1 /*M///////////////////////////////////////////////////////////////////////////////////////
3 // IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
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.
11 // For Open Source Computer Vision Library
13 // Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
14 // Copyright (C) 2009-2011, Willow Garage Inc., all rights reserved.
15 // Third party copyrights are property of their respective owners.
17 // Redistribution and use in source and binary forms, with or without modification,
18 // are permitted provided that the following conditions are met:
20 // * Redistribution's of source code must retain the above copyright notice,
21 // this list of conditions and the following disclaimer.
23 // * Redistribution's in binary form must reproduce the above copyright notice,
24 // this list of conditions and the following disclaimer in the documentation
25 // and/or other materials provided with the distribution.
27 // * The name of the copyright holders may not be used to endorse or promote products
28 // derived from this software without specific prior written permission.
30 // This software is provided by the copyright holders and contributors "as is" and
31 // any express or implied warranties, including, but not limited to, the implied
32 // warranties of merchantability and fitness for a particular purpose are disclaimed.
33 // In no event shall the Intel Corporation or contributors be liable for any direct,
34 // indirect, incidental, special, exemplary, or consequential damages
35 // (including, but not limited to, procurement of substitute goods or services;
36 // loss of use, data, or profits; or business interruption) however caused
37 // and on any theory of liability, whether in contract, strict liability,
38 // or tort (including negligence or otherwise) arising in any way out of
39 // the use of this software, even if advised of the possibility of such damage.
43 #include "precomp.hpp"
44 #include "opencl_kernels.hpp"
49 /****************************************************************************************\
51 \****************************************************************************************/
53 template<typename T> static void
54 split_( const T* src, T** dst, int len, int cn )
56 int k = cn % 4 ? cn % 4 : 4;
61 for( i = j = 0; i < len; i++, j += cn )
66 T *dst0 = dst[0], *dst1 = dst[1];
67 for( i = j = 0; i < len; i++, j += cn )
75 T *dst0 = dst[0], *dst1 = dst[1], *dst2 = dst[2];
76 for( i = j = 0; i < len; i++, j += cn )
85 T *dst0 = dst[0], *dst1 = dst[1], *dst2 = dst[2], *dst3 = dst[3];
86 for( i = j = 0; i < len; i++, j += cn )
88 dst0[i] = src[j]; dst1[i] = src[j+1];
89 dst2[i] = src[j+2]; dst3[i] = src[j+3];
93 for( ; k < cn; k += 4 )
95 T *dst0 = dst[k], *dst1 = dst[k+1], *dst2 = dst[k+2], *dst3 = dst[k+3];
96 for( i = 0, j = k; i < len; i++, j += cn )
98 dst0[i] = src[j]; dst1[i] = src[j+1];
99 dst2[i] = src[j+2]; dst3[i] = src[j+3];
104 template<typename T> static void
105 merge_( const T** src, T* dst, int len, int cn )
107 int k = cn % 4 ? cn % 4 : 4;
111 const T* src0 = src[0];
112 for( i = j = 0; i < len; i++, j += cn )
117 const T *src0 = src[0], *src1 = src[1];
118 for( i = j = 0; i < len; i++, j += cn )
126 const T *src0 = src[0], *src1 = src[1], *src2 = src[2];
127 for( i = j = 0; i < len; i++, j += cn )
136 const T *src0 = src[0], *src1 = src[1], *src2 = src[2], *src3 = src[3];
137 for( i = j = 0; i < len; i++, j += cn )
139 dst[j] = src0[i]; dst[j+1] = src1[i];
140 dst[j+2] = src2[i]; dst[j+3] = src3[i];
144 for( ; k < cn; k += 4 )
146 const T *src0 = src[k], *src1 = src[k+1], *src2 = src[k+2], *src3 = src[k+3];
147 for( i = 0, j = k; i < len; i++, j += cn )
149 dst[j] = src0[i]; dst[j+1] = src1[i];
150 dst[j+2] = src2[i]; dst[j+3] = src3[i];
155 static void split8u(const uchar* src, uchar** dst, int len, int cn )
157 split_(src, dst, len, cn);
160 static void split16u(const ushort* src, ushort** dst, int len, int cn )
162 split_(src, dst, len, cn);
165 static void split32s(const int* src, int** dst, int len, int cn )
167 split_(src, dst, len, cn);
170 static void split64s(const int64* src, int64** dst, int len, int cn )
172 split_(src, dst, len, cn);
175 static void merge8u(const uchar** src, uchar* dst, int len, int cn )
177 merge_(src, dst, len, cn);
180 static void merge16u(const ushort** src, ushort* dst, int len, int cn )
182 merge_(src, dst, len, cn);
185 static void merge32s(const int** src, int* dst, int len, int cn )
187 merge_(src, dst, len, cn);
190 static void merge64s(const int64** src, int64* dst, int len, int cn )
192 merge_(src, dst, len, cn);
195 typedef void (*SplitFunc)(const uchar* src, uchar** dst, int len, int cn);
196 typedef void (*MergeFunc)(const uchar** src, uchar* dst, int len, int cn);
198 static SplitFunc getSplitFunc(int depth)
200 static SplitFunc splitTab[] =
202 (SplitFunc)GET_OPTIMIZED(split8u), (SplitFunc)GET_OPTIMIZED(split8u), (SplitFunc)GET_OPTIMIZED(split16u), (SplitFunc)GET_OPTIMIZED(split16u),
203 (SplitFunc)GET_OPTIMIZED(split32s), (SplitFunc)GET_OPTIMIZED(split32s), (SplitFunc)GET_OPTIMIZED(split64s), 0
206 return splitTab[depth];
209 static MergeFunc getMergeFunc(int depth)
211 static MergeFunc mergeTab[] =
213 (MergeFunc)GET_OPTIMIZED(merge8u), (MergeFunc)GET_OPTIMIZED(merge8u), (MergeFunc)GET_OPTIMIZED(merge16u), (MergeFunc)GET_OPTIMIZED(merge16u),
214 (MergeFunc)GET_OPTIMIZED(merge32s), (MergeFunc)GET_OPTIMIZED(merge32s), (MergeFunc)GET_OPTIMIZED(merge64s), 0
217 return mergeTab[depth];
222 void cv::split(const Mat& src, Mat* mv)
224 int k, depth = src.depth(), cn = src.channels();
231 SplitFunc func = getSplitFunc(depth);
232 CV_Assert( func != 0 );
234 int esz = (int)src.elemSize(), esz1 = (int)src.elemSize1();
235 int blocksize0 = (BLOCK_SIZE + esz-1)/esz;
236 AutoBuffer<uchar> _buf((cn+1)*(sizeof(Mat*) + sizeof(uchar*)) + 16);
237 const Mat** arrays = (const Mat**)(uchar*)_buf;
238 uchar** ptrs = (uchar**)alignPtr(arrays + cn + 1, 16);
241 for( k = 0; k < cn; k++ )
243 mv[k].create(src.dims, src.size, depth);
244 arrays[k+1] = &mv[k];
247 NAryMatIterator it(arrays, ptrs, cn+1);
248 int total = (int)it.size, blocksize = cn <= 4 ? total : std::min(total, blocksize0);
250 for( size_t i = 0; i < it.nplanes; i++, ++it )
252 for( int j = 0; j < total; j += blocksize )
254 int bsz = std::min(total - j, blocksize);
255 func( ptrs[0], &ptrs[1], bsz, cn );
257 if( j + blocksize < total )
260 for( k = 0; k < cn; k++ )
261 ptrs[k+1] += bsz*esz1;
271 static bool ocl_split( InputArray _m, OutputArrayOfArrays _mv )
273 int type = _m.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type),
274 rowsPerWI = ocl::Device::getDefault().isIntel() ? 4 : 1;
276 String dstargs, processelem, indexdecl;
277 for (int i = 0; i < cn; ++i)
279 dstargs += format("DECLARE_DST_PARAM(%d)", i);
280 indexdecl += format("DECLARE_INDEX(%d)", i);
281 processelem += format("PROCESS_ELEM(%d)", i);
284 ocl::Kernel k("split", ocl::core::split_merge_oclsrc,
285 format("-D T=%s -D OP_SPLIT -D cn=%d -D DECLARE_DST_PARAMS=%s"
286 " -D PROCESS_ELEMS_N=%s -D DECLARE_INDEX_N=%s",
287 ocl::memopTypeToStr(depth), cn, dstargs.c_str(),
288 processelem.c_str(), indexdecl.c_str()));
292 Size size = _m.size();
293 _mv.create(cn, 1, depth);
294 for (int i = 0; i < cn; ++i)
295 _mv.create(size, depth, i);
297 std::vector<UMat> dst;
298 _mv.getUMatVector(dst);
300 int argidx = k.set(0, ocl::KernelArg::ReadOnly(_m.getUMat()));
301 for (int i = 0; i < cn; ++i)
302 argidx = k.set(argidx, ocl::KernelArg::WriteOnlyNoSize(dst[i]));
303 k.set(argidx, rowsPerWI);
305 size_t globalsize[2] = { size.width, (size.height + rowsPerWI - 1) / rowsPerWI };
306 return k.run(2, globalsize, NULL, false);
313 void cv::split(InputArray _m, OutputArrayOfArrays _mv)
