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 materials 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, 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                 const CvHaarFeature *feature =
 634                     &_cascade->stage_classifier[i].classifier[j].haar_feature[l];
 635                 GpuHidHaarTreeNode *hidnode = &stage_classifier[i].classifier[j].node[l];
 636
 637                 for( int k = 0; k < CV_HAAR_FEATURE_MAX; k++ )
 638                 {
 639                     const CvRect tr = feature->rect[k].r;
 640                     if (tr.width == 0)
 641                         break;
 642                     double correction_ratio = weight_scale * (!feature->tilted ? 1 : 0.5);
 643                     hidnode->p[k][0] = tr.x;
 644                     hidnode->p[k][1] = tr.y;
 645                     hidnode->p[k][2] = tr.width;
 646                     hidnode->p[k][3] = tr.height;
 647                     hidnode->weight[k] = (float)(feature->rect[k].weight * correction_ratio);
 648                 }
 649             } /* l */
 650         } /* j */
 651     }
 652 }
 653
 654 CvSeq *cv::ocl::OclCascadeClassifier::oclHaarDetectObjects( oclMat &gimg, CvMemStorage *storage, double scaleFactor,
 655         int minNeighbors, int flags, CvSize minSize, CvSize maxSize)
 656 {
 657     CvHaarClassifierCascade *cascade = oldCascade;
 658
 659     const double GROUP_EPS = 0.2;
 660     CvSeq *result_seq = 0;
 661
 662     cv::ConcurrentRectVector allCandidates;
 663     std::vector<cv::Rect> rectList;
 664     std::vector<int> rweights;
 665     double factor;
 666     int datasize=0;
 667     int totalclassifier=0;
 668
 669     GpuHidHaarClassifierCascade *gcascade;
 670     GpuHidHaarStageClassifier    *stage;
 671     GpuHidHaarClassifier         *classifier;
 672     GpuHidHaarTreeNode           *node;
 673
 674     int *candidate;
 675     cl_int status;
 676
 677     bool findBiggestObject = (flags & CV_HAAR_FIND_BIGGEST_OBJECT) != 0;
 678
 679     if( maxSize.height == 0 || maxSize.width == 0 )
 680     {
 681         maxSize.height = gimg.rows;
 682         maxSize.width = gimg.cols;
 683     }
 684
 685     if( !CV_IS_HAAR_CLASSIFIER(cascade) )
 686         CV_Error( !cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier cascade" );
 687
 688     if( !storage )
 689         CV_Error( CV_StsNullPtr, "Null storage pointer" );
 690
 691     if( CV_MAT_DEPTH(gimg.type()) != CV_8U )
 692         CV_Error( CV_StsUnsupportedFormat, "Only 8-bit images are supported" );
 693
 694     if( scaleFactor <= 1 )
 695         CV_Error( CV_StsOutOfRange, "scale factor must be > 1" );
 696
 697     if( findBiggestObject )
 698         flags &= ~CV_HAAR_SCALE_IMAGE;
 699
 700     if( !cascade->hid_cascade )
 701         gpuCreateHidHaarClassifierCascade(cascade, &datasize, &totalclassifier);
 702
 703     result_seq = cvCreateSeq( 0, sizeof(CvSeq), sizeof(CvAvgComp), storage );
 704
 705     if( CV_MAT_CN(gimg.type()) > 1 )
 706     {
 707         oclMat gtemp;
 708         cvtColor( gimg, gtemp, CV_BGR2GRAY );
 709         gimg = gtemp;
 710     }
 711
 712     if( findBiggestObject )
 713         flags &= ~(CV_HAAR_SCALE_IMAGE | CV_HAAR_DO_CANNY_PRUNING);
 714
 715     if( gimg.cols < minSize.width || gimg.rows < minSize.height )
 716         CV_Error(CV_StsError, "Image too small");
 717
 718     cl_command_queue qu = getClCommandQueue(Context::getContext());
 719     if( (flags & CV_HAAR_SCALE_IMAGE) )
 720     {
 721         CvSize winSize0 = cascade->orig_window_size;
 722         int totalheight = 0;
 723         int indexy = 0;
 724         CvSize sz;
 725         vector<CvSize> sizev;
 726         vector<float> scalev;
 727         for(factor = 1.f;; factor *= scaleFactor)
 728         {
 729             CvSize winSize = { cvRound(winSize0.width * factor), cvRound(winSize0.height * factor) };
 730             sz.width     = cvRound( gimg.cols / factor ) + 1;
 731             sz.height    = cvRound( gimg.rows / factor ) + 1;
 732             CvSize sz1     = { sz.width - winSize0.width - 1,      sz.height - winSize0.height - 1 };
 733
 734             if( sz1.width <= 0 || sz1.height <= 0 )
 735                 break;
 736             if( winSize.width > maxSize.width || winSize.height > maxSize.height )
 737                 break;
 738             if( winSize.width < minSize.width || winSize.height < minSize.height )
 739                 continue;
 740
 741             totalheight += sz.height;
 742             sizev.push_back(sz);
 743             scalev.push_back(factor);
 744         }
 745
 746         oclMat gimg1(gimg.rows, gimg.cols, CV_8UC1);
 747         oclMat gsum(totalheight + 4, gimg.cols + 1, CV_32SC1);
 748         oclMat gsqsum(totalheight + 4, gimg.cols + 1, CV_32FC1);
 749
 750         cl_mem stagebuffer;
 751         cl_mem nodebuffer;
 752         cl_mem candidatebuffer;
 753         cl_mem scaleinfobuffer;
 754         cv::Rect roi, roi2;
 755         cv::Mat imgroi, imgroisq;
 756         cv::ocl::oclMat resizeroi, gimgroi, gimgroisq;
 757         int grp_per_CU = 12;
 758
 759         size_t blocksize = 8;
 760         size_t localThreads[3] = { blocksize, blocksize , 1 };
 761         size_t globalThreads[3] = { grp_per_CU *(gsum.clCxt->getDeviceInfo().maxComputeUnits) *localThreads[0],
 762                                     localThreads[1], 1
 763                                   };
 764         int outputsz = 256 * globalThreads[0] / localThreads[0];
 765         int loopcount = sizev.size();
 766         detect_piramid_info *scaleinfo = (detect_piramid_info *)malloc(sizeof(detect_piramid_info) * loopcount);
 767
 768         for( int i = 0; i < loopcount; i++ )
 769         {
 770             sz = sizev[i];
 771             factor = scalev[i];
 772             roi = Rect(0, indexy, sz.width, sz.height);
 773             roi2 = Rect(0, 0, sz.width - 1, sz.height - 1);
 774             resizeroi = gimg1(roi2);
 775             gimgroi = gsum(roi);
 776             gimgroisq = gsqsum(roi);
 777             int width = gimgroi.cols - 1 - cascade->orig_window_size.width;
 778             int height = gimgroi.rows - 1 - cascade->orig_window_size.height;
 779             scaleinfo[i].width_height = (width << 16) | height;
 780
 781
 782             int grpnumperline = (width + localThreads[0] - 1) / localThreads[0];
 783             int totalgrp = ((height + localThreads[1] - 1) / localThreads[1]) * grpnumperline;
 784
 785             scaleinfo[i].grpnumperline_totalgrp = (grpnumperline << 16) | totalgrp;
 786             scaleinfo[i].imgoff = gimgroi.offset >> 2;
 787             scaleinfo[i].factor = factor;
 788             cv::ocl::resize(gimg, resizeroi, Size(sz.width - 1, sz.height - 1), 0, 0, INTER_LINEAR);
 789             cv::ocl::integral(resizeroi, gimgroi, gimgroisq);
 790             indexy += sz.height;
 791         }
 792
 793         gcascade   = (GpuHidHaarClassifierCascade *)cascade->hid_cascade;
 794         stage      = (GpuHidHaarStageClassifier *)(gcascade + 1);
 795         classifier = (GpuHidHaarClassifier *)(stage + gcascade->count);
 796         node       = (GpuHidHaarTreeNode *)(classifier->node);
 797
 798         int nodenum = (datasize - sizeof(GpuHidHaarClassifierCascade) -
 799                        sizeof(GpuHidHaarStageClassifier) * gcascade->count - sizeof(GpuHidHaarClassifier) * totalclassifier) / sizeof(GpuHidHaarTreeNode);
