Peter M. Groen / openbr

Commit 42c4fdadd5e1cb7937f1be3b66c5493e9098a2c1

Authored by Scott Klum
1 parent f0dc837e

Code simplification by removing short type optimization

Inline Side-by-side
Showing 1 changed file with 64 additions and 208 deletions
openbr/core/boost.cpp
  View file @42c4fda
... ... @@ -21,7 +21,6 @@ logRatio( double val )
21 21  
22 22 #define CV_CMP_NUM_IDX(i,j) (aux[i] < aux[j])
23 23 static CV_IMPLEMENT_QSORT_EX( icvSortIntAux, int, CV_CMP_NUM_IDX, const float* )
24   -static CV_IMPLEMENT_QSORT_EX( icvSortUShAux, unsigned short, CV_CMP_NUM_IDX, const float* )
25 24  
26 25 #define CV_THRESHOLD_EPS (0.00001F)
27 26  
... ... @@ -186,7 +185,7 @@ struct CascadeBoostTrainData : CvDTreeTrainData
186 185 return ((uint64_t)work_var_count + 1) * sample_count;
187 186 }
188 187  
189   - virtual const int* getBufferValues(CvDTreeNode* n, int* labelsBuf, uint64_t vi);
  188 + virtual const int* getBufferValues(CvDTreeNode* n, uint64_t vi);
190 189 virtual const int* get_class_labels(CvDTreeNode* n, int* labelsBuf);
191 190 virtual const int* get_cv_labels(CvDTreeNode* n, int* labelsBuf);
192 191 virtual const int* get_sample_indices(CvDTreeNode* n, int* indicesBuf);
... ... @@ -298,62 +297,29 @@ CvDTreeNode* CascadeBoostTrainData::subsample_data( const CvMat* _subsample_idx
298 297 int num_valid = data_root->get_num_valid(vi);
299 298 CV_Assert( num_valid == sample_count );
300 299  
301   - if (is_buf_16u)
  300 + int* idst_idx = buf->data.i + root->buf_idx*getLength() + (uint64_t)vi*sample_count + root->offset;
  301 + for( int i = 0; i < num_valid; i++ )
302 302 {
303   - unsigned short* udst_idx = (unsigned short*)(buf->data.s + root->buf_idx*getLength() + (uint64_t)vi*sample_count + data_root->offset);
304   - for( int i = 0; i < num_valid; i++ )
305   - {
306   - idx = src_idx[i];
307   - count_i = co[idx*2];
308   - if( count_i )
309   - for( cur_ofs = co[idx*2+1]; count_i > 0; count_i--, j++, cur_ofs++ )
310   - udst_idx[j] = (unsigned short)cur_ofs;
311   - }
312   - }
313   - else
314   - {
315   - int* idst_idx = buf->data.i + root->buf_idx*getLength() + (uint64_t)vi*sample_count + root->offset;
316   - for( int i = 0; i < num_valid; i++ )
317   - {
318   - idx = src_idx[i];
319   - count_i = co[idx*2];
320   - if( count_i )
321   - for( cur_ofs = co[idx*2+1]; count_i > 0; count_i--, j++, cur_ofs++ )
322   - idst_idx[j] = cur_ofs;
323   - }
  303 + idx = src_idx[i];
  304 + count_i = co[idx*2];
  305 + if( count_i )
  306 + for( cur_ofs = co[idx*2+1]; count_i > 0; count_i--, j++, cur_ofs++ )
  307 + idst_idx[j] = cur_ofs;
324 308 }
325 309 }
326 310  
327 311 // subsample cv_lables
328 312 const int* src_lbls = get_cv_labels(data_root, (int*)(uchar*)inn_buf);
329   - if (is_buf_16u)
330   - {
331   - unsigned short* udst = (unsigned short*)(buf->data.s + root->buf_idx*getLength() + (workVarCount-1)*sample_count + root->offset);
332   - for( int i = 0; i < count; i++ )
333   - udst[i] = (unsigned short)src_lbls[sidx[i]];
334   - }
335   - else
336   - {
337   - int* idst = buf->data.i + root->buf_idx*getLength() +
