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 //
  24 // Redistribution and use in source and binary forms, with or without modification,
  25 // are permitted provided that the following conditions are met:
  26 //
  27 //   * Redistribution's of source code must retain the above copyright notice,
  28 //     this list of conditions and the following disclaimer.
  29 //
  30 //   * Redistribution's in binary form must reproduce the above copyright notice,
  31 //     this list of conditions and the following disclaimer in the documentation
  32 //     and/or other materials provided with the distribution.
  33 //
  34 //   * The name of the copyright holders may not be used to endorse or promote products
  35 //     derived from this software without specific prior written permission.
  36 //
  37 // This software is provided by the copyright holders and contributors "as is" and
  38 // any express or implied warranties, including, but not limited to, the implied
  39 // warranties of merchantability and fitness for a particular purpose are disclaimed.
  40 // In no event shall the Intel Corporation or contributors be liable for any direct,
  41 // indirect, incidental, special, exemplary, or consequential damages
  42 // (including, but not limited to, procurement of substitute goods or services;
  43 // loss of use, data, or profits; or business interruption) however caused
  44 // and on any theory of liability, whether in contract, strict liability,
  45 // or tort (including negligence or otherwise) arising in any way out of
  46 // the use of this software, even if advised of the possibility of such damage.
  47 //
  48 //M*/
  49
  50 #include "precomp.hpp"
  51 #include "opencl_kernels.hpp"
  52
  53 using namespace cv;
  54 using namespace cv::ocl;
  55
  56 /* these settings affect the quality of detection: change with care */
  57 #define CV_ADJUST_FEATURES  1
  58 #define CV_ADJUST_WEIGHTS   0
  59 #define CV_HAAR_FEATURE_MAX 3
  60 typedef int sumtype;
  61 typedef double sqsumtype;
  62
  63 typedef struct CvHidHaarFeature
  64 {
  65     struct
  66     {
  67         sumtype *p0, *p1, *p2, *p3;
  68         float weight;
  69     }
  70     rect[CV_HAAR_FEATURE_MAX];
  71 }
  72 CvHidHaarFeature;
  73
  74
  75 typedef struct CvHidHaarTreeNode
  76 {
  77     CvHidHaarFeature feature;
  78     float threshold;
  79     int left;
  80     int right;
  81 }
  82 CvHidHaarTreeNode;
  83
  84
  85 typedef struct CvHidHaarClassifier
  86 {
  87     int count;
  88     //CvHaarFeature* orig_feature;
  89     CvHidHaarTreeNode *node;
  90     float *alpha;
  91 }
  92 CvHidHaarClassifier;
  93
  94
  95 typedef struct CvHidHaarStageClassifier
  96 {
  97     int  count;
  98     float threshold;
  99     CvHidHaarClassifier *classifier;
 100     int two_rects;
 101
 102     struct CvHidHaarStageClassifier *next;
 103     struct CvHidHaarStageClassifier *child;
 104     struct CvHidHaarStageClassifier *parent;
 105 }
 106 CvHidHaarStageClassifier;
 107
 108
 109 struct CvHidHaarClassifierCascade
 110 {
 111     int  count;
 112     int  is_stump_based;
 113     int  has_tilted_features;
 114     int  is_tree;
 115     double inv_window_area;
 116     CvMat sum, sqsum, tilted;
 117     CvHidHaarStageClassifier *stage_classifier;
 118     sqsumtype *pq0, *pq1, *pq2, *pq3;
 119     sumtype *p0, *p1, *p2, *p3;
 120
 121     void **ipp_stages;
 122 };
 123 typedef struct
 124 {
 125     int width_height;
 126     int grpnumperline_totalgrp;
 127     int imgoff;
 128     float factor;
 129 } detect_piramid_info;
 130 #ifdef _MSC_VER
 131 #define _ALIGNED_ON(_ALIGNMENT) __declspec(align(_ALIGNMENT))
 132
 133 typedef _ALIGNED_ON(128) struct  GpuHidHaarTreeNode
 134 {
 135     _ALIGNED_ON(64) int p[CV_HAAR_FEATURE_MAX][4];
 136     float weight[CV_HAAR_FEATURE_MAX] ;
 137     float threshold ;
 138     _ALIGNED_ON(16) float alpha[3] ;
 139     _ALIGNED_ON(4) int left ;
 140     _ALIGNED_ON(4) int right ;
 141 }
 142 GpuHidHaarTreeNode;
 143
 144
 145 typedef  _ALIGNED_ON(32) struct  GpuHidHaarClassifier
 146 {
 147     _ALIGNED_ON(4) int count;
 148     _ALIGNED_ON(8) GpuHidHaarTreeNode *node ;
 149     _ALIGNED_ON(8) float *alpha ;
 150 }
 151 GpuHidHaarClassifier;
 152
 153
 154 typedef _ALIGNED_ON(64) struct   GpuHidHaarStageClassifier
 155 {
 156     _ALIGNED_ON(4) int  count ;
 157     _ALIGNED_ON(4) float threshold ;
 158     _ALIGNED_ON(4) int two_rects ;
 159     _ALIGNED_ON(8) GpuHidHaarClassifier *classifier ;
 160     _ALIGNED_ON(8) struct GpuHidHaarStageClassifier *next;
 161     _ALIGNED_ON(8) struct GpuHidHaarStageClassifier *child ;
 162     _ALIGNED_ON(8) struct GpuHidHaarStageClassifier *parent ;
 163 }
 164 GpuHidHaarStageClassifier;
 165
 166