 800
 801         candidate = (int *)malloc(4 * sizeof(int) * outputsz);
 802
 803         gpuSetImagesForHaarClassifierCascade( cascade, 1., gsum.step / 4 );
 804
 805         stagebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(GpuHidHaarStageClassifier) * gcascade->count);
 806         openCLSafeCall(clEnqueueWriteBuffer(qu, stagebuffer, 1, 0, sizeof(GpuHidHaarStageClassifier)*gcascade->count, stage, 0, NULL, NULL));
 807
 808         nodebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, nodenum * sizeof(GpuHidHaarTreeNode));
 809
 810         openCLSafeCall(clEnqueueWriteBuffer(qu, nodebuffer, 1, 0, nodenum * sizeof(GpuHidHaarTreeNode),
 811                                             node, 0, NULL, NULL));
 812         candidatebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_WRITE_ONLY, 4 * sizeof(int) * outputsz);
 813
 814         scaleinfobuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(detect_piramid_info) * loopcount);
 815         openCLSafeCall(clEnqueueWriteBuffer(qu, scaleinfobuffer, 1, 0, sizeof(detect_piramid_info)*loopcount, scaleinfo, 0, NULL, NULL));
 816
 817         int startstage = 0;
 818         int endstage = gcascade->count;
 819         int startnode = 0;
 820         int pixelstep = gsum.step / 4;
 821         int splitstage = 3;
 822         int splitnode = stage[0].count + stage[1].count + stage[2].count;
 823         cl_int4 p, pq;
 824         p.s[0] = gcascade->p0;
 825         p.s[1] = gcascade->p1;
 826         p.s[2] = gcascade->p2;
 827         p.s[3] = gcascade->p3;
 828         pq.s[0] = gcascade->pq0;
 829         pq.s[1] = gcascade->pq1;
 830         pq.s[2] = gcascade->pq2;
 831         pq.s[3] = gcascade->pq3;
 832         float correction = gcascade->inv_window_area;
 833
 834         vector<pair<size_t, const void *> > args;
 835         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&stagebuffer ));
 836         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&scaleinfobuffer ));
 837         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&nodebuffer ));
 838         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsum.data ));
 839         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsqsum.data ));
 840         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&candidatebuffer ));
 841         args.push_back ( make_pair(sizeof(cl_int) , (void *)&pixelstep ));
 842         args.push_back ( make_pair(sizeof(cl_int) , (void *)&loopcount ));
 843         args.push_back ( make_pair(sizeof(cl_int) , (void *)&startstage ));
 844         args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitstage ));
 845         args.push_back ( make_pair(sizeof(cl_int) , (void *)&endstage ));
 846         args.push_back ( make_pair(sizeof(cl_int) , (void *)&startnode ));
 847         args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitnode ));
 848         args.push_back ( make_pair(sizeof(cl_int4) , (void *)&p ));
 849         args.push_back ( make_pair(sizeof(cl_int4) , (void *)&pq ));
 850         args.push_back ( make_pair(sizeof(cl_float) , (void *)&correction ));
 851
 852         if(gcascade->is_stump_based && gsum.clCxt->supportsFeature(FEATURE_CL_INTEL_DEVICE))
 853         {
 854             //setup local group size
 855             localThreads[0] = 8;
 856             localThreads[1] = 16;
 857             localThreads[2] = 1;
 858
 859             //init maximal number of workgroups
 860             int WGNumX = 1+(sizev[0].width /(localThreads[0]));
 861             int WGNumY = 1+(sizev[0].height/(localThreads[1]));
 862             int WGNumZ = loopcount;
 863             int WGNum = 0; //accurate number of non -empty workgroups
 864             oclMat      oclWGInfo(1,sizeof(cl_int4) * WGNumX*WGNumY*WGNumZ,CV_8U);
 865             {
 866                 cl_int4*    pWGInfo = (cl_int4*)clEnqueueMapBuffer(getClCommandQueue(oclWGInfo.clCxt),(cl_mem)oclWGInfo.datastart,true,CL_MAP_WRITE, 0, oclWGInfo.step, 0,0,0,&status);
 867                 openCLVerifyCall(status);
 868                 for(int z=0;z<WGNumZ;++z)
 869                 {
 870                     int     Width  = (scaleinfo[z].width_height >> 16)&0xFFFF;
 871                     int     Height = (scaleinfo[z].width_height >> 0 )& 0xFFFF;
 872                     for(int y=0;y<WGNumY;++y)
 873                     {
 874                         int     gy = y*localThreads[1];
 875                         if(gy>=(Height-cascade->orig_window_size.height))
 876                             continue; // no data to process
 877                         for(int x=0;x<WGNumX;++x)
 878                         {
 879                             int     gx = x*localThreads[0];
 880                             if(gx>=(Width-cascade->orig_window_size.width))
 881                                 continue; // no data to process
 882
 883                             // save no-empty workgroup info into array
 884                             pWGInfo[WGNum].s[0] = scaleinfo[z].width_height;
 885                             pWGInfo[WGNum].s[1] = (gx << 16) | gy;
 886                             pWGInfo[WGNum].s[2] = scaleinfo[z].imgoff;
 887                             memcpy(&(pWGInfo[WGNum].s[3]),&(scaleinfo[z].factor),sizeof(float));
 888                             WGNum++;
 889                         }
 890                     }
 891                 }
 892                 openCLSafeCall(clEnqueueUnmapMemObject(getClCommandQueue(oclWGInfo.clCxt),(cl_mem)oclWGInfo.datastart,pWGInfo,0,0,0));
 893                 pWGInfo = NULL;
 894             }
 895
 896             // setup global sizes to have linear array of workgroups with WGNum size
 897             globalThreads[0] = localThreads[0]*WGNum;
 898             globalThreads[1] = localThreads[1];
 899             globalThreads[2] = 1;
 900
 901 #define NODE_SIZE 12
 902             // pack node info to have less memory loads
 903             oclMat  oclNodesPK(1,sizeof(cl_int) * NODE_SIZE * nodenum,CV_8U);
 904             {
 905                 cl_int  status;
 906                 cl_int* pNodesPK = (cl_int*)clEnqueueMapBuffer(getClCommandQueue(oclNodesPK.clCxt),(cl_mem)oclNodesPK.datastart,true,CL_MAP_WRITE, 0, oclNodesPK.step, 0,0,0,&status);
 907                 openCLVerifyCall(status);
 908                 //use known local data stride to precalulate indexes
 909                 int DATA_SIZE_X = (localThreads[0]+cascade->orig_window_size.width);
 910                 // check that maximal value is less than maximal unsigned short
 911                 assert(DATA_SIZE_X*cascade->orig_window_size.height+cascade->orig_window_size.width < USHRT_MAX);
 912                 for(int i = 0;i<nodenum;++i)
 913                 {//process each node from classifier
 914                     struct NodePK
 915                     {
 916                         unsigned short  slm_index[3][4];
 917                         float           weight[3];
 918                         float           threshold;
 919                         float           alpha[2];
 920                     };
 921                     struct NodePK * pOut = (struct NodePK *)(pNodesPK + NODE_SIZE*i);
 922                     for(int k=0;k<3;++k)
 923                     {// calc 4 short indexes in shared local mem for each rectangle instead of 2 (x,y) pair.