315 CV_OCL_RUN(_m.dims() <= 2 && _mv.isUMatVector(),
325 CV_Assert( !_mv.fixedType() || _mv.empty() || _mv.type() == m.depth() );
327 Size size = m.size();
328 int depth = m.depth(), cn = m.channels();
329 _mv.create(cn, 1, depth);
330 for (int i = 0; i < cn; ++i)
331 _mv.create(size, depth, i);
333 std::vector<Mat> dst;
334 _mv.getMatVector(dst);
339 void cv::merge(const Mat* mv, size_t n, OutputArray _dst)
341 CV_Assert( mv && n > 0 );
343 int depth = mv[0].depth();
348 for( i = 0; i < n; i++ )
350 CV_Assert(mv[i].size == mv[0].size && mv[i].depth() == depth);
351 allch1 = allch1 && mv[i].channels() == 1;
352 cn += mv[i].channels();
355 CV_Assert( 0 < cn && cn <= CV_CN_MAX );
356 _dst.create(mv[0].dims, mv[0].size, CV_MAKETYPE(depth, cn));
357 Mat dst = _dst.getMat();
367 AutoBuffer<int> pairs(cn*2);
370 for( i = 0, j = 0; i < n; i++, j += ni )
372 ni = mv[i].channels();
373 for( k = 0; k < ni; k++ )
375 pairs[(j+k)*2] = j + k;
376 pairs[(j+k)*2+1] = j + k;
379 mixChannels( mv, n, &dst, 1, &pairs[0], cn );
383 size_t esz = dst.elemSize(), esz1 = dst.elemSize1();
384 int blocksize0 = (int)((BLOCK_SIZE + esz-1)/esz);
385 AutoBuffer<uchar> _buf((cn+1)*(sizeof(Mat*) + sizeof(uchar*)) + 16);
386 const Mat** arrays = (const Mat**)(uchar*)_buf;
387 uchar** ptrs = (uchar**)alignPtr(arrays + cn + 1, 16);
390 for( k = 0; k < cn; k++ )
391 arrays[k+1] = &mv[k];
393 NAryMatIterator it(arrays, ptrs, cn+1);
394 int total = (int)it.size, blocksize = cn <= 4 ? total : std::min(total, blocksize0);
395 MergeFunc func = getMergeFunc(depth);
397 for( i = 0; i < it.nplanes; i++, ++it )
399 for( int j = 0; j < total; j += blocksize )
401 int bsz = std::min(total - j, blocksize);
402 func( (const uchar**)&ptrs[1], ptrs[0], bsz, cn );
404 if( j + blocksize < total )
407 for( int t = 0; t < cn; t++ )
408 ptrs[t+1] += bsz*esz1;
418 static bool ocl_merge( InputArrayOfArrays _mv, OutputArray _dst )
420 std::vector<UMat> src, ksrc;
421 _mv.getUMatVector(src);
422 CV_Assert(!src.empty());
424 int type = src[0].type(), depth = CV_MAT_DEPTH(type),
425 rowsPerWI = ocl::Device::getDefault().isIntel() ? 4 : 1;
426 Size size = src[0].size();
428 for (size_t i = 0, srcsize = src.size(); i < srcsize; ++i)
430 int itype = src[i].type(), icn = CV_MAT_CN(itype), idepth = CV_MAT_DEPTH(itype),
431 esz1 = CV_ELEM_SIZE1(idepth);
435 CV_Assert(size == src[i].size() && depth == idepth);
437 for (int cn = 0; cn < icn; ++cn)
440 tsrc.offset += cn * esz1;
441 ksrc.push_back(tsrc);
444 int dcn = (int)ksrc.size();
446 String srcargs, processelem, cndecl, indexdecl;
447 for (int i = 0; i < dcn; ++i)
449 srcargs += format("DECLARE_SRC_PARAM(%d)", i);
450 processelem += format("PROCESS_ELEM(%d)", i);
451 indexdecl += format("DECLARE_INDEX(%d)", i);
452 cndecl += format(" -D scn%d=%d", i, ksrc[i].channels());
455 ocl::Kernel k("merge", ocl::core::split_merge_oclsrc,
456 format("-D OP_MERGE -D cn=%d -D T=%s -D DECLARE_SRC_PARAMS_N=%s"
457 " -D DECLARE_INDEX_N=%s -D PROCESS_ELEMS_N=%s%s",
458 dcn, ocl::memopTypeToStr(depth), srcargs.c_str(),
459 indexdecl.c_str(), processelem.c_str(), cndecl.c_str()));
463 _dst.create(size, CV_MAKE_TYPE(depth, dcn));
464 UMat dst = _dst.getUMat();
467 for (int i = 0; i < dcn; ++i)
468 argidx = k.set(argidx, ocl::KernelArg::ReadOnlyNoSize(ksrc[i]));
469 argidx = k.set(argidx, ocl::KernelArg::WriteOnly(dst));
470 k.set(argidx, rowsPerWI);
472 size_t globalsize[2] = { dst.cols, (dst.rows + rowsPerWI - 1) / rowsPerWI };
473 return k.run(2, globalsize, NULL, false);
480 void cv::merge(InputArrayOfArrays _mv, OutputArray _dst)
482 CV_OCL_RUN(_mv.isUMatVector() && _dst.isUMat(),
483 ocl_merge(_mv, _dst))
486 _mv.getMatVector(mv);
487 merge(!mv.empty() ? &mv[0] : 0, mv.size(), _dst);
490 /****************************************************************************************\
491 * Generalized split/merge: mixing channels *
492 \****************************************************************************************/
497 template<typename T> static void
498 mixChannels_( const T** src, const int* sdelta,
499 T** dst, const int* ddelta,
500 int len, int npairs )
503 for( k = 0; k < npairs; k++ )
507 int ds = sdelta[k], dd = ddelta[k];
510 for( i = 0; i <= len - 2; i += 2, s += ds*2, d += dd*2 )
512 T t0 = s[0], t1 = s[ds];
513 d[0] = t0; d[dd] = t1;
520 for( i = 0; i <= len - 2; i += 2, d += dd*2 )
529 static void mixChannels8u( const uchar** src, const int* sdelta,
530 uchar** dst, const int* ddelta,
531 int len, int npairs )
533 mixChannels_(src, sdelta, dst, ddelta, len, npairs);
536 static void mixChannels16u( const ushort** src, const int* sdelta,
537 ushort** dst, const int* ddelta,
538 int len, int npairs )
540 mixChannels_(src, sdelta, dst, ddelta, len, npairs);
543 static void mixChannels32s( const int** src, const int* sdelta,
544 int** dst, const int* ddelta,
545 int len, int npairs )
547 mixChannels_(src, sdelta, dst, ddelta, len, npairs);
550 static void mixChannels64s( const int64** src, const int* sdelta,
551 int64** dst, const int* ddelta,
552 int len, int npairs )
554 mixChannels_(src, sdelta, dst, ddelta, len, npairs);
557 typedef void (*MixChannelsFunc)( const uchar** src, const int* sdelta,
558 uchar** dst, const int* ddelta, int len, int npairs );
560 static MixChannelsFunc getMixchFunc(int depth)
562 static MixChannelsFunc mixchTab[] =
564 (MixChannelsFunc)mixChannels8u, (MixChannelsFunc)mixChannels8u, (MixChannelsFunc)mixChannels16u,
565 (MixChannelsFunc)mixChannels16u, (MixChannelsFunc)mixChannels32s, (MixChannelsFunc)mixChannels32s,
566 (MixChannelsFunc)mixChannels64s, 0
569 return mixchTab[depth];
574 void cv::mixChannels( const Mat* src, size_t nsrcs, Mat* dst, size_t ndsts, const int* fromTo, size_t npairs )
578 CV_Assert( src && nsrcs > 0 && dst && ndsts > 0 && fromTo && npairs > 0 );
580 size_t i, j, k, esz1 = dst[0].elemSize1();
581 int depth = dst[0].depth();
583 AutoBuffer<uchar> buf((nsrcs + ndsts + 1)*(sizeof(Mat*) + sizeof(uchar*)) + npairs*(sizeof(uchar*)*2 + sizeof(int)*6));
584 const Mat** arrays = (const Mat**)(uchar*)buf;
585 uchar** ptrs = (uchar**)(arrays + nsrcs + ndsts);
586 const uchar** srcs = (const uchar**)(ptrs + nsrcs + ndsts + 1);
587 uchar** dsts = (uchar**)(srcs + npairs);
588 int* tab = (int*)(dsts + npairs);
589 int *sdelta = (int*)(tab + npairs*4), *ddelta = sdelta + npairs;
591 for( i = 0; i < nsrcs; i++ )
593 for( i = 0; i < ndsts; i++ )
594 arrays[i + nsrcs] = &dst[i];
595 ptrs[nsrcs + ndsts] = 0;
597 for( i = 0; i < npairs; i++ )
599 int i0 = fromTo[i*2], i1 = fromTo[i*2+1];
602 for( j = 0; j < nsrcs; i0 -= src[j].channels(), j++ )
603 if( i0 < src[j].channels() )
605 CV_Assert(j < nsrcs && src[j].depth() == depth);
606 tab[i*4] = (int)j; tab[i*4+1] = (int)(i0*esz1);
607 sdelta[i] = src[j].channels();
611 tab[i*4] = (int)(nsrcs + ndsts); tab[i*4+1] = 0;
615 for( j = 0; j < ndsts; i1 -= dst[j].channels(), j++ )
616 if( i1 < dst[j].channels() )
618 CV_Assert(i1 >= 0 && j < ndsts && dst[j].depth() == depth);
619 tab[i*4+2] = (int)(j + nsrcs); tab[i*4+3] = (int)(i1*esz1);
620 ddelta[i] = dst[j].channels();
623 NAryMatIterator it(arrays, ptrs, (int)(nsrcs + ndsts));
624 int total = (int)it.size, blocksize = std::min(total, (int)((BLOCK_SIZE + esz1-1)/esz1));
625 MixChannelsFunc func = getMixchFunc(depth);
627 for( i = 0; i < it.nplanes; i++, ++it )
629 for( k = 0; k < npairs; k++ )
631 srcs[k] = ptrs[tab[k*4]] + tab[k*4+1];
632 dsts[k] = ptrs[tab[k*4+2]] + tab[k*4+3];
635 for( int t = 0; t < total; t += blocksize )
637 int bsz = std::min(total - t, blocksize);
638 func( srcs, sdelta, dsts, ddelta, bsz, (int)npairs );
640 if( t + blocksize < total )
641 for( k = 0; k < npairs; k++ )
643 srcs[k] += blocksize*sdelta[k]*esz1;
644 dsts[k] += blocksize*ddelta[k]*esz1;
654 static void getUMatIndex(const std::vector<UMat> & um, int cn, int & idx, int & cnidx)