 167 typedef _ALIGNED_ON(64) struct  GpuHidHaarClassifierCascade
 168 {
 169     _ALIGNED_ON(4) int  count ;
 170     _ALIGNED_ON(4) int  is_stump_based ;
 171     _ALIGNED_ON(4) int  has_tilted_features ;
 172     _ALIGNED_ON(4) int  is_tree ;
 173     _ALIGNED_ON(4) int pq0 ;
 174     _ALIGNED_ON(4) int pq1 ;
 175     _ALIGNED_ON(4) int pq2 ;
 176     _ALIGNED_ON(4) int pq3 ;
 177     _ALIGNED_ON(4) int p0 ;
 178     _ALIGNED_ON(4) int p1 ;
 179     _ALIGNED_ON(4) int p2 ;
 180     _ALIGNED_ON(4) int p3 ;
 181     _ALIGNED_ON(4) float inv_window_area ;
 182 } GpuHidHaarClassifierCascade;
 183 #else
 184 #define _ALIGNED_ON(_ALIGNMENT) __attribute__((aligned(_ALIGNMENT) ))
 185
 186 typedef struct _ALIGNED_ON(128) GpuHidHaarTreeNode
 187 {
 188     int p[CV_HAAR_FEATURE_MAX][4] _ALIGNED_ON(64);
 189     float weight[CV_HAAR_FEATURE_MAX];// _ALIGNED_ON(16);
 190     float threshold;// _ALIGNED_ON(4);
 191     float alpha[3] _ALIGNED_ON(16);
 192     int left _ALIGNED_ON(4);
 193     int right _ALIGNED_ON(4);
 194 }
 195 GpuHidHaarTreeNode;
 196
 197 typedef struct _ALIGNED_ON(32) GpuHidHaarClassifier
 198 {
 199     int count _ALIGNED_ON(4);
 200     GpuHidHaarTreeNode *node _ALIGNED_ON(8);
 201     float *alpha _ALIGNED_ON(8);
 202 }
 203 GpuHidHaarClassifier;
 204
 205
 206 typedef struct _ALIGNED_ON(64) GpuHidHaarStageClassifier
 207 {
 208     int  count _ALIGNED_ON(4);
 209     float threshold _ALIGNED_ON(4);
 210     int two_rects _ALIGNED_ON(4);
 211     GpuHidHaarClassifier *classifier _ALIGNED_ON(8);
 212     struct GpuHidHaarStageClassifier *next _ALIGNED_ON(8);
 213     struct GpuHidHaarStageClassifier *child _ALIGNED_ON(8);
 214     struct GpuHidHaarStageClassifier *parent _ALIGNED_ON(8);
 215 }
 216 GpuHidHaarStageClassifier;
 217
 218
 219 typedef struct _ALIGNED_ON(64) GpuHidHaarClassifierCascade
 220 {
 221     int  count _ALIGNED_ON(4);
 222     int  is_stump_based _ALIGNED_ON(4);
 223     int  has_tilted_features _ALIGNED_ON(4);
 224     int  is_tree _ALIGNED_ON(4);
 225     int pq0 _ALIGNED_ON(4);
 226     int pq1 _ALIGNED_ON(4);
 227     int pq2 _ALIGNED_ON(4);
 228     int pq3 _ALIGNED_ON(4);
 229     int p0 _ALIGNED_ON(4);
 230     int p1 _ALIGNED_ON(4);
 231     int p2 _ALIGNED_ON(4);
 232     int p3 _ALIGNED_ON(4);
 233     float inv_window_area _ALIGNED_ON(4);
 234 } GpuHidHaarClassifierCascade;
 235 #endif
 236
 237 const int icv_object_win_border = 1;
 238 const float icv_stage_threshold_bias = 0.0001f;
 239 double globaltime = 0;
 240
 241 /* create more efficient internal representation of haar classifier cascade */
 242 static GpuHidHaarClassifierCascade * gpuCreateHidHaarClassifierCascade( CvHaarClassifierCascade *cascade, int *size, int *totalclassifier)
 243 {
 244     GpuHidHaarClassifierCascade *out = 0;
 245
 246     int i, j, k, l;
 247     int datasize;
 248     int total_classifiers = 0;
 249     int total_nodes = 0;
 250     char errorstr[256];
 251
 252     GpuHidHaarStageClassifier *stage_classifier_ptr;
 253     GpuHidHaarClassifier *haar_classifier_ptr;
 254     GpuHidHaarTreeNode *haar_node_ptr;
 255
 256     CvSize orig_window_size;
 257     int has_tilted_features = 0;
 258
 259     if( !CV_IS_HAAR_CLASSIFIER(cascade) )
 260         CV_Error( !cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier pointer" );
 261
 262     if( cascade->hid_cascade )
 263         CV_Error( CV_StsError, "hid_cascade has been already created" );
 264
 265     if( !cascade->stage_classifier )
 266         CV_Error( CV_StsNullPtr, "" );
 267
 268     if( cascade->count <= 0 )
 269         CV_Error( CV_StsOutOfRange, "Negative number of cascade stages" );
 270
 271     orig_window_size = cascade->orig_window_size;
 272
 273     /* check input structure correctness and calculate total memory size needed for
 274     internal representation of the classifier cascade */
 275     for( i = 0; i < cascade->count; i++ )
 276     {
 277         CvHaarStageClassifier *stage_classifier = cascade->stage_classifier + i;
 278
 279         if( !stage_classifier->classifier ||
 280                 stage_classifier->count <= 0 )
 281         {
 282             sprintf( errorstr, "header of the stage classifier #%d is invalid "
 283                      "(has null pointers or non-positive classfier count)", i );
 284             CV_Error( CV_StsError, errorstr );
 285         }
 286
 287         total_classifiers += stage_classifier->count;
 288
 289         for( j = 0; j < stage_classifier->count; j++ )
 290         {
 291             CvHaarClassifier *classifier = stage_classifier->classifier + j;
 292
 293             total_nodes += classifier->count;
 294             for( l = 0; l < classifier->count; l++ )
 295             {
 296                 for( k = 0; k < CV_HAAR_FEATURE_MAX; k++ )