 924                         int* p = &(node[i].p[k][0]);
 925                         pOut->slm_index[k][0] = (unsigned short)(p[1]*DATA_SIZE_X+p[0]);
 926                         pOut->slm_index[k][1] = (unsigned short)(p[1]*DATA_SIZE_X+p[2]);
 927                         pOut->slm_index[k][2] = (unsigned short)(p[3]*DATA_SIZE_X+p[0]);
 928                         pOut->slm_index[k][3] = (unsigned short)(p[3]*DATA_SIZE_X+p[2]);
 929                     }
 930                     //store used float point values for each node
 931                     pOut->weight[0] = node[i].weight[0];
 932                     pOut->weight[1] = node[i].weight[1];
 933                     pOut->weight[2] = node[i].weight[2];
 934                     pOut->threshold = node[i].threshold;
 935                     pOut->alpha[0] = node[i].alpha[0];
 936                     pOut->alpha[1] = node[i].alpha[1];
 937                 }
 938                 openCLSafeCall(clEnqueueUnmapMemObject(getClCommandQueue(oclNodesPK.clCxt),(cl_mem)oclNodesPK.datastart,pNodesPK,0,0,0));
 939                 pNodesPK = NULL;
 940             }
 941             // add 2 additional buffers (WGinfo and packed nodes) as 2 last args
 942             args.push_back ( make_pair(sizeof(cl_mem) , (void *)&oclNodesPK.datastart ));
 943             args.push_back ( make_pair(sizeof(cl_mem) , (void *)&oclWGInfo.datastart ));
 944
 945             //form build options for kernel
 946             string  options = "-D PACKED_CLASSIFIER";
 947             options += format(" -D NODE_SIZE=%d",NODE_SIZE);
 948             options += format(" -D WND_SIZE_X=%d",cascade->orig_window_size.width);
 949             options += format(" -D WND_SIZE_Y=%d",cascade->orig_window_size.height);
 950             options += format(" -D STUMP_BASED=%d",gcascade->is_stump_based);
 951             options += format(" -D LSx=%d",localThreads[0]);
 952             options += format(" -D LSy=%d",localThreads[1]);
 953             options += format(" -D SPLITNODE=%d",splitnode);
 954             options += format(" -D SPLITSTAGE=%d",splitstage);
 955             options += format(" -D OUTPUTSZ=%d",outputsz);
 956
 957             // init candiate global count by 0
 958             int pattern = 0;
 959             openCLSafeCall(clEnqueueWriteBuffer(qu, candidatebuffer, 1, 0, 1 * sizeof(pattern),&pattern, 0, NULL, NULL));
 960             // execute face detector
 961             openCLExecuteKernel(gsum.clCxt, &haarobjectdetect, "gpuRunHaarClassifierCascadePacked", globalThreads, localThreads, args, -1, -1, options.c_str());
 962             //read candidate buffer back and put it into host list
 963             openCLReadBuffer( gsum.clCxt, candidatebuffer, candidate, 4 * sizeof(int)*outputsz );
 964             assert(candidate[0]<outputsz);
 965             //printf("candidate[0]=%d\n",candidate[0]);
 966             for(int i = 1; i <= candidate[0]; i++)
 967             {
 968                 allCandidates.push_back(Rect(candidate[4 * i], candidate[4 * i + 1],candidate[4 * i + 2], candidate[4 * i + 3]));
 969             }
 970         }
 971         else
 972         {
 973             const char * build_options = gcascade->is_stump_based ? "-D STUMP_BASED=1" : "-D STUMP_BASED=0";
 974
 975             openCLExecuteKernel(gsum.clCxt, &haarobjectdetect, "gpuRunHaarClassifierCascade", globalThreads, localThreads, args, -1, -1, build_options);
 976
 977             openCLReadBuffer( gsum.clCxt, candidatebuffer, candidate, 4 * sizeof(int)*outputsz );
 978
 979             for(int i = 0; i < outputsz; i++)
 980                 if(candidate[4 * i + 2] != 0)
 981                     allCandidates.push_back(Rect(candidate[4 * i], candidate[4 * i + 1],
 982                     candidate[4 * i + 2], candidate[4 * i + 3]));
 983         }
 984
 985         free(scaleinfo);
 986         free(candidate);
 987         openCLSafeCall(clReleaseMemObject(stagebuffer));
 988         openCLSafeCall(clReleaseMemObject(scaleinfobuffer));
 989         openCLSafeCall(clReleaseMemObject(nodebuffer));
 990         openCLSafeCall(clReleaseMemObject(candidatebuffer));
 991
 992     }
 993     else
 994     {
 995         CvSize winsize0 = cascade->orig_window_size;
 996         int n_factors = 0;
 997         oclMat gsum;
 998         oclMat gsqsum;
 999         cv::ocl::integral(gimg, gsum, gsqsum);
1000         CvSize sz;
1001         vector<CvSize> sizev;
1002         vector<float> scalev;
1003         gpuSetHaarClassifierCascade(cascade);
1004         gcascade   = (GpuHidHaarClassifierCascade *)cascade->hid_cascade;
1005         stage      = (GpuHidHaarStageClassifier *)(gcascade + 1);
1006         classifier = (GpuHidHaarClassifier *)(stage + gcascade->count);
1007         node       = (GpuHidHaarTreeNode *)(classifier->node);
1008         cl_mem stagebuffer;
1009         cl_mem nodebuffer;
1010         cl_mem candidatebuffer;
1011         cl_mem scaleinfobuffer;
1012         cl_mem pbuffer;
1013         cl_mem correctionbuffer;
1014         for( n_factors = 0, factor = 1;
1015                 cvRound(factor * winsize0.width) < gimg.cols - 10 &&
1016                 cvRound(factor * winsize0.height) < gimg.rows - 10;
1017                 n_factors++, factor *= scaleFactor )
1018         {
1019             CvSize winSize = { cvRound( winsize0.width * factor ),
1020                                cvRound( winsize0.height * factor )
1021                              };
1022             if( winSize.width < minSize.width || winSize.height < minSize.height )
1023             {
1024                 continue;
1025             }
1026             sizev.push_back(winSize);
1027             scalev.push_back(factor);
1028         }
1029         int loopcount = scalev.size();
1030         if(loopcount == 0)
1031         {
1032             loopcount = 1;
1033             n_factors = 1;
1034             sizev.push_back(minSize);
1035             scalev.push_back( std::min(cvRound(minSize.width / winsize0.width), cvRound(minSize.height / winsize0.height)) );
1036
1037         }
1038         detect_piramid_info *scaleinfo = (detect_piramid_info *)malloc(sizeof(detect_piramid_info) * loopcount);
1039         cl_int4 *p = (cl_int4 *)malloc(sizeof(cl_int4) * loopcount);
1040         float *correction = (float *)malloc(sizeof(float) * loopcount);
1041         int grp_per_CU = 12;
1042         size_t blocksize = 8;
1043         size_t localThreads[3] = { blocksize, blocksize , 1 };
1044         size_t globalThreads[3] = { grp_per_CU *gsum.clCxt->getDeviceInfo().maxComputeUnits *localThreads[0],
1045                                     localThreads[1], 1 };
1046         int outputsz = 256 * globalThreads[0] / localThreads[0];
1047         int nodenum = (datasize - sizeof(GpuHidHaarClassifierCascade) -
1048                        sizeof(GpuHidHaarStageClassifier) * gcascade->count - sizeof(GpuHidHaarClassifier) * totalclassifier) / sizeof(GpuHidHaarTreeNode);
1049         nodebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY,
1050                                         nodenum * sizeof(GpuHidHaarTreeNode));