656 int totalChannels = 0;
657 for (size_t i = 0, size = um.size(); i < size; ++i)
659 int ccn = um[i].channels();
660 totalChannels += ccn;
662 if (totalChannels == cn)
668 else if (totalChannels > cn)
671 cnidx = i == 0 ? cn : (cn - totalChannels + ccn);
679 static bool ocl_mixChannels(InputArrayOfArrays _src, InputOutputArrayOfArrays _dst,
680 const int* fromTo, size_t npairs)
682 std::vector<UMat> src, dst;
683 _src.getUMatVector(src);
684 _dst.getUMatVector(dst);
686 size_t nsrc = src.size(), ndst = dst.size();
687 CV_Assert(nsrc > 0 && ndst > 0);
689 Size size = src[0].size();
690 int depth = src[0].depth(), esz = CV_ELEM_SIZE(depth),
691 rowsPerWI = ocl::Device::getDefault().isIntel() ? 4 : 1;
693 for (size_t i = 1, ssize = src.size(); i < ssize; ++i)
694 CV_Assert(src[i].size() == size && src[i].depth() == depth);
695 for (size_t i = 0, dsize = dst.size(); i < dsize; ++i)
696 CV_Assert(dst[i].size() == size && dst[i].depth() == depth);
698 String declsrc, decldst, declproc, declcn, indexdecl;
699 std::vector<UMat> srcargs(npairs), dstargs(npairs);
701 for (size_t i = 0; i < npairs; ++i)
703 int scn = fromTo[i<<1], dcn = fromTo[(i<<1) + 1];
704 int src_idx, src_cnidx, dst_idx, dst_cnidx;
706 getUMatIndex(src, scn, src_idx, src_cnidx);
707 getUMatIndex(dst, dcn, dst_idx, dst_cnidx);
709 CV_Assert(dst_idx >= 0 && src_idx >= 0);
711 srcargs[i] = src[src_idx];
712 srcargs[i].offset += src_cnidx * esz;
714 dstargs[i] = dst[dst_idx];
715 dstargs[i].offset += dst_cnidx * esz;
717 declsrc += format("DECLARE_INPUT_MAT(%d)", i);
718 decldst += format("DECLARE_OUTPUT_MAT(%d)", i);
719 indexdecl += format("DECLARE_INDEX(%d)", i);
720 declproc += format("PROCESS_ELEM(%d)", i);
721 declcn += format(" -D scn%d=%d -D dcn%d=%d", i, src[src_idx].channels(), i, dst[dst_idx].channels());
724 ocl::Kernel k("mixChannels", ocl::core::mixchannels_oclsrc,
725 format("-D T=%s -D DECLARE_INPUT_MAT_N=%s -D DECLARE_OUTPUT_MAT_N=%s"
726 " -D PROCESS_ELEM_N=%s -D DECLARE_INDEX_N=%s%s",
727 ocl::memopTypeToStr(depth), declsrc.c_str(), decldst.c_str(),
728 declproc.c_str(), indexdecl.c_str(), declcn.c_str()));
733 for (size_t i = 0; i < npairs; ++i)
734 argindex = k.set(argindex, ocl::KernelArg::ReadOnlyNoSize(srcargs[i]));
735 for (size_t i = 0; i < npairs; ++i)
736 argindex = k.set(argindex, ocl::KernelArg::WriteOnlyNoSize(dstargs[i]));
737 argindex = k.set(argindex, size.height);
738 argindex = k.set(argindex, size.width);
739 k.set(argindex, rowsPerWI);
741 size_t globalsize[2] = { size.width, (size.height + rowsPerWI - 1) / rowsPerWI };
742 return k.run(2, globalsize, NULL, false);
749 void cv::mixChannels(InputArrayOfArrays src, InputOutputArrayOfArrays dst,
750 const int* fromTo, size_t npairs)
752 if (npairs == 0 || fromTo == NULL)
755 CV_OCL_RUN(dst.isUMatVector(),
756 ocl_mixChannels(src, dst, fromTo, npairs))
758 bool src_is_mat = src.kind() != _InputArray::STD_VECTOR_MAT &&
759 src.kind() != _InputArray::STD_VECTOR_VECTOR &&
760 src.kind() != _InputArray::STD_VECTOR_UMAT;
761 bool dst_is_mat = dst.kind() != _InputArray::STD_VECTOR_MAT &&
762 dst.kind() != _InputArray::STD_VECTOR_VECTOR &&
763 dst.kind() != _InputArray::STD_VECTOR_UMAT;
765 int nsrc = src_is_mat ? 1 : (int)src.total();
766 int ndst = dst_is_mat ? 1 : (int)dst.total();
768 CV_Assert(nsrc > 0 && ndst > 0);
769 cv::AutoBuffer<Mat> _buf(nsrc + ndst);
771 for( i = 0; i < nsrc; i++ )
772 buf[i] = src.getMat(src_is_mat ? -1 : i);
773 for( i = 0; i < ndst; i++ )
774 buf[nsrc + i] = dst.getMat(dst_is_mat ? -1 : i);
775 mixChannels(&buf[0], nsrc, &buf[nsrc], ndst, fromTo, npairs);
778 void cv::mixChannels(InputArrayOfArrays src, InputOutputArrayOfArrays dst,
779 const std::vector<int>& fromTo)
784 CV_OCL_RUN(dst.isUMatVector(),
785 ocl_mixChannels(src, dst, &fromTo[0], fromTo.size()>>1))
787 bool src_is_mat = src.kind() != _InputArray::STD_VECTOR_MAT &&
788 src.kind() != _InputArray::STD_VECTOR_VECTOR &&
789 src.kind() != _InputArray::STD_VECTOR_UMAT;
790 bool dst_is_mat = dst.kind() != _InputArray::STD_VECTOR_MAT &&
791 dst.kind() != _InputArray::STD_VECTOR_VECTOR &&
792 dst.kind() != _InputArray::STD_VECTOR_UMAT;
794 int nsrc = src_is_mat ? 1 : (int)src.total();
795 int ndst = dst_is_mat ? 1 : (int)dst.total();
797 CV_Assert(fromTo.size()%2 == 0 && nsrc > 0 && ndst > 0);
798 cv::AutoBuffer<Mat> _buf(nsrc + ndst);
800 for( i = 0; i < nsrc; i++ )
801 buf[i] = src.getMat(src_is_mat ? -1 : i);
802 for( i = 0; i < ndst; i++ )
803 buf[nsrc + i] = dst.getMat(dst_is_mat ? -1 : i);
804 mixChannels(&buf[0], nsrc, &buf[nsrc], ndst, &fromTo[0], fromTo.size()/2);
807 void cv::extractChannel(InputArray _src, OutputArray _dst, int coi)
809 int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
810 CV_Assert( 0 <= coi && coi < cn );
811 int ch[] = { coi, 0 };
813 if (ocl::useOpenCL() && _src.dims() <= 2 && _dst.isUMat())
815 UMat src = _src.getUMat();
816 _dst.create(src.dims, &src.size[0], depth);
817 UMat dst = _dst.getUMat();
818 mixChannels(std::vector<UMat>(1, src), std::vector<UMat>(1, dst), ch, 1);
822 Mat src = _src.getMat();
823 _dst.create(src.dims, &src.size[0], depth);
824 Mat dst = _dst.getMat();
825 mixChannels(&src, 1, &dst, 1, ch, 1);
828 void cv::insertChannel(InputArray _src, InputOutputArray _dst, int coi)
830 int stype = _src.type(), sdepth = CV_MAT_DEPTH(stype), scn = CV_MAT_CN(stype);
831 int dtype = _dst.type(), ddepth = CV_MAT_DEPTH(dtype), dcn = CV_MAT_CN(dtype);
832 CV_Assert( _src.sameSize(_dst) && sdepth == ddepth );
833 CV_Assert( 0 <= coi && coi < dcn && scn == 1 );
835 int ch[] = { 0, coi };
836 if (ocl::useOpenCL() && _src.dims() <= 2 && _dst.isUMat())
838 UMat src = _src.getUMat(), dst = _dst.getUMat();
839 mixChannels(std::vector<UMat>(1, src), std::vector<UMat>(1, dst), ch, 1);
843 Mat src = _src.getMat(), dst = _dst.getMat();
844 mixChannels(&src, 1, &dst, 1, ch, 1);
847 /****************************************************************************************\
848 * convertScale[Abs] *
849 \****************************************************************************************/
854 template<typename T, typename DT, typename WT>
855 struct cvtScaleAbs_SSE2
857 int operator () (const T *, DT *, int, WT, WT) const
866 struct cvtScaleAbs_SSE2<uchar, uchar, float>
868 int operator () (const uchar * src, uchar * dst, int width,
869 float scale, float shift) const
875 __m128 v_scale = _mm_set1_ps(scale), v_shift = _mm_set1_ps(shift),
876 v_zero_f = _mm_setzero_ps();
877 __m128i v_zero_i = _mm_setzero_si128();
879 for ( ; x <= width - 16; x += 16)
881 __m128i v_src = _mm_loadu_si128((const __m128i *)(src + x));
882 __m128i v_src12 = _mm_unpacklo_epi8(v_src, v_zero_i), v_src_34 = _mm_unpackhi_epi8(v_src, v_zero_i);
883 __m128 v_dst1 = _mm_add_ps(_mm_mul_ps(_mm_cvtepi32_ps(_mm_unpacklo_epi16(v_src12, v_zero_i)), v_scale), v_shift);
884 v_dst1 = _mm_max_ps(_mm_sub_ps(v_zero_f, v_dst1), v_dst1);
885 __m128 v_dst2 = _mm_add_ps(_mm_mul_ps(_mm_cvtepi32_ps(_mm_unpackhi_epi16(v_src12, v_zero_i)), v_scale), v_shift);
886 v_dst2 = _mm_max_ps(_mm_sub_ps(v_zero_f, v_dst2), v_dst2);
887 __m128 v_dst3 = _mm_add_ps(_mm_mul_ps(_mm_cvtepi32_ps(_mm_unpacklo_epi16(v_src_34, v_zero_i)), v_scale), v_shift);
888 v_dst3 = _mm_max_ps(_mm_sub_ps(v_zero_f, v_dst3), v_dst3);
889 __m128 v_dst4 = _mm_add_ps(_mm_mul_ps(_mm_cvtepi32_ps(_mm_unpackhi_epi16(v_src_34, v_zero_i)), v_scale), v_shift);
890 v_dst4 = _mm_max_ps(_mm_sub_ps(v_zero_f, v_dst4), v_dst4);
892 __m128i v_dst_i = _mm_packus_epi16(_mm_packs_epi32(_mm_cvtps_epi32(v_dst1), _mm_cvtps_epi32(v_dst2)),
893 _mm_packs_epi32(_mm_cvtps_epi32(v_dst3), _mm_cvtps_epi32(v_dst4)));
894 _mm_storeu_si128((__m128i *)(dst + x), v_dst_i);
903 struct cvtScaleAbs_SSE2<ushort, uchar, float>
905 int operator () (const ushort * src, uchar * dst, int width,