 297                 {
 298                     if( classifier->haar_feature[l].rect[k].r.width )
 299                     {
 300                         CvRect r = classifier->haar_feature[l].rect[k].r;
 301                         int tilted = classifier->haar_feature[l].tilted;
 302                         has_tilted_features |= tilted != 0;
 303                         if( r.width < 0 || r.height < 0 || r.y < 0 ||
 304                                 r.x + r.width > orig_window_size.width
 305                                 ||
 306                                 (!tilted &&
 307                                  (r.x < 0 || r.y + r.height > orig_window_size.height))
 308                                 ||
 309                                 (tilted && (r.x - r.height < 0 ||
 310                                             r.y + r.width + r.height > orig_window_size.height)))
 311                         {
 312                             sprintf( errorstr, "rectangle #%d of the classifier #%d of "
 313                                      "the stage classifier #%d is not inside "
 314                                      "the reference (original) cascade window", k, j, i );
 315                             CV_Error( CV_StsNullPtr, errorstr );
 316                         }
 317                     }
 318                 }
 319             }
 320         }
 321     }
 322
 323     // this is an upper boundary for the whole hidden cascade size
 324     datasize = sizeof(GpuHidHaarClassifierCascade)                   +
 325                sizeof(GpuHidHaarStageClassifier) * cascade->count    +
 326                sizeof(GpuHidHaarClassifier)      * total_classifiers +
 327                sizeof(GpuHidHaarTreeNode)        * total_nodes;
 328
 329     *totalclassifier = total_classifiers;
 330     *size = datasize;
 331     out = (GpuHidHaarClassifierCascade *)cvAlloc( datasize );
 332     memset( out, 0, sizeof(*out) );
 333
 334     /* init header */
 335     out->count = cascade->count;
 336     stage_classifier_ptr = (GpuHidHaarStageClassifier *)(out + 1);
 337     haar_classifier_ptr = (GpuHidHaarClassifier *)(stage_classifier_ptr + cascade->count);
 338     haar_node_ptr = (GpuHidHaarTreeNode *)(haar_classifier_ptr + total_classifiers);
 339
 340     out->is_stump_based = 1;
 341     out->has_tilted_features = has_tilted_features;
 342     out->is_tree = 0;
 343
 344     /* initialize internal representation */
 345     for( i = 0; i < cascade->count; i++ )
 346     {
 347         CvHaarStageClassifier *stage_classifier = cascade->stage_classifier + i;
 348         GpuHidHaarStageClassifier *hid_stage_classifier = stage_classifier_ptr + i;
 349
 350         hid_stage_classifier->count = stage_classifier->count;
 351         hid_stage_classifier->threshold = stage_classifier->threshold - icv_stage_threshold_bias;
 352         hid_stage_classifier->classifier = haar_classifier_ptr;
 353         hid_stage_classifier->two_rects = 1;
 354         haar_classifier_ptr += stage_classifier->count;
 355
 356         for( j = 0; j < stage_classifier->count; j++ )
 357         {
 358             CvHaarClassifier *classifier         = stage_classifier->classifier + j;
 359             GpuHidHaarClassifier *hid_classifier = hid_stage_classifier->classifier + j;
 360             int node_count = classifier->count;
 361
 362             float *alpha_ptr = &haar_node_ptr->alpha[0];
 363
 364             hid_classifier->count = node_count;
 365             hid_classifier->node = haar_node_ptr;
 366             hid_classifier->alpha = alpha_ptr;
 367
 368             for( l = 0; l < node_count; l++ )
 369             {
 370                 GpuHidHaarTreeNode *node     = hid_classifier->node + l;
 371                 CvHaarFeature      *feature = classifier->haar_feature + l;
 372
 373                 memset( node, -1, sizeof(*node) );
 374                 node->threshold = classifier->threshold[l];
 375                 node->left      = classifier->left[l];
 376                 node->right     = classifier->right[l];
 377
 378                 if( fabs(feature->rect[2].weight) < DBL_EPSILON ||
 379                         feature->rect[2].r.width == 0 ||
 380                         feature->rect[2].r.height == 0 )
 381                 {
 382                     node->p[2][0] = 0;
 383                     node->p[2][1] = 0;
 384                     node->p[2][2] = 0;
 385                     node->p[2][3] = 0;
 386                     node->weight[2] = 0;
 387                 }
 388                 else
 389                     hid_stage_classifier->two_rects = 0;
 390
 391                 memcpy( node->alpha, classifier->alpha, (node_count + 1)*sizeof(alpha_ptr[0]));
 392                 haar_node_ptr = haar_node_ptr + 1;