1051         openCLSafeCall(clEnqueueWriteBuffer(qu, nodebuffer, 1, 0,
1052                                             nodenum * sizeof(GpuHidHaarTreeNode),
1053                                             node, 0, NULL, NULL));
1054         cl_mem newnodebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_WRITE,
1055                                loopcount * nodenum * sizeof(GpuHidHaarTreeNode));
1056         int startstage = 0;
1057         int endstage = gcascade->count;
1058         for(int i = 0; i < loopcount; i++)
1059         {
1060             sz = sizev[i];
1061             factor = scalev[i];
1062             double ystep = std::max(2., factor);
1063             int equRect_x = cvRound(factor * gcascade->p0);
1064             int equRect_y = cvRound(factor * gcascade->p1);
1065             int equRect_w = cvRound(factor * gcascade->p3);
1066             int equRect_h = cvRound(factor * gcascade->p2);
1067             p[i].s[0] = equRect_x;
1068             p[i].s[1] = equRect_y;
1069             p[i].s[2] = equRect_x + equRect_w;
1070             p[i].s[3] = equRect_y + equRect_h;
1071             correction[i] = 1. / (equRect_w * equRect_h);
1072             int width = (gsum.cols - 1 - sz.width  + ystep - 1) / ystep;
1073             int height = (gsum.rows - 1 - sz.height + ystep - 1) / ystep;
1074             int grpnumperline = (width + localThreads[0] - 1) / localThreads[0];
1075             int totalgrp = ((height + localThreads[1] - 1) / localThreads[1]) * grpnumperline;
1076
1077             scaleinfo[i].width_height = (width << 16) | height;
1078             scaleinfo[i].grpnumperline_totalgrp = (grpnumperline << 16) | totalgrp;
1079             scaleinfo[i].imgoff = 0;
1080             scaleinfo[i].factor = factor;
1081             int startnodenum = nodenum * i;
1082             float factor2 = (float)factor;
1083
1084             vector<pair<size_t, const void *> > args1;
1085             args1.push_back ( make_pair(sizeof(cl_mem) , (void *)&nodebuffer ));
1086             args1.push_back ( make_pair(sizeof(cl_mem) , (void *)&newnodebuffer ));
1087             args1.push_back ( make_pair(sizeof(cl_float) , (void *)&factor2 ));
1088             args1.push_back ( make_pair(sizeof(cl_float) , (void *)&correction[i] ));
1089             args1.push_back ( make_pair(sizeof(cl_int) , (void *)&startnodenum ));
1090
1091             size_t globalThreads2[3] = {nodenum, 1, 1};
1092             openCLExecuteKernel(gsum.clCxt, &haarobjectdetect_scaled2, "gpuscaleclassifier", globalThreads2, NULL/*localThreads2*/, args1, -1, -1);
1093         }
1094
1095         int step = gsum.step / 4;
1096         int startnode = 0;
1097         int splitstage = 3;
1098         stagebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(GpuHidHaarStageClassifier) * gcascade->count);
1099         openCLSafeCall(clEnqueueWriteBuffer(qu, stagebuffer, 1, 0, sizeof(GpuHidHaarStageClassifier)*gcascade->count, stage, 0, NULL, NULL));
1100         candidatebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_WRITE_ONLY | CL_MEM_ALLOC_HOST_PTR, 4 * sizeof(int) * outputsz);
1101         scaleinfobuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(detect_piramid_info) * loopcount);
1102         openCLSafeCall(clEnqueueWriteBuffer(qu, scaleinfobuffer, 1, 0, sizeof(detect_piramid_info)*loopcount, scaleinfo, 0, NULL, NULL));
1103         pbuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(cl_int4) * loopcount);
1104         openCLSafeCall(clEnqueueWriteBuffer(qu, pbuffer, 1, 0, sizeof(cl_int4)*loopcount, p, 0, NULL, NULL));
1105         correctionbuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(cl_float) * loopcount);
1106         openCLSafeCall(clEnqueueWriteBuffer(qu, correctionbuffer, 1, 0, sizeof(cl_float)*loopcount, correction, 0, NULL, NULL));
1107
1108         vector<pair<size_t, const void *> > args;
1109         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&stagebuffer ));
1110         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&scaleinfobuffer ));
1111         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&newnodebuffer ));
1112         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsum.data ));
1113         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsqsum.data ));
1114         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&candidatebuffer ));
1115         args.push_back ( make_pair(sizeof(cl_int) , (void *)&gsum.rows ));
1116         args.push_back ( make_pair(sizeof(cl_int) , (void *)&gsum.cols ));
1117         args.push_back ( make_pair(sizeof(cl_int) , (void *)&step ));
1118         args.push_back ( make_pair(sizeof(cl_int) , (void *)&loopcount ));
1119         args.push_back ( make_pair(sizeof(cl_int) , (void *)&startstage ));
1120         args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitstage ));
1121         args.push_back ( make_pair(sizeof(cl_int) , (void *)&endstage ));
1122         args.push_back ( make_pair(sizeof(cl_int) , (void *)&startnode ));
1123         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&pbuffer ));
1124         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&correctionbuffer ));
1125         args.push_back ( make_pair(sizeof(cl_int) , (void *)&nodenum ));
1126         const char * build_options = gcascade->is_stump_based ? "-D STUMP_BASED=1" : "-D STUMP_BASED=0";
1127         openCLExecuteKernel(gsum.clCxt, &haarobjectdetect_scaled2, "gpuRunHaarClassifierCascade_scaled2", globalThreads, localThreads, args, -1, -1, build_options);
1128
1129         candidate = (int *)clEnqueueMapBuffer(qu, candidatebuffer, 1, CL_MAP_READ, 0, 4 * sizeof(int) * outputsz, 0, 0, 0, &status);
1130
1131         for(int i = 0; i < outputsz; i++)
1132         {
1133             if(candidate[4 * i + 2] != 0)
1134                 allCandidates.push_back(Rect(candidate[4 * i], candidate[4 * i + 1], candidate[4 * i + 2], candidate[4 * i + 3]));
1135         }
1136
1137         free(scaleinfo);
1138         free(p);
1139         free(correction);
1140         clEnqueueUnmapMemObject(qu, candidatebuffer, candidate, 0, 0, 0);
1141         openCLSafeCall(clReleaseMemObject(stagebuffer));
1142         openCLSafeCall(clReleaseMemObject(scaleinfobuffer));
1143         openCLSafeCall(clReleaseMemObject(nodebuffer));
1144         openCLSafeCall(clReleaseMemObject(newnodebuffer));
1145         openCLSafeCall(clReleaseMemObject(candidatebuffer));
1146         openCLSafeCall(clReleaseMemObject(pbuffer));
1147         openCLSafeCall(clReleaseMemObject(correctionbuffer));
1148     }
1149
1150     cvFree(&cascade->hid_cascade);
1151     rectList.resize(allCandidates.size());
1152     if(!allCandidates.empty())
1153         std::copy(allCandidates.begin(), allCandidates.end(), rectList.begin());
1154
1155     if( minNeighbors != 0 || findBiggestObject )
1156         groupRectangles(rectList, rweights, std::max(minNeighbors, 1), GROUP_EPS);
1157     else
1158         rweights.resize(rectList.size(), 0);
1159
1160     if( findBiggestObject && rectList.size() )
1161     {
1162         CvAvgComp result_comp = {{0, 0, 0, 0}, 0};
