||
 381                         feature->rect[2].r.height == 0 )
 382                 {
 383                     node->p[2][0] = 0;
 384                     node->p[2][1] = 0;
 385                     node->p[2][2] = 0;
 386                     node->p[2][3] = 0;
 387                     node->weight[2] = 0;
 388                 }
 389                 else
 390                     hid_stage_classifier->two_rects = 0;
 391
 392                 memcpy( node->alpha, classifier->alpha, (node_count + 1)*sizeof(alpha_ptr[0]));
 393                 haar_node_ptr = haar_node_ptr + 1;
 394             }
 395             out->is_stump_based &= node_count == 1;
 396         }
 397     }
 398
 399     cascade->hid_cascade = (CvHidHaarClassifierCascade *)out;
 400     assert( (char *)haar_node_ptr - (char *)out <= datasize );
 401
 402     return out;
 403 }
 404
 405
 406 #define sum_elem_ptr(sum,row,col)  \
 407     ((sumtype*)CV_MAT_ELEM_PTR_FAST((sum),(row),(col),sizeof(sumtype)))
 408
 409 #define sqsum_elem_ptr(sqsum,row,col)  \
 410     ((sqsumtype*)CV_MAT_ELEM_PTR_FAST((sqsum),(row),(col),sizeof(sqsumtype)))
 411
 412 #define calc_sum(rect,offset) \
 413     ((rect).p0[offset] - (rect).p1[offset] - (rect).p2[offset] + (rect).p3[offset])
 414
 415
 416 static void gpuSetImagesForHaarClassifierCascade( CvHaarClassifierCascade *_cascade,
 417                                       double scale,
 418                                       int step)
 419 {
 420     GpuHidHaarClassifierCascade *cascade;
 421     int coi0 = 0, coi1 = 0;
 422     int i;
 423     int datasize;
 424     int total;
 425     CvRect equRect;
 426     double weight_scale;
 427     GpuHidHaarStageClassifier *stage_classifier;
 428
 429     if( !CV_IS_HAAR_CLASSIFIER(_cascade) )
 430         CV_Error( !_cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier pointer" );
 431
 432     if( scale <= 0 )
 433         CV_Error( CV_StsOutOfRange, "Scale must be positive" );
 434
 435     if( coi0 || coi1 )
 436         CV_Error( CV_BadCOI, "COI is not supported" );
 437
 438     if( !_cascade->hid_cascade )
 439         gpuCreateHidHaarClassifierCascade(_cascade, &datasize, &total);
 440
 441     cascade = (GpuHidHaarClassifierCascade *) _cascade->hid_cascade;
 442     stage_classifier = (GpuHidHaarStageClassifier *) (cascade + 1);
 443
 444     _cascade->scale = scale;
 445     _cascade->real_window_size.width = cvRound( _cascade->orig_window_size.width * scale );
 446     _cascade->real_window_size.height = cvRound( _cascade->orig_window_size.height * scale );
 447
 448     equRect.x = equRect.y = cvRound(scale);
 449     equRect.width = cvRound((_cascade->orig_window_size.width - 2) * scale);
 450     equRect.height = cvRound((_cascade->orig_window_size.height - 2) * scale);
 451     weight_scale = 1. / (equRect.width * equRect.height);
 452     cascade->inv_window_area = weight_scale;
 453
 454     cascade->pq0 = equRect.x;
 455     cascade->pq1 = equRect.y;
 456     cascade->pq2 = equRect.x + equRect.width;
 457     cascade->pq3 = equRect.y + equRect.height;
 458
 459     cascade->p0 = equRect.x;
 460     cascade->p1 = equRect.y;
 461     cascade->p2 = equRect.x + equRect.width;
 462     cascade->p3 = equRect.y + equRect.height;
 463
 464
 465     /* init pointers in haar features according to real window size and
 466     given image pointers */
 467     for( i = 0; i < _cascade->count; i++ )
 468     {
 469         int j, k, l;
 470         for( j = 0; j < stage_classifier[i].count; j++ )
 471         {
 472             for( l = 0; l < stage_classifier[i].classifier[j].count; l++ )
 473             {
 474                 CvHaarFeature *feature =
 475                     &_cascade->stage_classifier[i].classifier[j].haar_feature[l];
 476                 GpuHidHaarTreeNode *hidnode = &stage_classifier[i].classifier[j].node[l];
 477                 double sum0 = 0, area0 = 0;
 478                 CvRect r[3];
 479
 480                 int base_w = -1, base_h = -1;
 481                 int new_base_w = 0, new_base_h = 0;
 482                 int kx, ky;
 483                 int flagx = 0, flagy = 0;
 484                 int x0 = 0, y0 = 0;
 485                 int nr;
 486
 487                 /* align blocks */
 488                 for( k = 0; k < CV_HAAR_FEATURE_MAX; k++ )
 489                 {
 490                     if(!hidnode->p[k][0])
 491                         break;
 492                     r[k] = feature->rect[k].r;
 493                     base_w = (int)CV_IMIN( (unsigned)base_w, (unsigned)(r[k].width - 1) );
 494                     base_w = (int)CV_IMIN( (unsigned)base_w, (unsigned)(r[k].x - r[0].x - 1) );
 495                     base_h = (int)CV_IMIN( (unsigned)base_h, (unsigned)(r[k].height - 1) );
 496                     base_h = (int)CV_IMIN( (unsigned)base_h, (unsigned)(r[k].y - r[0].y - 1) );
 497                 }
 498
 499                 nr = k;
 500                 base_w += 1;
 501                 base_h += 1;
 502                 if(base_w == 0)
 503                     base_w = 1;
 504                 kx = r[0].width / base_w;
 505                 if(base_h == 0)
 506                     base_h = 1;
 507                 ky = r[0].height / base_h;
 508
 509                 if( kx <= 0 )
 510                 {
 511                     flagx = 1;
 512                     new_base_w = cvRound( r[0].width * scale ) / kx;
 513                     x0 = cvRound( r[0].x * scale );
 514                 }
 515
 516                 if( ky <= 0 )
 517                 {
 518                     flagy = 1;
 519                     new_base_h = cvRound( r[0].height * scale ) / ky;
 520                     y0 = cvRound( r[0].y * scale );
 521                 }
 522
 523                 for( k = 0; k < nr; k++ )
 524                 {
 525                     CvRect tr;
 526                     double correction_ratio;
 527
 528                     if( flagx )
 529                     {
 530                         tr.x = (r[k].x - r[0].x) * new_base_w / base_w + x0;
 531                         tr.width = r[k].width * new_base_w / base_w;
 532                     }
 533                     else
 534                     {
 535                         tr.x = cvRound( r[k].x * scale );
 536                         tr.width = cvRound( r[k].width * scale );
 537                     }
 538
 539                     if( flagy )
 540                     {
 541                         tr.y = (r[k].y - r[0].y) * new_base_h / base_h + y0;
 542                         tr.height = r[k].height * new_base_h / base_h;
 543                     }
 544                     else
 545                     {
 546                         tr.y = cvRound( r[k].y * scale );
 547                         tr.height = cvRound( r[k].height * scale );
 548                     }
 549
 550 #if CV_ADJUST_WEIGHTS
 551                     {
 552                         // RAINER START
 553                         const float orig_feature_size =  (float)(feature->rect[k].r.width) * feature->rect[k].r.height;
 554                         const float orig_norm_size = (float)(_cascade->orig_window_size.width) * (_cascade->orig_window_size.height);
 555                         const float feature_size = float(tr.width * tr.height);
 556                         //const float normSize    = float(equRect.width*equRect.height);
 557                         float target_ratio = orig_feature_size / orig_norm_size;
 558                         //float isRatio = featureSize / normSize;