338   - (workVarCount-1)*sample_count + root->offset;
339   - for( int i = 0; i < count; i++ )
340   - idst[i] = src_lbls[sidx[i]];
341   - }
  313 + int* idst = buf->data.i + root->buf_idx*getLength() +
  314 + (workVarCount-1)*sample_count + root->offset;
  315 + for( int i = 0; i < count; i++ )
  316 + idst[i] = src_lbls[sidx[i]];
342 317  
343 318 // subsample sample_indices
344 319 const int* sample_idx_src = get_sample_indices(data_root, (int*)(uchar*)inn_buf);
345   - if (is_buf_16u)
346   - {
347   - unsigned short* sample_idx_dst = (unsigned short*)(buf->data.s + root->buf_idx*getLength() + workVarCount*sample_count + root->offset);
348   - for( int i = 0; i < count; i++ )
349   - sample_idx_dst[i] = (unsigned short)sample_idx_src[sidx[i]];
350   - }
351   - else
352   - {
353   - int* sample_idx_dst = buf->data.i + root->buf_idx*getLength() + workVarCount*sample_count + root->offset;
354   - for( int i = 0; i < count; i++ )
355   - sample_idx_dst[i] = sample_idx_src[sidx[i]];
356   - }
  320 + int* sample_idx_dst = buf->data.i + root->buf_idx*getLength() + workVarCount*sample_count + root->offset;
  321 + for( int i = 0; i < count; i++ )
  322 + sample_idx_dst[i] = sample_idx_src[sidx[i]];
357 323  
358 324 for( int vi = 0; vi < var_count; vi++ )
359 325 root->set_num_valid(vi, count);
... ... @@ -459,12 +425,10 @@ void CascadeBoostTrainData::setData( const FeatureEvaluator* _featureEvaluator,
459 425 sample_count = _numSamples;
460 426  
461 427 is_buf_16u = false;
462   - if (sample_count < 65536)
463   - is_buf_16u = true;
464 428  
465 429 // 1048576 is the number of bytes in a megabyte
466 430 numPrecalcVal = min( cvRound((double)_precalcValBufSize*1048576. / (sizeof(float)*sample_count)), var_count );
467   - numPrecalcIdx = min( cvRound((double)_precalcIdxBufSize*1048576. / ((is_buf_16u ? sizeof(unsigned short) : sizeof (int))*sample_count)), var_count );
  431 + numPrecalcIdx = min( cvRound((double)_precalcIdxBufSize*1048576. / (sizeof (int)*sample_count)), var_count );
468 432  
469 433 assert( numPrecalcIdx >= 0 && numPrecalcVal >= 0 );
470 434  
... ... @@ -510,13 +474,9 @@ void CascadeBoostTrainData::setData( const FeatureEvaluator* _featureEvaluator,
510 474 effective_buf_width *= buf_count;
511 475  
512 476 if (effective_buf_width * effective_buf_height != effective_buf_size)
513   - {
514 477 CV_Error(CV_StsBadArg, "The memory buffer cannot be allocated since its size exceeds integer fields limit");
515   - }
516   - if ( is_buf_16u )
517   - buf = cvCreateMat( effective_buf_height, effective_buf_width, CV_16UC(channels) );
518   - else
519   - buf = cvCreateMat( effective_buf_height, effective_buf_width, CV_32SC(channels) );
  478 +
  479 + buf = cvCreateMat( effective_buf_height, effective_buf_width, CV_32SC(channels) );
520 480  
521 481 cat_count = cvCreateMat( 1, cat_var_count + 1, CV_32SC(channels) );
522 482  
... ... @@ -542,10 +502,7 @@ void CascadeBoostTrainData::setData( const FeatureEvaluator* _featureEvaluator,
542 502 data_root = new_node( 0, sample_count, 0, 0 );
543 503  
544 504 // set sample labels