1163
1164         for( size_t i = 0; i < rectList.size(); i++ )
1165         {
1166             cv::Rect r = rectList[i];
1167             if( r.area() > cv::Rect(result_comp.rect).area() )
1168             {
1169                 result_comp.rect = r;
1170                 result_comp.neighbors = rweights[i];
1171             }
1172         }
1173         cvSeqPush( result_seq, &result_comp );
1174     }
1175     else
1176     {
1177         for( size_t i = 0; i < rectList.size(); i++ )
1178         {
1179             CvAvgComp c;
1180             c.rect = rectList[i];
1181             c.neighbors = rweights[i];
1182             cvSeqPush( result_seq, &c );
1183         }
1184     }
1185
1186     return result_seq;
1187 }
1188
1189
1190 struct getRect
1191 {
1192     Rect operator()(const CvAvgComp &e) const
1193     {
1194         return e.rect;
1195     }
1196 };
1197
1198 void cv::ocl::OclCascadeClassifier::detectMultiScale(oclMat &gimg, CV_OUT std::vector<cv::Rect>& faces,
1199                                                         double scaleFactor, int minNeighbors, int flags,
1200                                                         Size minSize, Size maxSize)
1201 {
1202     CvSeq* _objects;
1203     MemStorage storage(cvCreateMemStorage(0));
1204     _objects = oclHaarDetectObjects(gimg, storage, scaleFactor, minNeighbors, flags, minSize, maxSize);
1205     vector<CvAvgComp> vecAvgComp;
1206     Seq<CvAvgComp>(_objects).copyTo(vecAvgComp);
1207     faces.resize(vecAvgComp.size());
1208     std::transform(vecAvgComp.begin(), vecAvgComp.end(), faces.begin(), getRect());
1209 }
1210
1211 struct OclBuffers
1212 {
1213     cl_mem stagebuffer;
1214     cl_mem nodebuffer;
1215     cl_mem candidatebuffer;
1216     cl_mem scaleinfobuffer;
1217     cl_mem pbuffer;
1218     cl_mem correctionbuffer;
1219     cl_mem newnodebuffer;
1220 };
1221
1222
1223 void cv::ocl::OclCascadeClassifierBuf::detectMultiScale(oclMat &gimg, CV_OUT std::vector<cv::Rect>& faces,
1224                                                         double scaleFactor, int minNeighbors, int flags,
1225                                                         Size minSize, Size maxSize)
1226 {
1227     int blocksize = 8;
1228     int grp_per_CU = 12;
1229     size_t localThreads[3] = { blocksize, blocksize, 1 };
1230     size_t globalThreads[3] = { grp_per_CU * cv::ocl::Context::getContext()->getDeviceInfo().maxComputeUnits *localThreads[0],
1231         localThreads[1],
1232         1 };
1233     int outputsz = 256 * globalThreads[0] / localThreads[0];
1234
1235     Init(gimg.rows, gimg.cols, scaleFactor, flags, outputsz, localThreads, minSize, maxSize);
1236
1237     const double GROUP_EPS = 0.2;
1238
1239     cv::ConcurrentRectVector allCandidates;
1240     std::vector<cv::Rect> rectList;
1241     std::vector<int> rweights;
1242
1243     CvHaarClassifierCascade      *cascade = oldCascade;
1244     GpuHidHaarClassifierCascade  *gcascade;
1245     GpuHidHaarStageClassifier    *stage;
1246
1247     if( CV_MAT_DEPTH(gimg.type()) != CV_8U )
1248         CV_Error( CV_StsUnsupportedFormat, "Only 8-bit images are supported" );
1249
1250     if( CV_MAT_CN(gimg.type()) > 1 )
1251     {
1252         oclMat gtemp;
1253         cvtColor( gimg, gtemp, CV_BGR2GRAY );
1254         gimg = gtemp;
1255     }
1256
1257     int *candidate;
1258     cl_command_queue qu = getClCommandQueue(Context::getContext());
1259     if( (flags & CV_HAAR_SCALE_IMAGE) )
1260     {
1261         int indexy = 0;
1262         CvSize sz;
1263
1264         cv::Rect roi, roi2;
1265         cv::ocl::oclMat resizeroi, gimgroi, gimgroisq;
1266
1267         for( int i = 0; i < m_loopcount; i++ )
1268         {
1269             sz = sizev[i];
1270             roi = Rect(0, indexy, sz.width, sz.height);
1271             roi2 = Rect(0, 0, sz.width - 1, sz.height - 1);
1272             resizeroi = gimg1(roi2);
1273             gimgroi = gsum(roi);
1274             gimgroisq = gsqsum(roi);
1275
1276             cv::ocl::resize(gimg, resizeroi, Size(sz.width - 1, sz.height - 1), 0, 0, INTER_LINEAR);
1277             cv::ocl::integral(resizeroi, gimgroi, gimgroisq);
1278             indexy += sz.height;
1279         }
1280
1281         gcascade   = (GpuHidHaarClassifierCascade *)(cascade->hid_cascade);
1282         stage      = (GpuHidHaarStageClassifier *)(gcascade + 1);
1283
1284         int startstage = 0;
1285         int endstage = gcascade->count;
1286         int startnode = 0;
1287         int pixelstep = gsum.step / 4;
1288         int splitstage = 3;
1289         int splitnode = stage[0].count + stage[1].count + stage[2].count;
1290         cl_int4 p, pq;
1291         p.s[0] = gcascade->p0;
1292         p.s[1] = gcascade->p1;
1293         p.s[2] = gcascade->p2;
1294         p.s[3] = gcascade->p3;
1295         pq.s[0] = gcascade->pq0;
1296         pq.s[1] = gcascade->pq1;
1297         pq.s[2] = gcascade->pq2;
1298         pq.s[3] = gcascade->pq3;
1299         float correction = gcascade->inv_window_area;
1300
1301         vector<pair<size_t, const void *> > args;
1302         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->stagebuffer ));
1303         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->scaleinfobuffer ));
1304         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->nodebuffer ));
1305         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsum.data ));
1306         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsqsum.data ));
1307         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->candidatebuffer ));
1308         args.push_back ( make_pair(sizeof(cl_int) , (void *)&pixelstep ));
1309         args.push_back ( make_pair(sizeof(cl_int) , (void *)&m_loopcount ));
1310         args.push_back ( make_pair(sizeof(cl_int) , (void *)&startstage ));
1311         args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitstage ));
1312         args.push_back ( make_pair(sizeof(cl_int) , (void *)&endstage ));
1313         args.push_back ( make_pair(sizeof(cl_int) , (void *)&startnode ));
1314         args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitnode ));
1315         args.push_back ( make_pair(sizeof(cl_int4) , (void *)&p ));
1316         args.push_back ( make_pair(sizeof(cl_int4) , (void *)&pq ));
1317         args.push_back ( make_pair(sizeof(cl_float) , (void *)&correction ));
1318
1319         const char * build_options = gcascade->is_stump_based ? "-D STUMP_BASED=1" : "-D STUMP_BASED=0";
1320
1321         openCLExecuteKernel(gsum.clCxt, &haarobjectdetect, "gpuRunHaarClassifierCascade", globalThreads, localThreads, args, -1, -1, build_options);
1322
1323         candidate = (int *)malloc(4 * sizeof(int) * outputsz);
1324         memset(candidate, 0, 4 * sizeof(int) * outputsz);
1325
1326         openCLReadBuffer( gsum.clCxt, ((OclBuffers *)buffers)->candidatebuffer, candidate, 4 * sizeof(int)*outputsz );
1327
1328         for(int i = 0; i < outputsz; i++)
1329         {
1330             if(candidate[4 * i + 2] != 0)
1331             {
1332                 allCandidates.push_back(Rect(candidate[4 * i], candidate[4 * i + 1],