906 float scale, float shift) const
912 __m128 v_scale = _mm_set1_ps(scale), v_shift = _mm_set1_ps(shift),
913 v_zero_f = _mm_setzero_ps();
914 __m128i v_zero_i = _mm_setzero_si128();
916 for ( ; x <= width - 8; x += 8)
918 __m128i v_src = _mm_loadu_si128((const __m128i *)(src + x));
919 __m128 v_dst1 = _mm_add_ps(_mm_mul_ps(_mm_cvtepi32_ps(_mm_unpacklo_epi16(v_src, v_zero_i)), v_scale), v_shift);
920 v_dst1 = _mm_max_ps(_mm_sub_ps(v_zero_f, v_dst1), v_dst1);
921 __m128 v_dst2 = _mm_add_ps(_mm_mul_ps(_mm_cvtepi32_ps(_mm_unpackhi_epi16(v_src, v_zero_i)), v_scale), v_shift);
922 v_dst2 = _mm_max_ps(_mm_sub_ps(v_zero_f, v_dst2), v_dst2);
924 __m128i v_dst_i = _mm_packus_epi16(_mm_packs_epi32(_mm_cvtps_epi32(v_dst1), _mm_cvtps_epi32(v_dst2)), v_zero_i);
925 _mm_storel_epi64((__m128i *)(dst + x), v_dst_i);
934 struct cvtScaleAbs_SSE2<short, uchar, float>
936 int operator () (const short * src, uchar * dst, int width,
937 float scale, float shift) const
943 __m128 v_scale = _mm_set1_ps(scale), v_shift = _mm_set1_ps(shift),
944 v_zero_f = _mm_setzero_ps();
945 __m128i v_zero_i = _mm_setzero_si128();
947 for ( ; x <= width - 8; x += 8)
949 __m128i v_src = _mm_loadu_si128((const __m128i *)(src + x));
950 __m128 v_dst1 = _mm_add_ps(_mm_mul_ps(_mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(v_src, v_src), 16)), v_scale), v_shift);
951 v_dst1 = _mm_max_ps(_mm_sub_ps(v_zero_f, v_dst1), v_dst1);
952 __m128 v_dst2 = _mm_add_ps(_mm_mul_ps(_mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(v_src, v_src), 16)), v_scale), v_shift);
953 v_dst2 = _mm_max_ps(_mm_sub_ps(v_zero_f, v_dst2), v_dst2);
955 __m128i v_dst_i = _mm_packus_epi16(_mm_packs_epi32(_mm_cvtps_epi32(v_dst1), _mm_cvtps_epi32(v_dst2)), v_zero_i);
956 _mm_storel_epi64((__m128i *)(dst + x), v_dst_i);
965 struct cvtScaleAbs_SSE2<int, uchar, float>
967 int operator () (const int * src, uchar * dst, int width,
968 float scale, float shift) const
974 __m128 v_scale = _mm_set1_ps(scale), v_shift = _mm_set1_ps(shift),
975 v_zero_f = _mm_setzero_ps();
976 __m128i v_zero_i = _mm_setzero_si128();
978 for ( ; x <= width - 8; x += 4)
980 __m128i v_src = _mm_loadu_si128((const __m128i *)(src + x));
981 __m128 v_dst1 = _mm_add_ps(_mm_mul_ps(_mm_cvtepi32_ps(v_src), v_scale), v_shift);
982 v_dst1 = _mm_max_ps(_mm_sub_ps(v_zero_f, v_dst1), v_dst1);
984 __m128i v_dst_i = _mm_packus_epi16(_mm_packs_epi32(_mm_cvtps_epi32(v_dst1), v_zero_i), v_zero_i);
985 _mm_storel_epi64((__m128i *)(dst + x), v_dst_i);
994 struct cvtScaleAbs_SSE2<float, uchar, float>
996 int operator () (const float * src, uchar * dst, int width,
997 float scale, float shift) const
1003 __m128 v_scale = _mm_set1_ps(scale), v_shift = _mm_set1_ps(shift),
1004 v_zero_f = _mm_setzero_ps();
1005 __m128i v_zero_i = _mm_setzero_si128();
1007 for ( ; x <= width - 8; x += 4)
1009 __m128 v_dst = _mm_add_ps(_mm_mul_ps(_mm_loadu_ps(src + x), v_scale), v_shift);
1010 v_dst = _mm_max_ps(_mm_sub_ps(v_zero_f, v_dst), v_dst);
1012 __m128i v_dst_i = _mm_packs_epi32(_mm_cvtps_epi32(v_dst), v_zero_i);
1013 _mm_storel_epi64((__m128i *)(dst + x), _mm_packus_epi16(v_dst_i, v_zero_i));
1023 template<typename T, typename DT, typename WT> static void
1024 cvtScaleAbs_( const T* src, size_t sstep,
1025 DT* dst, size_t dstep, Size size,
1026 WT scale, WT shift )
1028 sstep /= sizeof(src[0]);
1029 dstep /= sizeof(dst[0]);
1030 cvtScaleAbs_SSE2<T, DT, WT> vop;
1032 for( ; size.height--; src += sstep, dst += dstep )
1034 int x = vop(src, dst, size.width, scale, shift);
1036 #if CV_ENABLE_UNROLLED
1037 for( ; x <= size.width - 4; x += 4 )
1040 t0 = saturate_cast<DT>(std::abs(src[x]*scale + shift));
1041 t1 = saturate_cast<DT>(std::abs(src[x+1]*scale + shift));
1042 dst[x] = t0; dst[x+1] = t1;
1043 t0 = saturate_cast<DT>(std::abs(src[x+2]*scale + shift));
1044 t1 = saturate_cast<DT>(std::abs(src[x+3]*scale + shift));
1045 dst[x+2] = t0; dst[x+3] = t1;
1048 for( ; x < size.width; x++ )
1049 dst[x] = saturate_cast<DT>(std::abs(src[x]*scale + shift));
1053 template<typename T, typename DT, typename WT> static void
1054 cvtScale_( const T* src, size_t sstep,
1055 DT* dst, size_t dstep, Size size,
1056 WT scale, WT shift )
1058 sstep /= sizeof(src[0]);
1059 dstep /= sizeof(dst[0]);
1061 for( ; size.height--; src += sstep, dst += dstep )
1064 #if CV_ENABLE_UNROLLED
1065 for( ; x <= size.width - 4; x += 4 )
1068 t0 = saturate_cast<DT>(src[x]*scale + shift);
1069 t1 = saturate_cast<DT>(src[x+1]*scale + shift);
1070 dst[x] = t0; dst[x+1] = t1;
1071 t0 = saturate_cast<DT>(src[x+2]*scale + shift);
1072 t1 = saturate_cast<DT>(src[x+3]*scale + shift);
1073 dst[x+2] = t0; dst[x+3] = t1;
1077 for( ; x < size.width; x++ )
1078 dst[x] = saturate_cast<DT>(src[x]*scale + shift);
1082 //vz optimized template specialization
1084 cvtScale_<short, short, float>( const short* src, size_t sstep,
1085 short* dst, size_t dstep, Size size,
1086 float scale, float shift )
1088 sstep /= sizeof(src[0]);
1089 dstep /= sizeof(dst[0]);
1091 for( ; size.height--; src += sstep, dst += dstep )
1097 __m128 scale128 = _mm_set1_ps (scale);
1098 __m128 shift128 = _mm_set1_ps (shift);
1099 for(; x <= size.width - 8; x += 8 )
1101 __m128i r0 = _mm_loadl_epi64((const __m128i*)(src + x));
1102 __m128i r1 = _mm_loadl_epi64((const __m128i*)(src + x + 4));
1103 __m128 rf0 =_mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(r0, r0), 16));
1104 __m128 rf1 =_mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(r1, r1), 16));
1105 rf0 = _mm_add_ps(_mm_mul_ps(rf0, scale128), shift128);
1106 rf1 = _mm_add_ps(_mm_mul_ps(rf1, scale128), shift128);
1107 r0 = _mm_cvtps_epi32(rf0);
1108 r1 = _mm_cvtps_epi32(rf1);
1109 r0 = _mm_packs_epi32(r0, r1);
1110 _mm_storeu_si128((__m128i*)(dst + x), r0);
1115 for(; x < size.width; x++ )
1116 dst[x] = saturate_cast<short>(src[x]*scale + shift);
1121 cvtScale_<short, int, float>( const short* src, size_t sstep,
1122 int* dst, size_t dstep, Size size,
1123 float scale, float shift )
1125 sstep /= sizeof(src[0]);
1126 dstep /= sizeof(dst[0]);
1128 for( ; size.height--; src += sstep, dst += dstep )
1135 __m128 scale128 = _mm_set1_ps (scale);
1136 __m128 shift128 = _mm_set1_ps (shift);
1137 for(; x <= size.width - 8; x += 8 )
1139 __m128i r0 = _mm_loadl_epi64((const __m128i*)(src + x));
1140 __m128i r1 = _mm_loadl_epi64((const __m128i*)(src + x + 4));
1141 __m128 rf0 =_mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(r0, r0), 16));
1142 __m128 rf1 =_mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(r1, r1), 16));
1143 rf0 = _mm_add_ps(_mm_mul_ps(rf0, scale128), shift128);
1144 rf1 = _mm_add_ps(_mm_mul_ps(rf1, scale128), shift128);
1145 r0 = _mm_cvtps_epi32(rf0);
1146 r1 = _mm_cvtps_epi32(rf1);
1148 _mm_storeu_si128((__m128i*)(dst + x), r0);
1149 _mm_storeu_si128((__m128i*)(dst + x + 4), r1);
1154 //We will wait Haswell
1157 if(USE_AVX)//2X - bad variant
1159 ////TODO:AVX implementation (optimization?) required
1160 __m256 scale256 = _mm256_set1_ps (scale);
1161 __m256 shift256 = _mm256_set1_ps (shift);
1162 for(; x <= size.width - 8; x += 8 )
1164 __m256i buf = _mm256_set_epi32((int)(*(src+x+7)),(int)(*(src+x+6)),(int)(*(src+x+5)),(int)(*(src+x+4)),(int)(*(src+x+3)),(int)(*(src+x+2)),(int)(*(src+x+1)),(int)(*(src+x)));
1165 __m256 r0 = _mm256_add_ps( _mm256_mul_ps(_mm256_cvtepi32_ps (buf), scale256), shift256);
1166 __m256i res = _mm256_cvtps_epi32(r0);
1167 _mm256_storeu_si256 ((__m256i*)(dst+x), res);
1172 for(; x < size.width; x++ )
1173 dst[x] = saturate_cast<int>(src[x]*scale + shift);
1177 template<typename T, typename DT> static void
1178 cvt_( const T* src, size_t sstep,
1179 DT* dst, size_t dstep, Size size )
1181 sstep /= sizeof(src[0]);
1182 dstep /= sizeof(dst[0]);
1184 for( ; size.height--; src += sstep, dst += dstep )