545   - if (is_buf_16u)
546   - udst = (unsigned short*)(buf->data.s + (uint64_t)work_var_count*sample_count);
547   - else
548   - idst = buf->data.i + (uint64_t)work_var_count*sample_count;
  505 + idst = buf->data.i + (uint64_t)work_var_count*sample_count;
549 506  
550 507 for (int si = 0; si < sample_count; si++)
551 508 {
... ... @@ -594,29 +551,21 @@ const int* CascadeBoostTrainData::get_class_labels( CvDTreeNode* n, int* labelsB
594 551 return labelsBuf;
595 552 }
596 553  
597   -const int* CascadeBoostTrainData::getBufferValues(CvDTreeNode* n, int* indicesBuf, uint64_t vi)
  554 +const int* CascadeBoostTrainData::getBufferValues(CvDTreeNode* n, uint64_t vi)
598 555 {
599   - const int* cat_values = 0;
600   - if (!is_buf_16u)
601   - cat_values = buf->data.i + n->buf_idx*getLength() + vi*sample_count + n->offset;
602   - else {
603   - const unsigned short* short_values = (const unsigned short*)(buf->data.s + n->buf_idx*getLength() + vi*sample_count + n->offset);
604   - for (int i = 0; i < n->sample_count; i++)
605   - indicesBuf[i] = short_values[i];
606   - cat_values = indicesBuf;
607   - }
608   -
609   - return cat_values;
  556 + return buf->data.i + n->buf_idx*getLength() + vi*sample_count + n->offset;
610 557 }
611 558  
612   -const int* CascadeBoostTrainData::get_sample_indices( CvDTreeNode* n, int* indicesBuf )
  559 +const int* CascadeBoostTrainData::get_sample_indices(CvDTreeNode* n, int* indicesBuf)
613 560 {
614   - return getBufferValues(n,indicesBuf,get_work_var_count());
  561 + Q_UNUSED(indicesBuf)
  562 + return getBufferValues(n,get_work_var_count());
615 563 }
616 564  
617   -const int* CascadeBoostTrainData::get_cv_labels( CvDTreeNode* n, int* indicesBuf )
  565 +const int* CascadeBoostTrainData::get_cv_labels(CvDTreeNode* n, int* indicesBuf)
618 566 {
619   - return getBufferValues(n,indicesBuf,get_work_var_count()-1);
  567 + Q_UNUSED(indicesBuf)
  568 + return getBufferValues(n,get_work_var_count()-1);
620 569 }
621 570  
622 571 void CascadeBoostTrainData::get_ord_var_data( CvDTreeNode* n, int vi, float* ordValuesBuf, int* sortedIndicesBuf, const float** ordValues, const int** sortedIndices, int* sampleIndicesBuf )
... ... @@ -629,14 +578,7 @@ void CascadeBoostTrainData::get_ord_var_data( CvDTreeNode* n, int vi, float* ord
629 578 // For this feature (this code refers to features as values, hence vi == value index),
630 579 // have we precalculated (presorted) the training samples by their feature response?
631 580 if (vi < numPrecalcIdx) {
632   - if (!is_buf_16u)
633   - *sortedIndices = buf->data.i + n->buf_idx*getLength() + (uint64_t)vi*sample_count + n->offset;
634   - else {
635   - const unsigned short* shortIndices = (const unsigned short*)(buf->data.s + n->buf_idx*getLength() + (uint64_t)vi*sample_count + n->offset );
636   - for (int i = 0; i < nodeSampleCount; i++)
637   - sortedIndicesBuf[i] = shortIndices[i];
638   - *sortedIndices = sortedIndicesBuf;
639   - }
  581 + *sortedIndices = getBufferValues(n, vi);
640 582  
641 583 // For this feature, have we precalculated all of the feature responses?