 559                         //correctionRatio = targetRatio / isRatio / normSize;
 560                         correction_ratio = target_ratio / feature_size;
 561                         // RAINER END
 562                     }
 563 #else
 564                     correction_ratio = weight_scale * (!feature->tilted ? 1 : 0.5);
 565 #endif
 566
 567                     if( !feature->tilted )
 568                     {
 569                         hidnode->p[k][0] = tr.x;
 570                         hidnode->p[k][1] = tr.y;
 571                         hidnode->p[k][2] = tr.x + tr.width;
 572                         hidnode->p[k][3] = tr.y + tr.height;
 573                     }
 574                     else
 575                     {
 576                         hidnode->p[k][2] = (tr.y + tr.width) * step + tr.x + tr.width;
 577                         hidnode->p[k][3] = (tr.y + tr.width + tr.height) * step + tr.x + tr.width - tr.height;
 578                         hidnode->p[k][0] = tr.y * step + tr.x;
 579                         hidnode->p[k][1] = (tr.y + tr.height) * step + tr.x - tr.height;
 580                     }
 581                     hidnode->weight[k] = (float)(feature->rect[k].weight * correction_ratio);
 582                     if( k == 0 )
 583                         area0 = tr.width * tr.height;
 584                     else
 585                         sum0 += hidnode->weight[k] * tr.width * tr.height;
 586                 }
 587                 hidnode->weight[0] = (float)(-sum0 / area0);
 588             } /* l */
 589         } /* j */
 590     }
 591 }
 592
 593 static void gpuSetHaarClassifierCascade( CvHaarClassifierCascade *_cascade)
 594 {
 595     GpuHidHaarClassifierCascade *cascade;
 596     int i;
 597     int datasize;
 598     int total;
 599     CvRect equRect;
 600     double weight_scale;
 601     GpuHidHaarStageClassifier *stage_classifier;
 602
 603     if( !CV_IS_HAAR_CLASSIFIER(_cascade) )
 604         CV_Error( !_cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier pointer" );
 605
 606     if( !_cascade->hid_cascade )
 607         gpuCreateHidHaarClassifierCascade(_cascade, &datasize, &total);
 608
 609     cascade = (GpuHidHaarClassifierCascade *) _cascade->hid_cascade;
 610     stage_classifier = (GpuHidHaarStageClassifier *) cascade + 1;
 611
 612     _cascade->scale = 1.0;
 613     _cascade->real_window_size.width =  _cascade->orig_window_size.width ;
 614     _cascade->real_window_size.height = _cascade->orig_window_size.height;
 615
 616     equRect.x = equRect.y = 1;
 617     equRect.width = _cascade->orig_window_size.width - 2;
 618     equRect.height = _cascade->orig_window_size.height - 2;
 619     weight_scale = 1;
 620     cascade->inv_window_area = weight_scale;
 621
 622     cascade->p0 = equRect.x;
 623     cascade->p1 = equRect.y;
 624     cascade->p2 = equRect.height;
 625     cascade->p3 = equRect.width ;
 626     for( i = 0; i < _cascade->count; i++ )
 627     {
 628         int j, k, l;
 629         for( j = 0; j < stage_classifier[i].count; j++ )
 630         {
 631             for( l = 0; l < stage_classifier[i].classifier[j].count; l++ )
 632             {
 633                 CvHaarFeature *feature =
 634                     &_cascade->stage_classifier[i].classifier[j].haar_feature[l];
 635                 GpuHidHaarTreeNode *hidnode = &stage_classifier[i].classifier[j].node[l];
 636                 CvRect r[3];
 637
 638
 639                 int nr;
 640
 641                 /* align blocks */
 642                 for( k = 0; k < CV_HAAR_FEATURE_MAX; k++ )
 643                 {
 644                     if(!hidnode->p[k][0])
 645                         break;
 646                     r[k] = feature->rect[k].r;
 647                }
 648
 649                 nr = k;
 650                 for( k = 0; k < nr; k++ )
 651                 {
 652                     CvRect tr;
 653                     double correction_ratio;
 654                     tr.x = r[k].x;
 655                     tr.width = r[k].width;
 656                     tr.y = r[k].y ;
 657                     tr.height = r[k].height;
 658                     correction_ratio = weight_scale * (!feature->tilted ? 1 : 0.5);
 659                     hidnode->p[k][0] = tr.x;
 660                     hidnode->p[k][1] = tr.y;
 661                     hidnode->p[k][2] = tr.width;
 662                     hidnode->p[k][3] = tr.height;
 663                     hidnode->weight[k] = (float)(feature->rect[k].weight * correction_ratio);
 664                 }
 665             } /* l */
 666         } /* j */
 667     }
 668 }
 669
 670 CvSeq *cv::ocl::OclCascadeClassifier::oclHaarDetectObjects( oclMat &gimg, CvMemStorage *storage, double scaleFactor,
 671         int minNeighbors, int flags, CvSize minSize, CvSize maxSize)
 672 {
 673     CvHaarClassifierCascade *cascade = oldCascade;
 674
 675     const double GROUP_EPS = 0.2;
 676     CvSeq *result_seq = 0;
 677
 678     cv::ConcurrentRectVector allCandidates;
 679     std::vector<cv::Rect> rectList;
 680     std::vector<int> rweights;
 681     double factor;
 682     int datasize=0;
 683     int totalclassifier=0;
 684
 685     GpuHidHaarClassifierCascade *gcascade;
 686     GpuHidHaarStageClassifier    *stage;
 687     GpuHidHaarClassifier         *classifier;
 688     GpuHidHaarTreeNode           *node;
 689
 690     int *candidate;
 691     cl_int status;
 692
 693     bool findBiggestObject = (flags & CV_HAAR_FIND_BIGGEST_OBJECT) != 0;
 694
 695     if( maxSize.height == 0 || maxSize.width == 0 )
 696     {
 697         maxSize.height = gimg.rows;
 698         maxSize.width = gimg.cols;
 699     }
 700
 701     if( !CV_IS_HAAR_CLASSIFIER(cascade) )
 702         CV_Error( !cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier cascade" );
 703
 704     if( !storage )
 705         CV_Error( CV_StsNullPtr, "Null storage pointer" );
 706
 707     if( CV_MAT_DEPTH(gimg.type()) != CV_8U )
 708         CV_Error( CV_StsUnsupportedFormat, "Only 8-bit images are supported" );
 709
 710     if( scaleFactor <= 1 )
 711         CV_Error( CV_StsOutOfRange, "scale factor must be > 1" );
 712
 713     if( findBiggestObject )
 714         flags &= ~CV_HAAR_SCALE_IMAGE;
 715
 716     if( !cascade->hid_cascade )
 717         gpuCreateHidHaarClassifierCascade(cascade, &datasize, &totalclassifier);
 718
 719     result_seq = cvCreateSeq( 0, sizeof(CvSeq), sizeof(CvAvgComp), storage );
 720
 721     if( CV_MAT_CN(gimg.type()) > 1 )
 722     {
 723         oclMat gtemp;
 724         cvtColor( gimg, gtemp, CV_BGR2GRAY );
 725         gimg = gtemp;
 726     }
 727
 728     if( findBiggestObject )
 729         flags &= ~(CV_HAAR_SCALE_IMAGE | CV_HAAR_DO_CANNY_PRUNING);
 730
 731     if( gimg.cols < minSize.width || gimg.rows < minSize.height )
 732         CV_Error(CV_StsError, "Image too small");
 733
 734     cl_command_queue qu = getClCommandQueue(Context::getContext());
 735     if( (flags & CV_HAAR_SCALE_IMAGE) )
 736     {
 737         CvSize winSize0 = cascade->orig_window_size;
 738         int totalheight = 0;
 739         int indexy = 0;
 740         CvSize sz;
 741         vector<CvSize> sizev;
 742         vector<float> scalev;
 743         for(factor = 1.f;; factor *= scaleFactor)
 744         {
 745             CvSize winSize = { cvRound(winSize0.width * factor), cvRound(winSize0.height * factor) };
 746             sz.width     = cvRound( gimg.cols / factor ) + 1;
 747             sz.height    = cvRound( gimg.rows / factor ) + 1;