1333                 candidate[4 * i + 2], candidate[4 * i + 3]));
1334             }
1335         }
1336         free((void *)candidate);
1337         candidate = NULL;
1338     }
1339     else
1340     {
1341         cv::ocl::integral(gimg, gsum, gsqsum);
1342
1343         gcascade   = (GpuHidHaarClassifierCascade *)cascade->hid_cascade;
1344
1345         int step = gsum.step / 4;
1346         int startnode = 0;
1347         int splitstage = 3;
1348
1349         int startstage = 0;
1350         int endstage = gcascade->count;
1351
1352         vector<pair<size_t, const void *> > args;
1353         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->stagebuffer ));
1354         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->scaleinfobuffer ));
1355         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->newnodebuffer ));
1356         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsum.data ));
1357         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&gsqsum.data ));
1358         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->candidatebuffer ));
1359         args.push_back ( make_pair(sizeof(cl_int) , (void *)&gsum.rows ));
1360         args.push_back ( make_pair(sizeof(cl_int) , (void *)&gsum.cols ));
1361         args.push_back ( make_pair(sizeof(cl_int) , (void *)&step ));
1362         args.push_back ( make_pair(sizeof(cl_int) , (void *)&m_loopcount ));
1363         args.push_back ( make_pair(sizeof(cl_int) , (void *)&startstage ));
1364         args.push_back ( make_pair(sizeof(cl_int) , (void *)&splitstage ));
1365         args.push_back ( make_pair(sizeof(cl_int) , (void *)&endstage ));
1366         args.push_back ( make_pair(sizeof(cl_int) , (void *)&startnode ));
1367         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->pbuffer ));
1368         args.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->correctionbuffer ));
1369         args.push_back ( make_pair(sizeof(cl_int) , (void *)&m_nodenum ));
1370
1371         const char * build_options = gcascade->is_stump_based ? "-D STUMP_BASED=1" : "-D STUMP_BASED=0";
1372         openCLExecuteKernel(gsum.clCxt, &haarobjectdetect_scaled2, "gpuRunHaarClassifierCascade_scaled2", globalThreads, localThreads, args, -1, -1, build_options);
1373
1374         candidate = (int *)clEnqueueMapBuffer(qu, ((OclBuffers *)buffers)->candidatebuffer, 1, CL_MAP_READ, 0, 4 * sizeof(int) * outputsz, 0, 0, 0, NULL);
1375
1376         for(int i = 0; i < outputsz; i++)
1377         {
1378             if(candidate[4 * i + 2] != 0)
1379                 allCandidates.push_back(Rect(candidate[4 * i], candidate[4 * i + 1],
1380                 candidate[4 * i + 2], candidate[4 * i + 3]));
1381         }
1382         clEnqueueUnmapMemObject(qu, ((OclBuffers *)buffers)->candidatebuffer, candidate, 0, 0, 0);
1383     }
1384     rectList.resize(allCandidates.size());
1385     if(!allCandidates.empty())
1386         std::copy(allCandidates.begin(), allCandidates.end(), rectList.begin());
1387
1388     if( minNeighbors != 0 || findBiggestObject )
1389         groupRectangles(rectList, rweights, std::max(minNeighbors, 1), GROUP_EPS);
1390     else
1391         rweights.resize(rectList.size(), 0);
1392
1393     GenResult(faces, rectList, rweights);
1394 }
1395
1396 void cv::ocl::OclCascadeClassifierBuf::Init(const int rows, const int cols,
1397     double scaleFactor, int flags,
1398     const int outputsz, const size_t localThreads[],
1399     CvSize minSize, CvSize maxSize)
1400 {
1401     if(initialized)
1402     {
1403         return; // we only allow one time initialization
1404     }
1405     CvHaarClassifierCascade      *cascade = oldCascade;
1406
1407     if( !CV_IS_HAAR_CLASSIFIER(cascade) )
1408         CV_Error( !cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier cascade" );
1409
1410     if( scaleFactor <= 1 )
1411         CV_Error( CV_StsOutOfRange, "scale factor must be > 1" );
1412
1413     if( cols < minSize.width || rows < minSize.height )
1414         CV_Error(CV_StsError, "Image too small");
1415
1416     int datasize=0;
1417     int totalclassifier=0;
1418
1419     if( !cascade->hid_cascade )
1420     {
1421         gpuCreateHidHaarClassifierCascade(cascade, &datasize, &totalclassifier);
1422     }
1423
1424     if( maxSize.height == 0 || maxSize.width == 0 )
1425     {
1426         maxSize.height = rows;
1427         maxSize.width = cols;
1428     }
1429
1430     findBiggestObject = (flags & CV_HAAR_FIND_BIGGEST_OBJECT) != 0;
1431     if( findBiggestObject )
1432         flags &= ~(CV_HAAR_SCALE_IMAGE | CV_HAAR_DO_CANNY_PRUNING);
1433
1434     CreateBaseBufs(datasize, totalclassifier, flags, outputsz);
1435     CreateFactorRelatedBufs(rows, cols, flags, scaleFactor, localThreads, minSize, maxSize);
1436
1437     m_scaleFactor = scaleFactor;
1438     m_rows = rows;
1439     m_cols = cols;
1440     m_flags = flags;
1441     m_minSize = minSize;
1442     m_maxSize = maxSize;
1443
1444     // initialize nodes
1445     GpuHidHaarClassifierCascade  *gcascade;
1446     GpuHidHaarStageClassifier    *stage;
1447     GpuHidHaarClassifier         *classifier;
1448     GpuHidHaarTreeNode           *node;
1449     cl_command_queue qu = getClCommandQueue(Context::getContext());
1450     if( (flags & CV_HAAR_SCALE_IMAGE) )
1451     {
1452         gcascade   = (GpuHidHaarClassifierCascade *)(cascade->hid_cascade);
1453         stage      = (GpuHidHaarStageClassifier *)(gcascade + 1);
1454         classifier = (GpuHidHaarClassifier *)(stage + gcascade->count);
1455         node       = (GpuHidHaarTreeNode *)(classifier->node);
1456
1457         gpuSetImagesForHaarClassifierCascade( cascade, 1., gsum.step / 4 );
1458
1459         openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->stagebuffer, 1, 0,
1460             sizeof(GpuHidHaarStageClassifier) * gcascade->count,
1461             stage, 0, NULL, NULL));
1462
1463         openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->nodebuffer, 1, 0,
1464                                             m_nodenum * sizeof(GpuHidHaarTreeNode),
1465                                             node, 0, NULL, NULL));
1466     }
1467     else
1468     {
1469         gpuSetHaarClassifierCascade(cascade);
1470
1471         gcascade   = (GpuHidHaarClassifierCascade *)cascade->hid_cascade;
1472         stage      = (GpuHidHaarStageClassifier *)(gcascade + 1);
1473         classifier = (GpuHidHaarClassifier *)(stage + gcascade->count);
1474         node       = (GpuHidHaarTreeNode *)(classifier->node);
1475
1476         openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->nodebuffer, 1, 0,
1477             m_nodenum * sizeof(GpuHidHaarTreeNode),
1478             node, 0, NULL, NULL));
1479
1480         cl_int4 *p = (cl_int4 *)malloc(sizeof(cl_int4) * m_loopcount);
1481         float *correction = (float *)malloc(sizeof(float) * m_loopcount);
1482         double factor;
1483         for(int i = 0; i < m_loopcount; i++)
1484         {
1485             factor = scalev[i];
1486             int equRect_x = (int)(factor * gcascade->p0 + 0.5);