642 584 if (vi < numPrecalcVal) {
... ... @@ -731,26 +673,20 @@ struct IndexPrecalc : Precalc
731 673 {
732 674 int* idst;
733 675 unsigned short* udst;
734   - bool isBufShort;
735 676  
736   - IndexPrecalc(const FeatureEvaluator* featureEvaluator, CvMat* buf, int sampleCount, bool isBufShort) :
  677 + IndexPrecalc(const FeatureEvaluator* featureEvaluator, CvMat* buf, int sampleCount) :
737 678 Precalc(featureEvaluator, sampleCount),
738   - isBufShort(isBufShort)
739   - {
740   - udst = (unsigned short*)buf->data.s;
741   - idst = buf->data.i;
742   - }
  679 + idst(buf->data.i)
  680 + {}
743 681  
744 682 void setBuffer(uint64_t fi, uint64_t si) const
745 683 {
746   - if (isBufShort) *(udst + fi*sampleCount + si) = (unsigned short)si;
747   - else *(idst + fi*sampleCount + si) = si;
  684 + *(idst + fi*sampleCount + si) = si;
748 685 }
749 686  
750 687 void sortBuffer(uint64_t fi, float *valCachePtr) const
751 688 {
752   - if (isBufShort) icvSortUShAux(udst + fi*sampleCount, sampleCount, valCachePtr);
753   - else icvSortIntAux(idst + fi*sampleCount, sampleCount, valCachePtr);
  689 + icvSortIntAux(idst + fi*sampleCount, sampleCount, valCachePtr);
754 690 }
755 691  
756 692 virtual void operator()(const Range& range) const
... ... @@ -771,8 +707,8 @@ struct FeatureAndIndexPrecalc : IndexPrecalc
771 707 {
772 708 Mat *valCache;
773 709  
774   - FeatureAndIndexPrecalc(const FeatureEvaluator* featureEvaluator, CvMat* buf, Mat* valCache, int sampleCount, bool isBufShort) :
775   - IndexPrecalc(featureEvaluator, buf, sampleCount, isBufShort),
  710 + FeatureAndIndexPrecalc(const FeatureEvaluator* featureEvaluator, CvMat* buf, Mat* valCache, int sampleCount) :
  711 + IndexPrecalc(featureEvaluator, buf, sampleCount),
776 712 valCache(valCache)
777 713 {}
778 714  
... ... @@ -816,11 +752,11 @@ void CascadeBoostTrainData::precalculate()
816 752  
817 753 // Compute features and sort training samples for feature indices we are not going to cache
818 754 parallel_for_(Range(numPrecalcVal, numPrecalcIdx),
819   - IndexPrecalc(featureEvaluator, buf, sample_count, is_buf_16u!=0));
  755 + IndexPrecalc(featureEvaluator, buf, sample_count));
820 756  
821 757 // Compute features and sort training samples for features indices we are going to cache
822 758 parallel_for_(Range(0, minPrecalc),
823   - FeatureAndIndexPrecalc(featureEvaluator, buf, &valCache, sample_count, is_buf_16u!=0));
  759 + FeatureAndIndexPrecalc(featureEvaluator, buf, &valCache, sample_count));
824 760  
825 761 // Compute feature values for feature indices for which we are not going to sort training samples
826 762 parallel_for_(Range(minPrecalc, numPrecalcVal),
... ... @@ -880,8 +816,7 @@ void CascadeBoostTree::split_node_data( CvDTreeNode* node )
880 816  
881 817 complete_node_dir(node);
882 818  
883   - for (uint64_t i = nLeft = nRight = 0; i < nodeSampleCount; i++)
884   - {
  819 + for (uint64_t i = nLeft = nRight = 0; i < nodeSampleCount; i++) {
885 820 int d = dir[i];
886 821 // initialize new indices for splitting ordered variables
887 822 newIdx[i] = (nLeft & (d-1)) | (nRight & -d); // d ? ri : li