 748             CvSize sz1     = { sz.width - winSize0.width - 1,      sz.height - winSize0.height - 1 };
 749
 750             if( sz1.width <= 0 || sz1.height <= 0 )
 751                 break;
 752             if( winSize.width > maxSize.width || winSize.height > maxSize.height )
 753                 break;
 754             if( winSize.width < minSize.width || winSize.height < minSize.height )
 755                 continue;
 756
 757             totalheight += sz.height;
 758             sizev.push_back(sz);
 759             scalev.push_back(factor);
 760         }
 761
 762         oclMat gimg1(gimg.rows, gimg.cols, CV_8UC1);
 763         oclMat gsum(totalheight + 4, gimg.cols + 1, CV_32SC1);
 764         oclMat gsqsum(totalheight + 4, gimg.cols + 1, CV_32FC1);
 765
 766         cl_mem stagebuffer;
 767         cl_mem nodebuffer;
 768         cl_mem candidatebuffer;
 769         cl_mem scaleinfobuffer;
 770         cv::Rect roi, roi2;
 771         cv::Mat imgroi, imgroisq;
 772         cv::ocl::oclMat resizeroi, gimgroi, gimgroisq;
 773         int grp_per_CU = 12;
 774
 775         size_t blocksize = 8;
 776         size_t localThreads[3] = { blocksize, blocksize , 1 };
 777         size_t globalThreads[3] = { grp_per_CU *(gsum.clCxt->getDeviceInfo().maxComputeUnits) *localThreads[0],
 778                                     localThreads[1], 1
 779                                   };
 780         int outputsz = 256 * globalThreads[0] / localThreads[0];
 781         int loopcount = sizev.size();
 782         detect_piramid_info *scaleinfo = (detect_piramid_info *)malloc(sizeof(detect_piramid_info) * loopcount);
 783
 784         for( int i = 0; i < loopcount; i++ )
 785         {
 786             sz = sizev[i];
 787             factor = scalev[i];
 788             roi = Rect(0, indexy, sz.width, sz.height);
 789             roi2 = Rect(0, 0, sz.width - 1, sz.height - 1);
 790             resizeroi = gimg1(roi2);
 791             gimgroi = gsum(roi);
 792             gimgroisq = gsqsum(roi);
 793             int width = gimgroi.cols - 1 - cascade->orig_window_size.width;
 794             int height = gimgroi.rows - 1 - cascade->orig_window_size.height;
 795             scaleinfo[i].width_height = (width << 16) | height;
 796
 797
 798             int grpnumperline = (width + localThreads[0] - 1) / localThreads[0];
 799             int totalgrp = ((height + localThreads[1] - 1) / localThreads[1]) * grpnumperline;
 800
 801             scaleinfo[i].grpnumperline_totalgrp = (grpnumperline << 16) | totalgrp;
 802             scaleinfo[i].imgoff = gimgroi.offset >> 2;
 803             scaleinfo[i].factor = factor;
 804             cv::ocl::resize(gimg, resizeroi, Size(sz.width - 1, sz.height - 1), 0, 0, INTER_LINEAR);
 805             cv::ocl::integral(resizeroi, gimgroi, gimgroisq);
 806             indexy += sz.height;
 807         }
 808
 809         gcascade   = (GpuHidHaarClassifierCascade *)cascade->hid_cascade;
 810         stage      = (GpuHidHaarStageClassifier *)(gcascade + 1);
 811         classifier = (GpuHidHaarClassifier *)(stage + gcascade->count);
 812         node       = (GpuHidHaarTreeNode *)(classifier->node);
 813
 814         int nodenum = (datasize - sizeof(GpuHidHaarClassifierCascade) -
 815                        sizeof(GpuHidHaarStageClassifier) * gcascade->count - sizeof(GpuHidHaarClassifier) * totalclassifier) / sizeof(GpuHidHaarTreeNode);
 816
 817         candidate = (int *)malloc(4 * sizeof(int) * outputsz);
 818
 819         gpuSetImagesForHaarClassifierCascade( cascade, 1., gsum.step / 4 );
 820
 821         stagebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(GpuHidHaarStageClassifier) * gcascade->count);
 822         openCLSafeCall(clEnqueueWriteBuffer(qu, stagebuffer, 1, 0, sizeof(GpuHidHaarStageClassifier)*gcascade->count, stage, 0, NULL, NULL));
 823
 824         nodebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, nodenum * sizeof(GpuHidHaarTreeNode));
 825
 826         openCLSafeCall(clEnqueueWriteBuffer(qu, nodebuffer, 1, 0, nodenum * sizeof(GpuHidHaarTreeNode),
 827                                             node, 0, NULL, NULL));
 828         candidatebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_WRITE_ONLY, 4 * sizeof(int) * outputsz);
 829
 830         scaleinfobuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(detect_piramid_info) * loopcount);
 831         openCLSafeCall(clEnqueueWriteBuffer(qu, scaleinfobuffer, 1, 0, sizeof(detect_piramid_info)*loopcount, scaleinfo, 0, NULL, NULL));
 832
 833         int startstage = 0;
 834         int endstage = gcascade->count;
 835         int startnode = 0;
 836         int pixelstep = gsum.step / 4;
 837         int splitstage = 3;
 838         int splitnode = stage[0].count + stage[1].count + stage[2].count;
 839         cl_int4 p, pq;
 840         p.s[0] = gcascade->p0;
 841         p.s[1] = gcascade->p1;
 842         p.s[2] = gcascade->p2;
 843         p.s[3] = gcascade->p3;
 844         pq.s[0] = gcascade->pq0;
 845         pq.s[1] = gcascade->pq1;
 846         pq.s[2] = gcascade->pq2;
 847         pq.s[3] = gcascade->pq3;
 848         float correction = gcascade->inv_window_area;
 849
 850         vector<pair<size_t, const void *> > args;
 851         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&stagebuffer ));
 852         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&scaleinfobuffer ));
 853         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&nodebuffer ));
 854         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsum.data ));
 855         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsqsum.data ));
 856         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&candidatebuffer ));
 857         args.push_back ( make_pair(sizeof(cl_int) , (void *)&pixelstep ));
 858         args.push_back ( make_pair(sizeof(cl_int) , (void *)&loopcount ));
 859         args.push_back ( make_pair(sizeof(cl_int) , (void *)&startstage ));
 860         args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitstage ));
 861         args.push_back ( make_pair(sizeof(cl_int) , (void *)&endstage ));
 862         args.push_back ( make_pair(sizeof(cl_int) , (void *)&startnode ));
 863         args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitnode ));
 864         args.push_back ( make_pair(sizeof(cl_int4) , (void *)&p ));
 865         args.push_back ( make_pair(sizeof(cl_int4) , (void *)&pq ));
 866         args.push_back ( make_pair(sizeof(cl_float) , (void *)&correction ));
 867
 868         const char * build_options = gcascade->is_stump_based ? "-D STUMP_BASED=1" : "-D STUMP_BASED=0";
 869
 870         openCLExecuteKernel(gsum.clCxt, &haarobjectdetect, "gpuRunHaarClassifierCascade", globalThreads, localThreads, args, -1, -1, build_options);
 871
 872         openCLReadBuffer( gsum.clCxt, candidatebuffer, candidate, 4 * sizeof(int)*outputsz );
 873
 874         for(int i = 0; i < outputsz; i++)
 875             if(candidate[4 * i + 2] != 0)
 876                 allCandidates.push_back(Rect(candidate[4 * i], candidate[4 * i + 1],
 877                 candidate[4 * i + 2], candidate[4 * i + 3]));
 878
 879         free(scaleinfo);
 880         free(candidate);
 881         openCLSafeCall(clReleaseMemObject(stagebuffer));
 882         openCLSafeCall(clReleaseMemObject(scaleinfobuffer));
 883         openCLSafeCall(clReleaseMemObject(nodebuffer));
 884         openCLSafeCall(clReleaseMemObject(candidatebuffer));
 885
 886     }
 887     else
 888     {