1187 #if CV_ENABLE_UNROLLED
1188 for( ; x <= size.width - 4; x += 4 )
1191 t0 = saturate_cast<DT>(src[x]);
1192 t1 = saturate_cast<DT>(src[x+1]);
1193 dst[x] = t0; dst[x+1] = t1;
1194 t0 = saturate_cast<DT>(src[x+2]);
1195 t1 = saturate_cast<DT>(src[x+3]);
1196 dst[x+2] = t0; dst[x+3] = t1;
1199 for( ; x < size.width; x++ )
1200 dst[x] = saturate_cast<DT>(src[x]);
1204 //vz optimized template specialization, test Core_ConvertScale/ElemWiseTest
1206 cvt_<float, short>( const float* src, size_t sstep,
1207 short* dst, size_t dstep, Size size )
1209 sstep /= sizeof(src[0]);
1210 dstep /= sizeof(dst[0]);
1212 for( ; size.height--; src += sstep, dst += dstep )
1217 for( ; x <= size.width - 8; x += 8 )
1219 __m128 src128 = _mm_loadu_ps (src + x);
1220 __m128i src_int128 = _mm_cvtps_epi32 (src128);
1222 src128 = _mm_loadu_ps (src + x + 4);
1223 __m128i src1_int128 = _mm_cvtps_epi32 (src128);
1225 src1_int128 = _mm_packs_epi32(src_int128, src1_int128);
1226 _mm_storeu_si128((__m128i*)(dst + x),src1_int128);
1230 for( ; x < size.width; x++ )
1231 dst[x] = saturate_cast<short>(src[x]);
1237 template<typename T> static void
1238 cpy_( const T* src, size_t sstep, T* dst, size_t dstep, Size size )
1240 sstep /= sizeof(src[0]);
1241 dstep /= sizeof(dst[0]);
1243 for( ; size.height--; src += sstep, dst += dstep )
1244 memcpy(dst, src, size.width*sizeof(src[0]));
1247 #define DEF_CVT_SCALE_ABS_FUNC(suffix, tfunc, stype, dtype, wtype) \
1248 static void cvtScaleAbs##suffix( const stype* src, size_t sstep, const uchar*, size_t, \
1249 dtype* dst, size_t dstep, Size size, double* scale) \
1251 tfunc(src, sstep, dst, dstep, size, (wtype)scale[0], (wtype)scale[1]); \
1254 #define DEF_CVT_SCALE_FUNC(suffix, stype, dtype, wtype) \
1255 static void cvtScale##suffix( const stype* src, size_t sstep, const uchar*, size_t, \
1256 dtype* dst, size_t dstep, Size size, double* scale) \
1258 cvtScale_(src, sstep, dst, dstep, size, (wtype)scale[0], (wtype)scale[1]); \
1261 #if defined(HAVE_IPP)
1262 #define DEF_CVT_FUNC_F(suffix, stype, dtype, ippFavor) \
1263 static void cvt##suffix( const stype* src, size_t sstep, const uchar*, size_t, \
1264 dtype* dst, size_t dstep, Size size, double*) \
1268 if (ippiConvert_##ippFavor(src, (int)sstep, dst, (int)dstep, ippiSize(size.width, size.height)) >= 0) \
1270 setIppErrorStatus(); \
1272 cvt_(src, sstep, dst, dstep, size); \
1275 #define DEF_CVT_FUNC_F2(suffix, stype, dtype, ippFavor) \
1276 static void cvt##suffix( const stype* src, size_t sstep, const uchar*, size_t, \
1277 dtype* dst, size_t dstep, Size size, double*) \
1281 if (ippiConvert_##ippFavor(src, (int)sstep, dst, (int)dstep, ippiSize(size.width, size.height), ippRndFinancial, 0) >= 0) \
1283 setIppErrorStatus(); \
1285 cvt_(src, sstep, dst, dstep, size); \
1288 #define DEF_CVT_FUNC_F(suffix, stype, dtype, ippFavor) \
1289 static void cvt##suffix( const stype* src, size_t sstep, const uchar*, size_t, \
1290 dtype* dst, size_t dstep, Size size, double*) \
1292 cvt_(src, sstep, dst, dstep, size); \
1294 #define DEF_CVT_FUNC_F2 DEF_CVT_FUNC_F
1297 #define DEF_CVT_FUNC(suffix, stype, dtype) \
1298 static void cvt##suffix( const stype* src, size_t sstep, const uchar*, size_t, \
1299 dtype* dst, size_t dstep, Size size, double*) \
1301 cvt_(src, sstep, dst, dstep, size); \
1304 #define DEF_CPY_FUNC(suffix, stype) \
1305 static void cvt##suffix( const stype* src, size_t sstep, const uchar*, size_t, \
1306 stype* dst, size_t dstep, Size size, double*) \
1308 cpy_(src, sstep, dst, dstep, size); \
1312 DEF_CVT_SCALE_ABS_FUNC(8u, cvtScaleAbs_, uchar, uchar, float)
1313 DEF_CVT_SCALE_ABS_FUNC(8s8u, cvtScaleAbs_, schar, uchar, float)
1314 DEF_CVT_SCALE_ABS_FUNC(16u8u, cvtScaleAbs_, ushort, uchar, float)
1315 DEF_CVT_SCALE_ABS_FUNC(16s8u, cvtScaleAbs_, short, uchar, float)
1316 DEF_CVT_SCALE_ABS_FUNC(32s8u, cvtScaleAbs_, int, uchar, float)
1317 DEF_CVT_SCALE_ABS_FUNC(32f8u, cvtScaleAbs_, float, uchar, float)
1318 DEF_CVT_SCALE_ABS_FUNC(64f8u, cvtScaleAbs_, double, uchar, float)
1320 DEF_CVT_SCALE_FUNC(8u, uchar, uchar, float)
1321 DEF_CVT_SCALE_FUNC(8s8u, schar, uchar, float)
1322 DEF_CVT_SCALE_FUNC(16u8u, ushort, uchar, float)
1323 DEF_CVT_SCALE_FUNC(16s8u, short, uchar, float)
1324 DEF_CVT_SCALE_FUNC(32s8u, int, uchar, float)
1325 DEF_CVT_SCALE_FUNC(32f8u, float, uchar, float)
1326 DEF_CVT_SCALE_FUNC(64f8u, double, uchar, float)
1328 DEF_CVT_SCALE_FUNC(8u8s, uchar, schar, float)
1329 DEF_CVT_SCALE_FUNC(8s, schar, schar, float)
1330 DEF_CVT_SCALE_FUNC(16u8s, ushort, schar, float)
1331 DEF_CVT_SCALE_FUNC(16s8s, short, schar, float)
1332 DEF_CVT_SCALE_FUNC(32s8s, int, schar, float)
1333 DEF_CVT_SCALE_FUNC(32f8s, float, schar, float)
1334 DEF_CVT_SCALE_FUNC(64f8s, double, schar, float)
1336 DEF_CVT_SCALE_FUNC(8u16u, uchar, ushort, float)
1337 DEF_CVT_SCALE_FUNC(8s16u, schar, ushort, float)
1338 DEF_CVT_SCALE_FUNC(16u, ushort, ushort, float)
1339 DEF_CVT_SCALE_FUNC(16s16u, short, ushort, float)
1340 DEF_CVT_SCALE_FUNC(32s16u, int, ushort, float)
1341 DEF_CVT_SCALE_FUNC(32f16u, float, ushort, float)
1342 DEF_CVT_SCALE_FUNC(64f16u, double, ushort, float)
1344 DEF_CVT_SCALE_FUNC(8u16s, uchar, short, float)
1345 DEF_CVT_SCALE_FUNC(8s16s, schar, short, float)
1346 DEF_CVT_SCALE_FUNC(16u16s, ushort, short, float)
1347 DEF_CVT_SCALE_FUNC(16s, short, short, float)
1348 DEF_CVT_SCALE_FUNC(32s16s, int, short, float)
1349 DEF_CVT_SCALE_FUNC(32f16s, float, short, float)
1350 DEF_CVT_SCALE_FUNC(64f16s, double, short, float)
1352 DEF_CVT_SCALE_FUNC(8u32s, uchar, int, float)
1353 DEF_CVT_SCALE_FUNC(8s32s, schar, int, float)
1354 DEF_CVT_SCALE_FUNC(16u32s, ushort, int, float)
1355 DEF_CVT_SCALE_FUNC(16s32s, short, int, float)
1356 DEF_CVT_SCALE_FUNC(32s, int, int, double)
1357 DEF_CVT_SCALE_FUNC(32f32s, float, int, float)
1358 DEF_CVT_SCALE_FUNC(64f32s, double, int, double)
1360 DEF_CVT_SCALE_FUNC(8u32f, uchar, float, float)
1361 DEF_CVT_SCALE_FUNC(8s32f, schar, float, float)
1362 DEF_CVT_SCALE_FUNC(16u32f, ushort, float, float)
1363 DEF_CVT_SCALE_FUNC(16s32f, short, float, float)
1364 DEF_CVT_SCALE_FUNC(32s32f, int, float, double)
1365 DEF_CVT_SCALE_FUNC(32f, float, float, float)
1366 DEF_CVT_SCALE_FUNC(64f32f, double, float, double)
1368 DEF_CVT_SCALE_FUNC(8u64f, uchar, double, double)
1369 DEF_CVT_SCALE_FUNC(8s64f, schar, double, double)
1370 DEF_CVT_SCALE_FUNC(16u64f, ushort, double, double)
1371 DEF_CVT_SCALE_FUNC(16s64f, short, double, double)
1372 DEF_CVT_SCALE_FUNC(32s64f, int, double, double)
1373 DEF_CVT_SCALE_FUNC(32f64f, float, double, double)
1374 DEF_CVT_SCALE_FUNC(64f, double, double, double)
1376 DEF_CPY_FUNC(8u, uchar)
1377 DEF_CVT_FUNC_F(8s8u, schar, uchar, 8s8u_C1Rs)
1378 DEF_CVT_FUNC_F(16u8u, ushort, uchar, 16u8u_C1R)
1379 DEF_CVT_FUNC_F(16s8u, short, uchar, 16s8u_C1R)
1380 DEF_CVT_FUNC_F(32s8u, int, uchar, 32s8u_C1R)
1381 DEF_CVT_FUNC_F2(32f8u, float, uchar, 32f8u_C1RSfs)
1382 DEF_CVT_FUNC(64f8u, double, uchar)
1384 DEF_CVT_FUNC_F2(8u8s, uchar, schar, 8u8s_C1RSfs)
1385 DEF_CVT_FUNC_F2(16u8s, ushort, schar, 16u8s_C1RSfs)
1386 DEF_CVT_FUNC_F2(16s8s, short, schar, 16s8s_C1RSfs)
1387 DEF_CVT_FUNC_F(32s8s, int, schar, 32s8s_C1R)
1388 DEF_CVT_FUNC_F2(32f8s, float, schar, 32f8s_C1RSfs)
1389 DEF_CVT_FUNC(64f8s, double, schar)
1391 DEF_CVT_FUNC_F(8u16u, uchar, ushort, 8u16u_C1R)
1392 DEF_CVT_FUNC_F(8s16u, schar, ushort, 8s16u_C1Rs)
1393 DEF_CPY_FUNC(16u, ushort)
1394 DEF_CVT_FUNC_F(16s16u, short, ushort, 16s16u_C1Rs)
1395 DEF_CVT_FUNC_F2(32s16u, int, ushort, 32s16u_C1RSfs)
1396 DEF_CVT_FUNC_F2(32f16u, float, ushort, 32f16u_C1RSfs)
1397 DEF_CVT_FUNC(64f16u, double, ushort)
1399 DEF_CVT_FUNC_F(8u16s, uchar, short, 8u16s_C1R)
1400 DEF_CVT_FUNC_F(8s16s, schar, short, 8s16s_C1R)
1401 DEF_CVT_FUNC_F2(16u16s, ushort, short, 16u16s_C1RSfs)
1402 DEF_CVT_FUNC_F2(32s16s, int, short, 32s16s_C1RSfs)