... ... @@ -900,59 +835,27 @@ void CascadeBoostTree::split_node_data( CvDTreeNode* node )
900 835 if( ci >= 0 || !splitInputData )
901 836 continue;
902 837  
903   - int n1 = node->get_num_valid(vi);
904   - float *src_val_buf = (float*)(tempBuf + nodeSampleCount);
905   - int *src_sorted_idx_buf = (int*)(src_val_buf + nodeSampleCount);
906   - int *src_sample_idx_buf = src_sorted_idx_buf + nodeSampleCount;
907   - const int* src_sorted_idx = 0;
908   - const float* src_val = 0;
909   -
910   - data->get_ord_var_data(node, vi, src_val_buf, src_sorted_idx_buf, &src_val, &src_sorted_idx, src_sample_idx_buf);
911   -
  838 + const int* src_sorted_idx = ((CascadeBoostTrainData*)data)->getBufferValues(node, vi);
912 839 for(uint64_t i = 0; i < nodeSampleCount; i++)
913 840 tempBuf[i] = src_sorted_idx[i];
914 841  
915   - if (data->is_buf_16u) {
916   - ushort *ldst, *rdst;
917   - ldst = (ushort*)(buf->data.s + left->buf_idx*length_buf_row + vi*sampleCount + left->offset);
918   - rdst = (ushort*)(ldst + nLeft);
919   -
920   - // split sorted
921   - for (int i = 0; i < n1; i++) {
922   - int idx = tempBuf[i];
923   - int d = dir[idx];
924   - idx = newIdx[idx];
925   - if (d) {
926   - *rdst = (ushort)idx;
927   - rdst++;
928   - } else {
929   - *ldst = (ushort)idx;
930   - ldst++;
931   - }
932   - }
933   - }
934   - else
935   - {
936   - int *ldst, *rdst;
937   - ldst = buf->data.i + left->buf_idx*length_buf_row + vi*sampleCount + left->offset;
938   - rdst = buf->data.i + right->buf_idx*length_buf_row + vi*sampleCount + right->offset;
  842 + int *ldst, *rdst;
  843 + ldst = buf->data.i + left->buf_idx*length_buf_row + vi*sampleCount + left->offset;
  844 + rdst = buf->data.i + right->buf_idx*length_buf_row + vi*sampleCount + right->offset;
939 845  
940   - // split sorted
941   - for( int i = 0; i < n1; i++ )
942   - {
943   - int idx = tempBuf[i];
944   - int d = dir[idx];
945   - idx = newIdx[idx];
946   - if (d)
947   - {
948   - *rdst = idx;
949   - rdst++;
950   - }
951   - else
952   - {
953   - *ldst = idx;
954   - ldst++;
955   - }
  846 + int n1 = node->get_num_valid(vi);
  847 +
  848 + // split sorted
  849 + for (int i = 0; i < n1; i++) {
  850 + int idx = tempBuf[i];
  851 + int d = dir[idx];
  852 + idx = newIdx[idx];
  853 + if (d) {
  854 + *rdst = idx;
  855 + rdst++;
  856 + } else {
  857 + *ldst = idx;
  858 + ldst++;
956 859 }
957 860 }
958 861 }
... ... @@ -964,37 +867,16 @@ void CascadeBoostTree::split_node_data( CvDTreeNode* node )
964 867 for(uint64_t i = 0; i < nodeSampleCount; i++)
965 868 tempBuf[i] = src_lbls[i];
966 869  
967   - if (data->is_buf_16u) {
968   - unsigned short *ldst = (unsigned short *)(buf->data.s + left->buf_idx*length_buf_row + (workVarCount-1)*sampleCount + left->offset);
969   - unsigned short *rdst = (unsigned short *)(buf->data.s + right->buf_idx*length_buf_row + (workVarCount-1)*sampleCount + right->offset);
970   -
971   - for( uint64_t i = 0; i < nodeSampleCount; i++ ) {
972   - int idx = tempBuf[i];
973   - if (dir[i]) {
974   - *rdst = (unsigned short)idx;
975   - rdst++;
976   - } else {
977   - *ldst = (unsigned short)idx;
978   - ldst++;
979   - }