 889         CvSize winsize0 = cascade->orig_window_size;
 890         int n_factors = 0;
 891         oclMat gsum;
 892         oclMat gsqsum;
 893         cv::ocl::integral(gimg, gsum, gsqsum);
 894         CvSize sz;
 895         vector<CvSize> sizev;
 896         vector<float> scalev;
 897         gpuSetHaarClassifierCascade(cascade);
 898         gcascade   = (GpuHidHaarClassifierCascade *)cascade->hid_cascade;
 899         stage      = (GpuHidHaarStageClassifier *)(gcascade + 1);
 900         classifier = (GpuHidHaarClassifier *)(stage + gcascade->count);
 901         node       = (GpuHidHaarTreeNode *)(classifier->node);
 902         cl_mem stagebuffer;
 903         cl_mem nodebuffer;
 904         cl_mem candidatebuffer;
 905         cl_mem scaleinfobuffer;
 906         cl_mem pbuffer;
 907         cl_mem correctionbuffer;
 908         for( n_factors = 0, factor = 1;
 909                 cvRound(factor * winsize0.width) < gimg.cols - 10 &&
 910                 cvRound(factor * winsize0.height) < gimg.rows - 10;
 911                 n_factors++, factor *= scaleFactor )
 912         {
 913             CvSize winSize = { cvRound( winsize0.width * factor ),
 914                                cvRound( winsize0.height * factor )
 915                              };
 916             if( winSize.width < minSize.width || winSize.height < minSize.height )
 917             {
 918                 continue;
 919             }
 920             sizev.push_back(winSize);
 921             scalev.push_back(factor);
 922         }
 923         int loopcount = scalev.size();
 924         if(loopcount == 0)
 925         {
 926             loopcount = 1;
 927             n_factors = 1;
 928             sizev.push_back(minSize);
 929             scalev.push_back( std::min(cvRound(minSize.width / winsize0.width), cvRound(minSize.height / winsize0.height)) );
 930
 931         }
 932         detect_piramid_info *scaleinfo = (detect_piramid_info *)malloc(sizeof(detect_piramid_info) * loopcount);
 933         cl_int4 *p = (cl_int4 *)malloc(sizeof(cl_int4) * loopcount);
 934         float *correction = (float *)malloc(sizeof(float) * loopcount);
 935         int grp_per_CU = 12;
 936         size_t blocksize = 8;
 937         size_t localThreads[3] = { blocksize, blocksize , 1 };
 938         size_t globalThreads[3] = { grp_per_CU *gsum.clCxt->getDeviceInfo().maxComputeUnits *localThreads[0],
 939                                     localThreads[1], 1 };
 940         int outputsz = 256 * globalThreads[0] / localThreads[0];
 941         int nodenum = (datasize - sizeof(GpuHidHaarClassifierCascade) -
 942                        sizeof(GpuHidHaarStageClassifier) * gcascade->count - sizeof(GpuHidHaarClassifier) * totalclassifier) / sizeof(GpuHidHaarTreeNode);
 943         nodebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY,
 944                                         nodenum * sizeof(GpuHidHaarTreeNode));
 945         openCLSafeCall(clEnqueueWriteBuffer(qu, nodebuffer, 1, 0,
 946                                             nodenum * sizeof(GpuHidHaarTreeNode),
 947                                             node, 0, NULL, NULL));
 948         cl_mem newnodebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_WRITE,
 949                                loopcount * nodenum * sizeof(GpuHidHaarTreeNode));
 950         int startstage = 0;
 951         int endstage = gcascade->count;
 952         for(int i = 0; i < loopcount; i++)
 953         {
 954             sz = sizev[i];
 955             factor = scalev[i];
 956             int ystep = cvRound(std::max(2., factor));
 957             int equRect_x = (int)(factor * gcascade->p0 + 0.5);
 958             int equRect_y = (int)(factor * gcascade->p1 + 0.5);
 959             int equRect_w = (int)(factor * gcascade->p3 + 0.5);
 960             int equRect_h = (int)(factor * gcascade->p2 + 0.5);
 961             p[i].s[0] = equRect_x;
 962             p[i].s[1] = equRect_y;
 963             p[i].s[2] = equRect_x + equRect_w;
 964             p[i].s[3] = equRect_y + equRect_h;
 965             correction[i] = 1. / (equRect_w * equRect_h);
 966             int width = (gsum.cols - 1 - sz.width  + ystep - 1) / ystep;
 967             int height = (gsum.rows - 1 - sz.height + ystep - 1) / ystep;
 968             int grpnumperline = (width + localThreads[0] - 1) / localThreads[0];
 969             int totalgrp = ((height + localThreads[1] - 1) / localThreads[1]) * grpnumperline;
 970
 971             scaleinfo[i].width_height = (width << 16) | height;
 972             scaleinfo[i].grpnumperline_totalgrp = (grpnumperline << 16) | totalgrp;
 973             scaleinfo[i].imgoff = 0;
 974             scaleinfo[i].factor = factor;
 975             int startnodenum = nodenum * i;
 976             float factor2 = (float)factor;
 977
 978             vector<pair<size_t, const void *> > args1;
 979             args1.push_back ( make_pair(sizeof(cl_mem) , (void *)&nodebuffer ));
 980             args1.push_back ( make_pair(sizeof(cl_mem) , (void *)&newnodebuffer ));
 981             args1.push_back ( make_pair(sizeof(cl_float) , (void *)&factor2 ));
 982             args1.push_back ( make_pair(sizeof(cl_float) , (void *)&correction[i] ));
 983             args1.push_back ( make_pair(sizeof(cl_int) , (void *)&startnodenum ));
 984
 985             size_t globalThreads2[3] = {nodenum, 1, 1};
 986             openCLExecuteKernel(gsum.clCxt, &haarobjectdetect_scaled2, "gpuscaleclassifier", globalThreads2, NULL/*localThreads2*/, args1, -1, -1);
 987         }
 988
 989         int step = gsum.step / 4;
 990         int startnode = 0;
 991         int splitstage = 3;
 992         stagebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(GpuHidHaarStageClassifier) * gcascade->count);
 993         openCLSafeCall(clEnqueueWriteBuffer(qu, stagebuffer, 1, 0, sizeof(GpuHidHaarStageClassifier)*gcascade->count, stage, 0, NULL, NULL));
 994         candidatebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_WRITE_ONLY | CL_MEM_ALLOC_HOST_PTR, 4 * sizeof(int) * outputsz);
 995         scaleinfobuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(detect_piramid_info) * loopcount);
 996         openCLSafeCall(clEnqueueWriteBuffer(qu, scaleinfobuffer, 1, 0, sizeof(detect_piramid_info)*loopcount, scaleinfo, 0, NULL, NULL));
 997         pbuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(cl_int4) * loopcount);
 998         openCLSafeCall(clEnqueueWriteBuffer(qu, pbuffer, 1, 0, sizeof(cl_int4)*loopcount, p, 0, NULL, NULL));
 999         correctionbuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(cl_float) * loopcount);
1000         openCLSafeCall(clEnqueueWriteBuffer(qu, correctionbuffer, 1, 0, sizeof(cl_float)*loopcount, correction, 0, NULL, NULL));
1001
1002         vector<pair<size_t, const void *> > args;
1003         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&stagebuffer ));
1004         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&scaleinfobuffer ));
1005         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&newnodebuffer ));
1006         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsum.data ));
1007         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsqsum.data ));
1008         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&candidatebuffer ));
1009         args.push_back ( make_pair(sizeof(cl_int) , (void *)&gsum.rows ));
1010         args.push_back ( make_pair(sizeof(cl_int) , (void *)&gsum.cols ));
1011         args.push_back ( make_pair(sizeof(cl_int) , (void *)&step ));