 393             }
 394             out->is_stump_based &= node_count == 1;
 395         }
 396     }
 397
 398     cascade->hid_cascade = (CvHidHaarClassifierCascade *)out;
 399     assert( (char *)haar_node_ptr - (char *)out <= datasize );
 400
 401     return out;
 402 }
 403
 404
 405 #define sum_elem_ptr(sum,row,col)  \
 406     ((sumtype*)CV_MAT_ELEM_PTR_FAST((sum),(row),(col),sizeof(sumtype)))
 407
 408 #define sqsum_elem_ptr(sqsum,row,col)  \
 409     ((sqsumtype*)CV_MAT_ELEM_PTR_FAST((sqsum),(row),(col),sizeof(sqsumtype)))
 410
 411 #define calc_sum(rect,offset) \
 412     ((rect).p0[offset] - (rect).p1[offset] - (rect).p2[offset] + (rect).p3[offset])
 413
 414
 415 static void gpuSetImagesForHaarClassifierCascade( CvHaarClassifierCascade *_cascade,
 416                                       double scale,
 417                                       int step)
 418 {
 419     GpuHidHaarClassifierCascade *cascade;
 420     int coi0 = 0, coi1 = 0;
 421     int i;
 422     int datasize;
 423     int total;
 424     CvRect equRect;
 425     double weight_scale;
 426     GpuHidHaarStageClassifier *stage_classifier;
 427
 428     if( !CV_IS_HAAR_CLASSIFIER(_cascade) )
 429         CV_Error( !_cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier pointer" );
 430
 431     if( scale <= 0 )
 432         CV_Error( CV_StsOutOfRange, "Scale must be positive" );
 433
 434     if( coi0 || coi1 )
 435         CV_Error( CV_BadCOI, "COI is not supported" );
 436
 437     if( !_cascade->hid_cascade )
 438         gpuCreateHidHaarClassifierCascade(_cascade, &datasize, &total);
 439
 440     cascade = (GpuHidHaarClassifierCascade *) _cascade->hid_cascade;
 441     stage_classifier = (GpuHidHaarStageClassifier *) (cascade + 1);
 442
 443     _cascade->scale = scale;
 444     _cascade->real_window_size.width = cvRound( _cascade->orig_window_size.width * scale );
 445     _cascade->real_window_size.height = cvRound( _cascade->orig_window_size.height * scale );
 446
 447     equRect.x = equRect.y = cvRound(scale);
 448     equRect.width = cvRound((_cascade->orig_window_size.width - 2) * scale);
 449     equRect.height = cvRound((_cascade->orig_window_size.height - 2) * scale);
 450     weight_scale = 1. / (equRect.width * equRect.height);
 451     cascade->inv_window_area = weight_scale;
 452
 453     cascade->pq0 = equRect.x;
 454     cascade->pq1 = equRect.y;
 455     cascade->pq2 = equRect.x + equRect.width;
 456     cascade->pq3 = equRect.y + equRect.height;
 457
 458     cascade->p0 = equRect.x;
 459     cascade->p1 = equRect.y;
 460     cascade->p2 = equRect.x + equRect.width;
 461     cascade->p3 = equRect.y + equRect.height;
 462
 463
 464     /* init pointers in haar features according to real window size and
 465     given image pointers */
 466     for( i = 0; i < _cascade->count; i++ )
 467     {
 468         int j, k, l;
 469         for( j = 0; j < stage_classifier[i].count; j++ )
 470         {
 471             for( l = 0; l < stage_classifier[i].classifier[j].count; l++ )
 472             {
 473                 CvHaarFeature *feature =
 474                     &_cascade->stage_classifier[i].classifier[j].haar_feature[l];
 475                 GpuHidHaarTreeNode *hidnode = &stage_classifier[i].classifier[j].node[l];
 476                 double sum0 = 0, area0 = 0;
 477                 CvRect r[3];
 478
 479                 int base_w = -1, base_h = -1;
 480                 int new_base_w = 0, new_base_h = 0;
 481                 int kx, ky;
 482                 int flagx = 0, flagy = 0;
 483                 int x0 = 0, y0 = 0;
 484                 int nr;
 485
 486                 /* align blocks */
 487                 for( k = 0; k < CV_HAAR_FEATURE_MAX; k++ )
 488                 {
 489                     if(!hidnode->p[k][0])
 490                         break;
 491                     r[k] = feature->rect[k].r;
 492                     base_w = (int)CV_IMIN( (unsigned)base_w, (unsigned)(r[k].width - 1) );
 493                     base_w = (int)CV_IMIN( (unsigned)base_w, (unsigned)(r[k].x - r[0].x - 1) );
 494                     base_h = (int)CV_IMIN( (unsigned)base_h, (unsigned)(r[k].height - 1) );
 495                     base_h = (int)CV_IMIN( (unsigned)base_h, (unsigned)(r[k].y - r[0].y - 1) );
 496                 }
 497
 498                 nr = k;
 499                 base_w += 1;
 500                 base_h += 1;
 501                 if(base_w == 0)
 502                     base_w = 1;
 503                 kx = r[0].width / base_w;
 504                 if(base_h == 0)
 505                     base_h = 1;
 506                 ky = r[0].height / base_h;
 507
 508                 if( kx <= 0 )
 509                 {
 510                     flagx = 1;