1403 DEF_CVT_FUNC_F2(32f16s, float, short, 32f16s_C1RSfs)
1404 DEF_CVT_FUNC(64f16s, double, short)
1406 DEF_CVT_FUNC_F(8u32s, uchar, int, 8u32s_C1R)
1407 DEF_CVT_FUNC_F(8s32s, schar, int, 8s32s_C1R)
1408 DEF_CVT_FUNC_F(16u32s, ushort, int, 16u32s_C1R)
1409 DEF_CVT_FUNC_F(16s32s, short, int, 16s32s_C1R)
1410 DEF_CPY_FUNC(32s, int)
1411 DEF_CVT_FUNC_F2(32f32s, float, int, 32f32s_C1RSfs)
1412 DEF_CVT_FUNC(64f32s, double, int)
1414 DEF_CVT_FUNC_F(8u32f, uchar, float, 8u32f_C1R)
1415 DEF_CVT_FUNC_F(8s32f, schar, float, 8s32f_C1R)
1416 DEF_CVT_FUNC_F(16u32f, ushort, float, 16u32f_C1R)
1417 DEF_CVT_FUNC_F(16s32f, short, float, 16s32f_C1R)
1418 DEF_CVT_FUNC_F(32s32f, int, float, 32s32f_C1R)
1419 DEF_CVT_FUNC(64f32f, double, float)
1421 DEF_CVT_FUNC(8u64f, uchar, double)
1422 DEF_CVT_FUNC(8s64f, schar, double)
1423 DEF_CVT_FUNC(16u64f, ushort, double)
1424 DEF_CVT_FUNC(16s64f, short, double)
1425 DEF_CVT_FUNC(32s64f, int, double)
1426 DEF_CVT_FUNC(32f64f, float, double)
1427 DEF_CPY_FUNC(64s, int64)
1429 static BinaryFunc getCvtScaleAbsFunc(int depth)
1431 static BinaryFunc cvtScaleAbsTab[] =
1433 (BinaryFunc)cvtScaleAbs8u, (BinaryFunc)cvtScaleAbs8s8u, (BinaryFunc)cvtScaleAbs16u8u,
1434 (BinaryFunc)cvtScaleAbs16s8u, (BinaryFunc)cvtScaleAbs32s8u, (BinaryFunc)cvtScaleAbs32f8u,
1435 (BinaryFunc)cvtScaleAbs64f8u, 0
1438 return cvtScaleAbsTab[depth];
1441 BinaryFunc getConvertFunc(int sdepth, int ddepth)
1443 static BinaryFunc cvtTab[][8] =
1446 (BinaryFunc)(cvt8u), (BinaryFunc)GET_OPTIMIZED(cvt8s8u), (BinaryFunc)GET_OPTIMIZED(cvt16u8u),
1447 (BinaryFunc)GET_OPTIMIZED(cvt16s8u), (BinaryFunc)GET_OPTIMIZED(cvt32s8u), (BinaryFunc)GET_OPTIMIZED(cvt32f8u),
1448 (BinaryFunc)GET_OPTIMIZED(cvt64f8u), 0
1451 (BinaryFunc)GET_OPTIMIZED(cvt8u8s), (BinaryFunc)cvt8u, (BinaryFunc)GET_OPTIMIZED(cvt16u8s),
1452 (BinaryFunc)GET_OPTIMIZED(cvt16s8s), (BinaryFunc)GET_OPTIMIZED(cvt32s8s), (BinaryFunc)GET_OPTIMIZED(cvt32f8s),
1453 (BinaryFunc)GET_OPTIMIZED(cvt64f8s), 0
1456 (BinaryFunc)GET_OPTIMIZED(cvt8u16u), (BinaryFunc)GET_OPTIMIZED(cvt8s16u), (BinaryFunc)cvt16u,
1457 (BinaryFunc)GET_OPTIMIZED(cvt16s16u), (BinaryFunc)GET_OPTIMIZED(cvt32s16u), (BinaryFunc)GET_OPTIMIZED(cvt32f16u),
1458 (BinaryFunc)GET_OPTIMIZED(cvt64f16u), 0
1461 (BinaryFunc)GET_OPTIMIZED(cvt8u16s), (BinaryFunc)GET_OPTIMIZED(cvt8s16s), (BinaryFunc)GET_OPTIMIZED(cvt16u16s),
1462 (BinaryFunc)cvt16u, (BinaryFunc)GET_OPTIMIZED(cvt32s16s), (BinaryFunc)GET_OPTIMIZED(cvt32f16s),
1463 (BinaryFunc)GET_OPTIMIZED(cvt64f16s), 0
1466 (BinaryFunc)GET_OPTIMIZED(cvt8u32s), (BinaryFunc)GET_OPTIMIZED(cvt8s32s), (BinaryFunc)GET_OPTIMIZED(cvt16u32s),
1467 (BinaryFunc)GET_OPTIMIZED(cvt16s32s), (BinaryFunc)cvt32s, (BinaryFunc)GET_OPTIMIZED(cvt32f32s),
1468 (BinaryFunc)GET_OPTIMIZED(cvt64f32s), 0
1471 (BinaryFunc)GET_OPTIMIZED(cvt8u32f), (BinaryFunc)GET_OPTIMIZED(cvt8s32f), (BinaryFunc)GET_OPTIMIZED(cvt16u32f),
1472 (BinaryFunc)GET_OPTIMIZED(cvt16s32f), (BinaryFunc)GET_OPTIMIZED(cvt32s32f), (BinaryFunc)cvt32s,
1473 (BinaryFunc)GET_OPTIMIZED(cvt64f32f), 0
1476 (BinaryFunc)GET_OPTIMIZED(cvt8u64f), (BinaryFunc)GET_OPTIMIZED(cvt8s64f), (BinaryFunc)GET_OPTIMIZED(cvt16u64f),
1477 (BinaryFunc)GET_OPTIMIZED(cvt16s64f), (BinaryFunc)GET_OPTIMIZED(cvt32s64f), (BinaryFunc)GET_OPTIMIZED(cvt32f64f),
1478 (BinaryFunc)(cvt64s), 0
1481 0, 0, 0, 0, 0, 0, 0, 0
1485 return cvtTab[CV_MAT_DEPTH(ddepth)][CV_MAT_DEPTH(sdepth)];
1488 static BinaryFunc getConvertScaleFunc(int sdepth, int ddepth)
1490 static BinaryFunc cvtScaleTab[][8] =
1493 (BinaryFunc)GET_OPTIMIZED(cvtScale8u), (BinaryFunc)GET_OPTIMIZED(cvtScale8s8u), (BinaryFunc)GET_OPTIMIZED(cvtScale16u8u),
1494 (BinaryFunc)GET_OPTIMIZED(cvtScale16s8u), (BinaryFunc)GET_OPTIMIZED(cvtScale32s8u), (BinaryFunc)GET_OPTIMIZED(cvtScale32f8u),
1495 (BinaryFunc)cvtScale64f8u, 0
1498 (BinaryFunc)GET_OPTIMIZED(cvtScale8u8s), (BinaryFunc)GET_OPTIMIZED(cvtScale8s), (BinaryFunc)GET_OPTIMIZED(cvtScale16u8s),
1499 (BinaryFunc)GET_OPTIMIZED(cvtScale16s8s), (BinaryFunc)GET_OPTIMIZED(cvtScale32s8s), (BinaryFunc)GET_OPTIMIZED(cvtScale32f8s),
1500 (BinaryFunc)cvtScale64f8s, 0
1503 (BinaryFunc)GET_OPTIMIZED(cvtScale8u16u), (BinaryFunc)GET_OPTIMIZED(cvtScale8s16u), (BinaryFunc)GET_OPTIMIZED(cvtScale16u),
1504 (BinaryFunc)GET_OPTIMIZED(cvtScale16s16u), (BinaryFunc)GET_OPTIMIZED(cvtScale32s16u), (BinaryFunc)GET_OPTIMIZED(cvtScale32f16u),
1505 (BinaryFunc)cvtScale64f16u, 0
1508 (BinaryFunc)GET_OPTIMIZED(cvtScale8u16s), (BinaryFunc)GET_OPTIMIZED(cvtScale8s16s), (BinaryFunc)GET_OPTIMIZED(cvtScale16u16s),
1509 (BinaryFunc)GET_OPTIMIZED(cvtScale16s), (BinaryFunc)GET_OPTIMIZED(cvtScale32s16s), (BinaryFunc)GET_OPTIMIZED(cvtScale32f16s),
1510 (BinaryFunc)cvtScale64f16s, 0
1513 (BinaryFunc)GET_OPTIMIZED(cvtScale8u32s), (BinaryFunc)GET_OPTIMIZED(cvtScale8s32s), (BinaryFunc)GET_OPTIMIZED(cvtScale16u32s),
1514 (BinaryFunc)GET_OPTIMIZED(cvtScale16s32s), (BinaryFunc)GET_OPTIMIZED(cvtScale32s), (BinaryFunc)GET_OPTIMIZED(cvtScale32f32s),
1515 (BinaryFunc)cvtScale64f32s, 0
1518 (BinaryFunc)GET_OPTIMIZED(cvtScale8u32f), (BinaryFunc)GET_OPTIMIZED(cvtScale8s32f), (BinaryFunc)GET_OPTIMIZED(cvtScale16u32f),
1519 (BinaryFunc)GET_OPTIMIZED(cvtScale16s32f), (BinaryFunc)GET_OPTIMIZED(cvtScale32s32f), (BinaryFunc)GET_OPTIMIZED(cvtScale32f),
1520 (BinaryFunc)cvtScale64f32f, 0
1523 (BinaryFunc)cvtScale8u64f, (BinaryFunc)cvtScale8s64f, (BinaryFunc)cvtScale16u64f,
1524 (BinaryFunc)cvtScale16s64f, (BinaryFunc)cvtScale32s64f, (BinaryFunc)cvtScale32f64f,
1525 (BinaryFunc)cvtScale64f, 0
1528 0, 0, 0, 0, 0, 0, 0, 0
1532 return cvtScaleTab[CV_MAT_DEPTH(ddepth)][CV_MAT_DEPTH(sdepth)];
1537 static bool ocl_convertScaleAbs( InputArray _src, OutputArray _dst, double alpha, double beta )
1539 const ocl::Device & d = ocl::Device::getDefault();
1540 int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type),
1541 kercn = ocl::predictOptimalVectorWidth(_src, _dst), rowsPerWI = d.isIntel() ? 4 : 1;
1542 bool doubleSupport = d.doubleFPConfig() > 0;
1544 if (depth == CV_32F || depth == CV_64F)
1548 int wdepth = std::max(depth, CV_32F);
1549 ocl::Kernel k("KF", ocl::core::arithm_oclsrc,
1550 format("-D OP_CONVERT_SCALE_ABS -D UNARY_OP -D dstT=%s -D srcT1=%s"
1551 " -D workT=%s -D wdepth=%d -D convertToWT1=%s -D convertToDT=%s"
1552 " -D workT1=%s -D rowsPerWI=%d%s",
1553 ocl::typeToStr(CV_8UC(kercn)),
1554 ocl::typeToStr(CV_MAKE_TYPE(depth, kercn)),
1555 ocl::typeToStr(CV_MAKE_TYPE(wdepth, kercn)), wdepth,
1556 ocl::convertTypeStr(depth, wdepth, kercn, cvt[0]),
1557 ocl::convertTypeStr(wdepth, CV_8U, kercn, cvt[1]),
1558 ocl::typeToStr(wdepth), rowsPerWI,
1559 doubleSupport ? " -D DOUBLE_SUPPORT" : ""));
1563 UMat src = _src.getUMat();
1564 _dst.create(src.size(), CV_8UC(cn));
1565 UMat dst = _dst.getUMat();
1567 ocl::KernelArg srcarg = ocl::KernelArg::ReadOnlyNoSize(src),
1568 dstarg = ocl::KernelArg::WriteOnly(dst, cn, kercn);
1570 if (wdepth == CV_32F)
1571 k.args(srcarg, dstarg, (float)alpha, (float)beta);
1572 else if (wdepth == CV_64F)
1573 k.args(srcarg, dstarg, alpha, beta);
1575 size_t globalsize[2] = { src.cols * cn / kercn, (src.rows + rowsPerWI - 1) / rowsPerWI };
1576 return k.run(2, globalsize, NULL, false);
1583 void cv::convertScaleAbs( InputArray _src, OutputArray _dst, double alpha, double beta )
1585 CV_OCL_RUN(_src.dims() <= 2 && _dst.isUMat(),
1586 ocl_convertScaleAbs(_src, _dst, alpha, beta))
1588 Mat src = _src.getMat();
1589 int cn = src.channels();
1590 double scale[] = {alpha, beta};
1591 _dst.create( src.dims, src.size, CV_8UC(cn) );
1592 Mat dst = _dst.getMat();
1593 BinaryFunc func = getCvtScaleAbsFunc(src.depth());
1594 CV_Assert( func != 0 );
1598 Size sz = getContinuousSize(src, dst, cn);
1599 func( src.data, src.step, 0, 0, dst.data, dst.step, sz, scale );
1603 const Mat* arrays[] = {&src, &dst, 0};
1605 NAryMatIterator it(arrays, ptrs);