1487             int equRect_y = (int)(factor * gcascade->p1 + 0.5);
1488             int equRect_w = (int)(factor * gcascade->p3 + 0.5);
1489             int equRect_h = (int)(factor * gcascade->p2 + 0.5);
1490             p[i].s[0] = equRect_x;
1491             p[i].s[1] = equRect_y;
1492             p[i].s[2] = equRect_x + equRect_w;
1493             p[i].s[3] = equRect_y + equRect_h;
1494             correction[i] = 1. / (equRect_w * equRect_h);
1495             int startnodenum = m_nodenum * i;
1496             float factor2 = (float)factor;
1497
1498             vector<pair<size_t, const void *> > args1;
1499             args1.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->nodebuffer ));
1500             args1.push_back ( make_pair(sizeof(cl_mem) , (void *)&((OclBuffers *)buffers)->newnodebuffer ));
1501             args1.push_back ( make_pair(sizeof(cl_float) , (void *)&factor2 ));
1502             args1.push_back ( make_pair(sizeof(cl_float) , (void *)&correction[i] ));
1503             args1.push_back ( make_pair(sizeof(cl_int) , (void *)&startnodenum ));
1504
1505             size_t globalThreads2[3] = {m_nodenum, 1, 1};
1506
1507             openCLExecuteKernel(Context::getContext(), &haarobjectdetect_scaled2, "gpuscaleclassifier", globalThreads2, NULL/*localThreads2*/, args1, -1, -1);
1508         }
1509         openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->stagebuffer, 1, 0, sizeof(GpuHidHaarStageClassifier)*gcascade->count, stage, 0, NULL, NULL));
1510         openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->pbuffer, 1, 0, sizeof(cl_int4)*m_loopcount, p, 0, NULL, NULL));
1511         openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->correctionbuffer, 1, 0, sizeof(cl_float)*m_loopcount, correction, 0, NULL, NULL));
1512
1513         free(p);
1514         free(correction);
1515     }
1516     initialized = true;
1517 }
1518
1519 void cv::ocl::OclCascadeClassifierBuf::CreateBaseBufs(const int datasize, const int totalclassifier,
1520                                                       const int flags, const int outputsz)
1521 {
1522     if (!initialized)
1523     {
1524         buffers = malloc(sizeof(OclBuffers));
1525
1526         size_t tempSize =
1527             sizeof(GpuHidHaarStageClassifier) * ((GpuHidHaarClassifierCascade *)oldCascade->hid_cascade)->count;
1528         m_nodenum = (datasize - sizeof(GpuHidHaarClassifierCascade) - tempSize - sizeof(GpuHidHaarClassifier) * totalclassifier)
1529             / sizeof(GpuHidHaarTreeNode);
1530
1531         ((OclBuffers *)buffers)->stagebuffer     = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_ONLY,  tempSize);
1532         ((OclBuffers *)buffers)->nodebuffer      = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_ONLY,  m_nodenum * sizeof(GpuHidHaarTreeNode));
1533     }
1534
1535     if (initialized
1536         && ((m_flags & CV_HAAR_SCALE_IMAGE) ^ (flags & CV_HAAR_SCALE_IMAGE)))
1537     {
1538         openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->candidatebuffer));
1539     }
1540
1541     if (flags & CV_HAAR_SCALE_IMAGE)
1542     {
1543         ((OclBuffers *)buffers)->candidatebuffer = openCLCreateBuffer(cv::ocl::Context::getContext(),
1544                                                         CL_MEM_WRITE_ONLY,
1545                                                         4 * sizeof(int) * outputsz);
1546     }
1547     else
1548     {
1549         ((OclBuffers *)buffers)->candidatebuffer = openCLCreateBuffer(cv::ocl::Context::getContext(),
1550                                                         CL_MEM_WRITE_ONLY | CL_MEM_ALLOC_HOST_PTR,
1551                                                         4 * sizeof(int) * outputsz);
1552     }
1553 }
1554
1555 void cv::ocl::OclCascadeClassifierBuf::CreateFactorRelatedBufs(
1556     const int rows, const int cols, const int flags,
1557     const double scaleFactor, const size_t localThreads[],
1558     CvSize minSize, CvSize maxSize)
1559 {
1560     if (initialized)
1561     {
1562         if ((m_flags & CV_HAAR_SCALE_IMAGE) && !(flags & CV_HAAR_SCALE_IMAGE))
1563         {
1564             gimg1.release();
1565             gsum.release();
1566             gsqsum.release();
1567         }
1568         else if (!(m_flags & CV_HAAR_SCALE_IMAGE) && (flags & CV_HAAR_SCALE_IMAGE))
1569         {
1570             openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->newnodebuffer));
1571             openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->correctionbuffer));
1572             openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->pbuffer));
1573         }
1574         else if ((m_flags & CV_HAAR_SCALE_IMAGE) && (flags & CV_HAAR_SCALE_IMAGE))
1575         {
1576             if (fabs(m_scaleFactor - scaleFactor) < 1e-6
1577                 && (rows == m_rows && cols == m_cols)
1578                 && (minSize.width == m_minSize.width)
1579                 && (minSize.height == m_minSize.height)
1580                 && (maxSize.width == m_maxSize.width)
1581                 && (maxSize.height == m_maxSize.height))
1582             {
1583                 return;
1584             }
1585         }
1586         else
1587         {
1588             if (fabs(m_scaleFactor - scaleFactor) < 1e-6
1589                 && (rows == m_rows && cols == m_cols)
1590                 && (minSize.width == m_minSize.width)
1591                 && (minSize.height == m_minSize.height)
1592                 && (maxSize.width == m_maxSize.width)
1593                 && (maxSize.height == m_maxSize.height))
1594             {
1595                 return;
1596             }
1597             else
1598             {
1599                 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->newnodebuffer));
1600                 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->correctionbuffer));
1601                 openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->pbuffer));
1602             }
1603         }
1604     }
1605
1606     int loopcount;
1607     int indexy = 0;
1608     int totalheight = 0;
1609     double factor;
1610     Rect roi;
1611     CvSize sz;
1612     CvSize winSize0 = oldCascade->orig_window_size;
1613     detect_piramid_info *scaleinfo;
1614     cl_command_queue qu = getClCommandQueue(Context::getContext());
1615     if (flags & CV_HAAR_SCALE_IMAGE)
1616     {
1617         for(factor = 1.f;; factor *= scaleFactor)
1618         {
1619             CvSize winSize = { cvRound(winSize0.width * factor), cvRound(winSize0.height * factor) };
1620             sz.width     = cvRound( cols / factor ) + 1;
1621             sz.height    = cvRound( rows / factor ) + 1;
1622             CvSize sz1     = { sz.width - winSize0.width - 1,      sz.height - winSize0.height - 1 };
1623
1624             if( sz1.width <= 0 || sz1.height <= 0 )
1625                 break;
1626             if( winSize.width > maxSize.width || winSize.height > maxSize.height )
1627                 break;
1628             if( winSize.width < minSize.width || winSize.height < minSize.height )