 511                     new_base_w = cvRound( r[0].width * scale ) / kx;
 512                     x0 = cvRound( r[0].x * scale );
 513                 }
 514
 515                 if( ky <= 0 )
 516                 {
 517                     flagy = 1;
 518                     new_base_h = cvRound( r[0].height * scale ) / ky;
 519                     y0 = cvRound( r[0].y * scale );
 520                 }
 521
 522                 for( k = 0; k < nr; k++ )
 523                 {
 524                     CvRect tr;
 525                     double correction_ratio;
 526
 527                     if( flagx )
 528                     {
 529                         tr.x = (r[k].x - r[0].x) * new_base_w / base_w + x0;
 530                         tr.width = r[k].width * new_base_w / base_w;
 531                     }
 532                     else
 533                     {
 534                         tr.x = cvRound( r[k].x * scale );
 535                         tr.width = cvRound( r[k].width * scale );
 536                     }
 537
 538                     if( flagy )
 539                     {
 540                         tr.y = (r[k].y - r[0].y) * new_base_h / base_h + y0;
 541                         tr.height = r[k].height * new_base_h / base_h;
 542                     }
 543                     else
 544                     {
 545                         tr.y = cvRound( r[k].y * scale );
 546                         tr.height = cvRound( r[k].height * scale );
 547                     }
 548
 549 #if CV_ADJUST_WEIGHTS
 550                     {
 551                         // RAINER START
 552                         const float orig_feature_size =  (float)(feature->rect[k].r.width) * feature->rect[k].r.height;
 553                         const float orig_norm_size = (float)(_cascade->orig_window_size.width) * (_cascade->orig_window_size.height);
 554                         const float feature_size = float(tr.width * tr.height);
 555                         //const float normSize    = float(equRect.width*equRect.height);
 556                         float target_ratio = orig_feature_size / orig_norm_size;
 557                         //float isRatio = featureSize / normSize;
 558                         //correctionRatio = targetRatio / isRatio / normSize;
 559                         correction_ratio = target_ratio / feature_size;
 560                         // RAINER END
 561                     }
 562 #else
 563                     correction_ratio = weight_scale * (!feature->tilted ? 1 : 0.5);
 564 #endif
 565
 566                     if( !feature->tilted )
 567                     {
 568                         hidnode->p[k][0] = tr.x;
 569                         hidnode->p[k][1] = tr.y;
 570                         hidnode->p[k][2] = tr.x + tr.width;
 571                         hidnode->p[k][3] = tr.y + tr.height;
 572                     }
 573                     else
 574                     {
 575                         hidnode->p[k][2] = (tr.y + tr.width) * step + tr.x + tr.width;
 576                         hidnode->p[k][3] = (tr.y + tr.width + tr.height) * step + tr.x + tr.width - tr.height;
 577                         hidnode->p[k][0] = tr.y * step + tr.x;
 578                         hidnode->p[k][1] = (tr.y + tr.height) * step + tr.x - tr.height;
 579                     }
 580                     hidnode->weight[k] = (float)(feature->rect[k].weight * correction_ratio);
 581                     if( k == 0 )
 582                         area0 = tr.width * tr.height;
 583                     else
 584                         sum0 += hidnode->weight[k] * tr.width * tr.height;
 585                 }
 586                 hidnode->weight[0] = (float)(-sum0 / area0);
 587             } /* l */
 588         } /* j */
 589     }
 590 }
 591
 592 static void gpuSetHaarClassifierCascade( CvHaarClassifierCascade *_cascade)
 593 {
 594     GpuHidHaarClassifierCascade *cascade;
 595     int i;
 596     int datasize;
 597     int total;
 598     CvRect equRect;
 599     double weight_scale;
 600     GpuHidHaarStageClassifier *stage_classifier;
 601
 602     if( !CV_IS_HAAR_CLASSIFIER(_cascade) )
 603         CV_Error( !_cascade ? CV_StsNullPtr : CV_StsBadArg, "Invalid classifier pointer" );
 604
 605     if( !_cascade->hid_cascade )
 606         gpuCreateHidHaarClassifierCascade(_cascade, &datasize, &total);
 607
 608     cascade = (GpuHidHaarClassifierCascade *) _cascade->hid_cascade;
 609     stage_classifier = (GpuHidHaarStageClassifier *) cascade + 1;
 610
 611     _cascade->scale = 1.0;
 612     _cascade->real_window_size.width =  _cascade->orig_window_size.width ;
 613     _cascade->real_window_size.height = _cascade->orig_window_size.height;
 614
 615     equRect.x = equRect.y = 1;