980   - }
981   -
982   - }
983   - else
984 870 {
985 871 int *ldst = buf->data.i + left->buf_idx*length_buf_row + (workVarCount-1)*sampleCount + left->offset;
986 872 int *rdst = buf->data.i + right->buf_idx*length_buf_row + (workVarCount-1)*sampleCount + right->offset;
987 873  
988   - for( uint64_t i = 0; i < nodeSampleCount; i++ )
989   - {
  874 + for (uint64_t i = 0; i < nodeSampleCount; i++) {
990 875 int idx = tempBuf[i];
991   - if (dir[i])
992   - {
  876 + if (dir[i]) {
993 877 *rdst = idx;
994 878 rdst++;
995   - }
996   - else
997   - {
  879 + } else {
998 880 *ldst = idx;
999 881 ldst++;
1000 882 }
... ... @@ -1008,35 +890,15 @@ void CascadeBoostTree::split_node_data( CvDTreeNode* node )
1008 890 for(uint64_t i = 0; i < nodeSampleCount; i++)
1009 891 tempBuf[i] = sampleIdx_src[i];
1010 892  
1011   - if (data->is_buf_16u) {
1012   - unsigned short* ldst = (unsigned short*)(buf->data.s + left->buf_idx*length_buf_row + workVarCount*sampleCount + left->offset);
1013   - unsigned short* rdst = (unsigned short*)(buf->data.s + right->buf_idx*length_buf_row + workVarCount*sampleCount + right->offset);
1014   - for (uint64_t i = 0; i < nodeSampleCount; i++) {
1015   - unsigned short idx = (unsigned short)tempBuf[i];
1016   - if (dir[i]) {
1017   - *rdst = idx;
1018   - rdst++;
1019   - } else {
1020   - *ldst = idx;
1021   - ldst++;
1022   - }
1023   - }
1024   - }
1025   - else
1026 893 {
1027 894 int* ldst = buf->data.i + left->buf_idx*length_buf_row + workVarCount*sampleCount + left->offset;
1028 895 int* rdst = buf->data.i + right->buf_idx*length_buf_row + workVarCount*sampleCount + right->offset;
1029   - for (uint64_t i = 0; i < nodeSampleCount; i++)
1030   - {
  896 + for (uint64_t i = 0; i < nodeSampleCount; i++) {
1031 897 int idx = tempBuf[i];
1032   - if (dir[i])
1033   - {
1034   -
  898 + if (dir[i]) {
1035 899 *rdst = idx;
1036 900 rdst++;
1037   - }
1038   - else
1039   - {
  901 + } else {
1040 902 *ldst = idx;
1041 903 ldst++;
1042 904 }
... ... @@ -1167,15 +1029,9 @@ void CascadeBoost::update_weights(CvBoostTree* tree)
1167 1029  
1168 1030 // set the labels to find (from within weak tree learning proc)
1169 1031 // the particular sample weight, and where to store the response.
1170   - if (data->is_buf_16u) {
1171   - unsigned short* labels = (unsigned short*)(buf->data.s + data->data_root->buf_idx*length_buf_row + data->data_root->offset + (uint64_t)(data->work_var_count-1)*data->sample_count);
1172   - for (int i = 0; i < n; i++)
1173   - labels[i] = (unsigned short)i;
1174   - } else {
1175   - int* labels = buf->data.i + data->data_root->buf_idx*length_buf_row + data->data_root->offset + (uint64_t)(data->work_var_count-1)*data->sample_count;
1176   - for( int i = 0; i < n; i++ )
1177   - labels[i] = i;
1178   - }
  1032 + int* labels = buf->data.i + data->data_root->buf_idx*length_buf_row + data->data_root->offset + (uint64_t)(data->work_var_count-1)*data->sample_count;
  1033 + for( int i = 0; i < n; i++ )
  1034 + labels[i] = i;
1179 1035  
1180 1036 for (int i = 0; i < n; i++) {
1181 1037 // save original categorical responses {0,1}, convert them to {-1,1}
... ...