1606 Size sz((int)it.size*cn, 1);
1608 for( size_t i = 0; i < it.nplanes; i++, ++it )
1609 func( ptrs[0], 0, 0, 0, ptrs[1], 0, sz, scale );
1613 void cv::Mat::convertTo(OutputArray _dst, int _type, double alpha, double beta) const
1615 bool noScale = fabs(alpha-1) < DBL_EPSILON && fabs(beta) < DBL_EPSILON;
1618 _type = _dst.fixedType() ? _dst.type() : type();
1620 _type = CV_MAKETYPE(CV_MAT_DEPTH(_type), channels());
1622 int sdepth = depth(), ddepth = CV_MAT_DEPTH(_type);
1623 if( sdepth == ddepth && noScale )
1631 BinaryFunc func = noScale ? getConvertFunc(sdepth, ddepth) : getConvertScaleFunc(sdepth, ddepth);
1632 double scale[] = {alpha, beta};
1633 int cn = channels();
1634 CV_Assert( func != 0 );
1638 _dst.create( size(), _type );
1639 Mat dst = _dst.getMat();
1640 Size sz = getContinuousSize(src, dst, cn);
1641 func( src.data, src.step, 0, 0, dst.data, dst.step, sz, scale );
1645 _dst.create( dims, size, _type );
1646 Mat dst = _dst.getMat();
1647 const Mat* arrays[] = {&src, &dst, 0};
1649 NAryMatIterator it(arrays, ptrs);
1650 Size sz((int)(it.size*cn), 1);
1652 for( size_t i = 0; i < it.nplanes; i++, ++it )
1653 func(ptrs[0], 1, 0, 0, ptrs[1], 1, sz, scale);
1657 /****************************************************************************************\
1659 \****************************************************************************************/
1664 template<typename T> static void
1665 LUT8u_( const uchar* src, const T* lut, T* dst, int len, int cn, int lutcn )
1669 for( int i = 0; i < len*cn; i++ )
1670 dst[i] = lut[src[i]];
1674 for( int i = 0; i < len*cn; i += cn )
1675 for( int k = 0; k < cn; k++ )
1676 dst[i+k] = lut[src[i+k]*cn+k];
1680 static void LUT8u_8u( const uchar* src, const uchar* lut, uchar* dst, int len, int cn, int lutcn )
1682 LUT8u_( src, lut, dst, len, cn, lutcn );
1685 static void LUT8u_8s( const uchar* src, const schar* lut, schar* dst, int len, int cn, int lutcn )
1687 LUT8u_( src, lut, dst, len, cn, lutcn );
1690 static void LUT8u_16u( const uchar* src, const ushort* lut, ushort* dst, int len, int cn, int lutcn )
1692 LUT8u_( src, lut, dst, len, cn, lutcn );
1695 static void LUT8u_16s( const uchar* src, const short* lut, short* dst, int len, int cn, int lutcn )
1697 LUT8u_( src, lut, dst, len, cn, lutcn );
1700 static void LUT8u_32s( const uchar* src, const int* lut, int* dst, int len, int cn, int lutcn )
1702 LUT8u_( src, lut, dst, len, cn, lutcn );
1705 static void LUT8u_32f( const uchar* src, const float* lut, float* dst, int len, int cn, int lutcn )
1707 LUT8u_( src, lut, dst, len, cn, lutcn );
1710 static void LUT8u_64f( const uchar* src, const double* lut, double* dst, int len, int cn, int lutcn )
1712 LUT8u_( src, lut, dst, len, cn, lutcn );
1715 typedef void (*LUTFunc)( const uchar* src, const uchar* lut, uchar* dst, int len, int cn, int lutcn );
1717 static LUTFunc lutTab[] =
1719 (LUTFunc)LUT8u_8u, (LUTFunc)LUT8u_8s, (LUTFunc)LUT8u_16u, (LUTFunc)LUT8u_16s,
1720 (LUTFunc)LUT8u_32s, (LUTFunc)LUT8u_32f, (LUTFunc)LUT8u_64f, 0
1725 static bool ocl_LUT(InputArray _src, InputArray _lut, OutputArray _dst)
1727 int lcn = _lut.channels(), dcn = _src.channels(), ddepth = _lut.depth();
1729 UMat src = _src.getUMat(), lut = _lut.getUMat();
1730 _dst.create(src.size(), CV_MAKETYPE(ddepth, dcn));
1731 UMat dst = _dst.getUMat();
1732 int kercn = lcn == 1 ? std::min(4, ocl::predictOptimalVectorWidth(_dst)) : dcn;
1734 ocl::Kernel k("LUT", ocl::core::lut_oclsrc,
1735 format("-D dcn=%d -D lcn=%d -D srcT=%s -D dstT=%s", kercn, lcn,
1736 ocl::typeToStr(src.depth()), ocl::memopTypeToStr(ddepth)));
1740 k.args(ocl::KernelArg::ReadOnlyNoSize(src), ocl::KernelArg::ReadOnlyNoSize(lut),
1741 ocl::KernelArg::WriteOnly(dst, dcn, kercn));
1743 size_t globalSize[2] = { dst.cols * dcn / kercn, (dst.rows + 3) / 4 };
1744 return k.run(2, globalSize, NULL, false);
1749 #if defined(HAVE_IPP)
1752 #if 0 // there are no performance benefits (PR #2653)
1753 class IppLUTParallelBody_LUTC1 : public ParallelLoopBody
1761 typedef IppStatus (*IppFn)(const Ipp8u* pSrc, int srcStep, void* pDst, int dstStep,
1762 IppiSize roiSize, const void* pTable, int nBitSize);
1767 IppLUTParallelBody_LUTC1(const Mat& src, const Mat& lut, Mat& dst, bool* _ok)
1768 : ok(_ok), src_(src), lut_(lut), dst_(dst)
1770 width = dst.cols * dst.channels();
1772 size_t elemSize1 = CV_ELEM_SIZE1(dst.depth());
1775 elemSize1 == 1 ? (IppFn)ippiLUTPalette_8u_C1R :
1776 elemSize1 == 4 ? (IppFn)ippiLUTPalette_8u32u_C1R :
1782 void operator()( const cv::Range& range ) const
1787 const int row0 = range.start;
1788 const int row1 = range.end;
1790 Mat src = src_.rowRange(row0, row1);
1791 Mat dst = dst_.rowRange(row0, row1);
1793 IppiSize sz = { width, dst.rows };
1795 CV_DbgAssert(fn != NULL);
1796 if (fn(src.data, (int)src.step[0], dst.data, (int)dst.step[0], sz, lut_.data, 8) < 0)
1798 setIppErrorStatus();
1803 IppLUTParallelBody_LUTC1(const IppLUTParallelBody_LUTC1&);
1804 IppLUTParallelBody_LUTC1& operator=(const IppLUTParallelBody_LUTC1&);
1808 class IppLUTParallelBody_LUTCN : public ParallelLoopBody
1821 IppLUTParallelBody_LUTCN(const Mat& src, const Mat& lut, Mat& dst, bool* _ok)
1822 : ok(_ok), src_(src), lut_(lut), dst_(dst), lutBuffer(NULL)
1824 lutcn = lut.channels();
1825 IppiSize sz256 = {256, 1};
1827 size_t elemSize1 = dst.elemSize1();
1828 CV_DbgAssert(elemSize1 == 1);
1829 lutBuffer = (uchar*)ippMalloc(256 * (int)elemSize1 * 4);
1830 lutTable[0] = lutBuffer + 0;
1831 lutTable[1] = lutBuffer + 1 * 256 * elemSize1;
1832 lutTable[2] = lutBuffer + 2 * 256 * elemSize1;
1833 lutTable[3] = lutBuffer + 3 * 256 * elemSize1;
1835 CV_DbgAssert(lutcn == 3 || lutcn == 4);
1838 IppStatus status = ippiCopy_8u_C3P3R(lut.data, (int)lut.step[0], lutTable, (int)lut.step[0], sz256);
1841 setIppErrorStatus();
1845 else if (lutcn == 4)
1847 IppStatus status = ippiCopy_8u_C4P4R(lut.data, (int)lut.step[0], lutTable, (int)lut.step[0], sz256);
1850 setIppErrorStatus();
1858 ~IppLUTParallelBody_LUTCN()
1860 if (lutBuffer != NULL)
1866 void operator()( const cv::Range& range ) const
1871 const int row0 = range.start;
1872 const int row1 = range.end;
1874 Mat src = src_.rowRange(row0, row1);
1875 Mat dst = dst_.rowRange(row0, row1);
1879 if (ippiLUTPalette_8u_C3R(
1880 src.data, (int)src.step[0], dst.data, (int)dst.step[0],
1881 ippiSize(dst.size()), lutTable, 8) >= 0)
1884 else if (lutcn == 4)
1886 if (ippiLUTPalette_8u_C4R(
1887 src.data, (int)src.step[0], dst.data, (int)dst.step[0],
1888 ippiSize(dst.size()), lutTable, 8) >= 0)
1891 setIppErrorStatus();
1895 IppLUTParallelBody_LUTCN(const IppLUTParallelBody_LUTCN&);
1896 IppLUTParallelBody_LUTCN& operator=(const IppLUTParallelBody_LUTCN&);
1901 class LUTParallelBody : public ParallelLoopBody
1911 LUTParallelBody(const Mat& src, const Mat& lut, Mat& dst, bool* _ok)
1912 : ok(_ok), src_(src), lut_(lut), dst_(dst)
1914 func = lutTab[lut.depth()];
1915 *ok = (func != NULL);
1918 void operator()( const cv::Range& range ) const
1922 const int row0 = range.start;
1923 const int row1 = range.end;
1925 Mat src = src_.rowRange(row0, row1);
1926 Mat dst = dst_.rowRange(row0, row1);
1928 int cn = src.channels();
1929 int lutcn = lut_.channels();
1931 const Mat* arrays[] = {&src, &dst, 0};
1933 NAryMatIterator it(arrays, ptrs);
1934 int len = (int)it.size;
1936 for( size_t i = 0; i < it.nplanes; i++, ++it )
1937 func(ptrs[0], lut_.data, ptrs[1], len, cn, lutcn);
1940 LUTParallelBody(const LUTParallelBody&);
1941 LUTParallelBody& operator=(const LUTParallelBody&);
1946 void cv::LUT( InputArray _src, InputArray _lut, OutputArray _dst )
1948 int cn = _src.channels(), depth = _src.depth();
1949 int lutcn = _lut.channels();
1951 CV_Assert( (lutcn == cn || lutcn == 1) &&
1952 _lut.total() == 256 && _lut.isContinuous() &&
1953 (depth == CV_8U || depth == CV_8S) );