 616     equRect.width = _cascade->orig_window_size.width - 2;
 617     equRect.height = _cascade->orig_window_size.height - 2;
 618     weight_scale = 1;
 619     cascade->inv_window_area = weight_scale;
 620
 621     cascade->p0 = equRect.x;
 622     cascade->p1 = equRect.y;
 623     cascade->p2 = equRect.height;
 624     cascade->p3 = equRect.width ;
 625     for( i = 0; i < _cascade->count; i++ )
 626     {
 627         int j, l;
 628         for( j = 0; j < stage_classifier[i].count; j++ )
 629         {
 630             for( l = 0; l < stage_classifier[i].classifier[j].count; l++ )
 631             {
 632                 const CvHaarFeature *feature =
 633                     &_cascade->stage_classifier[i].classifier[j].haar_feature[l];
 634                 GpuHidHaarTreeNode *hidnode = &stage_classifier[i].classifier[j].node[l];
 635
 636                 for( int k = 0; k < CV_HAAR_FEATURE_MAX; k++ )
 637                 {
 638                     const CvRect tr = feature->rect[k].r;
 639                     if (tr.width == 0)
 640                         break;
 641                     double correction_ratio = weight_scale * (!feature->tilted ? 1 : 0.5);
 642                     hidnode->p[k][0] = tr.x;
 643                     hidnode->p[k][1] = tr.y;
 644                     hidnode->p[k][2] = tr.width;
 645                     hidnode->p[k][3] = tr.height;
 646                     hidnode->weight[k] = (float)(feature->rect[k].weight * correction_ratio);
 647                 }
 648             } /* l */
 649         } /* j */
 650     }
 651 }
 652 void OclCascadeClassifier::detectMultiScale(oclMat &gimg, CV_OUT std::vector<cv::Rect>& faces,
 653                                             double scaleFactor, int minNeighbors, int flags,
 654                                             Size minSize, Size maxSize)
 655 //CvSeq *cv::ocl::OclCascadeClassifier::oclHaarDetectObjects( oclMat &gimg, CvMemStorage *storage, double scaleFactor,
 656 //        int minNeighbors, int flags, CvSize minSize, CvSize maxSize)
 657 {
 658     CvHaarClassifierCascade *cascade = oldCascade;
 659
 660     const double GROUP_EPS = 0.2;
 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, COLOR_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         std::vector<CvSize> sizev;
 726         std::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         std::vector<std::pair<size_t, const void *> > args;
 835         args.push_back ( std::make_pair(sizeof(cl_mem) , (void *)&stagebuffer ));
 836         args.push_back ( std::make_pair(sizeof(cl_mem) , (void *)&scaleinfobuffer ));
 837         args.push_back ( std::make_pair(sizeof(cl_mem) , (void *)&nodebuffer ));
 838         args.push_back ( std::make_pair(sizeof(cl_mem) , (void *)&gsum.data ));
 839         args.push_back ( std::make_pair(sizeof(cl_mem) , (void *)&gsqsum.data ));
 840         args.push_back ( std::make_pair(sizeof(cl_mem) , (void *)&candidatebuffer ));
 841         args.push_back ( std::make_pair(sizeof(cl_int) , (void *)&pixelstep ));
 842         args.push_back ( std::make_pair(sizeof(cl_int) , (void *)&loopcount ));
 843         args.push_back ( std::make_pair(sizeof(cl_int) , (void *)&startstage ));
 844         args.push_back ( std::make_pair(sizeof(cl_int) , (void *)&splitstage ));
 845         args.push_back ( std::make_pair(sizeof(cl_int) , (void *)&endstage ));
 846         args.push_back ( std::make_pair(sizeof(cl_int) , (void *)&startnode ));
 847         args.push_back ( std::make_pair(sizeof(cl_int) , (void *)&splitnode ));
 848         args.push_back ( std::make_pair(sizeof(cl_int4) , (void *)&p ));
 849         args.push_back ( std::make_pair(sizeof(cl_int4) , (void *)&pq ));
 850         args.push_back ( std::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 ( std::make_pair(sizeof(cl_mem) , (void *)&oclNodesPK.datastart ));
 943             args.push_back ( std::make_pair(sizeof(cl_mem) , (void *)&oclWGInfo.datastart ));
 944
 945             //form build options for kernel
 946             String  options = "-D PACKED_CLASSIFIER";
 947             options = options + format(" -D NODE_SIZE=%d",NODE_SIZE);
 948             options = options + format(" -D WND_SIZE_X=%d",cascade->orig_window_size.width);
 949             options = options + format(" -D WND_SIZE_Y=%d",cascade->orig_window_size.height);
 950             options = options + format(" -D STUMP_BASED=%d",gcascade->is_stump_based);
 951             options = options + format(" -D LSx=%d",localThreads[0]);
 952             options = options + format(" -D LSy=%d",localThreads[1]);
 953             options = options + format(" -D SPLITNODE=%d",splitnode);
 954             options = options + format(" -D SPLITSTAGE=%d",splitstage);