1012         args.push_back ( make_pair(sizeof(cl_int) , (void *)&loopcount ));
1013         args.push_back ( make_pair(sizeof(cl_int) , (void *)&startstage ));
1014         args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitstage ));
1015         args.push_back ( make_pair(sizeof(cl_int) , (void *)&endstage ));
1016         args.push_back ( make_pair(sizeof(cl_int) , (void *)&startnode ));
1017         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&pbuffer ));
1018         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&correctionbuffer ));
1019         args.push_back ( make_pair(sizeof(cl_int) , (void *)&nodenum ));
1020         const char * build_options = gcascade->is_stump_based ? "-D STUMP_BASED=1" : "-D STUMP_BASED=0";
1021         openCLExecuteKernel(gsum.clCxt, &haarobjectdetect_scaled2, "gpuRunHaarClassifierCascade_scaled2", globalThreads, localThreads, args, -1, -1, build_options);
1022
1023         candidate = (int *)clEnqueueMapBuffer(qu, candidatebuffer, 1, CL_MAP_READ, 0, 4 * sizeof(int) * outputsz, 0, 0, 0, &status);
1024
1025         for(int i = 0; i < outputsz; i++)
1026         {
1027             if(candidate[4 * i + 2] != 0)
1028                 allCandidates.push_back(Rect(candidate[4 * i], candidate[4 * i + 1], candidate[4 * i + 2], candidate[4 * i + 3]));
1029         }
1030
1031         free(scaleinfo);
1032         free(p);
1033         free(correction);
1034         clEnqueueUnmapMemObject(qu, candidatebuffer, candidate, 0, 0, 0);
1035         openCLSafeCall(clReleaseMemObject(stagebuffer));
1036         openCLSafeCall(clReleaseMemObject(scaleinfobuffer));
1037         openCLSafeCall(clReleaseMemObject(nodebuffer));
1038         openCLSafeCall(clReleaseMemObject(newnodebuffer));
1039         openCLSafeCall(clReleaseMemObject(candidatebuffer));
1040         openCLSafeCall(clReleaseMemObject(pbuffer));
1041         openCLSafeCall(clReleaseMemObject(correctionbuffer));
1042     }
1043
1044     cvFree(&cascade->hid_cascade);
1045     rectList.resize(allCandidates.size());
1046     if(!allCandidates.empty())
1047         std::copy(allCandidates.begin(), allCandidates.end(), rectList.begin());
1048
1049     if( minNeighbors != 0 || findBiggestObject )
1050         groupRectangles(rectList, rweights, std::max(minNeighbors, 1), GROUP_EPS);
1051     else
1052         rweights.resize(rectList.size(), 0);
1053
1054     if( findBiggestObject && rectList.size() )
1055     {
1056         CvAvgComp result_comp = {{0, 0, 0, 0}, 0};
1057
1058         for( size_t i = 0; i < rectList.size(); i++ )
1059         {
1060             cv::Rect r = rectList[i];
1061             if( r.area() > cv::Rect(result_comp.rect).area() )
1062             {
1063                 result_comp.rect = r;
1064                 result_comp.neighbors = rweights[i];
1065             }
1066         }
1067         cvSeqPush( result_seq, &result_comp );
1068     }
1069     else
1070     {
1071         for( size_t i = 0; i < rectList.size(); i++ )
1072         {
1073             CvAvgComp c;
1074             c.rect = rectList[i];
1075             c.neighbors = rweights[i];
1076             cvSeqPush( result_seq, &c );
1077         }
1078     }
1079
1080     return result_seq;
1081 }
1082
1083 struct OclBuffers
1084 {
1085     cl_mem stagebuffer;
1086     cl_mem nodebuffer;
1087     cl_mem candidatebuffer;
1088     cl_mem scaleinfobuffer;
1089     cl_mem pbuffer;
1090     cl_mem correctionbuffer;
1091     cl_mem newnodebuffer;
1092 };
1093
1094 struct getRect
1095 {
1096     Rect operator()(const CvAvgComp &e) const
1097     {
1098         return e.rect;
1099     }
1100 };
1101
1102 void cv::ocl::OclCascadeClassifierBuf::detectMultiScale(oclMat &gimg, CV_OUT std::vector<cv::Rect>& faces,
1103                                                         double scaleFactor, int minNeighbors, int flags,
1104                                                         Size minSize, Size maxSize)
1105 {
1106     int blocksize = 8;
1107     int grp_per_CU = 12;
1108     size_t localThreads[3] = { blocksize, blocksize, 1 };
1109     size_t globalThreads[3] = { grp_per_CU * cv::ocl::Context::getContext()->getDeviceInfo().maxComputeUnits *localThreads[0],
1110         localThreads[1],
1111         1 };
1112     int outputsz = 256 * globalThreads[0] / localThreads[0];
1113
1114     Init(gimg.rows, gimg.cols, scaleFactor, flags, outputsz, localThreads, minSize, maxSize);
1115
1116     const double GROUP_EPS = 0.2;
1117
1118     cv::ConcurrentRectVector allCandidates;
1119     std::vector<cv::Rect> rectList;
1120     std::vector<int> rweights;
1121
1122     CvHaarClassifierCascade      *cascade = oldCascade;
1123     GpuHidHaarClassifierCascade  *gcascade;
1124     GpuHidHaarStageClassifier    *stage;
1125
1126     if( CV_MAT_DEPTH(gimg.type()) != CV_8U )
1127         CV_Error( CV_StsUnsupportedFormat, "Only 8-bit images are supported" );
1128
1129     if( CV_MAT_CN(gimg.type()) > 1 )
1130     {
1131         oclMat gtemp;
1132         cvtColor( gimg, gtemp, CV_BGR2GRAY );
1133         gimg = gtemp;
1134     }
1135
1136     int *candidate;
1137     cl_command_queue qu = getClCommandQueue(Context::getContext());
1138     if( (flags & CV_HAAR_SCALE_IMAGE) )
1139     {
1140         int indexy = 0;
1141         CvSize sz;
1142
1143         cv::Rect roi, roi2;
1144         cv::ocl::oclMat resizeroi, gimgroi, gimgroisq;
1145
1146         for( int i = 0; i < m_loopcount; i++ )
1147         {
1148             sz = sizev[i];
1149             roi = Rect(0, indexy, sz.width, sz.height);
1150             roi2 = Rect(0, 0, sz.width - 1, sz.height - 1);
1151             resizeroi = gimg1(roi2);
1152             gimgroi = gsum(roi);
1153             gimgroisq = gsqsum(roi);
1154
1155             cv::ocl::resize(gimg, resizeroi, Size(sz.width - 1, sz.height - 1), 0, 0, INTER_LINEAR);
1156             cv::ocl::integral(resizeroi, gimgroi, gimgroisq);
1157             indexy += sz.height;
1158         }
1159
1160         gcascade   = (GpuHidHaarClassifierCascade *)(cascade->hid_cascade);
1161         stage      = (GpuHidHaarStageClassifier *)(gcascade + 1);
1162
1163         int startstage = 0;
1164         int endstage = gcascade->count;
1165         int startnode = 0;
1166         int pixelstep = gsum.step / 4;
1167         int splitstage = 3;
1168         int splitnode = stage[0].count + stage[1].count + stage[2].count;
1169         cl_int4 p, pq;
1170         p.s[0] = gcascade->p0;
1171         p.s[1] = gcascade->p1;
1172         p.s[2] = gcascade->p2;
1173         p.s[3] = gcascade->p3;
1174         pq.s[0] = gcascade->pq0;
1175         pq.s[1] = gcascade->pq1;
1176         pq.s[2] = gcascade->pq2;
1177         pq.s[3] = gcascade->pq3;
1178         float correction = gcascade->inv_window_area;
1179
1180         vector<pair<size_t, const void *> > args;
1181         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->stagebuffer ));
1182         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->scaleinfobuffer ));
1183         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->nodebuffer ));
1184         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsum.data ));
1185         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsqsum.data ));
1186         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->candidatebuffer ));
1187         args.push_back ( make_pair(sizeof(cl_int) , (void *)&pixelstep ));
1188         args.push_back ( make_pair(sizeof(cl_int) , (void *)&m_loopcount ));