1955 CV_OCL_RUN(_dst.isUMat() && _src.dims() <= 2,
1956 ocl_LUT(_src, _lut, _dst))
1958 Mat src = _src.getMat(), lut = _lut.getMat();
1959 _dst.create(src.dims, src.size, CV_MAKETYPE(_lut.depth(), cn));
1960 Mat dst = _dst.getMat();
1962 if (_src.dims() <= 2)
1965 Ptr<ParallelLoopBody> body;
1966 #if defined(HAVE_IPP)
1967 size_t elemSize1 = CV_ELEM_SIZE1(dst.depth());
1968 #if 0 // there are no performance benefits (PR #2653)
1971 ParallelLoopBody* p = new ipp::IppLUTParallelBody_LUTC1(src, lut, dst, &ok);
1976 if ((lutcn == 3 || lutcn == 4) && elemSize1 == 1)
1978 ParallelLoopBody* p = new ipp::IppLUTParallelBody_LUTCN(src, lut, dst, &ok);
1982 if (body == NULL || ok == false)
1985 ParallelLoopBody* p = new LUTParallelBody(src, lut, dst, &ok);
1988 if (body != NULL && ok)
1990 Range all(0, dst.rows);
1991 if (dst.total()>>18)
1992 parallel_for_(all, *body, (double)std::max((size_t)1, dst.total()>>16));
2000 LUTFunc func = lutTab[lut.depth()];
2001 CV_Assert( func != 0 );
2003 const Mat* arrays[] = {&src, &dst, 0};
2005 NAryMatIterator it(arrays, ptrs);
2006 int len = (int)it.size;
2008 for( size_t i = 0; i < it.nplanes; i++, ++it )
2009 func(ptrs[0], lut.data, ptrs[1], len, cn, lutcn);
2016 static bool ocl_normalize( InputArray _src, InputOutputArray _dst, InputArray _mask, int dtype,
2017 double scale, double delta )
2019 UMat src = _src.getUMat();
2022 src.convertTo( _dst, dtype, scale, delta );
2023 else if (src.channels() <= 4)
2025 const ocl::Device & dev = ocl::Device::getDefault();
2027 int stype = _src.type(), sdepth = CV_MAT_DEPTH(stype), cn = CV_MAT_CN(stype),
2028 ddepth = CV_MAT_DEPTH(dtype), wdepth = std::max(CV_32F, std::max(sdepth, ddepth)),
2029 rowsPerWI = dev.isIntel() ? 4 : 1;
2031 float fscale = static_cast<float>(scale), fdelta = static_cast<float>(delta);
2032 bool haveScale = std::fabs(scale - 1) > DBL_EPSILON,
2033 haveZeroScale = !(std::fabs(scale) > DBL_EPSILON),
2034 haveDelta = std::fabs(delta) > DBL_EPSILON,
2035 doubleSupport = dev.doubleFPConfig() > 0;
2037 if (!haveScale && !haveDelta && stype == dtype)
2039 _src.copyTo(_dst, _mask);
2044 _dst.setTo(Scalar(delta), _mask);
2048 if ((sdepth == CV_64F || ddepth == CV_64F) && !doubleSupport)
2052 String opts = format("-D srcT=%s -D dstT=%s -D convertToWT=%s -D cn=%d -D rowsPerWI=%d"
2053 " -D convertToDT=%s -D workT=%s%s%s%s -D srcT1=%s -D dstT1=%s",
2054 ocl::typeToStr(stype), ocl::typeToStr(dtype),
2055 ocl::convertTypeStr(sdepth, wdepth, cn, cvt[0]), cn,
2056 rowsPerWI, ocl::convertTypeStr(wdepth, ddepth, cn, cvt[1]),
2057 ocl::typeToStr(CV_MAKE_TYPE(wdepth, cn)),
2058 doubleSupport ? " -D DOUBLE_SUPPORT" : "",
2059 haveScale ? " -D HAVE_SCALE" : "",
2060 haveDelta ? " -D HAVE_DELTA" : "",
2061 ocl::typeToStr(sdepth), ocl::typeToStr(ddepth));
2063 ocl::Kernel k("normalizek", ocl::core::normalize_oclsrc, opts);
2067 UMat mask = _mask.getUMat(), dst = _dst.getUMat();
2069 ocl::KernelArg srcarg = ocl::KernelArg::ReadOnlyNoSize(src),
2070 maskarg = ocl::KernelArg::ReadOnlyNoSize(mask),
2071 dstarg = ocl::KernelArg::ReadWrite(dst);
2076 k.args(srcarg, maskarg, dstarg, fscale, fdelta);
2078 k.args(srcarg, maskarg, dstarg, fscale);
2083 k.args(srcarg, maskarg, dstarg, fdelta);
2085 k.args(srcarg, maskarg, dstarg);
2088 size_t globalsize[2] = { src.cols, (src.rows + rowsPerWI - 1) / rowsPerWI };
2089 return k.run(2, globalsize, NULL, false);
2094 src.convertTo( temp, dtype, scale, delta );
2095 temp.copyTo( _dst, _mask );
2105 void cv::normalize( InputArray _src, InputOutputArray _dst, double a, double b,
2106 int norm_type, int rtype, InputArray _mask )
2108 double scale = 1, shift = 0;
2109 if( norm_type == CV_MINMAX )
2111 double smin = 0, smax = 0;
2112 double dmin = MIN( a, b ), dmax = MAX( a, b );
2113 minMaxLoc( _src, &smin, &smax, 0, 0, _mask );
2114 scale = (dmax - dmin)*(smax - smin > DBL_EPSILON ? 1./(smax - smin) : 0);
2115 shift = dmin - smin*scale;
2117 else if( norm_type == CV_L2 || norm_type == CV_L1 || norm_type == CV_C )
2119 scale = norm( _src, norm_type, _mask );
2120 scale = scale > DBL_EPSILON ? a/scale : 0.;
2124 CV_Error( CV_StsBadArg, "Unknown/unsupported norm type" );
2126 int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
2128 rtype = _dst.fixedType() ? _dst.depth() : depth;
2129 _dst.createSameSize(_src, CV_MAKETYPE(rtype, cn));
2131 CV_OCL_RUN(_dst.isUMat(),
2132 ocl_normalize(_src, _dst, _mask, rtype, scale, shift))
2134 Mat src = _src.getMat(), dst = _dst.getMat();
2136 src.convertTo( dst, rtype, scale, shift );
2140 src.convertTo( temp, rtype, scale, shift );
2141 temp.copyTo( dst, _mask );
2146 cvSplit( const void* srcarr, void* dstarr0, void* dstarr1, void* dstarr2, void* dstarr3 )
2148 void* dptrs[] = { dstarr0, dstarr1, dstarr2, dstarr3 };
2149 cv::Mat src = cv::cvarrToMat(srcarr);
2151 for( i = 0; i < 4; i++ )
2152 nz += dptrs[i] != 0;
2153 CV_Assert( nz > 0 );
2154 std::vector<cv::Mat> dvec(nz);
2155 std::vector<int> pairs(nz*2);
2157 for( i = j = 0; i < 4; i++ )
2161 dvec[j] = cv::cvarrToMat(dptrs[i]);
2162 CV_Assert( dvec[j].size() == src.size() );
2163 CV_Assert( dvec[j].depth() == src.depth() );
2164 CV_Assert( dvec[j].channels() == 1 );
2165 CV_Assert( i < src.channels() );
2171 if( nz == src.channels() )
2172 cv::split( src, dvec );
2175 cv::mixChannels( &src, 1, &dvec[0], nz, &pairs[0], nz );
2181 cvMerge( const void* srcarr0, const void* srcarr1, const void* srcarr2,
2182 const void* srcarr3, void* dstarr )
2184 const void* sptrs[] = { srcarr0, srcarr1, srcarr2, srcarr3 };
2185 cv::Mat dst = cv::cvarrToMat(dstarr);
2187 for( i = 0; i < 4; i++ )
2188 nz += sptrs[i] != 0;
2189 CV_Assert( nz > 0 );
2190 std::vector<cv::Mat> svec(nz);
2191 std::vector<int> pairs(nz*2);
2193 for( i = j = 0; i < 4; i++ )
2197 svec[j] = cv::cvarrToMat(sptrs[i]);
2198 CV_Assert( svec[j].size == dst.size &&
2199 svec[j].depth() == dst.depth() &&
2200 svec[j].channels() == 1 && i < dst.channels() );
2207 if( nz == dst.channels() )
2208 cv::merge( svec, dst );
2211 cv::mixChannels( &svec[0], nz, &dst, 1, &pairs[0], nz );
2217 cvMixChannels( const CvArr** src, int src_count,
2218 CvArr** dst, int dst_count,
2219 const int* from_to, int pair_count )
2221 cv::AutoBuffer<cv::Mat> buf(src_count + dst_count);
2224 for( i = 0; i < src_count; i++ )
2225 buf[i] = cv::cvarrToMat(src[i]);
2226 for( i = 0; i < dst_count; i++ )
2227 buf[i+src_count] = cv::cvarrToMat(dst[i]);
2228 cv::mixChannels(&buf[0], src_count, &buf[src_count], dst_count, from_to, pair_count);
2232 cvConvertScaleAbs( const void* srcarr, void* dstarr,
2233 double scale, double shift )
2235 cv::Mat src = cv::cvarrToMat(srcarr), dst = cv::cvarrToMat(dstarr);
2236 CV_Assert( src.size == dst.size && dst.type() == CV_8UC(src.channels()));
2237 cv::convertScaleAbs( src, dst, scale, shift );
2241 cvConvertScale( const void* srcarr, void* dstarr,
2242 double scale, double shift )
2244 cv::Mat src = cv::cvarrToMat(srcarr), dst = cv::cvarrToMat(dstarr);
2246 CV_Assert( src.size == dst.size && src.channels() == dst.channels() );
2247 src.convertTo(dst, dst.type(), scale, shift);
2250 CV_IMPL void cvLUT( const void* srcarr, void* dstarr, const void* lutarr )
2252 cv::Mat src = cv::cvarrToMat(srcarr), dst = cv::cvarrToMat(dstarr), lut = cv::cvarrToMat(lutarr);
2254 CV_Assert( dst.size() == src.size() && dst.type() == CV_MAKETYPE(lut.depth(), src.channels()) );
2255 cv::LUT( src, lut, dst );
2258 CV_IMPL void cvNormalize( const CvArr* srcarr, CvArr* dstarr,
2259 double a, double b, int norm_type, const CvArr* maskarr )
2261 cv::Mat src = cv::cvarrToMat(srcarr), dst = cv::cvarrToMat(dstarr), mask;
2263 mask = cv::cvarrToMat(maskarr);
2264 CV_Assert( dst.size() == src.size() && src.channels() == dst.channels() );
2265 cv::normalize( src, dst, a, b, norm_type, dst.type(), mask );