 955             options = 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         std::vector<CvSize> sizev;
1002         std::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 ), cvRound( winsize0.height * factor ) );
1020             if( winSize.width < minSize.width || winSize.height < minSize.height )
1021             {
1022                 continue;
1023             }
1024             sizev.push_back(winSize);
1025             scalev.push_back(factor);
1026         }
1027         int loopcount = scalev.size();
1028         if(loopcount == 0)
1029         {
1030             loopcount = 1;
1031             n_factors = 1;
1032             sizev.push_back(minSize);
1033             scalev.push_back( std::min(cvRound(minSize.width / winsize0.width), cvRound(minSize.height / winsize0.height)) );
1034         }
1035         detect_piramid_info *scaleinfo = (detect_piramid_info *)malloc(sizeof(detect_piramid_info) * loopcount);
1036         cl_int4 *p = (cl_int4 *)malloc(sizeof(cl_int4) * loopcount);
1037         float *correction = (float *)malloc(sizeof(float) * loopcount);
1038         int grp_per_CU = 12;
1039         size_t blocksize = 8;
1040         size_t localThreads[3] = { blocksize, blocksize , 1 };
1041         size_t globalThreads[3] = { grp_per_CU *gsum.clCxt->getDeviceInfo().maxComputeUnits *localThreads[0],
1042                                     localThreads[1], 1 };
1043         int outputsz = 256 * globalThreads[0] / localThreads[0];
1044         int nodenum = (datasize - sizeof(GpuHidHaarClassifierCascade) -
1045                        sizeof(GpuHidHaarStageClassifier) * gcascade->count - sizeof(GpuHidHaarClassifier) * totalclassifier) / sizeof(GpuHidHaarTreeNode);
1046         nodebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY,
1047                                         nodenum * sizeof(GpuHidHaarTreeNode));
1048         openCLSafeCall(clEnqueueWriteBuffer(qu, nodebuffer, 1, 0,
1049                                             nodenum * sizeof(GpuHidHaarTreeNode),
1050                                             node, 0, NULL, NULL));
1051         cl_mem newnodebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_WRITE,
1052                                loopcount * nodenum * sizeof(GpuHidHaarTreeNode));
1053         int startstage = 0;
1054         int endstage = gcascade->count;
1055         for(int i = 0; i < loopcount; i++)
1056         {
1057             sz = sizev[i];
1058             factor = scalev[i];
1059             double ystep = std::max(2., factor);
1060             int equRect_x = cvRound(factor * gcascade->p0);
1061             int equRect_y = cvRound(factor * gcascade->p1);
1062             int equRect_w = cvRound(factor * gcascade->p3);
1063             int equRect_h = cvRound(factor * gcascade->p2);
1064             p[i].s[0] = equRect_x;
1065             p[i].s[1] = equRect_y;
1066             p[i].s[2] = equRect_x + equRect_w;
1067             p[i].s[3] = equRect_y + equRect_h;
1068             correction[i] = 1. / (equRect_w * equRect_h);
1069             int width = (gsum.cols - 1 - sz.width  + ystep - 1) / ystep;
1070             int height = (gsum.rows - 1 - sz.height + ystep - 1) / ystep;
1071             int grpnumperline = (width + localThreads[0] - 1) / localThreads[0];
1072             int totalgrp = ((height + localThreads[1] - 1) / localThreads[1]) * grpnumperline;
1073
1074             scaleinfo[i].width_height = (width << 16) | height;
1075             scaleinfo[i].grpnumperline_totalgrp = (grpnumperline << 16) | totalgrp;
1076             scaleinfo[i].imgoff = 0;
1077             scaleinfo[i].factor = factor;
1078             int startnodenum = nodenum * i;
1079             float factor2 = (float)factor;
1080
1081             std::vector<std::pair<size_t, const void *> > args1;
1082             args1.push_back ( std::make_pair(sizeof(cl_mem) , (void *)&nodebuffer ));
1083             args1.push_back ( std::make_pair(sizeof(cl_mem) , (void *)&newnodebuffer ));
1084             args1.push_back ( std::make_pair(sizeof(cl_float) , (void *)&factor2 ));
1085             args1.push_back ( std::make_pair(sizeof(cl_float) , (void *)&correction[i] ));
1086             args1.push_back ( std::make_pair(sizeof(cl_int) , (void *)&startnodenum ));
1087
1088             size_t globalThreads2[3] = {nodenum, 1, 1};
1089             openCLExecuteKernel(gsum.clCxt, &haarobjectdetect_scaled2, "gpuscaleclassifier", globalThreads2, NULL/*localThreads2*/, args1, -1, -1);
1090         }
1091
1092         int step = gsum.step / 4;
1093         int startnode = 0;
1094         int splitstage = 3;
1095         stagebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(GpuHidHaarStageClassifier) * gcascade->count);