1189         args.push_back ( make_pair(sizeof(cl_int) , (void *)&startstage ));
1190         args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitstage ));
1191         args.push_back ( make_pair(sizeof(cl_int) , (void *)&endstage ));
1192         args.push_back ( make_pair(sizeof(cl_int) , (void *)&startnode ));
1193         args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitnode ));
1194         args.push_back ( make_pair(sizeof(cl_int4) , (void *)&p ));
1195         args.push_back ( make_pair(sizeof(cl_int4) , (void *)&pq ));
1196         args.push_back ( make_pair(sizeof(cl_float) , (void *)&correction ));
1197
1198         const char * build_options = gcascade->is_stump_based ? "-D STUMP_BASED=1" : "-D STUMP_BASED=0";
1199
1200         openCLExecuteKernel(gsum.clCxt, &haarobjectdetect, "gpuRunHaarClassifierCascade", globalThreads, localThreads, args, -1, -1, build_options);
1201
1202         candidate = (int *)malloc(4 * sizeof(int) * outputsz);
1203         memset(candidate, 0, 4 * sizeof(int) * outputsz);
1204
1205         openCLReadBuffer( gsum.clCxt, ((OclBuffers *)buffers)->candidatebuffer, candidate, 4 * sizeof(int)*outputsz );
1206
1207         for(int i = 0; i < outputsz; i++)
1208         {
1209             if(candidate[4 * i + 2] != 0)
1210             {
1211                 allCandidates.push_back(Rect(candidate[4 * i], candidate[4 * i + 1],
1212                 candidate[4 * i + 2], candidate[4 * i + 3]));
1213             }
1214         }
1215         free((void *)candidate);
1216         candidate = NULL;
1217     }
1218     else
1219     {
1220         cv::ocl::integral(gimg, gsum, gsqsum);
1221
1222         gcascade   = (GpuHidHaarClassifierCascade *)cascade->hid_cascade;
1223
1224         int step = gsum.step / 4;
1225         int startnode = 0;
1226         int splitstage = 3;
1227
1228         int startstage = 0;
1229         int endstage = gcascade->count;
1230
1231         vector<pair<size_t, const void *> > args;
1232         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->stagebuffer ));
1233         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->scaleinfobuffer ));
1234         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->newnodebuffer ));
1235         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsum.data ));
1236         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsqsum.data ));
1237         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->candidatebuffer ));
1238         args.push_back ( make_pair(sizeof(cl_int) , (void *)&gsum.rows ));
1239         args.push_back ( make_pair(sizeof(cl_int) , (void *)&gsum.cols ));
1240         args.push_back ( make_pair(sizeof(cl_int) , (void *)&step ));
1241         args.push_back ( make_pair(sizeof(cl_int) , (void *)&m_loopcount ));
1242         args.push_back ( make_pair(sizeof(cl_int) , (void *)&startstage ));
1243         args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitstage ));
1244         args.push_back ( make_pair(sizeof(cl_int) , (void *)&endstage ));
1245         args.push_back ( make_pair(sizeof(cl_int) , (void *)&startnode ));
1246         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->pbuffer ));
1247         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->correctionbuffer ));
1248         args.push_back ( make_pair(sizeof(cl_int) , (void *)&m_nodenum ));
1249
1250         const char * build_options = gcascade->is_stump_based ? "-D STUMP_BASED=1" : "-D STUMP_BASED=0";
1251         openCLExecuteKernel(gsum.clCxt, &haarobjectdetect_scaled2, "gpuRunHaarClassifierCascade_scaled2", globalThreads, localThreads, args, -1, -1, build_options);
1252
1253         candidate = (int *)clEnqueueMapBuffer(qu, ((OclBuffers *)buffers)->candidatebuffer, 1, CL_MAP_READ, 0, 4 * sizeof(int) * outputsz, 0, 0, 0, NULL);
1254
1255         for(int i = 0; i < outputsz; i++)
1256         {
1257             if(candidate[4 * i + 2] != 0)
1258                 allCandidates.push_back(Rect(candidate[4 * i], candidate[4 * i + 1],
1259                 candidate[4 * i + 2], candidate[4 * i + 3]));
1260         }
1261         clEnqueueUnmapMemObject(qu, ((OclBuffers *)buffers)->candidatebuffer, candidate, 0, 0, 0);
1262     }
1263     rectList.resize(allCandidates.size());
1264     if(!allCandidates.empty())
1265         std::copy(allCandidates.begin(), allCandidates.end(), rectList.begin());
1266
1267     if( minNeighbors != 0 || findBiggestObject )
1268         groupRectangles(rectList, rweights, std::max(minNeighbors, 1), GROUP_EPS);
1269     else
1270         rweights.resize(rectList.size(), 0);
1271
1272     GenResult(faces, rectList, rweights);
1273 }
1274
1275 void cv::ocl::OclCascadeClassifierBuf::Init(const int rows, const int cols,
1276     double scaleFactor, int flags,
1277     const int outputsz, const size_t localThreads[],
1278     CvSize minSize, CvSize maxSize)
1279 {
1280     if(initialized)
1281     {
1282         return; // we only allow one time initialization
1283     }
1284     CvHaarClassifierCascade      *cascade = oldCascade;
1285
1286     if( !CV_IS_HAAR_CLASSIFIER(cascade) )
1287         CV_Error( !cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier cascade" );
1288
1289     if( scaleFactor <= 1 )
1290         CV_Error( CV_StsOutOfRange, "scale factor must be > 1" );
1291
1292     if( cols < minSize.width || rows < minSize.height )
1293         CV_Error(CV_StsError, "Image too small");
1294
1295     int datasize=0;
1296     int totalclassifier=0;
1297
1298     if( !cascade->hid_cascade )
1299     {
1300         gpuCreateHidHaarClassifierCascade(cascade, &datasize, &totalclassifier);
1301     }
1302
1303     if( maxSize.height == 0 || maxSize.width == 0 )
1304     {
1305         maxSize.height = rows;
1306         maxSize.width = cols;
1307     }
1308
1309     findBiggestObject = (flags & CV_HAAR_FIND_BIGGEST_OBJECT) != 0;
1310     if( findBiggestObject )
1311         flags &= ~(CV_HAAR_SCALE_IMAGE | CV_HAAR_DO_CANNY_PRUNING);
1312
1313     CreateBaseBufs(datasize, totalclassifier, flags, outputsz);
1314     CreateFactorRelatedBufs(rows, cols, flags, scaleFactor, localThreads, minSize, maxSize);
1315
1316     m_scaleFactor = scaleFactor;
1317     m_rows = rows;
1318     m_cols = cols;
1319     m_flags = flags;
1320     m_minSize = minSize;
1321     m_maxSize = maxSize;
1322
1323     // initialize nodes
1324     GpuHidHaarClassifierCascade  *gcascade;
1325     GpuHidHaarStageClassifier    *stage;
1326     GpuHidHaarClassifier         *classifier;
1327     GpuHidHaarTreeNode           *node;
1328     cl_command_queue qu = getClCommandQueue(Context::getContext());
1329     if( (flags & CV_HAAR_SCALE_IMAGE) )
1330     {
1331         gcascade   = (GpuHidHaarClassifierCascade *)(cascade->hid_cascade);
1332         stage      = (GpuHidHaarStageClassifier *)(gcascade + 1);
1333         classifier = (GpuHidHaarClassifier *)(stage + gcascade->count);
1334         node       = (GpuHidHaarTreeNode *)(classifier->node);
1335
1336         gpuSetImagesForHaarClassifierCascade( cascade, 1., gsum.step / 4 );
1337
1338         openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->stagebuffer, 1, 0,
1339             sizeof(GpuHidHaarStageClassifier) * gcascade->count,
1340             stage, 0, NULL, NULL));
1341
1342         openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->nodebuffer, 1, 0,
1343                                             m_nodenum * sizeof(GpuHidHaarTreeNode),
1344                                             node, 0, NULL, NULL));
1345     }
1346     else