1629                 continue;
1630
1631             totalheight += sz.height;
1632             sizev.push_back(sz);
1633             scalev.push_back(static_cast<float>(factor));
1634         }
1635
1636         loopcount = sizev.size();
1637         gimg1.create(rows, cols, CV_8UC1);
1638         gsum.create(totalheight + 4, cols + 1, CV_32SC1);
1639         gsqsum.create(totalheight + 4, cols + 1, CV_32FC1);
1640
1641         scaleinfo = (detect_piramid_info *)malloc(sizeof(detect_piramid_info) * loopcount);
1642         for( int i = 0; i < loopcount; i++ )
1643         {
1644             sz = sizev[i];
1645             roi = Rect(0, indexy, sz.width, sz.height);
1646             int width = sz.width - 1 - oldCascade->orig_window_size.width;
1647             int height = sz.height - 1 - oldCascade->orig_window_size.height;
1648             int grpnumperline = (width + localThreads[0] - 1) / localThreads[0];
1649             int totalgrp = ((height + localThreads[1] - 1) / localThreads[1]) * grpnumperline;
1650
1651             ((detect_piramid_info *)scaleinfo)[i].width_height = (width << 16) | height;
1652             ((detect_piramid_info *)scaleinfo)[i].grpnumperline_totalgrp = (grpnumperline << 16) | totalgrp;
1653             ((detect_piramid_info *)scaleinfo)[i].imgoff = gsum(roi).offset >> 2;
1654             ((detect_piramid_info *)scaleinfo)[i].factor = scalev[i];
1655
1656             indexy += sz.height;
1657         }
1658     }
1659     else
1660     {
1661         for(factor = 1;
1662             cvRound(factor * winSize0.width) < cols - 10 && cvRound(factor * winSize0.height) < rows - 10;
1663             factor *= scaleFactor)
1664         {
1665             CvSize winSize = { cvRound( winSize0.width * factor ), cvRound( winSize0.height * factor ) };
1666             if( winSize.width < minSize.width || winSize.height < minSize.height )
1667             {
1668                 continue;
1669             }
1670             sizev.push_back(winSize);
1671             scalev.push_back(factor);
1672         }
1673
1674         loopcount = scalev.size();
1675         if(loopcount == 0)
1676         {
1677             loopcount = 1;
1678             sizev.push_back(minSize);
1679             scalev.push_back( std::min(cvRound(minSize.width / winSize0.width), cvRound(minSize.height / winSize0.height)) );
1680         }
1681
1682         ((OclBuffers *)buffers)->pbuffer = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_ONLY,
1683             sizeof(cl_int4) * loopcount);
1684         ((OclBuffers *)buffers)->correctionbuffer = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_ONLY,
1685             sizeof(cl_float) * loopcount);
1686         ((OclBuffers *)buffers)->newnodebuffer = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_WRITE,
1687             loopcount * m_nodenum * sizeof(GpuHidHaarTreeNode));
1688
1689         scaleinfo = (detect_piramid_info *)malloc(sizeof(detect_piramid_info) * loopcount);
1690         for( int i = 0; i < loopcount; i++ )
1691         {
1692             sz = sizev[i];
1693             factor = scalev[i];
1694             double ystep = cv::max(2.,factor);
1695             int width = cvRound((cols - 1 - sz.width  + ystep - 1) / ystep);
1696             int height = cvRound((rows - 1 - sz.height + ystep - 1) / ystep);
1697             int grpnumperline = (width + localThreads[0] - 1) / localThreads[0];
1698             int totalgrp = ((height + localThreads[1] - 1) / localThreads[1]) * grpnumperline;
1699
1700             ((detect_piramid_info *)scaleinfo)[i].width_height = (width << 16) | height;
1701             ((detect_piramid_info *)scaleinfo)[i].grpnumperline_totalgrp = (grpnumperline << 16) | totalgrp;
1702             ((detect_piramid_info *)scaleinfo)[i].imgoff = 0;
1703             ((detect_piramid_info *)scaleinfo)[i].factor = factor;
1704         }
1705     }
1706
1707     if (loopcount != m_loopcount)
1708     {
1709         if (initialized)
1710         {
1711             openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->scaleinfobuffer));
1712         }
1713         ((OclBuffers *)buffers)->scaleinfobuffer = openCLCreateBuffer(cv::ocl::Context::getContext(), CL_MEM_READ_ONLY, sizeof(detect_piramid_info) * loopcount);
1714     }
1715
1716     openCLSafeCall(clEnqueueWriteBuffer(qu, ((OclBuffers *)buffers)->scaleinfobuffer, 1, 0,
1717         sizeof(detect_piramid_info)*loopcount,
1718         scaleinfo, 0, NULL, NULL));
1719     free(scaleinfo);
1720
1721     m_loopcount = loopcount;
1722 }
1723
1724 void cv::ocl::OclCascadeClassifierBuf::GenResult(CV_OUT std::vector<cv::Rect>& faces,
1725                                                  const std::vector<cv::Rect> &rectList,
1726                                                  const std::vector<int> &rweights)
1727 {
1728     MemStorage tempStorage(cvCreateMemStorage(0));
1729     CvSeq *result_seq = cvCreateSeq( 0, sizeof(CvSeq), sizeof(CvAvgComp), tempStorage );
1730
1731     if( findBiggestObject && rectList.size() )
1732     {
1733         CvAvgComp result_comp = {{0, 0, 0, 0}, 0};
1734
1735         for( size_t i = 0; i < rectList.size(); i++ )
1736         {
1737             cv::Rect r = rectList[i];
1738             if( r.area() > cv::Rect(result_comp.rect).area() )
1739             {
1740                 result_comp.rect = r;
1741                 result_comp.neighbors = rweights[i];
1742             }
1743         }
1744         cvSeqPush( result_seq, &result_comp );
1745     }
1746     else
1747     {
1748         for( size_t i = 0; i < rectList.size(); i++ )
1749         {
1750             CvAvgComp c;
1751             c.rect = rectList[i];
1752             c.neighbors = rweights[i];
1753             cvSeqPush( result_seq, &c );
1754         }
1755     }
1756
1757     vector<CvAvgComp> vecAvgComp;
1758     Seq<CvAvgComp>(result_seq).copyTo(vecAvgComp);
1759     faces.resize(vecAvgComp.size());
1760     std::transform(vecAvgComp.begin(), vecAvgComp.end(), faces.begin(), getRect());
1761 }
1762
1763 void cv::ocl::OclCascadeClassifierBuf::release()
1764 {
1765     if(initialized)
1766     {
1767         openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->stagebuffer));
1768         openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->scaleinfobuffer));
1769         openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->nodebuffer));
1770         openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->candidatebuffer));
1771
1772         if( (m_flags & CV_HAAR_SCALE_IMAGE) )
1773         {
1774             cvFree(&oldCascade->hid_cascade);
1775         }
1776         else
1777         {
1778             openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->newnodebuffer));
1779             openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->correctionbuffer));
1780             openCLSafeCall(clReleaseMemObject(((OclBuffers *)buffers)->pbuffer));
1781         }
1782
1783         free(buffers);
1784         buffers = NULL;
1785         initialized = false;
1786     }
1787 }
1788
1789 #ifndef _MAX_PATH
1790 #define _MAX_PATH 1024
1791 #endif