1096         openCLSafeCall(clEnqueueWriteBuffer(qu, stagebuffer, 1, 0, sizeof(GpuHidHaarStageClassifier)*gcascade->count, stage, 0, NULL, NULL));
1097         candidatebuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_WRITE_ONLY | CL_MEM_ALLOC_HOST_PTR, 4 * sizeof(int) * outputsz);
1098         scaleinfobuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(detect_piramid_info) * loopcount);
1099         openCLSafeCall(clEnqueueWriteBuffer(qu, scaleinfobuffer, 1, 0, sizeof(detect_piramid_info)*loopcount, scaleinfo, 0, NULL, NULL));
1100         pbuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(cl_int4) * loopcount);
1101         openCLSafeCall(clEnqueueWriteBuffer(qu, pbuffer, 1, 0, sizeof(cl_int4)*loopcount, p, 0, NULL, NULL));
1102         correctionbuffer = openCLCreateBuffer(gsum.clCxt, CL_MEM_READ_ONLY, sizeof(cl_float) * loopcount);
1103         openCLSafeCall(clEnqueueWriteBuffer(qu, correctionbuffer, 1, 0, sizeof(cl_float)*loopcount, correction, 0, NULL, NULL));
1104
1105         std::vector<std::pair<size_t, const void *> > args;
1106         args.push_back ( std::make_pair(sizeof(cl_mem) , (void *)&stagebuffer ));
1107         args.push_back ( std::make_pair(sizeof(cl_mem) , (void *)&scaleinfobuffer ));
1108         args.push_back ( std::make_pair(sizeof(cl_mem) , (void *)&newnodebuffer ));
1109         args.push_back ( std::make_pair(sizeof(cl_mem) , (void *)&gsum.data ));
1110         args.push_back ( std::make_pair(sizeof(cl_mem) , (void *)&gsqsum.data ));
1111         args.push_back ( std::make_pair(sizeof(cl_mem) , (void *)&candidatebuffer ));
1112         args.push_back ( std::make_pair(sizeof(cl_int) , (void *)&gsum.rows ));
1113         args.push_back ( std::make_pair(sizeof(cl_int) , (void *)&gsum.cols ));
1114         args.push_back ( std::make_pair(sizeof(cl_int) , (void *)&step ));
1115         args.push_back ( std::make_pair(sizeof(cl_int) , (void *)&loopcount ));
1116         args.push_back ( std::make_pair(sizeof(cl_int) , (void *)&startstage ));
1117         args.push_back ( std::make_pair(sizeof(cl_int) , (void *)&splitstage ));
1118         args.push_back ( std::make_pair(sizeof(cl_int) , (void *)&endstage ));
1119         args.push_back ( std::make_pair(sizeof(cl_int) , (void *)&startnode ));
1120         args.push_back ( std::make_pair(sizeof(cl_mem) , (void *)&pbuffer ));
1121         args.push_back ( std::make_pair(sizeof(cl_mem) , (void *)&correctionbuffer ));
1122         args.push_back ( std::make_pair(sizeof(cl_int) , (void *)&nodenum ));
1123         const char * build_options = gcascade->is_stump_based ? "-D STUMP_BASED=1" : "-D STUMP_BASED=0";
1124         openCLExecuteKernel(gsum.clCxt, &haarobjectdetect_scaled2, "gpuRunHaarClassifierCascade_scaled2", globalThreads, localThreads, args, -1, -1, build_options);
1125
1126         candidate = (int *)clEnqueueMapBuffer(qu, candidatebuffer, 1, CL_MAP_READ, 0, 4 * sizeof(int) * outputsz, 0, 0, 0, &status);
1127
1128         for(int i = 0; i < outputsz; i++)
1129         {
1130             if(candidate[4 * i + 2] != 0)
1131                 allCandidates.push_back(Rect(candidate[4 * i], candidate[4 * i + 1], candidate[4 * i + 2], candidate[4 * i + 3]));
1132         }
1133
1134         free(scaleinfo);
1135         free(p);
1136         free(correction);
1137         clEnqueueUnmapMemObject(qu, candidatebuffer, candidate, 0, 0, 0);
1138         openCLSafeCall(clReleaseMemObject(stagebuffer));
1139         openCLSafeCall(clReleaseMemObject(scaleinfobuffer));
1140         openCLSafeCall(clReleaseMemObject(nodebuffer));
1141         openCLSafeCall(clReleaseMemObject(newnodebuffer));
1142         openCLSafeCall(clReleaseMemObject(candidatebuffer));
1143         openCLSafeCall(clReleaseMemObject(pbuffer));
1144         openCLSafeCall(clReleaseMemObject(correctionbuffer));
1145     }
1146
1147     cvFree(&cascade->hid_cascade);
1148     rectList.resize(allCandidates.size());
1149     if(!allCandidates.empty())
1150         std::copy(allCandidates.begin(), allCandidates.end(), rectList.begin());
1151
1152     if( minNeighbors != 0 || findBiggestObject )
1153         groupRectangles(rectList, rweights, std::max(minNeighbors, 1), GROUP_EPS);
1154     else
1155         rweights.resize(rectList.size(), 0);
1156
1157     faces.clear();
1158     if( findBiggestObject && rectList.size() )
1159     {
1160         Rect result_comp(0, 0, 0, 0);
1161         for( size_t i = 0; i < rectList.size(); i++ )
1162         {
1163             cv::Rect r = rectList[i];
1164             if( r.area() > result_comp.area() )
1165             {
1166                 result_comp = r;
1167             }
1168         }
1169         faces.push_back(result_comp);
1170     }
1171     else
1172     {
1173         faces = rectList;
1174     }
1175 }