1347     {
1348         gpuSetHaarClassifierCascade(cascade);
1349
1350         gcascade   = (GpuHidHaarClassifierCascade *)cascade->hid_cascade;
1351         stage      = (GpuHidHaarStageClassifier *)(gcascade + 1);
1352         classifier = (GpuHidHaarClassifier *)(stage + gcascade->count);
1353         node       = (GpuHidHaarTreeNode *)(classifier->node);
1354
1355         openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->nodebuffer, 1, 0,
1356             m_nodenum * sizeof(GpuHidHaarTreeNode),
1357             node, 0, NULL, NULL));
1358
1359         cl_int4 *p = (cl_int4 *)malloc(sizeof(cl_int4) * m_loopcount);
1360         float *correction = (float *)malloc(sizeof(float) * m_loopcount);
1361         double factor;
1362         for(int i = 0; i < m_loopcount; i++)
1363         {
1364             factor = scalev[i];
1365             int equRect_x = (int)(factor * gcascade->p0 + 0.5);
1366             int equRect_y = (int)(factor * gcascade->p1 + 0.5);
1367             int equRect_w = (int)(factor * gcascade->p3 + 0.5);
1368             int equRect_h = (int)(factor * gcascade->p2 + 0.5);
1369             p[i].s[0] = equRect_x;
1370             p[i].s[1] = equRect_y;
1371             p[i].s[2] = equRect_x + equRect_w;
1372             p[i].s[3] = equRect_y + equRect_h;
1373             correction[i] = 1. / (equRect_w * equRect_h);
1374             int startnodenum = m_nodenum * i;
1375             float factor2 = (float)factor;
1376
1377             vector<pair<size_t, const void *> > args1;
1378             args1.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->nodebuffer ));
1379             args1.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->newnodebuffer ));
1380             args1.push_back ( make_pair(sizeof(cl_float) , (void *)&factor2 ));
1381             args1.push_back ( make_pair(sizeof(cl_float) , (void *)&correction[i] ));
1382             args1.push_back ( make_pair(sizeof(cl_int) , (void *)&startnodenum ));
1383
1384             size_t globalThreads2[3] = {m_nodenum, 1, 1};
1385
1386             openCLExecuteKernel(Context::getContext(), &haarobjectdetect_scaled2, "gpuscaleclassifier", globalThreads2, NULL/*localThreads2*/, args1, -1, -1);
1387         }
1388         openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->stagebuffer, 1, 0, sizeof(GpuHidHaarStageClassifier)*gcascade->count, stage, 0, NULL, NULL));
1389         openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->pbuffer, 1, 0, sizeof(cl_int4)*m_loopcount, p, 0, NULL, NULL));
1390         openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->correctionbuffer, 1, 0, sizeof(cl_float)*m_loopcount, correction, 0, NULL, NULL));
1391
1392         free(p);
1393         free(correction);
1394     }
1395     initialized = true;
1396 }
1397
1398 void cv::ocl::OclCascadeClassifierBuf::CreateBaseBufs(const int datasize, const int totalclassifier,
1399                                                       const int flags, const int outputsz)
1400 {
1401     if (!initialized)
1402     {
1403         buffers = malloc(sizeof(OclBuffers));
1404
1405         size_t tempSize =
1406             sizeof(GpuHidHaarStageClassifier) * ((GpuHidHaarClassifierCascade *)oldCascade->hid_cascade)->count;
1407         m_nodenum = (datasize - sizeof(GpuHidHaarClassifierCascade) - tempSize - sizeof(GpuHidHaarClassifier) * totalclassifier)
1408             / sizeof(GpuHidHaarTreeNode);
1409
1410         ((OclBuffers *)buffers)->stagebuffer     = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_ONLY,  tempSize);
1411         ((OclBuffers *)buffers)->nodebuffer      = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_ONLY,  m_nodenum * sizeof(GpuHidHaarTreeNode));
1412     }
1413
1414     if (initialized
1415         && ((m_flags & CV_HAAR_SCALE_IMAGE) ^ (flags & CV_HAAR_SCALE_IMAGE)))
1416     {
1417         openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->candidatebuffer));
1418     }
1419
1420     if (flags & CV_HAAR_SCALE_IMAGE)
1421     {
1422         ((OclBuffers *)buffers)->candidatebuffer = openCLCreateBuffer(cv::ocl::Context::getContext(),
1423                                                         CL_MEM_WRITE_ONLY,
1424                                                         4 * sizeof(int) * outputsz);
1425     }
1426     else
1427     {
1428         ((OclBuffers *)buffers)->candidatebuffer = openCLCreateBuffer(cv::ocl::Context::getContext(),
1429                                                         CL_MEM_WRITE_ONLY | CL_MEM_ALLOC_HOST_PTR,
1430                                                         4 * sizeof(int) * outputsz);
1431     }
1432 }
1433
1434 void cv::ocl::OclCascadeClassifierBuf::CreateFactorRelatedBufs(
1435     const int rows, const int cols, const int flags,
1436     const double scaleFactor, const size_t localThreads[],
1437     CvSize minSize, CvSize maxSize)
1438 {
1439     if (initialized)
1440     {
1441         if ((m_flags & CV_HAAR_SCALE_IMAGE) && !(flags & CV_HAAR_SCALE_IMAGE))
1442         {
1443             gimg1.release();
1444             gsum.release();
1445             gsqsum.release();
1446         }
1447         else if (!(m_flags & CV_HAAR_SCALE_IMAGE) && (flags & CV_HAAR_SCALE_IMAGE))
1448         {
1449             openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->newnodebuffer));
1450             openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->correctionbuffer));
1451             openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->pbuffer));
1452         }
1453         else if ((m_flags & CV_HAAR_SCALE_IMAGE) && (flags & CV_HAAR_SCALE_IMAGE))
1454         {
1455             if (fabs(m_scaleFactor - scaleFactor) < 1e-6
1456                 && (rows == m_rows && cols == m_cols)
1457                 && (minSize.width == m_minSize.width)
1458                 && (minSize.height == m_minSize.height)
1459                 && (maxSize.width == m_maxSize.width)
1460                 && (maxSize.height == m_maxSize.height))
1461             {
1462                 return;
1463             }
1464         }
1465         else
1466         {
1467             if (fabs(m_scaleFactor - scaleFactor) < 1e-6
1468                 && (rows == m_rows && cols == m_cols)
1469                 && (minSize.width == m_minSize.width)
1470                 && (minSize.height == m_minSize.height)
1471                 && (maxSize.width == m_maxSize.width)
1472                 && (maxSize.height == m_maxSize.height))
1473             {
1474                 return;
1475             }
1476             else
1477             {
1478                 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->newnodebuffer));
1479                 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->correctionbuffer));
1480                 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->pbuffer));
1481             }
1482         }
1483     }
1484
1485     int loopcount;
1486     int indexy = 0;
1487     int totalheight = 0;
1488     double factor;
1489     Rect roi;
1490     CvSize sz;
1491     CvSize winSize0 = oldCascade->orig_window_size;
1492     detect_piramid_info *scaleinfo;
1493     cl_command_queue qu = getClCommandQueue(Context::getContext());
1494     if (flags & CV_HAAR_SCALE_IMAGE)
1495     {
1496         for(factor = 1.f;; factor *= scaleFactor)
1497         {
1498             CvSize winSize = { cvRound(winSize0.width * factor), cvRound(winSize0.height * factor) };
1499             sz.width     = cvRound( cols / factor ) + 1;
1500             sz.height    = cvRound( rows / factor ) + 1;
1501             CvSize sz1     = { sz.width - winSize0.width - 1,      sz.height - winSize0.height - 1 };
1502
1503             if( sz1.width <= 0 || sz1.height <= 0 )
1504                 break;
1505             if( winSize.width > maxSize.width || winSize.height > maxSize.height )
1506                 break;
1507             if( winSize.width < minSize.width || winSize.height < minSize.height )
1508                 continue;
1509
1510             totalheight += sz.height;
1511             sizev.push_back(sz);
1512             scalev.push_back(static_cast<float>(factor));
1513         }
1514
1515         loopcount = sizev.size();
1516         gimg1.create(rows, cols, CV_8UC1);
1517         gsum.create(totalheight + 4, cols + 1, CV_32SC1);
1518         gsqsum.create(totalheight + 4, cols + 1, CV_32FC1);
1519
1520         scaleinfo = (detect_piramid_info *)malloc(sizeof(detect_piramid_info) * loopcount);
1521         for( int i = 0; i < loopcount; i++ )
1522         {
1523             sz = sizev[i];
1524             roi = Rect(0, indexy, sz.width, sz.height);
1525             int width = sz.width - 1 - oldCascade->orig_window_size.width;
1526             int height = sz.height - 1 - oldCascade->orig_window_size.height;
1527             int grpnumperline = (width + localThreads[0] - 1) / localThreads[0];
1528             int totalgrp = ((height + localThreads[1] - 1) / localThreads[1]) * grpnumperline;
1529
1530             ((detect_piramid_info *)scaleinfo)[i].width_height = (width << 16) | height;
1531             ((detect_piramid_info *)scaleinfo)[i].grpnumperline_totalgrp = (grpnumperline << 16) | totalgrp;
1532             ((detect_piramid_info *)scaleinfo)[i].imgoff = gsum(roi).offset >> 2;
1533             ((detect_piramid_info *)scaleinfo)[i].factor = scalev[i];
1534
1535             indexy += sz.height;
1536         }
1537     }
1538     else
1539     {
1540         for(factor = 1;
1541             cvRound(factor * winSize0.width) < cols - 10 && cvRound(factor * winSize0.height) < rows - 10;
1542             factor *= scaleFactor)
1543         {
1544             CvSize winSize = { cvRound( winSize0.width * factor ), cvRound( winSize0.height * factor ) };
1545             if( winSize.width < minSize.width || winSize.height < minSize.height )
1546             {
1547                 continue;
1548             }
1549             sizev.push_back(winSize);
1550             scalev.push_back(factor);
1551         }
1552
1553         loopcount = scalev.size();
1554         if(loopcount == 0)
1555         {
1556             loopcount = 1;
1557             sizev.push_back(minSize);
1558             scalev.push_back( std::min(cvRound(minSize.width / winSize0.width), cvRound(minSize.height / winSize0.height)) );
1559         }
1560
1561         ((OclBuffers *)buffers)->pbuffer = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_ONLY,
1562             sizeof(cl_int4) * loopcount);
1563         ((OclBuffers *)buffers)->correctionbuffer = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_ONLY,
1564             sizeof(cl_float) * loopcount);
1565         ((OclBuffers *)buffers)->newnodebuffer = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_WRITE,
1566             loopcount * m_nodenum * sizeof(GpuHidHaarTreeNode));
1567
1568         scaleinfo = (detect_piramid_info *)malloc(sizeof(detect_piramid_info) * loopcount);
1569         for( int i = 0; i < loopcount; i++ )
1570         {
1571             sz = sizev[i];
1572             factor = scalev[i];
1573             int ystep = cvRound(std::max(2., factor));
1574             int width = (cols - 1 - sz.width  + ystep - 1) / ystep;
1575             int height = (rows - 1 - sz.height + ystep - 1) / ystep;
1576             int grpnumperline = (width + localThreads[0] - 1) / localThreads[0];
1577             int totalgrp = ((height + localThreads[1] - 1) / localThreads[1]) * grpnumperline;
1578
1579             ((detect_piramid_info *)scaleinfo)[i].width_height = (width << 16) | height;
1580             ((detect_piramid_info *)scaleinfo)[i].grpnumperline_totalgrp = (grpnumperline << 16) | totalgrp;
1581             ((detect_piramid_info *)scaleinfo)[i].imgoff = 0;
1582             ((detect_piramid_info *)scaleinfo)[i].factor = factor;
1583         }
1584     }
1585
1586     if (loopcount != m_loopcount)
1587     {
1588         if (initialized)
1589         {
1590             openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->scaleinfobuffer));
1591         }
1592         ((OclBuffers *)buffers)->scaleinfobuffer = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_ONLY, sizeof(detect_piramid_info) * loopcount);
1593     }
1594
1595     openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->scaleinfobuffer, 1, 0,
1596         sizeof(detect_piramid_info)*loopcount,
1597         scaleinfo, 0, NULL, NULL));
1598     free(scaleinfo);
1599
1600     m_loopcount = loopcount;
1601 }
1602
1603 void cv::ocl::OclCascadeClassifierBuf::GenResult(CV_OUT std::vector<cv::Rect>& faces,
1604                                                  const std::vector<cv::Rect> &rectList,
1605                                                  const std::vector<int> &rweights)
1606 {
1607     MemStorage tempStorage(cvCreateMemStorage(0));
1608     CvSeq *result_seq = cvCreateSeq( 0, sizeof(CvSeq), sizeof(CvAvgComp), tempStorage );
1609
1610     if( findBiggestObject && rectList.size() )
1611     {
1612         CvAvgComp result_comp = {{0, 0, 0, 0}, 0};
1613
1614         for( size_t i = 0; i < rectList.size(); i++ )
1615         {
1616             cv::Rect r = rectList[i];
1617             if( r.area() > cv::Rect(result_comp.rect).area() )
1618             {
1619                 result_comp.rect = r;
1620                 result_comp.neighbors = rweights[i];
1621             }
1622         }
1623         cvSeqPush( result_seq, &result_comp );
1624     }
1625     else
1626     {
1627         for( size_t i = 0; i < rectList.size(); i++ )
1628         {
1629             CvAvgComp c;
1630             c.rect = rectList[i];
1631             c.neighbors = rweights[i];
1632             cvSeqPush( result_seq, &c );
1633         }
1634     }
1635
1636     vector<CvAvgComp> vecAvgComp;
1637     Seq<CvAvgComp>(result_seq).copyTo(vecAvgComp);
1638     faces.resize(vecAvgComp.size());
1639     std::transform(vecAvgComp.begin(), vecAvgComp.end(), faces.begin(), getRect());
1640 }
1641
1642 void cv::ocl::OclCascadeClassifierBuf::release()
1643 {
1644     if(initialized)
1645     {
1646         openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->stagebuffer));
1647         openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->scaleinfobuffer));
1648         openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->nodebuffer));
1649         openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->candidatebuffer));
1650
1651         if( (m_flags & CV_HAAR_SCALE_IMAGE) )
1652         {
1653             cvFree(&oldCascade->hid_cascade);
1654         }
1655         else
1656         {
1657             openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->newnodebuffer));
1658             openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->correctionbuffer));
1659             openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->pbuffer));
1660         }
1661
1662         free(buffers);
1663         buffers = NULL;
1664         initialized = false;
1665     }
1666 }
1667
1668 #ifndef _MAX_PATH
1669 #define _MAX_PATH 1024
1670 #endif